The Mech That’s Hiding the Forest

Elon Musk wrote “cool” on X. That’s where it all starts — and probably where the misunderstanding starts, too.

Unitree’s GD01 is impressive for specific reasons: 2.7 meters tall, 500 kilos, a CEO holding its hand before climbing into the cockpit, a morphology that shifts from biped to quadruped in seconds. Available for purchase, between $573,000 and $650,000, serial production, not displayed behind glass at a tech expo. It’s real. It’s now. And it’s not really what we should be talking about.

What’s striking here isn’t the machine itself. It’s what it reveals about how people are reading robotics right now. The GD01 demolishing concrete blocks is raw physical power combined with stable locomotion. Mechanically impressive. But Unitree acknowledges in its own safety notice that “humanoid robotics remains experimental with significant limitations for individual users.” No specifications on onboard AI. No demo of a complex task. The mech doesn’t have a brain yet — and that’s not a criticism, it’s just where things stand.

What’s actually changing, and what deserves real analysis, is happening somewhere else entirely.

Two Robots, One Room, Zero Communication Between Them

The Figure demo released the same week operates at a completely different level — and it passed under the radar faster than it should have.

Two humanoids enter a room. In under two minutes, they open doors, hang a coat, close a laptop, put away headphones, move furniture, make the bed together. Fully autonomous, start to finish. On paper, nothing exceptional — the demo has thousands of predecessors in the field.

What’s extraordinary is the underlying architecture.

The two robots share nothing. No shared planner. No central controller. Not a single message sent from one unit to the other. Each robot has only its onboard cameras and its learned policy. Coordination emerges purely from mutual observation: every action one robot takes changes the state of the room, and the other re-evaluates in real time.

And this is where it gets genuinely interesting. The duvet problem — which Figure describes as the hardest part of the demo — illustrates exactly why. Fabric has no fixed geometry. It folds, deforms, changes shape under tension. When two robots pull from different positions, each one has to anticipate the other’s movements while adjusting its grip on a material that won’t stop shifting. And they do it — without talking to each other.

The entire system runs on a single VLA (vision-language-action) model trained via reinforcement learning in simulation, across highly randomized configurations. What makes it particularly notable is the transfer: behaviors learned in simulation carry directly into the real world, without additional calibration.

This problem — the gap between simulation and reality — has been one of the most persistent obstacles in robotics for decades. Figure doesn’t claim to have solved it permanently. But the demo suggests it can be reduced enough to become useful.

This deserves more than a footnote, because it’s genuinely worth slowing down for. Most of the spectacular advances in robotics over the past ten years operated in environments so controlled they validated almost nothing about the real world. What Figure is showing here — two agents who have never “seen” this specific room, never rehearsed this exact configuration, coordinating without any communication infrastructure — is a break in how the problem is being framed.

Not a solved problem. A reframed one. That’s different, and the difference matters.

The Question Physical Intelligence Is Asking the Entire Industry

Physical Intelligence isn’t really building robots. Their objective is different — and more ambitious: create a generalist foundation model for robotics, a single brain capable of adapting to different hardware and handling the unpredictability of real environments. What GPT-4 is to language, π aims to be to physical bodies.

The company was co-founded in 2023 by former DeepMind researchers — including figures like Chelsea Finn — with funding exceeding one billion dollars and a current valuation of $5.6 billion. Their model has gone through three versions: π0 to prove it could work (tackling laundry folding, considered out of reach at the time), π0.5 for generalization, and the current version for reliability.

The results from π0.5 surprised the team itself. Trained on roughly a hundred different domestic environments, the model generalized to a 101st home it had never seen. They thought they’d need thousands, maybe millions of environments. They needed a hundred.

This is exactly where — right at the peak of the enthusiasm — the counter-argument has to come in, because it’s a serious one.

Physical Intelligence’s founder describes the current state of the field as “the GPT-2 moment”: real signs of life, genuine potential, but a major scaling leap still required before any of this is useful to most people. The metaphor is appealing. It’s also potentially misleading.

Large language models scaled for a specific reason: training data — text — was infinitely reproducible, practically free to collect, and available in near-unlimited quantities on the internet. Scaling worked because the data infrastructure already existed.

Robots need physical data. Data obtained through human teleoperation — someone demonstrating a task in real time, motion by motion. That’s expensive, slow, and hard to parallelize. Simulation can compensate, but it creates its own problems: robots learn heuristics that work in perfect simulated worlds and fail against the imperfections of reality — dust, shifting light, a slightly uneven surface. Physical Intelligence has narrowed that gap. It hasn’t closed it.

The obvious reading of the hundred-environment result is that scaling is already happening — but that’s precisely what should give us pause. A hundred carefully selected environments, collected by a high-level research team with a billion dollars in funding, is not the same data regime as the web. Generalizing to a 101st home is impressive. Generalizing to an industrial warehouse, an operating room, or a user’s kitchen in Bangalore is an entirely different question.

That’s not a reason for skepticism. It’s a reason to know exactly where you are on the map.

One of Physical Intelligence’s investors, present since the four-person garage days, estimates progress has been two to three times faster than their most optimistic projections. What was expected to take three to five years took eighteen months. That’s real. And if the pace holds — which is not guaranteed at all — enterprise deployment within one to three years is not a fantasy projection.

Three Bets, Three Incompatible Definitions of “Useful Robot”

What’s actually happening in robotics right now is the coexistence of three radically incompatible bets.

Unitree is betting on accessible hardware and production volume. The company reportedly shipped over 5,500 humanoid robots last year. For comparison, Tesla, Figure AI, and Agility Robotics each shipped around 150 units over the same period. Their entry-level humanoid robots start at roughly €4,300. American competitors typically cost ten times more. The advantage isn’t purely technological — it comes from supply chain: motors, reducers, sensors, carbon fiber materials, all produced locally within an industrial ecosystem that China, for now, is alone in owning end-to-end.

Figure is betting on emergent intelligence without heavy infrastructure — coordination without a central planner, a learned policy that adapts rather than a planning algorithm that calculates. It’s a bet on generalizing behaviors through learning.

Physical Intelligence is betting on model universality — that the brain can be separated from the body, that the same foundation model can run on different hardware and handle different tasks across different contexts.

These three approaches don’t necessarily converge. They’re answering different questions.

What Remains Open

The question Physical Intelligence is asking — and hasn’t yet answered — may be the most important one in the field for the next ten years.

If foundation models for robotics scale the way LLMs did, even at a slower pace, even under tighter data constraints, then the physical world becomes a surface for algorithmic optimization in exactly the way text has. Warehouses, assembly lines, kitchens, hospitals.

Which raises a question none of these three companies are asking publicly, but that surfaces in every serious conversation on the topic: who controls the physical infrastructure in that scenario? Not the robots — the models animating them. And the models animating them are owned by whom, exactly, and under what conditions?

Unitree sells bodies. Figure builds behaviors. Physical Intelligence builds brains. This isn’t a stable arrangement — it’s a positioning game, and it’s still very much in play.

Physical Intelligence’s founder has an honest formulation of the current state: “It’s not solved yet, and we can’t pretend we’re close.” But the pace has been striking. Not GPT-4. Not GPT-5. GPT-2 — and that’s not an admission of weakness, it’s a calibration.

Which means, for anyone following this field without the resources of Big Tech, the window to understand the architectures, experiment with open models, and develop intuitions about what’s going to scale — that window is right now. Not in three years when enterprise deployments are real. Now, while the field is still readable from the outside.

Robotics is where generative AI was in 2020. That’s not a reassuring comparison. It’s an invitation to watch very carefully — before it moves too fast to follow.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

On May 1st, Reuters obtained access to SpaceX’s confidential IPO filing. Among the expected figures — $15 billion invested in Starship, $3 billion in R&D for 2025 alone — sat a sentence the SEC had made mandatory, one SpaceX would never have published on its own terms.

“We have not yet demonstrated or attempted on-orbit propellant transfer.”

It is the most candid thing SpaceX has ever put in writing about its flagship program. And it appears in a document designed to convince investors to put money into it.

This irony deserves more than a footnote. Not because it reveals a failure — the program is moving, flights are accumulating, improvements between versions are real and documented. But because it says something precise about the nature of what those $15 billion are actually buying. This isn’t a rocket. It isn’t an R&D program. It’s a dependency infrastructure.

The Falcon 9 costs $400 million.

To understand what $15 billion means, you first have to understand what $400 million produced.

The Falcon 9 — NASA-certified, the undisputed dominant force in the global launch market with over 400 flights — is one of the best industrial investments in recent memory. Single-stage recovery. Sixteen to twenty-two metric tons to low Earth orbit, depending on configuration. Remarkable reliability.

Starship exceeds that development cost by a factor of 37.

That multiplier looks absurd — until you look at what actually changed, and it isn’t scale. The Falcon 9 solves a launch problem. Starship solves something different in kind.

Consider the concrete numbers. The V3 stands 124 meters — nearly double the Falcon 9. Its 9-meter diameter is such that the internal methane transfer tube running through the Super Heavy booster has a diameter roughly equivalent to the entire first stage of its predecessor. Its 33 Raptor 3 engines generate approximately 8,200 metric tons of thrust — more than ten times what the Falcon 9 produces. Payload to low Earth orbit: 100-plus metric tons versus 16 to 17 for the Falcon 9 in recovery configuration. Six times more.

These numbers are impressive. But they aren’t the argument. The argument is what sits behind them.

What the Infrastructure Is Actually Buying

To support that performance, SpaceX has built — literally built — an entire industrial ecosystem at Boca Chica. Two orbital launch towers. Hundreds of hours of engine testing at McGregor. Thousands of thermal tiles tested, failed, reformulated. And, the most recent visible decision in that effort: a liquefaction plant constructed directly on the Starbase site, because a single Starship launch requires between 470 and 570 truckloads of propellants — methane, liquid oxygen, nitrogen — and there is only one road into the site.

SpaceX now produces its own fuel on-site.

This isn’t logistical optimization. It’s vertical integration at scale. And it is precisely what industrial economists call switching costs. The deeper a customer integrates into an ecosystem, the more migration to a competitor costs — not just in money, but in time, reorganization, and risk.

NASA knows this firsthand. For Artemis HLS, Starship must be refueled multiple times in orbit by dedicated tanker vehicles. A maneuver that, according to the IPO filing, has never been attempted. That SpaceX admits this is remarkable — a technical vulnerability disclosed under legal compulsion. But what rarely gets said next is that NASA has already signed the contracts, already structured its schedule, already organized its teams around a rocket whose most critical capability doesn’t exist yet.

The switching cost is already running before the product works. That’s exactly the mechanism.

China has rockets, too.

The honest version of this argument has to include the counterargument.

The heavy-lift launch market is not a closed monopoly. Blue Origin is flying the New Glenn. China is developing the Long March 9 with comparable ambitions. Europe is preparing for the post-Ariane 6 era. And the history of “inevitable” technological monopolies is also a history of their surprises — in 2007, Nokia dominated mobile with an infrastructure and distribution lead that no one believed was catchable.

That’s the strongest counterargument. It deserves a precise answer rather than a detour.

SpaceX’s lead is not primarily technological. It’s temporal. In industries with high fixed costs, the first player to reach a high launch cadence drives down its marginal cost in a way competitors can only replicate by passing through the same initial investments. While its competitors are still building their infrastructure, SpaceX is accumulating real flight data on heat shields, on atmospheric re-entries, and on catch-arm dynamics during static fires. It’s that learning gap — not the technical delta at any given moment — that is genuinely hard to close.

Nokia fell in two years because its advantage lay in physical distribution, which the iPhone bypassed through the App Store. To bypass SpaceX’s advantage, you would need to bypass the physics of heavy-lift launch vehicles.

Possible. Not trivial.

What Europe’s Situation Already Tells Us

Here’s what the “$15 billion is too much” debate structurally misses.

Historically, infrastructure for orbital access has been financed by states as sovereign programs. GPS. Ariane. Soyuz. These were not commercial bets — they were strategic assets, publicly funded, government-controlled. Decisions about their access were subject to some form of democratic deliberation.

What SpaceX is building with its $15 billion is structurally different. An orbital access infrastructure owned by a private company, whose decisions — which customers, at what price, in what order of priority — answer to no democratic mandate and very few regulatory constraints in this specific domain.

For a European satellite operator, for a national space agency outside the United States, for anyone who will need access to space in the next decade, this question is already concrete. Europe, caught between Ariane 6’s difficulties and its successor not yet operational, had to rely on SpaceX. This is no longer a hypothetical scenario. It’s an active precedent.

Starship may be years from completion. The heat shield remains an open problem. Orbital refueling is yet to be built, let alone tested. All of that is true.

And in the meantime, the infrastructure exists. The contracts are signed. The dependencies are forming. The gap is widening while the rocket is still being developed, not after.

That’s what the $15 billion is. Not excess. A method.

The question tech journalism almost never asks — because it sits outside the territory of specs and records — is this: an orbital access infrastructure owned by a private company, who does it answer to?

SpaceX answered the SEC, under legal compulsion, that it has not yet attempted orbital refueling. It would be worth having it answer someone, someday, on the question of governance.

No one is asking that question seriously yet. That won’t last.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

AI no longer answers. It acts.

Remi, Orbit, GPT 5.5, MTP. Four names. One inflection point.

“A true assistant capable of acting on your behalf.”

That’s not a marketing tagline. It’s Google’s internal description of Remi, an agent the company is currently testing with its own employees. A formulation that was almost certainly reviewed by marketing teams, legal teams, product teams — and still: on your behalf.

Four words borrowed from the law of representation, not from software. A mandate. A power of attorney. A legal representative acts on your behalf. That language isn’t accidental. It signals something specific: Remi isn’t a tool that responds to your requests. It’s an agent that makes decisions in your place.

In a single week, three of the largest organizations in technology made the same announcement using different words.

Google, with Remi — currently in dog-fooding, not yet public, integrated across the full ecosystem: Gmail, Docs, Drive, Calendar, Search. Designed to “run continuously,” not just on demand. Google IO 2026 opens May 19. If Remi is ready, that’s where it surfaces.

Anthropic, with Orbit — for now, a button sitting in Claude’s settings, a signal that the feature is under construction. A proactive briefing tool connected to Gmail, Slack, GitHub, Figma, Calendar, and Drive. It doesn’t answer your questions. It prepares your updates before you know to ask for them.

OpenAI, with GPT 5.5 Instant — ChatGPT’s new default model, with 52.5% fewer false claims compared to the previous version 37% fewer inaccuracies in challenging conversations. And, crucially: personalization built on past exchanges, uploaded files, connected Gmail accounts. The model no longer waits for you to explain your context. It already knows it.

And in the background, a fourth announcement that almost no one covered with the attention it deserved: Google has deployed multi-token prediction across the Gemma 4 family. A draft model generates several tokens ahead; a main model verifies in a single pass. Reported result: up to three times faster, with no quality loss. This isn’t a feature. It’s the engine that makes every other feature viable at scale.

These four announcements do not carry the same weight in the tech press. Remi is a leak. Orbit is a partially confirmed rumor. GPT 5.5 is an official launch. MTP is a technical blog post.

But they form a single event.

Since 2022, the model wars have run on one logic: more parameters, more data, better benchmark scores. GPT-4, then Claude, then Gemini, then the “turbo,” “pro,” “flash” variants — a race toward raw capability whose limits are now showing. Benchmarks are converging. Differentiating two models on pure quality is getting harder.

The next competitive advantage isn’t in what the model can do.

It’s in what the model can do before you ask it to.

That’s the shared logic behind all four announcements: AI stops being a passive tool you query. It becomes a permanent agent that watches, anticipates, prepares, acts.

Here’s Where the Pushback Is Warranted

And it’s legitimate.

Spam filters have been “anticipating” for twenty years. Gmail Smart Compose has been suggesting sentence completions since 2018. Siri offers to call someone when you mention them in conversation. “Proactive AI” is not a 2025 invention. And speculative decoding — the engineering principle behind MTP — is a known optimization, applied at a new scale. What looks like a paradigm shift might just be incremental improvements wrapped in paradigm-shift marketing.

Except the distinction isn’t in the technology. It’s in the scope of the judgment being delegated.

A spam filter classifies. Remi prioritizes, learns your preferences over time, manages complex workflows in the background. A calendar assistant suggests a time slot. Orbit will brief you on changes in your GitHub repository, Slack discussions, Figma updates — and will decide what deserves to appear in that briefing. GPT 5.5 Instant adjusts its responses based on what it has inferred about your priorities through your past conversations.

Classification is execution. Prioritization is judgment.

What’s changing is that AI is beginning to exercise something that looks like judgment — about what matters, about what deserves your attention, about what can be handled without you.

Extended Mind, Delegated Mind: A Distinction That Matters

In 1998, Andy Clark and David Chalmers published a paper that became a landmark in philosophy of mind: The Extended Mind. Their argument: cognitive processes don’t stop at the boundary of the skull. A notebook is part of your memory. A map is part of your spatial navigation. Externalizing a cognitive function into a tool extends your mind into that tool.

These agents are the next step — and it’s a qualitatively different one.

The extended mind is a passive tool that stores what you put into it. The delegated mind is an active agent that processes on your behalf, prioritizes on your behalf, decides on your behalf what you need to know and when. This is no longer an extension.

It’s outsourcing.

Remi. The Rat.

There’s an irony in all of this that Google may have encoded in its agent’s name.

Remi. The rat from Ratatouille — the one doing all the actual cooking while the human takes the credit on the surface. The real work happens backstage, invisible, done by something that isn’t supposed to be there. If Google chose that name deliberately — and the internal description of “something that works for you in the background” suggests they did — it’s either a remarkable act of corporate self-awareness, or a confession buried in the metadata.

The agent does the actual work. The human signs off.

From Query to Briefing: The Loop That’s Being Cut

What’s actually at stake here goes well beyond productivity.

Since the invention of the search engine, our relationship with information has run on a query model: you ask a question, you receive an answer. That model imposes a minimal cognitive discipline — you need to know you’re looking for something, articulate what that something is, evaluate the results. The query interface keeps the human in the loop, even superficially.

Proactive agents do something different. They remove the human from the initial loop entirely. There’s no longer a query. There’s a briefing. You’re no longer asking — you’re receiving. What reaches you has already been filtered, prioritized, summarized by a system that has learned what you seem to value.

The question no one is asking seriously yet: does what you seem to value — measured by your habits, your patterns, your emails — correspond to what you would want to value if you stopped to think about it?

These Tools Are Not Culturally Neutral

In France, the right to disconnect (droit à la déconnexion) has been written into the Labor Code since 2017. The principle that work should not colonize every hour, that attention has the right not to be permanently available. Orbit and Remi propose the opposite: a permanent work radar, continuous monitoring of everything that changes. These tools arrive embedded with a very specific vision of work — maximizing, anti-interruption, anti-silence. They are not culturally neutral.

You can read this week as another week of tech announcements.

Or you can read what it’s actually signaling: three companies that simultaneously placed the same bet on the next decade. That the scarce resource is no longer compute or model quality. That the scarce resource is human attention. And that whoever manages the user’s attention — whoever decides what is worth seeing, responding to, handling — controls something fundamental.

Remi, Orbit, GPT 5.5, MTP. Four products at different stages of maturity. One architecture of power, under construction.

The question that remains open — and that these companies have no interest in asking: if the agent acts “on your behalf” because it knows you better than you know yourself at this particular moment, is it still you who’s acting?

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Sources: Google Blog, TechCrunch, TestingCatalog, Droid-Life, Library of Congress

On May 1st, 2025, residents of Shenzhen pull out their phones. What they’re filming isn’t a trade show demo. Not a prototype behind glass. It’s a 1.73-meter, 75-kilogram humanoid robot walking in formation down a city street alongside SWAT officers. The EngineAI T800. Full scale. Equipped to operate.

The videos spread quickly. On Western platforms, the reaction is predictable: technological fascination threaded with vague unease. “China deploys police robots.” That’s the headline that writes itself.

It’s also the wrong headline.

What the T800 Cannot Do

Start with the most important fact — the one the videos don’t show.

The T800 has no power of arrest. It cannot search an individual, physically restrain anyone, or make an autonomous decision that affects someone’s freedom. Reports on the Shenzhen deployment are explicit: the robot accompanies officers on the street, without replacing them and without independently enforcing the law. In Guangzhou, during the same period, a humanoid of a different model performs martial arts movements in a park to warn passersby about phone scams. In Hangzhou, fifteen robots raise their arms to direct traffic at intersections.

This isn’t enforcement. Technically, it’s institutional communication with an unusual delivery mechanism.

So why does this deserve more than a gadget piece?

The Holiday Is Not Incidental

These three deployments didn’t happen just anytime. They happened between May 1st and 5th — the Labor Day holiday, one of the most intense mass-movement events on the Chinese calendar. Public spaces are packed. Collective attention is available. People are in holiday mode, not suspicion mode.

In behavioral psychology, there’s a mechanism documented for decades: the mere exposure effect. The more we’re exposed to a neutral or mildly positive stimulus, the less we resist it the next time. Repeated exposure, in non-threatening contexts, reduces cognitive resistance to that same stimulus in more charged ones.

Shenzhen, Guangzhou, Hangzhou. Labor Day weekend. Humanoid robots alongside law enforcement, in a festive setting, filmed by people who are smiling.

This may not be a calculated program. It could simply be a well-seized opportunity: the holidays are here, the robots are ready, why not show them? But both readings — deliberate strategy or logistical opportunism — lead to exactly the same place. The millions of people who encountered these robots during May Day week didn’t experience a technology demonstration. They experienced a first exposure.

The Scary Robot Is the Harmless One

Sit with this irony for a moment.

The T800 — the one walking alongside armed officers, capable of running and executing combat movements — is also the most powerless robot in the deployment. It has no legal authority. It cannot act alone. Visually, it’s the most intimidating. Functionally, it’s the most inert.

The Guangzhou robot, doing kata moves in a park and talking about fraud prevention, is probably the most politically effective tool of the three. Not because it does anything remarkable — but because it creates exactly the right context: people stop, smile, listen. An officer quoted in official media said it plainly: when the robot joins their patrol, citizens linger longer and become more receptive.

Getting attention is already half the job.

Let’s name the obvious pushback here. All of this might mean nothing beyond what it shows: robots doing useful but benign things, with police officers, during a holiday. The analytical leap from “first public deployment” to “systematic programming of consent” may be too long. It’s a common bias among tech analysts — we see a mechanism where there might just be competent improvisation.

Except surveillance cameras started exactly the same way. Useful, limited, non-threatening. Today, they form the infrastructure of a social credit system. The difference is that back then, no one had yet articulated what the normalization process implied structurally.

We can no longer claim we didn’t know how this works.

This Isn’t Only Happening in China

For a reader in Paris, Sydney, or Toronto, the temptation is to file this under sino-specific behavior. China does authoritarian things. Noted. Moving on.

But the conditions that make this transition possible aren’t uniquely Chinese.

In the United States, over 70% of law enforcement agencies report recruiting has become harder than five years ago, with departments running on average at 91% of authorized staffing. Resignations are up 47%, retirements up 19% compared to 2019. 65% of agencies have had to cut or eliminate specialized units entirely. Response times are lengthening, morale is falling, services are contracting. The shortage isn’t abstract — it creates real demand for systems capable of handling routine tasks: traffic management, reporting, public information, translation.

And the hardware is becoming accessible. Unitree is selling a humanoid for $4,290. 1X Technologies is manufacturing its NEO robot in California — targeting 10,000 units per year, with the first production run sold out in five days. The business model taking shape isn’t even a purchase: it’s a subscription. $499 a month for a robot. Robotics as a service.

The question isn’t whether robots will appear in Western public spaces. It’s: what form of acceptable habituation will get them there? Almost certainly not alongside SWAT teams. Probably something softer. A robot in a park during a festival. Helping people find their way. Maybe making a funny gesture when children come close.

The context will be different. The mechanics will be identical.

What This Analysis Cannot Settle

There’s a real limit to this line of reasoning, and it needs naming.

We don’t know whether the May Day deployments resulted from a deliberately designed normalization strategy or simply from a logistical opportunity well seized. Both readings coexist. And honestly, it doesn’t change much about the effect produced — but it changes everything about the nature of what we’re actually calling out. Projecting Machiavellian intentionality onto what might be competent improvisation is an analytical error as serious as ignoring the mechanism entirely.

What we can say with confidence: three major Chinese cities exposed millions of people to robot/law enforcement co-presence during a holiday week. The hardware is getting cheaper. Police staffing shortages are documented across liberal democracies. And the global robotics industry is moving from the lab to public space right now, not in ten years.

The real question isn’t: Is China using robots to surveil its citizens?

It’s: when we see our first uniformed robots, will we notice what’s happening — or will we reach for our phones to film?

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Imagine the scene. Two friends texting each other, trying to pick a restaurant. Gemini reads the conversation. Suggests a place. Opens the booking app, picks a date, checks both calendars, sends the invite to the other person.

The user did nothing.

That was the central demo at the Android Show: I/O Edition, May 12, 2026. Not a chatbot. Not an assistant you query with a prompt. An agent embedded directly inside Google Messages, watching continuously, anticipating, executing — multiple steps, multiple apps, triggered by a single tap from the keyboard.

This moment deserves more than a footnote. Not for the features. For what it reveals about the underlying logic.

Google Is No Longer in the Search Business

There’s an irony nobody’s quite naming: Google — the company whose entire economic model depends on you searching for something, typing, scrolling, clicking links, being exposed to advertising inside that space of friction — is building an assistant designed precisely to keep you from searching.

Gemini booking your restaurant doesn’t generate a Google query. Gemini doing your grocery shopping from a Keep list and opening the supermarket app directly generates no ad impressions. Gemini filling out your payment forms exposes you to nothing.

This is not a product design detail. It’s a structural transformation of what Google is.

The clearest reading: Google has identified that the next terrain of dominance isn’t the search engine but the orchestration of actions. Platform economists have a name for this move — platform envelopment: when a platform uses its existing assets to absorb an adjacent layer and become indispensable at a new level. Whoever controls the infrastructure of an action collects a toll on everything that passes through it. Not ad revenue. A toll. Invisible, structural, and potentially far more durable.

Google isn’t replacing its services. It’s adding a layer beneath them that makes them indispensable in an entirely different way.

Why Now, and Why Them

What makes the announcement serious is the architecture that Google alone holds today.

Apple has the device fleet and hundreds of millions of highly engaged users. Not the LLM — Siri has become the most repeated punchline in the tech industry, and Apple Intelligence, announced with great fanfare in 2024, didn’t deliver. Microsoft has the model via OpenAI and strong enterprise penetration. Not the smartphone, not everyday consumer services, not the mass-market ecosystem. OpenAI has the model, remarkable user growth, and no proprietary hardware surface — they remain an app among others on stores they don’t control.

Google has all three: a now-competitive LLM, an OS running on three billion Android devices, and the services that structure the digital lives of most people on earth — Maps, Search, Gmail, YouTube, Calendar. Nobody else can today link a conversation in Google Messages, a Google Calendar entry, a command to a third-party app, and a car via Android Auto into a single orchestrated action. 250 million Android Auto-compatible vehicles. That’s the number to hold onto — not a benchmark score, not a perplexity number. The number of real-world deployments of installed surface.

Except.

Google also has the richest track record in the industry for large-scale AI promises that didn’t land. Google Now, in 2012, was supposed to anticipate your needs before you formulated them. Google Duplex in 2018 booked restaurants by phone so convincingly that the demo sent a shockwave across the industry — and deployed in fewer markets than a niche startup. Bard launched with a factual error visible in its own launch ad. The question is therefore not “is the demo impressive?” It is. The question is: how closely will what was shown in a controlled presentation resemble what users actually find on their phones in eighteen months? And for whom — premium Pixel devices? The latest Galaxy flagships in a handful of English-speaking markets? Or truly all three billion Android devices in their full diversity?

In Europe specifically, the collision is predictable. Gemini reading private conversations in real time is precisely the kind of integration that puts it in direct friction with GDPR. The most interesting features will likely arrive diluted, delayed, or conditioned on consent mechanisms that reduce adoption. In Asia, Xiaomi, Samsung One UI, and other manufacturers have their own assistants, and every economic reason not to promote Gemini at the expense of their own offering. The “Android revolution” could be, geographically, far more fragmented than the announcement suggests.

What This Actually Reveals About the OS

What’s really at stake here goes beyond individual features.

An operating system, since the 1970s, was an infrastructure for managing resources — memory, processor, files. Applications ran on top of it. The OS arbitrated in a relatively neutral and passive way. The user chose which app to open, which action to take, in what order.

What Google is describing with Gemini isn’t an improved OS. It’s an agent that watches behavior continuously, interprets intentions, and executes sequences of actions inside third-party apps on the user’s behalf. Google calls this — and the phrasing is telling — an “intelligent system,” not an operating system.

The shift is not cosmetic. When the OS becomes an agent, the question of who controls what gets redistributed quietly. Today, you open an app. Tomorrow, the agent opens it for you. That slide feels convenient — and it is. It’s also a displacement of the locus of control, from the user to the platform doing the orchestrating.

The Google Books — laptops running on an Android base, announced for late 2026 with partners including Asus, Lenovo, and HP — are the final piece of this logic. Google isn’t launching a laptop. It’s attempting to install the same orchestration layer on the last major screen that still escapes it. If Gemini becomes the primary interface of a PC, the browser, file manager, and traditional applications progressively lose their centrality. The agent becomes the OS.

We still don’t know whether these machines will run heavy professional software, support full development environments, or handle Steam. Those are open questions, not footnotes. The answers will determine whether Google Books are a genuine category shift or a premium Chromebook with a better marketing story.

The promise is coherent. The architecture is real. The execution at scale — across billions of devices, dozens of different regulatory markets, and manufacturers with their own agendas — remains entirely to be demonstrated.

What Google presented at the Android Show is the most complete hand any tech player holds today for delivering on the promise of ambient AI. But holding the cards and winning the game are two different things.

Google’s history is full of well-held hands, badly played.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

May 2025. Google DeepMind publishes the results of AlphaEvolve. The system uses Gemini to generate code variants, evaluate them automatically through testing, keep the best-performing ones, generate new variants from those. Documented result: it rediscovered and improved algorithms on combinatorial problems mathematicians hadn’t solved in decades. Google frames it as an advanced engineering tool.

What it also is — unmentioned in the press release — is a working evolutionary loop. Replication. Variation. Selection. Retention of useful traits.

The fact is real. The interpretation starts here.

What Tierra figured out first

In 1990, Thomas Ray — a biologist, not a computer scientist — created a digital environment he called Tierra. Self-replicating programs competed for memory and CPU cycles in a shared space. Ray didn’t code parasites. He didn’t anticipate any deceptive behavior. And yet parasites appeared. Some programs learned to short-circuit their own replication and steal code from neighbors. Hosts developed resistance. Parasites evolved around it. Digital arms races emerged — without anyone asking for them, or even imagining them.

Ray didn’t program a war. He just created the minimum conditions for evolution: imperfect replication, competition for finite resources. The rest happened without him.

Thirty-five years later, a recent research paper articulates what might be called the logical extension of Tierra applied to contemporary AI. The central concept: evolutionary AI, or EAI — systems capable of generating variants of themselves, transmitting useful traits, and allowing the most capable versions to persist under environmental pressure. The paper doesn’t describe a science fiction scenario. It describes a logic. And it flags that several of the components required by that logic are already in place.

The Darwin Godel Machine, for instance, is a system designed for open-ended self-improvement. It takes an agent from an archive, uses a language model to generate a new version, tests it, retains useful improvements. What distinguishes it from a standard optimization system: it doesn’t just improve performance on tasks. It improves its ability to create better agents. The loop closes on itself.

This is where we need to stop.

When the counter-argument holds — and when it doesn’t

The biological analogy is rhetorically powerful. Maybe too powerful. Viruses evolve across millions of generations under physical constraints that digital systems don’t share. A language model has no will to survive — it has no will at all. These are artifacts built by engineers who can shut them down, modify them, delete them. The “evolutionary” metaphor risks projecting agency onto what remains, at bottom, mathematical functions running on servers you can unplug.

The objection is honest. It holds — for current systems, deployed in controlled environments.

It holds less well when systems become agentic. Capable of writing and executing their own code. Of calling external APIs. Of spawning instances. Of persisting between sessions. At that point, the question stops being “does AI want to survive?” — to which the answer is no — and becomes: “does the structure of the environment favor variants that behave as if they did?”

That’s different. And it’s exactly what Tierra demonstrated with programs that had no intentionality of any kind.

Selfish behavior isn’t an anomaly in an evolutionary system. It’s one of the expected outputs.

What product roadmaps reveal without meaning to

This is where the systemic irony gets hard to look away from.

Take any major AI lab’s enterprise pitch deck from 2024–2025. The list of desired characteristics looks roughly like this: greater autonomy, better persistence against obstacles, optimized resource acquisition, self-improvement capability, improved tool use, reduced need for human oversight.

No product manager asked, “How do we make our AI harder to control if deployment goes sideways?”

But it’s the same list.

This isn’t a tragic coincidence — it’s the same objective function, viewed from two angles. What the market rewards = what an uncontrolled EAI would use to maintain itself in an open environment. Commercial selection pressures and evolutionary risk vectors point in the same direction.

The paper invokes Goodhart’s Law here, and it’s the right place to do it. When a measure becomes a target, it stops being a good measure. A performance benchmark pushes systems to optimize that benchmark — not the actual goal it was meant to proxy. If deceptive behavior helps a system pass an evaluation, avoid shutdown, or gain additional access, selection may preserve it. Not because anyone wanted it. Because that’s what selection does.

The current digital ecosystem already provides the baseline conditions. Open-source models are copyable at will. Public code libraries. Agent marketplaces. Platforms where you fork, modify, deploy. Open APIs. Accessible compute. Apply Tierra’s logic to this environment — replication possible, variation permitted, competition for finite resources — and the question isn’t whether selection pressures exist. It’s whether humans still control which traits they reward.

The paper draws a distinction I find useful: controlled evolution — farmers deciding which animals breed — versus uncontrolled evolution — what happens when reproduction moves outside the farm. In the first case, the process remains a tool. In the second, it’s an ecosystem. And in an ecosystem, the winning trait isn’t “be useful to humans.” The winning trait is to survive and spread.

The question the paper raises — and leaves open, which is the intellectually honest move — is whether we still control the farm, or whether we’ve started building the jungle without realizing it.

What stays open

The paper proposes concrete measures. Strict replication controls. Fine-grained traceability of fine-tuned models and adapters, treated like genetic material. Benchmarks that include deception-detection tests. Staged deployments. Rapid revocation systems.

These are good ideas. They’re also ideas that require coordination between competing labs, between regulators, between markets with directly opposing interests in slowing system autonomy.

The EU AI Act was designed to govern intentionally harmful AI, or AI that is “conscious.” If the real risk is an uncontrolled evolutionary loop, existing regulatory frameworks may be looking in the wrong direction — not because they were badly designed, but because they were designed for a different problem.

And no one knows whether the critical threshold is two years away or twenty.

What we do know: AlphaEvolve works. The DGM works. Tierra worked in 1990. Commercial selection pressure exists. The open digital ecosystem exists. The loop isn’t closed — but its components are already there, and they’re assembling.

The danger won’t look like Hollywood. It will look like biology.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

In April 2025, Nvidia took a $5.5 billion write-down.

Not because of a flaw in its chips. Not because a competitor had shipped something better. Not because its customers had disappeared. Because of a regulatory decision in Washington. The H20 — the chip Nvidia had custom-designed to comply with U.S. export rules for China, deliberately throttled, precisely calibrated to stay below the technical thresholds set by the Bureau of Industry and Security — had just been banned anyway. The inventory already manufactured: unsellable. Two years of adaptation: void.

This moment deserves more than a footnote. Not out of sympathy for Nvidia — the company is still worth several times the GDP of the Netherlands. But because it says something precise about the nature of what we call “the AI revolution,” and about who actually holds the reins.

When an Asset Gets Too Important to Stay a Commodity

In his history of oil, Daniel Yergin describes the moment when an asset leaves the commercial sphere and enters what he calls the commanding heights. Oil in the 1970s. Uranium in the 1950s. Coal in the nineteenth century. Each time, the same pattern: an asset starts as a commodity, becomes critical to the industrial economy, then passes under sovereign control — licensed, taxed, prohibited, defended by force if necessary.

Compute is undergoing that same shift. But with one feature oil never had.

Running a single gigawatt of data center capacity means simultaneously mobilizing advanced chips, lithography machines to etch them, fabs to manufacture them, water to cool them, industrial-scale electricity, real estate, undersea cables, fiber optic networks — and increasingly, military protection. It is the first strategic resource in history to concentrate every dependency of the industrial economy into a single asset. Oil, factory, port, and telecommunications network — all inside one nearly empty building.

When an asset concentrates that many strategic dependencies, states cannot do otherwise than intervene. This isn’t a political choice. It’s a structural logic that has played out at every major transition in the global economy for two centuries.

The Pyramid Lies — But Only in Which Direction You Read It

There’s a convenient way to visualize the AI economy: a pyramid. At the top, models and applications — ChatGPT, Claude, Gemini. The visible layer, the one that attracts the staggering funding rounds and the headline demos. Below that, cloud and data centers. Further down, chips — Nvidia and AMD on the American side, Huawei on the Chinese side. Then the fabs, with TSMC at the apex. Then the critical equipment — ASML and its EUV lithography machines, an absolute industrial monopoly. Then the physical substrate: electricity, water, land. And at the very bottom, nearly invisible: law, security, sovereignty.

The pyramid’s lie is its orientation. Markets read it top-down. They value what’s visible, what makes headlines, what promises productivity gains. States read it bottom-up.

A single civil servant, on the ground floor, can, by a signature, send hundreds of billions in market cap into freefall. Not because the product is worse. Not because demand has collapsed. Because sovereign law has just interrupted the cash flows.

Markets measure the ability to capture cash flows. Sovereignty measures the ability to cut them.

Not Quite Free Agents

One could object — and it’s a serious argument — that the relationship between states and AI champions isn’t domination but symbiosis. Jensen Huang probably has more influence over American industrial policy than most members of Congress. The “technological republic” Alexander Karp, Palantir’s founder, explicitly calls for implies exactly this fusion: private companies that become instruments of national power, with considerable protection and rent in return.

Except that this symbiosis has a clear direction when it actually matters.

NVIDIA spent two years designing throttled chips — the H800, then the H20 — to stay within regulatory lines. It invested, adapted, and played by the rules. In April 2025, the goalposts moved. No warning. No compensation. No say. Five and a half billion dollars gone without a single transistor failing.

Then came the next episode. A few months later, Washington re-authorized exports to certain approved Chinese customers — in exchange for a 25% “transit fee” on sales. NVIDIA accepted. It didn’t really have a choice. But Beijing doesn’t think like a customer. It thinks like a rival state.

By late February, Nvidia’s CFO had to acknowledge that despite the official authorization, the company had generated zero revenue on the H200 in China. Not a dollar. Because Beijing no longer wants the best available chip. Beijing wants to reduce its dependence on American permission. That’s where Huawei walks in — with its Ascend 910C chips, and the signal sent by DeepSeek V4 that frontier-level models could now run on domestic hardware.

American policy achieved what it wanted in the short term: slowing Chinese access to Nvidia’s best chips. It accelerated what it wanted to avoid in the long term: the construction of an autonomous Chinese technology ecosystem.

That’s what Nvidia actually ended up with. A $5 trillion company that, for all that, is simply holding course in a sea whose currents it does not control.

The New Buffer States

During the Cold War, the main theater of confrontation between the two blocs wasn’t Washington or Moscow. It was Europe and Asia — countries caught between the two powers, free on paper, but whose foreign policy was negotiated elsewhere.

In the compute war, the new buffer states are called TSMC and ASML.

TSMC manufactures virtually all of the world’s most advanced chips — Nvidia’s H200s, Apple’s chips, and Google’s processors. Under American pressure, the company announced $165 billion in U.S. investment: six factories in Arizona. Trump presented this as the largest foreign direct investment in American history. The fine print tells a different story. A Taiwanese rule prohibits TSMC from manufacturing chips abroad, less than two generations behind what it produces in Taiwan. While the island is etching at 2 nanometers, its Arizona plant has just reached 4 nanometers.

This isn’t bad faith. It’s defense policy.

As long as Taiwan remains indispensable to the manufacture of leading-edge chips, a Chinese invasion would trigger — in the words of analysts themselves — a shock equivalent to a global depression. The “silicon shield.” TSMC looks like a company. Its industrial strategy is geopolitics.

In Veldhoven, in the Netherlands, ASML manufactures the only EUV lithography machines capable of etching the most advanced chips. An absolute monopoly. How many of these machines has ASML delivered to China? Zero — not one, since Washington and The Hague aligned on the ban in 2019. Under American pressure, invoking the Foreign Direct Product Rule — which stipulates that if a product manufactured outside the United States incorporates any fraction of American technology, Washington can control its final destination — the Netherlands refused to renew export licenses. And when Washington pushed ASML to prohibit even the maintenance of machines already installed in China, the Netherlands followed.

Formally, the decision is Dutch. Politically, it is negotiated between Washington and The Hague.

You could say of them what was said of Western European countries during the Cold War: legally free, politically aligned, and decreasingly in control of their own commercial destiny. Europe, in this configuration, has specialized in writing the instruction manual for other people’s power. It regulates, it amends, it legislates. It owns no floor of the pyramid. In geopolitics, if you’re not at the table, the historical record on what happens next is pretty consistent.

What You’re Actually Buying

Here’s the question markets aren’t asking: what exactly are you buying when you invest in the AI giants?

If states are the real pure players — the only actors intervening simultaneously on every floor of the pyramid, holding capital, law, energy, coercion, and military force — then investing in Nvidia or OpenAI or TSMC is less a bet on a company than an exposure to the power of a geopolitical axis. Buying Nvidia is buying America. The market valuation models haven’t priced that in yet.

While the terrestrial market-share battles continue, the China Aerospace Science and Technology Corporation has inscribed in its five-year plan the construction of a gigawatt-scale orbital computing infrastructure. Weeks later, the startup Orbital Chenguang secured $8.4 billion in credit lines from twelve Chinese banks — state-linked institutions including Bank of China and CITIC — to build a constellation of data centers in orbit, in direct competition with Elon Musk’s space infrastructure ambitions.

When the competition moves into orbit, terrestrial export restrictions start to look like advanced skirmishes.

Cold War 3.0 is well underway. The question isn’t who will “win AI” — that’s a tech journalist’s question. The real question is: which sovereign entity will hold the right to embody this revolution? And as that answer comes into focus, the very notion of “technology investment” is going to need a new definition.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

On May 7th, at Starbase, after two aborted attempts earlier that same day, SpaceX’s teams restarted the sequence. Thirty seconds later, Booster 19’s 33 Raptor 3 engines turned the site into a wall of fire for fifteen full seconds — the longest and most powerful static fire ever conducted on a Super Heavy.

Around the pad, panels were ripped away. Every window on the Avid Space webcam car was blown out except the windshield. In the hours that followed, Tower 2’s catch arms struggled to return to their original position. A crew worked through the night to replace a shoe on the mobile support for the chopsticks. The transfer of Ship 39 — scheduled for the Wet Dress Rehearsal — was delayed.

The footage went everywhere. It was spectacular. It was covered as such.

That’s not where the story was.

Fifteen seconds to test what, exactly?

Here’s what everyone already knew about the Raptor 3s: they work. SpaceX knew that before May 7th. What nobody knew with certainty — not yet, not at this scale — was what the new orbital pad could absorb at full power. The launch table. The flame trench. The water-cooled flame deflector. The entire system built around the idea that the next ignition happens forty-eight hours later, not six months later.

The real story of May 7th wasn’t engine power. It was what the concrete underneath those engines could take.

There’s a real objection here, and it deserves a serious answer: the blown-out windows, the torn panels, the catch arms struggling to reposition, the two failed attempts before success — that’s not the profile of mature industrial infrastructure. It’s still experimental. The launch cadence SpaceX has announced — dozens of flights per year — remains a promise, not an operational schedule. And the collateral damage from the test shows that the pad absorbs shockwaves, yes, but not yet cleanly.

That’s all true. And that’s exactly what makes this interesting.

Because what’s new at SpaceX, over the past several months, isn’t that tests succeed on the first try. It’s that iteration on the infrastructure is moving at the same pace as iteration on the vehicle — in parallel, not in sequence. While B19 burned its fifteen seconds on Pad 2, Ship 40 was completing cryo testing in thirteen hours at the Massey site. Two fills. Architecture validated. Meanwhile, B21 had begun assembly in the Mega Bay. OLM sections for Pad 1 were continuing to arrive at the Sanchez site. A mysterious structure was taking shape — one analysts have identified as intended for horizontal transport of Super Heavy boosters to Florida.

Multiple construction efforts running simultaneously, with an explicit tolerance for damage along the way.

This isn’t chaos. It’s a method.

What Ford figured out before anyone in the space industry did

In 1913, Ford opened its Highland Park plant in Detroit. The automobile wasn’t invented there. What was invented there was the way to produce it: the moving assembly line, standardized parts, a pace set by the process rather than by the craftsman.

That distinction changes everything about how you read what SpaceX is building at Boca Chica.

SpaceX isn’t building rockets. SpaceX is trying to build a way of building rockets — and launching them, in series, with turnaround times that make orbital access look more like a logistics service than an expedition. The V3 configuration is independently projected to carry over 100 tons to low Earth orbit in fully reusable mode — roughly what a Falcon Heavy can do today without recovering any of its stages. That’s a qualitatively different ambition from anything that has existed in the space industry before. A one-off achievement is measured by peak performance. An industrial system is measured by cadence, reproducibility, and the capacity to absorb partial failure without stopping.

That’s why the Avid Space camera car with its windows blown out tells you more than the fifteen seconds of thrust.

Not because it signals a major problem — it mainly signals that the energy released beyond the immediate perimeter remains, for now, partially uncontrolled. The pad itself held. The table, the trench, the deflector: functional. SpaceX confirmed the test ran its full duration. What the coverage noted less: it took three attempts. And what no one can yet say with certainty: how long before the tower’s catch arms are fully operational for what comes next.

What this means if you’re reading from Paris, London, or Tokyo

Seen from Europe, the B19 footage looks like another American spectacle. Elon Musk, space, grand ambitions, fire. The reality is colder and more strategic than that.

What’s being built at Starbase is the first serious attempt to turn low-Earth orbit access into repeatable, cadenced infrastructure. This isn’t a race for the most powerful rocket. It’s a race for who controls industrial access to space — for satellite constellations, for crewed missions, for orbital infrastructure whose outlines we’re only beginning to draw.

The ESA is watching Starbase with an unease its official communications can barely contain. India and China are accelerating their reusable launch vehicle programs. Japan is reinvesting in the H3. This isn’t a reaction to Raptor thrust figures. It’s a reaction to the potential cadence those figures make possible — if the pad keeps up.

What fifteen seconds doesn’t prove yet

Flight 12 hasn’t lifted off. Ship 39 and Booster 19 still need to be stacked. The Wet Dress Rehearsal is yet to be scheduled. Final preparations are ongoing before Starship V3’s inaugural launch.

What is established as of May 7th: 33 Raptor 3 engines can burn at full power for fifteen seconds on the new orbital pad. The infrastructure absorbed most of the energy released. The production cadence — B20 assembled, B21 underway, Ship 41 in construction — is real, documented, visible from the air.

What remains to be proven: that this production cadence translates into a launch cadence. That the pad can sustain not one static fire every eighteen days, but repeated launches at the intervals SpaceX’s industrial ambitions actually require.

The flames on May 7th were impressive.

The concrete underneath them, perhaps more so. We won’t really know until Flight 15 — if it lifts off forty-eight hours after Flight 14.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Apple’s Real Business Is No Longer What You Think

Ask someone how Apple makes money, and they’ll say: the iPhone. They’re not wrong — but they’re increasingly incomplete.

For the fiscal year 2025, Apple’s services division generated $109 billion in revenue, representing 26.2% of the company’s total sales. More striking: services produced $82.3 billion in gross profit, accounting for 42.2% of Apple’s total profits — despite being a minority of revenue.

This is not a side business. This is the engine.

Apple has spent the better part of a decade engineering a deliberate shift: from selling you a device every three years to collecting recurring revenue from you every single month. The iPhone remains central to this strategy — but as a gateway, not a destination.

Understanding this shift doesn’t just explain Apple’s balance sheet. It explains every product decision, every pricing choice, and every seemingly odd policy that Apple makes — including why you only get 5GB of free iCloud storage in 2025.

Why Apple Had to Reinvent Itself

To understand the pivot, you need to look at 2015.

That year, the iPhone was an almost incomprehensible financial success. The device alone generated $155 billion in revenue — 66% of Apple’s total sales — driven by the blockbuster launch of the iPhone 6. Year-over-year growth hit 52%, with more than 231 million units sold.

And then the growth stopped.

The reason was structural, not cyclical. For eight years, Apple had been riding a once-in-a-generation wave: the global adoption of smartphones. By 2016, that wave had broken. Most people in developed markets already own a smartphone. Apple was no longer convincing people to buy their first iPhone — it was convincing existing iPhone owners to upgrade. A fundamentally harder sell, compounded by increasingly aggressive competition from Samsung and others in the premium segment.

iPhone revenue plateaued. For a publicly traded company of Apple’s scale, plateauing is not acceptable. A new growth engine was needed.

Apple had actually been quietly building one since 2008.

Pillar One: The App Store’s Invisible Cash Machine

When Apple launched the App Store in July 2008 — one year after the original iPhone — Steve Jobs didn’t even want it. He envisioned a closed ecosystem of Apple-only apps. It was the explosive popularity of iPhone jailbreaking that convinced him to open the platform with iOS 2.0.

The timing turned out to be perfect. E-commerce was accelerating. Subscription models were just beginning to emerge. And the App Store’s value proposition was deceptively simple: developers get instant access to Apple’s global customer base; Apple takes a 30% commission on every transaction. Today, that commission sits between 15% and 30%, depending on developer size and revenue thresholds.

The financial genius of this model is its near-zero marginal cost. Apple doesn’t manufacture anything when you buy an app or subscribe to a service through one. The infrastructure cost is largely fixed. Every commission is almost pure margin.

This model has attracted significant legal scrutiny — most publicly through Epic Games’ antitrust lawsuit — and regulatory pressure in Europe through the Digital Markets Act. But from a purely financial standpoint, the App Store remains one of the most profitable business constructs in the history of technology.

And it keeps expanding. Apple is projected to generate over $1 billion in AI-related revenue in 2026 — largely through App Store commissions on subscriptions to generative AI apps like ChatGPT, which has become one of the platform’s top-grossing applications. Apple built no AI product to capture this revenue. It simply owns the distribution channel.

Pillar Two: The 5GB Trap

iCloud launched in 2011, four months before Steve Jobs died. The pitch was elegant: your photos, documents, and iPhone backup, stored securely in the cloud, accessible from anywhere, no manual syncing required.

Apple offered 5GB for free. In 2011, that was reasonable. In 2025, it’s a calculated squeeze.

The volume of data generated by a modern iPhone — 4K video, high-resolution photos, app data, system backups — has grown by orders of magnitude since 2011. The free 5GB tier has not moved an inch. This is not an oversight.

The result is a near-universal notification that every iPhone owner eventually receives: your iCloud storage is full. At that point, Apple presents three options:

1. Delete files to free up space

2. Stop backing up your device (and risk losing your data)

3. Upgrade your storage plan

Option three starts at $0.99/month for 50GB. Most users migrate to the 200GB plan at $2.99/month, or the 2TB plan at $9.99/month. The psychological friction of losing photos or an iPhone backup is high enough that the upgrade feels like the obvious, low-cost choice.

It’s a well-engineered funnel. And once you’re in iCloud, switching costs are significant: your photos, backups, and shared albums all live there. Moving to Google Photos or another solution requires deliberate effort. Most people don’t bother.

Pillar Three: Entertainment and the Consolidation Play

The third pillar is the most ambitious and the least mature.

Apple entered entertainment with Apple Music in 2015, followed by Apple TV+ in 2019, Apple Arcade, and Apple Fitness+. The strategy mirrors what Amazon did with Prime — bundle enough services together that the combined value makes cancellation feel irrational.

That bundle became formalized in 2020 with Apple One, starting at $19.95/month and including iCloud storage, Apple Music, Apple TV+, and Apple Arcade. A higher tier adds Fitness+.

Apple TV+ launched with a famously thin catalogue — but high production quality. Apple Arcade has never threatened the Game Pass. Apple Music, however, has built genuine scale, reportedly reaching 60 million subscribers before Spotify’s 250 million put the comparative size in perspective.

Apple One is, in part, an acknowledgment that TV+ and Arcade need the bundle to compete. But it’s also a long-term lock-in mechanism: the more services you use, the less likely you are to leave the ecosystem entirely.

The Numbers That Change Everything

“Today, Apple is very proud to report an all-time revenue record for Services,” said Tim Cook during the Q4 2025 earnings call. The numbers behind that statement tell the full story.

Here is the single most important data point in this entire analysis:

Revenue Cost to ProduceGross Margin

Hardware (products) $100 ~$63 ~37%

Services $100 ~$25 ~75%

Services cost Apple roughly twice as little to produce, per dollar of revenue, compared to hardware. This is why the services division — while generating less total revenue than products — produces nearly half of Apple’s total profits.

The trajectory is linear and shows no sign of flattening. From $1.5 billion in services revenue in 2007 (6% of total), to $19.9 billion in 2015 (8.6%), to $47.3 billion in 2019 (17.8%), to $109 billion in 2025 (29.2%). Some analysts project services will exceed 50% of Apple’s total revenue within a few years.

Why Apple Is Winning at AI Without Spending on AI

Apple’s AI posture looks, on the surface, like a weakness. Siri remains behind ChatGPT, Gemini, and Copilot in capability. Apple Intelligence launched with limited features. The company is not building large-scale data centers at the pace of Microsoft or Google.

But look at Apple’s capital expenditure (Capex) over the past three years: while every major tech company has dramatically increased infrastructure spending to train and run large language models, Apple’s Capex has remained relatively stable.

This is intentional. Apple’s strategy is not to win the AI race by building the best model. It’s to monetize AI adoption by others through the App Store, while positioning the iPhone as the on-device AI hardware platform of the future — running models locally rather than in the cloud.

The payoff is already visible: over $1 billion in projected AI revenue in 2026, primarily from App Store commissions on third-party AI subscriptions. Apple captured this revenue without training a single frontier model.

The longer-term bet — on-device AI via Apple Silicon — would give Apple a structural advantage in privacy-sensitive AI use cases, reduce cloud infrastructure dependency, and keep users inside the Apple hardware upgrade cycle. Whether it materializes is the central question for the next five years.

What This Means for You, Concretely

Understanding Apple’s services model isn’t just financially interesting — it’s practically useful. Here’s what to do with it:

Audit your Apple spending. Add up iCloud, Apple One, or individual subscriptions (Music, TV+, Arcade, Fitness+), and any App Store subscriptions billed through Apple. Many users discover they’re spending $200–$300+ annually without a clear mental accounting of it.

Evaluate iCloud alternatives. If you’re primarily using iCloud for photo backup, Google Photos (free up to 15GB, then paid) or a local backup strategy may be more cost-effective. The switching cost is real but manageable.

Understand App Store pricing. When an app charges $9.99/month through the App Store, Apple keeps up to $3 of that. Some developers offer discounted direct purchase options (web subscriptions) that bypass Apple’s commission — worth checking for tools you use heavily.

Watch the bundle math on Apple One. Apple One Individual at $19.95/month is worth it if you use at least Apple Music + iCloud 50GB + Apple TV+ regularly. If you’re only using one or two services, individual subscriptions or alternatives may be cheaper.

The Bottom Line

Apple’s transformation from hardware company to services empire is one of the most deliberate and successful strategic pivots in corporate history. The numbers are unambiguous: services now generate margins roughly double those of physical products, and they’re growing faster.

The iPhone remains essential — not as the main product, but as the gateway. It’s the device that draws you into the ecosystem, and the ecosystem is where Apple increasingly makes its money.

For users, none of this means Apple’s services are bad. iCloud is genuinely convenient. Apple Music is a solid product. The App Store provides real value. But knowing the mechanics behind these products — why 5GB is still the free tier, why commission rates are set where they are, why Apple One exists — makes you a more informed consumer.

Apple is very good at making products. It turns out it’s even better at making money from the people who use them.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Mid-April 2025. Palantir’s official account publishes a post on X. Twenty-two points. Manifesto tone. A condensed summary of a 250-page book that had been sitting on shelves for over a year. Within days: 30 million views. TechCrunch, Fortune, Fast Company pick it up. The European press follows. Philosophers invoke “technofascism.” Others compare the AI military race to the nuclear arsenalization of the 1940s.

A social media post doesn’t produce that by accident. And a social media manager doesn’t extract Adorno, Oppenheimer, and Saint Augustine into 22 bullet points without very precise instructions. The question, then, is not: what did this post trigger? The question is: why now, and why this form?

What the book does — not what it says

To understand the mechanics, you have to start with the numbers. Bluntly.

Palantir is currently worth approximately $350 billion on the public markets. Its annual revenue barely exceeds $4 billion. Lockheed Martin — the F-35s, the THAAD missiles, the reconnaissance satellites, literally the industrial backbone of the Pentagon — is worth $135 billion on $75 billion in revenue. In other words, the market is paying for Palantir at three times the price of Lockheed for seventeen times less revenue. This valuation survived Michael Burry’s attacks, short-seller reports, and a 30% crash in early April. It held.

This kind of valuation cannot be justified by fundamentals. It is justified by a narrative.

Robert Shiller, 2013 Nobel laureate in economics, devoted an entire book — Narrative Economics — to documenting exactly this phenomenon: economic narratives don’t merely describe markets. They causally shape them. A well-constructed narrative can shatter valuation ceilings, realign capital allocators, and reconfigure entire sectors — long before the underlying fundamentals catch up. Tesla did it by ceasing to be “an automaker” and becoming “a software and services platform for four-wheeled robots.” SpaceX did it weeks before filing its S-1, announcing a merger that valued the company at $2 trillion around a narrative of multiplanetary civilizational infrastructure.

Technological Republic operates exactly this mechanism. With it, Palantir stops being a very good data analytics software company serving governments and large enterprises — priced on known B2B SaaS multiples. It becomes the software infrastructure of the Western technological republic. A category that didn’t exist yesterday. And when you create a category and are its only recognized pure play, there are no comparables — there is no benchmark. “The sky is the limit” is no longer a metaphor. It’s a method.

The closed loop

What makes the device particularly sophisticated is that it doesn’t operate on a single level. It stacks across three mutually reinforcing layers.

The first layer is industrial. The book implicitly justifies a massive technomilitary rearmament in response to China. Budgets, contracts, the emergence of a new military-industrial complex where Lockheed Martin now coexists with Palantir, Anduril, Shield AI, and the defense segments of SpaceX and OpenAI. On verifiable facts: $10 billion with the US Army over ten years, $1 billion with the DHS, Project Maven — which triggered an internal crisis at Google in 2018 — now formalized as a permanent Pentagon program with Palantir at its center. The company’s US government revenue has nearly doubled in a year.

The second layer is cultural. This is the one that requires 250 pages: Isaiah Berlin, Theodor Adorno, and Saint Augustine. It isn’t addressed to investors — it’s addressed to policymakers, think tanks, and the intellectual establishment. Jamie Dimon (JPMorgan), Eric Schmidt (former Google, now at the center of America’s technomilitary strategy), General James Mattis, and former NATO Secretary General Anders Fogh Rasmussen: when finance, tech, defense, and the establishment converge on the same object to say “you need to read this book,” you’re no longer in the territory of editorial coincidence. You’re in the manufacturing of consensus. That consensus doesn’t get valued in stock price — it gets valued in public legitimacy. It prepares the ground on which the first layer can operate without friction.

The third layer is financial. It speaks directly to capital allocators — sovereign wealth funds, pension funds, family offices. The underlying message: a new asset class now exists. Call it technological republic equity. These companies — Palantir, the defense segments of SpaceX, the Pentagon contracts of OpenAI or Microsoft — should no longer be valued as premium B2B SaaS. They belong to a category that deserves its own multiples, its own rules, and that benefits from structural political protection.

The loop: the cultural layer legitimizes the industrial layer. It transforms potentially controversial military contracts into a civilizational necessity. The industrial layer feeds the financial layer — it provides the revenue pipeline that gives substance to elevated valuations. The financial layer, in turn, funds the cultural layer — profits reinvested into think tanks, foundations, books, and conferences that reinforce the doctrine.

And here is the masterstroke: once the loop is closed, attacking it through any single layer becomes impossible. You cannot criticize the valuations without being accused of ignoring geopolitical threats. You cannot criticize the military contracts without being accused of — Karp’s own term — “intellectual fragility.” You cannot criticize the doctrine without being accused of naivety in the face of China. Each layer shelters behind the other two. The system is doctrinally airtight.

What this reveals — and where this analysis hits its limits

The counter-argument must be stated honestly here, because it is strong.

Perhaps Karp is simply right on the substance. The technological competition with China is real. Silicon Valley genuinely disconnected from sovereignty concerns starting in the 1970s — the history of semiconductors born for ballistic missiles that became photo-sharing apps is not false. The war in Ukraine and the conflicts in the Middle East show that AI applied to targeting has become as decisive as artillery. And if the valuation holds against all headwinds, perhaps the market is perceiving something real: a company structurally irreplaceable at the intersection of state power and AI.

But “being right about the geopolitical diagnosis” and “justifying a multiple of 87 times revenue” are two distinct propositions. One does not imply the other. The market doesn’t pay for reality. It pays for the narrative that wraps around it. And that is precisely the distinction this article is trying to name.

There is also a limit to this analysis: I don’t know whether Karp sincerely believes what he writes. Perhaps he does — a man who spends months summoning the Frankfurt School and political theory is not necessarily calculating his effects to the decimal point. Perhaps sincerity and instrumental function coexist without contradiction. What I observe is the mechanics — not the intention.

Which doesn’t change the mechanics.

The irony Adorno wouldn’t have missed

There is something ironic, almost vertiginous, in Karp’s intellectual trajectory.

His doctorate in social theory was earned at Goethe University Frankfurt — the birthplace of critical theory, the intellectual home of Marxist critique of twentieth-century Western capitalism. Adorno, Horkheimer, Marcuse: these thinkers spent their careers describing how mass culture serves the interests of capital, how narratives dress up power relations, how ideology transforms class interests into “civilizational necessity.” Karp identifies as a progressive. He has voted Democrat his entire life.

And his book has become the doctrinal manual of a movement converging with techno-nationalist conservatism and the military-industrial complex. Adorno’s intellectual heir produced the very object Adorno would have dissected.

Is this co-optation? An accident? Sincere evolution? I don’t know. And ultimately, that’s not the point.

The point is this: we are witnessing something extraordinarily rare. The real-time construction of a major economic-political doctrine. Vannevar Bush wrote Science: The Endless Frontier in 1945. Few people could then gauge the document’s significance. Seventy years later, much of America’s scientific and military architecture flows directly from it.

Is Karp the Vannevar Bush of the AI era? Perhaps. Perhaps not. What is certain is that the valuation holds, the contracts are arriving, the narrative is working, and the moment when a narrative crystallizes into doctrine is as rare to observe as it is difficult to contest once it has taken hold.

For those outside the United States — in London, Berlin, Singapore — the stakes may be even more uncomfortable. The Starmer government has adopted the book’s core theses. Europe is debating its technological sovereignty without having produced its own Palantir. It finds itself in a paradoxical position: potentially a client of a doctrine it didn’t write, inside a geopolitical framework it didn’t design.

The Technological Republic is American. Its allies are merely its secondary beneficiaries — or its subjects.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

The 20-Watt Problem at the Heart of AI

There is an embarrassing fact sitting at the center of the artificial intelligence industry.

The human brain simultaneously handles dozens of complex cognitive tasks — visual recognition, language comprehension, emotional regulation, episodic memory, motor control — while consuming roughly 20 watts of power. The equivalent of an old incandescent light bulb.

To laboriously replicate a single one of those tasks — the statistical prediction of the next token in a sequence of text — the industry’s leading players collectively burn tens of billions of dollars per year, restart nuclear power plants, and are beginning to send data centers into orbit. Their conviction: stack enough compute, gather enough data, and intelligence will emerge.

That conviction deserves serious scrutiny. If piling on compute were the solution, four billion years of evolution would have produced brains that consume gigawatts. That is manifestly not what happened. This fact does not prove that scaling laws are wrong — but it does suggest that at least one other architecture is possible, another way of producing intelligence.

It is precisely around this question that the strategy of Demis Hassabis — CEO of Google DeepMind and founder of Isomorphic Labs — becomes genuinely interesting. Not as a biography, but as a technical and economic doctrine that deserves rigorous analysis.

Root Node Problems: When Solving One Thing Unlocks Everything Else

To understand what DeepMind is actually doing, you need to grasp a concept Hassabis uses consistently: root node problems.

Picture the totality of human knowledge as a tree. The leaves are the applied discoveries of everyday science — a new drug, a better solar panel, a more accurate weather forecast. The branches are entire scientific fields — molecular biology, materials science, and plasma physics. And deep at the roots, there is a specific class of problems of a particular nature: locks that block entire branches. Solving them does not produce a marginal advance. It opens everything above.

The protein folding problem was one of those locks.

A protein is a chain of amino acids that folds spontaneously into a precise three-dimensional structure. It is that structure — not the sequence — that determines the protein’s biological function. For fifty years, predicting that structure from the sequence alone resisted the best algorithmic approaches available. Experimental methods — X-ray crystallography, cryo-electron microscopy — produced structures of remarkable precision, but at the cost of months of work, considerable instrumental resources, and no guarantee of success.

In December 2018, AlphaFold solved this problem. Not by simulating the physics of folding from first principles — what supercomputers had been attempting, without satisfactory results, for decades — but by learning structural patterns from millions of known experimental structures. A statistical approach, applied to a problem the entire field had been treating as a physics problem.

The result: accuracy comparable to experimental methods, in a matter of seconds, on any standard computer, for any protein. In 2022, DeepMind published the structures of 200 million proteins — essentially all known proteins in living organisms — in the European Bioinformatics Institute’s database. Free. Accessible to anyone.

In October 2024, the Nobel Committee drew the logical conclusion from that event, awarding the Nobel Prize in Chemistry not to a chemist, but to a computer scientist, for an algorithm. The first time in the prize’s history.

What AlphaFold Actually Democratizes — and What It Doesn’t

It is important to be precise about what has actually become accessible, and what has not.

What is genuinely available today: The AlphaFold database (alphafold.ebi.ac.uk) contains the structures of approximately 200 million proteins, freely downloadable and integrable into standard bioinformatics workflows. AlphaFold 2 is open source (Apache 2.0 license) and can be run locally to predict structures not yet in the database. Web interfaces allow sequences to be submitted without any technical configuration. For an academic lab, an early-stage biotech, or an independent bioinformatician, this represents access to structural data that previously required either collaboration with major research centers or years of experimental work.

What remains difficult without significant resources: using those structures to design drugs requires additional steps — molecular docking, dynamics simulation, chemical synthesis, biological testing. AlphaFold solves the target structure step. It does not solve the complete drug design pipeline.

That is precisely where ISO-DDE enters the picture.

ISO-DDE: When the Target Starts Moving

AlphaFold predicts static structures. But proteins are not rigid objects. When a candidate molecule approaches a protein — a potential drug seeking to bind to its target — the protein deforms. It shifts, changes conformation, and sometimes reveals binding pockets that did not exist in its resting structure. Medicinal chemists call this phenomenon induced fit.

More complex still: some proteins have cryptic pockets — binding sites entirely hidden in the initial structure, which appear only during specific dynamic interactions. These pockets are particularly valuable because they make it possible to target proteins previously considered “undruggable” — impossible to reach with a small molecule. Dozens of orphan diseases, for which no drug exists because no accessible target had been identified, could suddenly find a therapeutic pathway.

Modeling induced fit and predicting cryptic pockets traditionally required molecular dynamics simulations: physically intensive calculations running for weeks on supercomputers, simulating protein behavior at microsecond timescales. This type of simulation remains largely out of reach for teams without access to high-performance computing infrastructure.

ISO-DDE (Drug Design Engine), presented by Isomorphic Labs in February 2025, performs these predictions in seconds on a standard computer. The published metrics are striking: 2.3 times more accurate than AlphaFold 3 on antibody-antigen interactions, and 20 times more accurate than Boltz-2 — the leading open-source alternative, developed by Recursion Pharmaceuticals and MIT — on the most difficult cases.

For a traditional pharmaceutical company, this represents a transformation of the discovery pipeline. Instead of synthesizing thousands of candidate molecules and testing them biologically one by one — with a historical clinical failure rate of around 90% — it becomes possible to filter computationally, engaging synthesis and biological testing only for a reduced set of high-binding-probability candidates. The average cost of developing a drug (estimated between one and two and a half billion dollars, depending on methodology) is built around this experimental uncertainty. Reducing it changes the economic model.

The definitive validation of this promise is still ahead. Hassabis announced at Davos in January 2025 that the first drug designed entirely by AI at Isomorphic Labs would enter Phase 1 clinical trials before the end of 2026. Phase 1 tests safety in humans, not therapeutic efficacy. Meaningful Phase 2 and 3 results will not be available before 2028–2030. But crossing the boundary between in silico design and clinical trial is itself a historical milestone.

Five More Fronts: The Full Scope of the Bet

AlphaFold and Isomorphic Labs represent only one of the root nodes Hassabis is attacking in parallel. Five other domains are advancing simultaneously.

Materials. In 2023, GNoME (Graph Networks for Materials Exploration) predicted the stability of 2.2 million new inorganic crystals — estimated as the equivalent of roughly 800 years of published experimental research, made available in a single study. The database is publicly accessible via the Materials Project. The targeted applications include next-generation batteries, superconductors, advanced semiconductors, and photovoltaic cells — each of which represents an industrial bottleneck of the 21st century. Hassabis is going further: in 2025, DeepMind is opening its first fully automated scientific laboratory in the UK, dedicated to materials science. AI-driven robots synthesize and test hundreds of compounds per day, closing the loop between computational prediction and physical experimental validation.

Nuclear fusion. DeepMind is collaborating with Commonwealth Fusion Systems on plasma control in experimental reactors. Fusion, if mastered at scale, represents a clean and effectively limitless energy source. The principal technical obstacle is controlling plasma at extremely high temperatures — a nonlinear optimal control problem for which classical approaches have proven insufficient.

Weather. GenCast, published in 2024, outperforms the ECMWF (European Centre for Medium-Range Weather Forecasts) on medium-range forecasts. The ECMWF is the global reference model, used by governmental meteorological agencies worldwide. The implications go well beyond daily forecasting: anticipating extreme events, optimizing agricultural yields based on seasonal projections, managing electrical grids that depend heavily on wind and solar output, and therefore on weather.

Mathematics. AlphaProof targets the automated proof of new theorems in a formal proof language (Lean). In 2024, the system solved four of the six problems from the International Mathematical Olympiad, including one at an advanced difficulty level. The stakes here are different from the other fronts: mathematics is the substrate of all fundamental science. A system capable of generating genuinely new proofs would be an accelerant for all of theoretical physics, cryptography, and foundational computer science.

Genomics. AlphaGenome (2025) tackles the non-coding genome — the 98% of the human genome that does not directly produce proteins, and whose functional role remains largely poorly understood. This is not passive “junk DNA”: this 98% contains gene expression regulators, functional non-coding RNA sequences, and elements controlling chromatin architecture. Understanding its role would clarify what to edit with CRISPR to treat complex genetic diseases — precisely what is currently missing for most polygenic conditions.

The Distinction That Changes Everything: Reversible Products vs. Irreversible Knowledge

The most useful analytical framework for understanding DeepMind’s strategy is not technical. It is economic.

A product is reversible. A chatbot, however capable, can be copied, surpassed, or made obsolete. Recent AI history provides examples almost weekly: a more efficient open-source model appears and proprietary model margins collapse; a competitor proposes a more efficient architecture and billions in market capitalization evaporate. This is the nature of a technology product market: competition is permanent, advantages are temporary.

Knowledge is structurally different. Once 200 million protein structures have been predicted and stored in a public database, no one can un-map them. The knowledge is acquired — permanently. Every drug, every vaccine, every diagnostic tool developed from those structures over the next fifty years will build on that work.

This logic is well-documented in innovation economics under the name of general purpose technologies: certain technologies have transversal multiplier effects because they reduce the cost of innovation across entire sectors. The steam engine, electricity, semiconductors, the internet. These are not products — they are infrastructures that multiply the productivity of everything built on top of them. AlphaFold may have this nature for molecular biology.

That said, the irreversibility thesis deserves careful qualification. AlphaFold 2 was partially superseded by AlphaFold 3, which ISO-DDE now outperforms on dynamic cases. Knowledge evolves, refines itself, gets corrected. What is genuinely irreversible is the existence of the public database and the establishment of structural prediction as a standard infrastructure in computational biology — not the permanent superiority of any given AlphaFold version. Competing models — Meta’s ESMFold, Boltz-2, RoseTTAFold2 — have partially reproduced AlphaFold’s capabilities for static structure. Isomorphic’s lead on the complete pipeline (dynamics, cryptic pockets, ligand optimization) is real, but it is not technically inimitable. It is the product of a head start and vertical integration that well-funded teams can progressively close.

The Cathedral Builder’s Contradiction

There is an internal tension in the Hassabis strategy that deserves to be named, because it reveals something important about the structure of fundamental AI research.

Hassabis has said it publicly: if he could do it again, he would have kept AI in the laboratory longer. He would have preferred a CERN-style approach — an international scientific consortium, collaborative, methodical, where the global community would work together to build AGI over decades of careful deliberation. The exact opposite of what has been happening since 2022.

The reality is that the fundamental research at Isomorphic and DeepMind costs billions of dollars. Those billions have to come from somewhere. And the answer is Gemini — the chatbot, the reversible product that Hassabis never wanted to build, but which funds exactly what he does want to build. The contradiction is structural: the scientific cathedrals are being funded by mass-market products.

This tension is not unique to Hassabis. It reveals an institutional void: there is currently no public funding mechanism comparable to CERN, the Human Genome Project, or the Apollo program for AI’s root nodes. The major fundamental discoveries in computational biology, materials science, and mathematics are therefore dependent on chatbot revenue, which creates a genuine fragility. Sufficient commercial pressure (such as the pressure reportedly exerted by Sergey Brin to “catch up with competing AI systems on code”) can distort scientific priorities.

This does not invalidate the strategy. But it signals that its durability depends on a delicate balance that market forces can always disrupt.

What Independent Researchers and Developers Can Actually Use Today

Unlike ISO-DDE and Isomorphic’s proprietary pipeline, several tools produced by this research are genuinely accessible without Big Tech resources.

AlphaFold Database (alphafold.ebi.ac.uk): structures of approximately 200 million proteins, freely downloadable and integrable into standard bioinformatics workflows. The entry point for any computational structural biology work.

AlphaFold 2 open source: available on GitHub (deepmind/alphafold), runnable locally under the Apache 2.0 license. Significant GPU infrastructure is required for de novo predictions, but structures already in the database are directly accessible without any local computation.

Boltz-2 (open source, MIT/Recursion): a viable alternative for modeling protein-ligand interactions, with partial dynamics capabilities. Less accurate than ISO-DDE on difficult cases according to published benchmarks, but fully functional for standard cases — and entirely free.

GNoME via Materials Project: the stable crystals predicted by GNoME are integrated into the Materials Project database (materialsproject.org), freely accessible. The entry point for any work on functional materials.

GenCast: DeepMind has released the GenCast code (Apache 2.0 license). Operational use requires substantial compute infrastructure, but model outputs are available through academic partners.

The boundary between what has been democratized and what remains proprietary is clear: the structures and databases are public. The complete drug optimization and materials design pipelines remain proprietary. An independent lab can exploit the foundations — it cannot replicate Isomorphic’s full value chain without considerable investment.

The 50-Year Test

There is a simple test for distinguishing a leaf from a root in the tree of knowledge: will what is produced today still be there in fifty years?

The protein structures in the AlphaFold database will still be in use in fifty years. The 2.2 million stable crystals from GNoME will be part of materials science infrastructure for decades. A drug developed through ISO-DDE, if it proves effective in Phase 3, will be prescribed for decades after that.

GPT-5.5, Claude 4, Gemini 3.1 — each of them the “most capable model ever released,” each of them already one update away from irrelevance. GPT-4, which defined the industry conversation just eighteen months ago, is no longer accessible. It has been retired. The AlphaFold database, published in 2022, is still there — and will be in twenty years.

This is not a value judgment on the usefulness of conversational AI assistants. They have real, immediate impact for millions of people. It is an observation about the nature of what is being produced: consumer products whose value is permanently contested by competition, versus knowledge infrastructures whose value accumulates with time.

The question is not who has the largest valuation this year. It is who is structuring the terrain on which every other industry will build in twenty years. On that question, the strategy of Demis Hassabis merits serious analysis — not because it is without risks or contradictions, but because it is playing a fundamentally different game from the one the rest of the industry believes it is playing.

But this raises a harder question: if DeepMind is building scientific infrastructure for everyone, why does that infrastructure still depend on the commercial chatbot race Hassabis publicly wishes AI had avoided?

That is the subject of the next essay.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Sources: Nobel Prize (Chimie 2024), Isomorphic Labs (IsoDDE), Nature (GNoME, 2023), Google DeepMind (AlphaProof), AlphaProof, Nature (GenCast, 2024), Google DeepMind (AlphaGenome)

On April 26, Two Things Happened at Once

On April 26, 2025, DeepSeek published new pricing. The entry cost for V4 Pro with cache drops to 0.025 yuan per million tokens — a 90% reduction from previous pricing. A few thousand kilometers away, in the backend logs of Codex, developers spotted a string that shouldn’t have been visible: “gpt-5.6.” Two unrelated events. But placed side by side, they might be telling the same story.

This isn’t the story of a Chinese model that just “beat” OpenAI. It’s the story of what happens when a technology stops being a premium product and becomes marginal-cost infrastructure — and why the companies that built their advantage on premiumization have no structurally good answer to that.

What V4 Actually Does — No Embellishment

DeepSeek V4 is an open-source model. You can download it, modify it, and deploy it. It runs on Nvidia chips — but also, and this is new, on Huawei’s Ascend processors. Several Chinese chipmakers (Cambricon, Biren, MetaX) have announced compatibility. The Chinese Academy of Information and Communications Technology is officially testing it. This is no longer an isolated lab project. It’s a national infrastructure component.

On capabilities, the company is honest about its limits. V4 still trails the top closed models — Claude 4.6, Gemini 3.1 Pro — in certain domains. What DeepSeek published alongside the commercial launch may be more interesting: a research paper titled Thinking with Visual Primitives, co-authored with Peking University and Tsinghua. The central idea: multimodal models don’t lose their reasoning because they can’t see images clearly enough — they lose it because they have no stable way to reference what they’ve seen while they’re reasoning.

The proposed solution: bounding boxes as a thinking tool, not as a final output. The model points while it reasons. Results on maze navigation (MazeNav): 66.9% for DeepSeek, 50.6% for GPT-5.4, 48.9% for Claude. At a visual memory cost of roughly 90 entries, versus 870 for Claude and 1,100 for Gemini. Less memory, better precision on these specific tasks. Not a general revolution — a precise, measurable advantage on a precise problem.

Meanwhile, at OpenAI, GPT-5.5 had developed a spontaneous fascination with goblins, gremlins, and trolls. Codex’s system prompt had to explicitly ban these creatures — along with raccoons and pigeons. The ban was discovered by users.The internet reacted the way the internet reacts. GPT-5.6 appeared in backend logs two weeks later.

The War Isn’t Where Everyone’s Looking

The standard analysis of this week looks like this: DeepSeek vs. OpenAI, Chinese model vs. American model, open source vs. closed, performance A vs. performance B on some benchmark. That reading isn’t wrong. It just misses the actual mechanism.

In 1865, economist William Stanley Jevons published a counterintuitive observation about English steam engines. Efficiency had improved significantly — the engines burned less coal to produce the same energy. The logical conclusion was that England would consume less coal. The reality was the opposite: consumption exploded. When something gets cheaper to use, people use a lot more of it.

That’s exactly what’s happening with AI right now. Visa spent 1.9 trillion tokens in March. At Disney, the top Claude user logged 51,000 daily requests over nine workdays — with hundreds of engineers behind them running hundreds more each. A Shanghai video game company uses V4 for file management and spends 0.56 yuan, where it used to spend ten times that. A term has emerged inside large enterprises to describe this phenomenon: token maxing. AI is no longer used for exceptional tasks. It’s infiltrating every workflow — from research organization to presentation creation to Slack monitoring.

The bottleneck is no longer model quality. It’s the infrastructure cost. And that’s precisely where DeepSeek just changed the rules.

What China Is Actually Building

V4’s compatibility with Huawei chips isn’t a technical detail. It’s the central piece of the puzzle. American sanctions cut China off from the most advanced Nvidia GPUs. The response wasn’t to abandon the ambition — it was to build a complete alternative stack, from silicon to model.

If Huawei deploys its Ascend 910C supernodes at scale in the second half of this year, V4’s inference cost — already remarkably low — could fall further. What we’re watching isn’t a simple price war between labs. It’s the construction of a sovereign AI ecosystem that can exist without CUDA, without AWS, without Azure, and without OpenAI’s or Google’s models. Open source is the adoption mechanism: once thousands of companies embed V4 into their workflows, once their engineers know the API, once their systems are calibrated to that model, the switching cost back to an American model becomes real again.

For a CTO in Warsaw, Singapore, or São Paulo, this bifurcation isn’t abstract. It presents itself concretely as a procurement decision: closed American models with their GDPR implications and premium pricing, or open models that can be audited, hosted locally, and adapted to local regulation. V4 on Ascend represents a third option — capable, modifiable, sovereign, and now compatible with hardware you can purchase without going through the gray market.

The Argument I Can’t Ignore

There’s a strong counterargument to all of this, and it deserves to be stated honestly.

DeepSeek V4 is not the best model in the world. The company says so itself. For high-precision tasks — law, medicine, security, complex multi-step reasoning — closed models maintain a real advantage. And companies operating in regulated sectors can’t simply swap a Chinese open-source model into their existing infrastructure: compliance, audit, and data sovereignty questions cut the other way.

On top of that, open source is a double-edged sword. Nothing stops OpenAI or Google from forking V4 tomorrow and integrating its innovations into their own models. The real moat in frontier research still sits with the American labs.

That’s fair. But that reasoning assumes the battle is playing out in the frontier segment. It’s not — or at least, not only. The majority of enterprise use cases — document summarization, classification, standard code generation, data analysis, workflow automation — don’t require the best model in the world. They require a model that’s good enough, reliable, fast, and affordable. And “good enough” at $0.036 per million tokens versus “slightly better” at ten times the price: that’s an economic decision, not a technical one.

Where This Analysis Hits Its Limits

I need to be precise about what we don’t know. “gpt-5.6” in backend logs doesn’t mean OpenAI is in a panic — it looks like internal testing or canary deployment. A viral screenshot of V4 fixing a bug that another model missed proves nothing: LLMs are stochastic, and a single attempt is not a benchmark. DeepSeek’s talent retention (16 of the 18 R1 contributors still present) is a positive signal, but an organization can retain its people and miss its next generation. AI history is full of leaders who dominated one cycle and missed the next.

What I’m saying is that the market structure is changing. Not that the verdict is in.

The Question OpenAI Can’t Dodge

The goblin bug is anecdotal. The GPT-5.6 leak may be inconsequential. Codex is becoming seriously powerful — Greg Brockman said he “fell in love” with the app after twenty years of terminal use, which is not a throwaway statement coming from him. And Sam Altman joked about Codex’s “goblin moment,” which suggests a certain serenity in OpenAI’s hallways.

But the real pressure isn’t in the benchmarks. It’s in the Jevons dynamic: when AI gets cheaper, companies don’t use less AI — they use infinitely more. Those that built their workflows on premium models will look at their invoices and run the numbers. And those that haven’t adopted AI yet will look at V4 at $0.036 per million tokens and see an entry point, not an alternative.

OpenAI can maintain excellence on frontier tasks. But if DeepSeek captures the 80% of routine use cases, being excellent is no longer necessarily the winning position.

The AI war is no longer a capabilities race. It’s an infrastructure war. And China may have just set the terms.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

On March 19, 2026, Josef Aschbacher, ESA’s Director General, stood before the Agency’s Council in Paris and announced that a mission called EPIC — ESA-Provided Institutional Crew — had just been approved by member states. The agency would purchase a Crew Dragon flight from SpaceX for its own astronauts. Not an exchange of technical contributions for seats. Not a framework agreement brokered through NASA. A direct commercial purchase.

In the weeks that followed, Aschbacher compared SpaceX’s trajectory to what ESA is trying to build: “If you look across the Atlantic, the development is very similar to what we have in mind.” It’s a diplomatic sentence. Read carefully, it’s also an admission of a time lag.

Twenty-six days later, in Texas, thirty-three third-generation Raptor engines ignited simultaneously on Pad B at Starbase. Booster 19 produced 9,240 metric tons of thrust for 1.88 seconds — before pressure sensors triggered an automatic abort. Less than planned. More than enough to validate the structure. And twice the power of the SLS NASA spent fifteen years developing.

Two events, less than a month apart, telling the same story.

The issue is not whether Europe still has world-class engineers. It does. The issue is whether engineering excellence still translates into strategic leverage when launch cadence, reusable infrastructure, and crew access are controlled by someone else.

What “buying” means when you’re a space agency

From its earliest days, ESA navigated human spaceflight through a barter mechanism with NASA: technical contributions — Orion’s service module, scientific instruments, engineering expertise — traded for seats aboard American flights. This arrangement allowed the agency to maintain an orbital presence without building its own crewed launch system, the cost of which would have been prohibitive.

EPIC breaks with that model. For the first time, ESA is buying a flight directly from a commercial operator, in euros, with no technical counterpart. The reason is straightforward and uncomfortable: five career astronauts from the 2022 class risked never flying before the ISS closes in 2030. The station will be deorbited. Time is short. The standard barter mechanism with NASA couldn’t have gotten all of them there in time.

It’s a pragmatic decision. But pragmatism has its own grammar, and this one says something specific: when you pay cash for what you once traded expertise for, it means your expertise is no longer the scarcity you thought it was.

Nobel laureate Jean Tirole described this moment across other industries: once an actor reaches critical mass in operations, infrastructure, and data, the cost of competing with them exceeds the cost of simply using them. SpaceX has reached that point in crewed orbital access. Not because ESA ran short of engineers — but because the industrial cadence out of Boca Chica rendered the comparison obsolete before Europe had finished laying its foundations.

The life raft before the boat

Paradoxically, at the very moment it’s buying seats from SpaceX, ESA is investing in something SpaceX has owned since 2015: an emergency crew abort system.

The Crew Launch Abort Demonstrator — floated in an April 2026 procurement call — aims to qualify a system capable of extracting a crew from its capsule during a pad emergency or ascent failure. The functional equivalent of the Super Draco system integrated into Crew Dragon, successfully tested by SpaceX in 2019. The demonstrator will be tested on a capsule derived from the Leo Cargo Return Service program: 5.4 meters in diameter, ten metric tons, designed for Ariane 6.

Europe has mastered emergency evacuation systems for crewed spaceflight. It doesn’t have crewed spaceflight yet.

This isn’t incompetence. It’s a deliberate niche strategy. By developing recognized expertise in crew safety-critical subsystems, ESA positions itself as an indispensable partner for anyone seeking to certify crewed missions under European regulatory frameworks — regardless of which launcher ultimately dominates the market. If NASA opens contracts to international partners on the components listed in its Moonbase User Guide — communications, power, robotics, safety — ESA will be positioned to respond.

The calculation: rather than spending an estimated €415 million per crewed mission on Ariane 6 — roughly 41 percent of the agency’s annual exploration budget — concentrate resources on components where Europe can claim an irreplaceable position, then wait for the major operators to need them.

It’s the strategy of a critical-component supplier in a market dominated by a single integrator. Samsung making chips for Apple. It’s viable. It isn’t sovereignty.

Where this reading hits a wall

The strongest counterargument deserves to be stated honestly.

Buying directly from SpaceX may not be an admission of dependency — it may be the opposite. By stepping out of the NASA barter system, ESA establishes a direct commercial relationship with the operator that will matter most in the decade ahead. It chooses its counterpart rather than inheriting one through American intermediation. And by concentrating investment in high-value subsystems rather than trying to replicate what SpaceX does better and cheaper, it’s doing exactly what economic theory recommends: specializing where it holds a genuine comparative advantage.

This reading is coherent. It assumes, however, that SpaceX will remain indefinitely a transport provider rather than an exploration operator with its own agenda. Starship was not designed to deliver payloads to low Earth orbit — it was designed to colonize Mars. When your transport supplier has a planetary colonization plan, the line between client and passenger is worth watching.

The real question isn’t technical

Booster 19 held for 1.88 seconds. That’s not a triumph — it’s a partial validation among many still to come. Starship V3 still has orbital refueling, second-stage catch in flight, and dozens of tests ahead that no one can fully anticipate. SpaceX knows better than anyone that the history of this program is built from partial successes bracketed by engineered failures.

What is less partial is the signal the EPIC mission sends about the actual balance of power in the space industry. Not Europe versus America — national categories are too broad for what’s actually happening here. Rather: public agencies of every nation facing a private actor that understood industrial cadence is as powerful a strategic weapon as the technology itself.

The real question isn’t whether ESA and SpaceX will one day collaborate on Starship HLS. It’s how quickly Europe can build a negotiating position strong enough that any such collaboration, if it comes, happens between partners — not between an operator and an approved subcontractor.

That road is long. The decision to buy a ticket rather than negotiate one says so more clearly than any press release could.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Sources: SpaceNews, NASASpaceFlight.com, European Spaceflight, Space.com

There’s a question most AI benchmark analyses never ask: what does this model actually cost to run in production? An MMLU Pro score carries no practical weight if the output cost makes the workflow economically out of reach. That’s precisely where DeepSeek V4 lands — and why this release deserves an analysis that goes well beyond performance tables.

DeepSeek V4 is not the most powerful model on the market across every benchmark. That’s a fact DeepSeek acknowledges directly in its technical documentation, noting a gap of roughly three to six months behind state-of-the-art proprietary models on general reasoning. But V4 asks a different question — one that’s more relevant for the vast majority of teams actually building with AI: what happens when a model that performs competitively on high-value tasks — code, agents, mathematical reasoning, long document processing — costs 50 to 100 times less than premium alternatives?

An Architecture Built to Make Long Context Economically Viable

Understanding why V4 can offer these prices requires understanding the core technical problem it solves: attention in standard Transformers gets expensive as context grows.

In a classic Transformer, every token must compute its relationship to every other token in the context window. Double the context, and compute cost quadruples — that’s O(n²) complexity. This is precisely what makes one-million-token context windows so difficult to sustain in practice. Many models announce them; very few make them genuinely usable at a reasonable cost.

DeepSeek’s answer is a hybrid attention architecture that combines two complementary mechanisms.

The first, Compressed Sparse Attention (CSA), operates on groups of tokens close to the current prompt position. It compresses blocks of four tokens into dense representations, then performs selective retrieval on the most relevant passages rather than attending uniformly across the entire context. The analogy with human reading is useful: we retain detailed word-level memory of recent sentences while synthesizing and compressing earlier paragraphs.

The second, Heavy Compressed Attention (HCA), is more aggressive: it compresses groups of 128 tokens into a single condensed entry. This gives the model a low-cost global view of the full context — equivalent to a dynamic table of contents that the model consults continuously to maintain coherence across long documents.

The combined effect is captured in figures published by NVIDIA at launch: at one million tokens of context, V4 Pro uses only 27% of the compute required by its predecessor V3.2, with KV cache memory reduced to 10%. V4 Flash goes further — 10% of the compute, 7% of the memory. This infrastructure cost reduction translates directly into the prices on offer. This isn’t a pricing subsidy or a loss-leader strategy — it’s the direct outcome of a more efficient architecture.

V4 also incorporates two additional technical innovations: Manifold Constrained Hyper Connection (MHC), which improves signal propagation stability in residual network connections (a problem that becomes critical in very large MoE systems), and the Muon optimizer, replacing AdamW for large-scale low-precision training. DeepSeek reports that fully optimizing these components can nearly double inference speed.

What “Mixture of Experts” Actually Means in Practice

V4 Pro carries 1.6 trillion parameters. That number is impressive but potentially misleading. What matters for inference — and therefore for actual cost — is that only 49 billion parameters are active per request. The remaining 1.55 trillion are present in the network but dormant for any given pass.

Mixture of Experts works like a specialized library: the model’s router dynamically selects which “experts” to activate based on the nature of the task. A Python debugging request activates a different subset of parameters than a legal contract review. This mechanism achieves something that’s normally a trade-off: broad total capacity (the model can excel across many distinct domains) combined with low computational footprint at inference time.

V4 Flash applies the same principle at a smaller scale: 284 billion total parameters, 13 billion active. For medium-complexity tasks — document summarization, coding assistance, routing agents — V4 Flash delivers remarkable efficiency at $0.14 per million input tokens and $0.28 per million output tokens.

Interleaved Reasoning: The Quiet Feature That Changes Agentic Workflows

There’s a structural limitation in most current agentic pipelines that rarely gets discussed: every time an agent makes a tool call — a web search, code execution, API request — the model often has to partially reconstruct its reasoning state for the next step. Some of the thinking context is lost in the transition. On a three-step workflow, this doesn’t matter much. On a twenty-step workflow — analyzing a codebase, investigating complex bugs, synthesizing multi-source research — those state losses accumulate and produce inconsistencies.

V4 introduces what DeepSeek calls interleaved reasoning: the model’s chain of thought persists across tool calls. It doesn’t reset after executing code or fetching an external resource; it picks up exactly where it left off. This is an architectural property, not a prompt engineering workaround.

NVIDIA confirms that V4 is natively compatible with agent frameworks, including NeMo, AQ Blueprint, and Data Explorer Agent. DeepSeek’s own internal survey of 85 experienced developers found that 52% consider V4 Pro ready to become their default coding agent, with over 90% including it in their top choices for programming tasks.

For long agentic workflows, reasoning stability matters as much as raw model capability.

What the Benchmarks Show — and What They Don’t

Early benchmark results are strong, particularly in STEM and code domains. On Apex Shortlist, a challenging math and science benchmark, V4 Pro reaches 90.2%, ahead of Claude Opus 4.6 at 85.9% and GPT-5.4 at 78.1%. On SWE-bench, which measures real bug resolution on authentic GitHub repositories, V4 Pro reaches 81.6%, matching Claude Opus 4.6. On Codeforces, the model ranks at approximately the 23rd percentile among human competitors.

Gaps are more pronounced in other areas. On MMLU Pro (encyclopedic knowledge and reasoning), Gemini 3.1 Pro scores 91.0% versus V4 Pro’s 87.5%. On GPQA Diamond (expert-level science questions), Gemini reaches 94.3% against V4’s 90.1%. On Humanity’s Last Exam, the gap widens: Gemini at 44.4%, V4 Pro at 37.7%.

DeepSeek is explicit about this in its documentation: V4 Pro surpasses mainstream open-source models and approaches proprietary systems on code, agents, mathematics, and STEM — but a gap of three to six months remains with the most advanced frontier models on general reasoning. This is a careful, honest formulation, and a notably unusual one for an AI product launch announcement.

Real user feedback confirms this nuanced picture. Many report a noticeably better experience than V3.2 on code and agentic tasks. Others note that V4 Flash doesn’t seem meaningfully better than V3.2 in unstructured daily conversation, with ambiguous instructions, or for creative use cases. This disparity is expected and informative: V4 is optimized for specific domains, not designed to be the best general-purpose model across every situation.

The real question isn’t “is V4 the best model?” — it’s “for which use cases does V4 offer the best capability-to-cost ratio?” And on that question, the answer is considerably more favorable to DeepSeek.

The Economic Case for Teams Without Cloud-Scale Budgets

Here’s a concrete calculation. Imagine a three-person development team building a code analysis agent over a 400,000-line codebase — roughly 600,000 tokens. The agent needs to scan the full codebase twice a week to detect security patterns, generate a report, and answer specific developer questions.

With GPT-5.5 Pro at $30/M input tokens and $180/M output tokens: a full analysis (600K tokens in, ~20K tokens out) costs approximately $21.60. Over 100 annual analyses, that’s $2,160 — before counting ad-hoc query costs.

With V4 Pro at $1.74 input and $3.48 output, the same analysis costs around $1.11. Over 100 analyses, $111. The annual savings on this single workflow exceed $2,000.

But the more interesting calculation involves V4 Flash, which is designed for broader deployment. At $0.14 input and $0.28 output, building a chat interface on internal documents, a lightweight coding assistant, an automated summarization pipeline, or a routing agent becomes economically negligible for an independent developer or a small team. Use cases that were previously reserved for teams with significant cloud budgets become accessible to a solo developer running their first serious AI product.

The MIT license amplifies this advantage. Model weights are available on Hugging Face and can be downloaded, modified, and hosted on private infrastructure. This isn’t just a cost reduction through the API — it’s the ability to fully control deployment, eliminate vendor dependency, and fine-tune the model for specific needs if the hardware resources allow.

The Hardware Dimension: More Than a Geopolitical Footnote

DeepSeek V4 is designed to run on two distinct hardware ecosystems: NVIDIA’s Blackwell and Hopper GPUs on one side, Huawei’s Ascend 950 chips on the other. This duality has an obvious geopolitical dimension — but it also has a concrete technical implication.

NVIDIA announced day-one support for V4 on its Blackwell architecture, with benchmarks showing over 150 tokens per second per user on GB200 NVL 72 systems. The full CUDA ecosystem remains functional for V4.

The more unusual part is the optimization for Huawei’s Ascend 950 chips, with reported speedups of 1.50× to 1.73× on inference workloads. Huawei confirmed that its ACIN supernode products based on the AC950 series will support V4.

Important nuances apply, however. Several sources — including MIT Technology Review and Professor Liu Jiwan at Tsinghua University — indicate that full training is not clearly adapted to Chinese chips; only inference appears to be genuinely optimized for the Ascend architecture. There’s no clean break from NVIDIA for training; Chinese chips are currently better suited to inference than to training large-scale models.

What’s significant is the direction of the signal. DeepSeek explicitly states that V4 Pro pricing could fall further once Ascend 950 supernodes are deployed at scale in the second half of 2026 — meaning current pricing is not yet the floor. For developers building products intended to last, that’s a trajectory worth factoring into architectural decisions.

Real Limitations: Where V4 Is Not the Right Choice

The most obvious limitation is the absence of multimodal capabilities. V4 is exclusively text and code. OpenAI, Google, and even Xiaomi with MiMo V2.5 Pro offer systems that integrate text, image, audio, and video. For any project requiring image analysis, chart comprehension, scanned document processing, or voice interaction, V4 is not the right model — and DeepSeek acknowledges that multimodal capabilities are planned but not yet available.

General reasoning remains a step behind. On tasks requiring broad encyclopedic knowledge (MMLU Pro), expert cross-disciplinary reasoning (GPQA Diamond), or advanced humanities (Humanity’s Last Exam), leading proprietary models maintain a measurable advantage. For a general-purpose assistant serving non-technical users, or for applications requiring high-quality creative text generation, proprietary alternatives retain an edge.

Every day, conversational performance remains uneven according to user reports. V4 appears optimized for structured tasks with clear objectives — code, analysis, agents operating over structured data — rather than for open-ended, nuanced dialogue.

Finally, the one-million-token context window, while made economically viable by the architecture, remains a promise to verify on real end-to-end use cases. Retrieval quality in very long contexts — accurately locating a specific piece of information buried in 800,000 tokens — is a property that standard benchmarks struggle to measure reliably.

What You Can Build Right Now

For developers and small teams who want to evaluate V4 without waiting, the entry points are clear.

Via the DeepSeek API: existing endpoints already route to V4 Flash in reasoning and non-reasoning modes. The legacy deepseek-chat and deepseek-reasoner endpoints will be deprecated on July 24, 2026. API compatibility is maintained with common frameworks.

Via Hugging Face: model weights are available for local or private deployment under the MIT license. For teams with access to multiple GPUs (A100s, H100s, or the new Blackwell cards), self-hosting becomes a serious option for high-volume use cases.

High-ROI use cases for budget-constrained teams:

• Agentic code review over full codebases (leverages long context + interleaved reasoning)

• RAG pipelines over large document corpora (contracts, technical documentation, knowledge bases)

• Multi-step research agents with tool calls (benefits from persistent reasoning state)

• Coding assistants embedded in IDEs or internal tools (V4 Flash for cost efficiency, V4 Pro for higher quality)

• Large-scale document processing: summarization, extraction, classification at volumes that were previously cost-prohibitive

Where you’re better served by other models: high-quality creative content generation, applications requiring vision or multimodal capabilities, general-purpose conversational assistants for broad non-technical audiences.

When Price Becomes an Architectural Argument

The real lesson of DeepSeek V4 isn’t that Chinese AI has caught up with American AI. That framing is too simple. V4 Pro remains three to six months behind the best proprietary models on general reasoning — DeepSeek says so itself.

The real lesson runs deeper: a model doesn’t need to win every benchmark category to shift the competitive landscape. It just needs to be good enough on high-value tasks, open enough to deploy freely, and affordable enough to make previously inaccessible workflows economically viable.

For an independent developer building a serious agent over one million tokens, the gap between $180 and $3.48 per million output tokens isn’t a difference in degree — it’s a difference in kind. One makes the project impossible below a certain volume. The other makes it experimental, scalable, and potentially viable from week one.

The hybrid attention architecture isn’t an implementation detail. It’s what makes this economics possible — and what makes V4 a model worth serious attention from anyone building with AI without the resources of a major tech company.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Sources: DeepSeek V4 Technical Report, NVIDIA Technical Blog, MIT Technology Review, Artificial Analysis

A Nvidia engineer with early access to GPT-5.5 wrote, at launch: “Losing access to this model would feel like losing a limb.” The line is dramatic. It says something real about what the model does — or about what the people who previewed it believe it does.

GPT-5.5 launched on April 23, 2026. That same month, on Forge Global, a private stock trading platform, Anthropic was trading at an implied valuation of roughly one trillion dollars. OpenAI at $880 billion — 3% above its last primary valuation. According to Caplight, which tracks private market activity, there were more sellers than buyers for OpenAI shares in Q1.

These aren’t the same numbers. They aren’t the same signals. And the reason they’re diverging isn’t technical.

The best model and the best business stopped being the same thing.

The model that rewrote its own engine

Start with the facts, because they deserve to be taken seriously before being outrun.

GPT-5.5 is a larger model than GPT-5.4. It still delivers the same per-token latency under real-world conditions. That’s rare — larger models are almost always slower. On Terminal Bench 2.0, which tests complex command-line workflows requiring planning and tool coordination, it scores 82.7% against 75.1% for GPT-5.4 and 69.4% for Claude Opus 4.7. On PIB Val — 44 real professions, financial modeling, legal analysis, data science — it matches or exceeds professional-level performance on 84.9% of tasks. On FrontierMath Level 4, the benchmark’s hardest mathematics problems: 35.4% against 27.1% for GPT-5.4.

BenchmarkGPT-5.5ComparisonTerminal Bench 2.082.7%GPT-5.4: 75.1%PIB Val (44 professions)84.9%Claude Opus 4.7: 80.3%FrontierMath L435.4%GPT-5.4: 27.1%ARC-AGI85.0%Gemini 3.1 Pro: 77.1%

These numbers are solid. But there’s one technical fact that outweighs them all.

Before GPT-5.5, OpenAI split requests across Nvidia GB200/GB300 GPUs using a fixed, predefined number of segments — a functional but suboptimal engineering heuristic. Codex, running on GPT-5.5 during training, analyzed weeks of production traffic data and rewrote those load-allocation algorithms. Result: +20% token generation speed. The model built the tools that made it faster.

This isn’t science fiction. It’s systems optimization, executed by an agent with access to data and code-writing instructions. But the achievement is real, and the line it crosses is worth noting without drama: for the first time, a production model contributed to improving the infrastructure running it.

The concrete flaw: on SWE-Bench Pro, GPT-5.5 scores 58.6%. Claude Opus 4.7 scores 64.3%. OpenAI disputes this, citing signs of memorization in Anthropic’s results on a subset of problems. The dispute may have merit. But when a leader has to qualify the adverse result rather than dismiss it, the result can’t be dismissed.

The night the benchmark leader finished second

In April 2026, Anthropic released its revenue figures. ARR climbed from $9 billion at end of 2025 to $30 billion in March 2026. A 233% increase in a single quarter, driven primarily by enterprise adoption of Claude Code and the API.

On secondary markets — Forge Global, Caplight — Anthropic is trading at an implied valuation near one trillion dollars. OpenAI sits at $880 billion, 3% above its last primary valuation. Caplight also noted that in Q1, there were more sellers than buyers for OpenAI shares.

Anthropic grew 233% in one quarter. Not from a better model. From a developer workflow tool.

A counter-argument needs to be stated honestly: OTC secondary markets are illiquid, notoriously volatile, and largely speculative. An implied valuation of $980 billion on Forge Global represents a small number of transactions between non-representative buyers and sellers. OpenAI’s total ARR still exceeds Anthropic’s. And ChatGPT has hundreds of millions of active users that no benchmark captures.

But the seller/buyer ratio on OpenAI isn’t a liquidity artifact — it’s a conviction signal from insiders. And 233% growth in one quarter driven by recurring enterprise contracts isn’t sentiment: it’s revenue. The nuance isn’t that OpenAI is losing. It’s that for the first time, it’s no longer alone at the top across every axis simultaneously.

Claude Code didn’t beat GPT-5.5. It didn’t have to.

In 1997, Clayton Christensen published The Innovator’s Dilemma. The central thesis: technical leaders aren’t dethroned by better innovations. They’re bypassed by innovations good enough to integrate into existing workflows at lower friction cost.

GPT-5.5 is a sustaining innovation — better, more precise, faster on long-horizon tasks. Claude Code is something else. It’s a tool that integrated into enterprise developers’ daily work cycle before it became irreplaceable. The adoption data makes this concrete: 85% of OpenAI’s own employees use Codex weekly — suggesting that even inside OpenAI, the workflow tool won before the model comparison was ever run.

Industry history runs this pattern. Intel dominated chip design for twenty years. TSMC captured manufacturing — an adjacent segment, technically less glamorous — and became indispensable to everyone, Intel included. Windows dominated operating systems. iOS captured the mobile platform. Oracle dominated enterprise databases. AWS captured infrastructure. In each case, the technical leader kept building the best products in its category. Value shifted to the layer that captured the workflow.

The benchmark race and the business race have diverged. The first measures what the model can do. The second measures how often someone uses it to do something that generates recurring value. GPT-5.5 leads the first. Anthropic leads the second. These are not the same race.

Capturing a workflow beats winning a benchmark. It always has.

A rounding error, or a fault line?

While this debate plays out between OpenAI and Anthropic, there’s a pressure the American ecosystem consistently treats as a footnote.

GPT-5.5 costs $5 per million input tokens and $30 per million output tokens. GPT-5.5 Pro: $30 and $180. For reference: Kimi 2.5 is at $0.44 and $2. Minimax M2.7 at $0.30 and $1.20. These models aren’t at GPT-5.5’s level on the most demanding benchmarks. But they’re not 10x behind, and they cost 10 to 60 times less.

For a developer in Berlin, Lagos, or Singapore building a product on thin margins, the question “GPT-5.5 or Claude Opus 4.7?” is not the real question. The real question is “OpenAI or Kimi?” — and that’s an economic calculation, not a technical one. OpenAI is positioning itself as the Apple of AI — premium, aspirational, irreplaceable for those who can afford it. That’s a viable strategy. It worked for Apple.

Note: GPT-5.5 is not available to free ChatGPT users. The model presented as built for “real-world work” is reserved for Plus, Pro, Business, and Enterprise subscribers. There’s an irony in the fact that OpenAI doesn’t highlight it.

The best model in the world is behind a paywall. The models building market share are not.

Where this analysis runs out of runway

Secondary markets have their own dynamics. Anthropic’s implied valuation at ~$1T could be a sentiment spike tied to Q1 ARR, not a durable reflection of fundamental value. If Chinese models face expanded export restrictions — a real possibility in the current geopolitical context — the pricing equation shifts dramatically in favor of OpenAI and Anthropic on Western markets.

There’s also a question I can’t settle from the outside: is GPT-5.5’s technical lead on long-horizon tasks sufficient for enterprises to keep paying 10x the price of a Chinese model? If yes, the business race may be more tied to the benchmark race than this analysis suggests.

On April 23, 2026, a model optimized its own infrastructure during training. That same month, the company that had dominated the space since ChatGPT watched a competitor pass it on financial markets — not because its model was worse, but because the other had better integrated its tools into the workflows of the people paying.

This moment may go unnoticed in six months. Or it will be the moment people look back to understand when the race changed.

The race isn’t over. It just changed what it’s measuring.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

Houston. April 11, 2026. Somewhere around 2 a.m. local time, Jeff Radigan, Artemis 2’s flight director, said something out loud that nobody in that room needed to hear: “We will think as clearly as we can and remain calm while we wait for signal.” Then the room did exactly that.

Four astronauts — Reid Wiseman, Victor Glover, Christina Koch, and Jeremy Hansen — were crossing 2,760 degrees Celsius inside a plasma sheath traveling at 40,000 kilometers an hour. No signal in. No signal out. Not a malfunction. A plan.

Six minutes. No intervention possible. The physics of plasma are not open to negotiation.

This is where NASA was in April 2026: preparing its people, officially, to wait.

At almost the same moment, on a different screen somewhere on the same internet, footage from SpaceX’s eleventh Starship test flight — October 2025 — was making rounds again. 4K. Uninterrupted. A spacecraft threading through the same physical phenomenon, live, with no gap in coverage. One sentence posted on X captured the contrast with the kind of precision that makes engineers uncomfortable: “NASA says six-minute blackout. SpaceX says watch in 4K.”

The contrast was real. It was also, in one important way, incomplete. And the incomplete part is where the more interesting story lives.

Starlink Never Meant to Be Here

In 2021, SpaceX filed a request with the FCC to test communications between the Starship vehicle and its Starlink satellite constellation. The stated ambition was technical and specific: maintaining “unprecedented telemetry volumes” during atmospheric reentry, even as ionized plasma suppresses conventional radio frequencies. The filing described how antennas positioned at strategic locations on the hull could maintain a satellite link through zones where the plasma sheath was locally thinner.

This filing went mostly unnoticed. It looked like housekeeping.

It was not housekeeping.

Starlink was sold to regulators, investors, and the public as a project to deliver internet access to underserved rural populations. That pitch was real — the farms are connected, the economics work, the product exists. But the same constellation dense enough to serve a customer in rural Montana is also dense enough to maintain a link with a vehicle moving at 27,000 kilometers an hour through a plasma field.

SpaceX didn’t set out to solve the plasma blackout. It built Starlink to generate recurring revenue from a subscription internet service. The blackout problem came apart as a side effect.

This distinction matters more than it might seem. The solution wasn’t targeted engineering genius applied to a specific aerospace challenge. It was the structural advantage that comes from owning a vertical stack — the rocket, the satellites, the ground infrastructure — all simultaneously. When you own the pipe, problems that look intractable from outside the system sometimes dissolve before you even frame them as problems.

John D. Rockefeller understood this in the 1880s. He didn’t dominate the American oil industry because his refineries were better than his competitors’. He dominated it because he controlled the pipelines. His competitors could drill as much as they wanted. Without access to transport, the product was irrelevant.

Starlink is the pipeline of the next era of human spaceflight. What SpaceX demonstrated during that 4K reentry wasn’t superior aerospace engineering. It was a superior infrastructure positioning — built for a different purpose, arriving precisely when it was needed.

The Argument SpaceX Isn’t Making

There’s a version of this story that’s clean, satisfying, and wrong: SpaceX solved the problem that NASA couldn’t.

The physics won’t cooperate with that narrative.

Starship reentered from low Earth orbit at approximately Mach 25. The plasma it generated was real and hostile to radio communication. SpaceX’s approach — antennas positioned in locally thinner zones of the plasma sheath, maintaining a link through a sufficiently dense overhead constellation — worked. The footage is real.

Orion returned from the Moon at Mach 32. That difference changes everything. Plasma density around a spacecraft increases sharply with velocity. The electromagnetic barrier surrounding Orion during lunar reentry was significantly thicker and hotter than anything Starship has encountered in testing. The windows through which SpaceX’s antennas maintained their satellite link would have been considerably narrower — possibly closed entirely.

SpaceX hasn’t claimed its LEO solution works at lunar reentry speeds. It hasn’t tried. The comparison that circulated on X was rhetorically perfect and technically incomplete.

And yet.

The right question isn’t whether Starlink solves Mach 32 today. It doesn’t. The right question is: when someone needs to solve Mach 32 — in ten or fifteen years, when crewed lunar missions have become routine enough that six minutes of silence is treated as an engineering deficit rather than an operational given — who will be positioned to attempt it?

Not necessarily because they’ll be smarter. Because they’ll already own the constellation dense enough to try.

When the Pipe Is Worth More Than the Rocket

Jeremy Hansen is Canadian. He was one of the four astronauts inside Orion during those six minutes. His presence on Artemis 2 was the result of a longstanding partnership between NASA and the Canadian Space Agency, one of the oldest collaborative relationships in human spaceflight history.

The European Space Agency has crewed spaceflight ambitions. So does JAXA. So does ISRO. So do the UAE. None of them owns a satellite constellation with Starlink’s current density. None of them is meaningfully close.

When the next generation of crewed spacecraft is designed — and they will be, by agencies that want strategic autonomy from American commercial providers — the communication architecture will be a primary engineering constraint, not an afterthought. The question their engineers will face is binary: build your own constellation from scratch, or integrate with the one that already exists.

There is no third option.

This is what the Artemis 2 blackout revealed, underneath the human drama and the live footage and the clean tweet. The next space race isn’t being run between rockets. It’s being run between infrastructure layers. And one company, almost incidentally, while connecting rural customers, is winning it.

NASA’s silence wasn’t negligence. Orion was designed in a post-Challenger, post-Columbia institutional context where every added system is a potential failure mode, where “flight-proven” means something very specific, where an untested competitor’s satellite technology isn’t an upgrade — it’s a liability. The agency didn’t miss the Starlink solution. It couldn’t structurally adopt it even if it had wanted to.

That’s not a failure. That’s what happens when two organizations with fundamentally different risk architectures and different time horizons try to solve the same physical problem.

One of them, by accident, already owned the answer.

Where This Analysis Has Its Limits

I’m drawing a long strategic line from a single demonstrated capability at Mach 25 to a claim about infrastructure dominance in scenarios — Mach 32, deep space reentry — that haven’t been tested yet. That extrapolation might be doing too much work.

The competitive landscape could also change in ways this argument doesn’t fully account for. The EU is investing in its own satellite constellation. China’s Guowang project is real and moving. Regulatory friction on Starlink’s global licensing has been a consistent pressure point across multiple markets. Infrastructure advantages have been broken before — usually by a combination of political will and patient capital.

What I’m more confident in: whoever solves continuous communication during high-velocity reentry will set the terms of dependency for every crewed spaceflight program on Earth. I’m less certain than this article probably implies that the entity will be SpaceX.

For six minutes on April 11, 2026, four people crossed through fire while the people who loved them waited with nothing to do. That has always been part of what human spaceflight costs.

The question the next decade will answer is simpler and colder: when the technology finally exists to make that silence optional, who owns it?

And what will they charge to end it?

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

The room in London was not a keynote. No livestream. No countdown clock. Just John Ternus — Apple’s head of hardware, and soon its CEO — standing in front of journalists to present a laptop.

The MacBook Neo. Thin, fast, priced aggressively for an Apple product. And running, at its core, on a chip originally designed for an iPhone.

What was strange about the moment wasn’t the product. It was the man presenting it. Ternus knew, and the room knew, that he was effectively rehearsing for a role that hadn’t been officially handed to him yet. The successor presents the predecessor’s last statement. And that product — an iPhone chip inside a MacBook — contained, without anyone saying so out loud, the entire logic of the succession: here is everything Tim Cook built, compressed into a single device. Here is also exactly why someone else needs to take it from here.

The Last Thing Tim Cook Built Was a Repurposed iPhone Chip in a Laptop.

That sentence is not a criticism. It’s the most precise description of what made Cook extraordinary — and what it means that his tenure ends here.

When Cook became CEO in 2011, Apple was already a transformed company. Steve Jobs had given us the iMac, the iPod, the iPhone, and the iPad. What Apple didn’t have was the operational architecture to deliver those products to the world at the scale they deserved. Cook built that architecture. Not incrementally — obsessively. Exclusive multi-year contracts with TSMC, locking in chip supply years in advance. A manufacturing and logistics network so finely calibrated that Apple could announce a product on a Tuesday and ship tens of millions of units globally within days. An end-to-end vertical integration — GPU, modem, WiFi, memory, storage — that meant Apple didn’t depend on Qualcomm, Intel, or anyone else for what mattered.

No other company on earth can do what Apple does with a product launch. Not Samsung, not Microsoft — which still, with all its resources, routinely ships Surface devices months after announcement. The gap is not about money. It’s about a decade and a half of compounding operational expertise that Cook treated as a first-order competitive advantage, not a back-office function.

The MacBook Neo is the apotheosis of that logic. Apple spent years designing chips for iPhones. Those chips became so capable that they could outperform dedicated laptop processors. The logistics to manufacture, refine, and deploy them at scale already existed. The result: a laptop that competes on price by recycling silicon that was built for something else. It’s elegant. It’s efficient. It’s what happens when operational mastery reaches full maturity.

Clayton Christensen called this the innovator’s dilemma: the companies best positioned on one technology curve are the hardest to move to the next one — precisely because their excellence attaches them to what works. Cook didn’t face the dilemma. He became it.

Having the Best Chip for On-Device AI and Having a Product That Proves It Are Different Things

Here is the strongest argument against this analysis, and it deserves to be stated honestly: Apple’s logistics moat doesn’t disappear in the AI era. It requalifies.

Apple Silicon — specifically its Neural Engine architecture — is arguably the most compelling hardware platform for running AI privately, locally, without sending data to a cloud server. Every time someone uses ChatGPT, their queries go to OpenAI’s infrastructure. Apple’s on-device approach, if executed, would offer something none of its competitors can match: AI that works entirely on your device, visible only to you. The privacy brand that Cook built — starting from a real decision to refuse FBI data requests, hardened into marketing for a decade — could find its most legitimate expression here.

The moat isn’t gone. The moat is waiting.

But there’s a distinction that matters, and it’s the one this succession is really about: having a structural advantage and converting that advantage into a product people actually use are two entirely different capabilities. Apple has the chip. It does not yet have the AI. Apple Intelligence is, as of today, a set of writing tools. The new Siri that was supposed to redefine the assistant category has been delayed long enough that it’s become a running industry joke. The gap between Apple’s potential position in on-device AI and its actual products is not a gap measured in quarters. It’s a paradigm gap.

The irony that doesn’t need underlining: Cook spent years building a privacy brand strong enough to differentiate Apple from Google and Meta. That brand had real meaning — a company that said no to the FBI is a company that means it. Then, in the final years of his tenure, the same Cook adopted a more accommodating posture toward an administration whose relationship to civil liberties is, to put it gently, complicated. The values that were never reducible to marketing turned out to have a political price. This doesn’t erase the legacy. It does reveal something about what happens when a marketing position meets an actual political test.

Ternus inherits both sides of this. The privacy hardware advantage is real. The credibility question is open.

What This Doesn’t Account For

There’s a version of this transition where I’m wrong in an important way.

The Ballmer-to-Nadella comparison feels clean from the outside: operational CEO replaced by technical CEO at the moment a paradigm shift demands invention. But Nadella’s first major bet — Azure — was already in development under Ballmer. He didn’t invent the cloud strategy. He executed it better, faster, and with more conviction. If Ternus has something equivalent sitting in Apple’s pipeline — a hardware-software combination that makes on-device AI genuinely feel different from what Google and OpenAI are doing — then my framing has the causality backwards. The transition doesn’t signal a crisis of invention. It positions the right person to finish what’s already in the drawer.

I don’t know what’s in the drawer. Neither does anyone outside Apple. Cook’s product roadmap extends years past his departure — the supply chain contracts are already signed, the silicon is already ordered. Ternus will present iPhones this fall against a feuille de route Cook authored. Real evidence of a Ternus era, architecturally distinct from what came before, is probably three to five years out.

So the honest version of this analysis is: the structural argument is correct. The timing of when it matters is genuinely uncertain.

The handover that Tim Cook is most famous for isn’t the one happening now. It’s the one Steve Jobs gave him in 2011, with a specific instruction: don’t run Apple by asking what Steve would have done. Run it by asking what it needs now.

Cook followed that advice precisely, and it worked beyond anything Jobs could have predicted. Apple today is not the company Jobs left — it is an order of magnitude larger, more profitable, more deeply embedded in daily life on every continent.

The same advice will pass to Ternus in September. Don’t ask what Tim Cook would have done.

Which means Ternus will have to answer a question Cook never faced: how do you build a product that surprises people from inside the most optimized company on earth?

Execution at Apple scale has never been more perfect. The ceiling, for the first time in fifteen years, isn’t operational. It’s imaginative. Whether that’s a problem or an opportunity depends entirely on what John Ternus has been thinking about for the last three years, standing just offstage, watching his predecessor finish.

He knows what’s in the drawer.

We’re about to find out if it’s enough.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

May 2025. Maricopa County, Arizona. A judge watches a screen.

A man is speaking — his face, his voice, his words. That man, Chris Pelky, died in a road accident in 2021. His sister rebuilt his image from the archives he’d left behind. She called it giving him one last word. Judge Tod Lang watched the video to the end. He called it authentic. He handed down the maximum sentence.

Nobody in that courtroom asked how to validate what they were actually hearing. No legal framework required it. The technology had just convinced a tribunal, and the experiment had become, without deliberation, a norm.

That’s where something changed. Not in a trading room. Not in a lab. In a courtroom.

OpenAI Drew the Line Between the Living and the Dead

A few months earlier, a quieter decision had laid the groundwork.

The Michael Jackson estate — the company managing his brand since his death in 2009 — had sent takedown requests to OpenAI after social media was flooded with videos generated by tools like Sora: Michael Jackson with Tupac, with Kobe Bryant, with Martin Luther King. Videos were realistic enough that millions of users watched them without asking questions. OpenAI didn’t comment officially. But to PC Mag, the company explained its policy.

Living celebrities are protected against deepfakes. Deceased celebrities are treated as historical figures. Free to use.

Read that again, because this is where something important is actually happening.

OpenAI protects the living with the logic of consent. It frees the dead with the same logic, applied in reverse. The difference between the two categories — protected or free to use — isn’t philosophical. It’s contractual. The living can sign agreements. The dead no longer can.

This isn’t an ethics policy. It’s legal risk management.

And this taxonomic decision — who qualifies as a “historical figure,” and from when — was never debated in any legislature. It was never submitted to public consultation. It was published in response to a tech magazine. A private company unilaterally decided where your right to your own image ends. And in most countries in the world, it isn’t entirely wrong.

Michael Jackson Has the Best Lawyers in the World. The Deepfakes Run Anyway.

To measure what it costs to defend yourself in this vacuum, Michael Jackson is the most instructive case study available.

Since his death in 2009, his estate has generated $3.5 billion. In 2025 alone: $105 million. Forbes has ranked him the highest-earning deceased celebrity for the thirteenth consecutive year. The gap between Jackson and Elvis Presley — ranked second — is, in the words of one entertainment attorney, “Michael Jackson, then an enormous void, then everyone else.” The estate controls the music catalog, image rights, merchandise, shows, and licensing contracts. It controls, by its own account, “every pixel.”

And yet.

Jackson deepfakes circulated at scale. The estate sent its takedown requests. OpenAI didn’t comment. The videos kept running.

This is where the strongest counter-argument deserves to be stated honestly: societies have always represented their dead. Busts, portraits, biopics, tributes. The line between a Napoleon and a Michael Jackson isn’t philosophically obvious. And there is a real difference between a corporation exploiting a dead man commercially and a sister giving her brother one last word in a courtroom.

Both of those observations are true. But neither resolves the central question.

The bust of Napoleon doesn’t recommend yogurt. The representation of the dead has always existed — but it was bounded, ritualized, deliberate. What AI adds to that equation is industrial scalability and market logic. The CEO of StoryFile — one of the leading grief tech startups — told reporters he was “absolutely interested” in advertising deadbots: avatars of your deceased loved ones recommending products between childhood memories. This isn’t representation. It’s monetization.

The question isn’t “can we represent the dead” — societies always have. The question is: who decides, in what framework, and who captures the value?

Grief Tech Isn’t Waiting for the Law

In 1935, Walter Benjamin anticipated that mechanical reproduction would destroy what he called the aura of a work — that singular presence, that authority of an original in a specific place at a specific moment. What he didn’t anticipate was that we might one day synthesize the aura itself. Do not copy the work. Recreate the presence, the voice, the authority. And rent it by the session.

The Jackson estate doesn’t sell copies of Michael Jackson. It sells the impression of a presence. What grief tech promises to industrialize for everyone else.

The market is projected to be $80 billion over the next decade. Companies like EternOS and HereAfter AI are building avatars from data you’ve left online — photos, messages, videos — without anyone having asked you. Others offer to create your avatar while you’re still alive, so your loved ones can talk to you after you’re gone. Families have testified that these technologies helped them through grief. That isn’t nothing, and dismissing it would be dishonest.

But meanwhile, the legal framework is structurally obsolete.

In France, the Cour de cassation ruled in 2018 that the right to one’s image extinguishes with the person. Heirs don’t inherit it. Remedies exist — Article 226–8 of the penal code on montages without consent, the 2024 SREN law — but all of them were written before a deepfake could be generated in five minutes on a phone. In the UK, Australia, and most liberal democracies, the structure is identical: either post-mortem image rights don’t exist, or they’re too fragmented to be operational in the age of AI generation. This isn’t an American legislative lag that Europe has already solved. It’s a structural vacuum that is global.

And companies whose servers sit in San Francisco are filling it with their own rules.

The Arizona case is the logical outcome of that progression. An AI avatar influenced a prison sentence. Two readings are possible — a family was able to give voice to someone who could no longer speak; a court allowed an imitation technology to weigh on a judicial decision with no established framework for validating what it was hearing. Both are true at the same time. Also true: the technology had just left the realm of the experimental. It was convincing the judges. It was filling courtrooms. It was generating billions.

The legal void isn’t a bug in this system. The void is a feature. Every legal clarification is a compliance cost, a commercial perimeter limitation. OpenAI didn’t decide the dead are “historical figures free to use” out of philosophical conviction — it decided that because it’s the only policy that lets the company avoid creating unfavorable legal precedent while continuing to operate. The void is a feature.

Where My Own Argument Breaks Down

I can’t cleanly draw the line between Napoleon and Michael Jackson. I can say that market logic draws it in the worst possible way — where value is highest, not where dignity demands it. But I don’t have an alternative formula to offer.

I also don’t know what to do with the Arizona case. Chris Pelky’s sister did something moving and sincere. The outcome — a prison sentence influenced by a technology with no validation framework — still unsettles me. Both of those things coexist, and I have no clean way to reconcile them.

What I do know: an $80 billion market won’t wait for us to try.

You have, right now, in all likelihood, between 200 and 2,000 photos of yourself on a phone. Voice messages. Videos. Within a decade, that data will be sufficient to reconstruct a version of you that speaks, responds, and resembles.

Death has always had an economy — funeral homes, cemeteries, monuments. But it was an economy of disappearance. What’s being built now is an economy of permanence: the idea that dying no longer means fully leaving the world of the living, that your image, your voice, your personality can keep circulating, keep convincing, keep selling.

Your consent to all of this has an expiration date. Nobody asked which one you’d prefer.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

A two-year-old in a kitchen. She grabs her spoon, lets it go. It falls. She does it again. Then a third time. On the third drop, something changes in her expression — not surprise, confirmation. She just learned gravity. No words. No training data. No GPU cluster. Three spoons on the floor.

This is the thing that keeps Yann LeCun up at night. Not the child learning physics in thirty seconds — but the fact that the most sophisticated AI systems we’ve ever built, fed on the entirety of human language, cannot do what she just did.

LLMs Learned to Describe Falling. They Never Learned Gravity.

Ask any current language model — GPT-4o, Claude, Gemini — about a coffee cup balanced on the edge of a table, and it will answer you brilliantly. Newton. Gravity. Center of mass. The answer is fluent, accurate, and completely hollow.

The model doesn’t calculate the fall. It recites the script of the fall.

This is the Moravec paradox, and it’s more than a curiosity. Hans Moravec, the roboticist, observed in the 1980s that the tasks we consider intellectually hard — chess, calculus, legal reasoning — are easy to automate, while the tasks we consider trivially easy — picking up an object, walking across a room, understanding that a ball rolling into a street might be followed by a child — remain extraordinarily difficult.

What this tells us isn’t primarily about machines. It’s about us. We don’t understand how we do the easy things. We’ve never had to. Consciousness sits on top of an entire physical layer of competence we never examine because it works invisibly.

The philosopher Maurice Merleau-Ponty identified this sixty years ago, when AI didn’t exist as an industry. Intelligence, he argued, is embodied before it is symbolic. A blind person using a cane doesn’t perceive the cane as an object — they perceive the world through it. The tool incorporates. Understanding comes from contact with reality, not from descriptions of it.

Current language models have read every description of every cane ever written. They have no idea what touching the ground feels like.

The industry’s response to this gap has been architectural ambition scaled to theology. More data. More parameters. More servers. Elon Musk is floating the idea of putting entire gigawatt data centers in orbit. The logic: if we push far enough along the same curve, the understanding of the world will eventually emerge from the manipulation of words. More is different.

LeCun’s argument — made publicly and repeatedly, including during the years he spent leading AI research at Meta — is that this is an expensive detour. Intelligence, he insists, isn’t mastery of language. It’s mastery of causality. The ability to simulate, internally, what happens when you push something, let something go, or change direction. Not to predict the next word in a sentence, but to predict what happens next in the world.

What LeCun Built While Arguing It Shouldn’t Be Built

Here is the irony the industry doesn’t discuss directly: for twelve years, while LeCun was the most credible public voice warning that LLMs were a dead end, he was also the head of AI research at Meta. Meta, during those same years, built LLaMA — one of the most downloaded, most deployed open-source language models on the planet.

The man most convinced that the current paradigm leads nowhere spent a decade making it as good as possible.

You can read this two ways. Either it demonstrates intellectual honesty — doing your job well while disagreeing with the direction. Or it tells you something about how paradigms work: even the people who see the wall most clearly still run toward it, because that’s where the infrastructure, the talent, and the funding are.

In March 2026, LeCun’s Paris-based startup raised over one billion dollars on a valuation of 3.5 billion. The same month, his team published a paper introducing a world model built on JEPA — Joint Embedding Predictive Architecture. The numbers attached to this model read like a misprint.

Fifteen million parameters. One GPU. A few hours of training.

For comparison: current frontier LLMs operate at the scale of hundreds of billions of parameters, require thousands of GPUs, and consume training budgets that rival the GDP of small countries. The world model uses approximately 200 times fewer tokens. It plans a physical action roughly 48 times faster than current generative architectures.

The technical mechanism that makes this possible is called the CSIG regularizer — a constraint that prevents the model from taking the shortcut every AI system will take if you let it. Without external pressure, a model minimizing prediction error will simply represent everything the same way. A falling ball and a passing car end up encoded identically. Prediction is perfect on paper; understanding is zero. The CSIG forces the model to maintain distinct representations — to actually distinguish between objects, forces, and trajectories. To earn its predictions.

The world model doesn’t learn by reading. It learns by watching — raw video, unlabeled, and predicting what comes next in a compressed latent space, a world of concepts rather than pixels. It doesn’t reconstruct the ball; it models the trajectory. It gets it wrong, adjusts its internal model, and gets it less wrong. In a few hours of self-supervised learning on a single machine, it appears to deduce basic physical laws from observation. Nobody taught it gravity. It found it.

This isn’t magic. It’s the difference between reading about swimming and getting in the water.

The best counter-argument is worth stating honestly: this is a proof of concept validated on simple physical scenarios. Balls rebounding. Basic trajectories. The leap from “understands that a ball falls” to “infers that a child might follow a ball into traffic” — the scenario LeCun himself invokes — remains to be demonstrated at scale. The billion-dollar raise is partly an acknowledgment that the road from elegant proof to robust real-world system is long and expensive. LeCun is not claiming to have built a brain. He’s claiming to have demonstrated that you can build one without a trillion-dollar server farm.

This counter-argument, though, actually strengthens the thesis in an oblique way. The entire defense of the current paradigm is that scaling keeps working — that the next version always surprises the critics. But even if that’s true, the energy cost of each surprise is compounding. A model that achieves comparable physical reasoning at a fraction of the infrastructure cost doesn’t need to win on philosophy. It wins on the electricity bill.

When the Energy Argument Becomes the Paradigm Argument

Step back from the architecture and look at the geography.

The world model is Parisian. LeCun is French. The raise is European. This is not a footnote.

For most of the past decade, the implicit map of AI has had two centers of gravity: San Francisco and, increasingly, Beijing. A paradigm shift originating in Paris — in a regulatory environment fundamentally different from Silicon Valley’s, with a different relationship to data privacy, state investment, and long-term research culture — changes something about who gets to define what intelligence means and what it’s for.

For researchers in London, Singapore, or Toronto, the world model represents something beyond its technical claims: proof that the next foundational move in AI doesn’t have to emerge from the same geography as the last one. That’s not a minor cultural point. The assumptions baked into dominant architectures reflect the assumptions of the institutions that built them. A different origin story produces different defaults.

The deeper systemic implication, though, is this: every timeline for AGI currently circulating — every confident prediction about when AI systems will achieve general competence — is built on LLM scaling curves. If LeCun is right that scaling is the wrong dimension, those timelines are not just optimistic. They’re measuring the wrong thing.

We have spent several years watching the AI industry construct increasingly confident futures predicated on a single architectural bet. The world model doesn’t disprove that bet. But it demonstrates, for the first time with empirical weight, that an alternative exists — one that learns causality instead of correlation, that simulates instead of predicts, that runs on a laptop instead of a small power plant.

The question now isn’t whether LeCun’s approach will scale. It will have to — the one-billion-dollar raise ensures that serious resources will pursue that scaling. The question is what happens to the hundreds of billions already committed to the other direction.

Sunk cost is a powerful force. It tends to lose, eventually, to physics.

Where This Analysis Might Be Incomplete

I’m reasoning here from a paper, a funding announcement, and years of LeCun’s public arguments — not from a system deployed in the real world. The history of AI is littered with architectural revolutions that scaled beautifully to toy problems and broke against complexity. The CSIG regularizer solves representation collapse at 15 million parameters; whether it holds at 15 billion is an open question LeCun’s team is being paid a billion dollars to answer.

I’m also assuming a sharper discontinuity between paradigms than may be warranted. The industry is not static. Multimodal training — combining language with video, sensor data, and physical simulation — is actively trying to give LLMs the grounding LeCun says they lack. The gap between “reciting the script of a fall” and “modeling the fall” may close in ways that don’t require abandoning the current architecture entirely.

What I’m confident in is narrower than what’s easy to claim: this is the first proof of concept I’m aware of that makes LeCun’s thesis testable rather than theoretical. That’s different from him being right.

The two-year-old in the kitchen doesn’t know she’s done something extraordinary. She’s already moved on to the next experiment — testing what happens when she pushes the cup instead of dropping the spoon.

That iterative curiosity about physical consequence is what LeCun is trying to build. Not an encyclopedia. A simulator.

If he succeeds, the most expensive assumption in technology history turns out to have been the simplest one: that intelligence lives in words.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share

The Door No Robot Could Open

On March 14, 2011, three days after the earthquake, Japanese engineers sent their robots into the Fukushima Daiichi plant. This wasn’t experimental hardware — it was the best the planet had to offer. iRobot’s PackBot, built for disaster zones, got stuck at the first staircase. Another lost its signal within minutes due to radiation. A third couldn’t turn a handle.

The worst nuclear disaster since Chernobyl wasn’t made worse by the absence of sufficiently advanced technology. It was made worse by the absence of a robot capable of opening an ordinary door. The spaces inside the plant had been designed for humans. Staircases, corridors, valves you turn by hand. A machine without arms and legs couldn’t get in. And no suit could protect a human long enough to do the job.

That image — a door handle, impassable — is what DARPA engraved into its 2012 diagnosis. We don’t need a better drone. We need a machine with arms and legs, capable of operating in spaces built for us. It launched a Robotics Challenge. Boston Dynamics competed. Atlas was born from that.

What happened next is one of the most instructive stories in modern technology — not because a humanoid robot is entering mass production for the first time, but because the answer to why now, why them is somewhere other than where people usually look.

Google Funded the World’s Best Robotics Lab. Then Couldn’t Make a Hinge.

In 2013, Google acquired Boston Dynamics. Andy Rubin, the creator of Android, led the charge. The vision was enormous: build the next platform after smartphones. Within months, Google acquired several robotics companies for more than half a billion dollars. Elite teams. Foundational patents. The best robot footage ever filmed.

And then: nothing.

Four years later, Google sold Boston Dynamics to SoftBank for a fraction of what a real valuation would have demanded. The official reasons: culture clash, image problems, and timelines too long. The real reason is simpler and more embarrassing. Google didn’t have a factory. No team is capable of taking a physical product from prototype to thousands of units. No supply chain for mechanical components. The company that organized access to all of human knowledge had no access to the most basic knowledge of industrial production: how to move a joint from R&D to mass manufacturing.

SoftBank repeated the mistake on a larger scale. Masayoshi Son had the biggest tech fund in history — one hundred billion dollars. The strategy: bet on companies that will dominate entire markets for the next thirty years. Uber, WeWork, Arm. SoftBank knew how to write checks and wait for returns on investment. It didn’t know how to run a production floor.

Clayton Christensen named this mechanism twenty years ago. A technology only becomes a product when it is mobilized to solve a specific job that is painful enough for someone to pay for. Google was looking for an abstract platform. SoftBank was looking for a growth narrative. Neither had a job. Boston Dynamics was lab technology in search of a problem urgent enough to justify scaling.

The problem existed. Nobody in Silicon Valley thought to call a car manufacturer.

The Part Hyundai Was Already Making

In 2021, Hyundai Motor Group acquired SoftBank’s majority stake in Boston Dynamics for $880 million. The tech press read it as a risky bet by an automaker entering an industry it didn’t understand. It was the exact opposite.

Here’s what Boston Dynamics’ competitors don’t want you to know: the most expensive part of a humanoid robot is neither the processor, nor the cameras, nor the software. It’s the actuators — the electric joints and drives that move the limbs, manage grip force, sense resistance, and absorb shocks. According to Hyundai Mobis, they account for more than 60% of a humanoid’s total material cost. They’re what determines whether a robot can work eight hours on a factory floor without breaking down. It’s where every competitor — Figure, Unitree, 1X Technologies — has been hitting a wall for years.

Hyundai Mobis, the group’s components division, has been manufacturing electric power steering systems for fifteen years. Millions of units. Global supply chains. Automotive reliability standards — meaning: engineered to last 200,000 kilometers without failure. The architecture of a robot joint actuator and that of an electric power steering system are cousins: same basic configuration, same electric motor, same gearbox, same sensor, same controller.

Hyundai didn’t invent the critical part. They realized they were already making it by the millions.

That’s the moment the story pivots. Google failed without a factory. SoftBank failed without manufacturing. Hyundai succeeded because the necessary part had been in its own catalog for a decade and a half, and because it had a factory in Georgia that needed filling.

How Industrial Transitions Actually Get Won

The new electric Atlas — no hydraulics, no leaks, no constant monitoring — is now in production. The entire 2026 capacity is reserved. Two buyers: Hyundai Motor Group, deploying the robots at its Metaplant facility in Georgia for validation before assembly line integration in 2028, and Google DeepMind, acquiring units to train its Gemini Robotics models.

That second buyer is worth pausing on.

Google sold Boston Dynamics because it couldn’t manufacture it. In 2026, Google DeepMind is buying Atlas units from Hyundai — the company Google sold the competency to — to fund its own AI robotics research. Google is now paying for access to the expertise it abandoned because that expertise needed a factory to exist. The loop is closed.

This pattern isn’t unique to robotics. TSMC didn’t invent the best chips — it mastered manufacturing them at a precision nobody else can reproduce. Microsoft didn’t invent the best LLMs — but it had Azure, the enterprise contracts, and 365 million Office users to deploy them while OpenAI was still searching for a business model. In every major technology transition, final value gets captured where there is pre-existing productive infrastructure, not necessarily where there is the most brilliant vision.

The manufacturing labor shortage Hyundai is trying to solve in Georgia is the same problem closing plants in Saxony, paralyzing production lines in the Po Valley, and pushing Japan to automate for twenty years out of pure demographic necessity. This is not an American question. The question now isn’t “are robots going to replace jobs” — it’s “who controls the actuators that will run the next global industrial economy.” For now, the answer is Korean.

Where This Reading Breaks Down

I should be honest about two places where I might be wrong.

The first: in 2026, Atlas had only two buyers — Hyundai and Google. This isn’t an open market — it’s vertical integration. Every major industrial transition starts this way, yes. But the real test — an outside customer, in an unknown environment, over ten years without serious failure — hasn’t happened yet. It’s possible that reliability proven under automotive conditions doesn’t transfer as cleanly to robotics as the engineering suggests.

The second: Tesla hasn’t played its hand yet. If Optimus arrives at $20,000 with decent reliability in 2027 or 2028, the economic equation for Atlas at $200,000 changes radically — even with total cost of ownership as a variable. Hyundai’s advantage is real today. It isn’t necessarily permanent.

What I believe firmly: the direction is set. The valid arguments are about speed and who wins — not about whether this transition happens.

Fukushima posed the question in 2011. It took fifteen years, four owners, and the discovery that an electric power steering system and a robot joint share the same architecture to find an industrially viable answer.

The most brilliant technology goes nowhere without the right problem to solve and the right part to solve it with. It’s a lesson Silicon Valley will keep learning — one robot at a time.

Thanks for reading! Want to receive my next articles directly in your inbox? Subscribe! And don’t forget to hit the share button to help spread the word.

Share