The AI Bubble Can Burst Without AI Failing

What began as a question about the AI bubble turned into a research piece in five parts: through China and distillation, collapsing token prices, the Jevons effect, Europe and Mistral, Japan and Korea, the welfare state — all the way to AGI and the question of when enough is enough.

A technological revolution can still be a financial bubble

The debate about a possible AI bubble is usually reduced to two extremes.

On one side stand the optimists. They believe artificial intelligence will transform almost every industry, sharply raise productivity and eventually trigger a new economic revolution.

On the other side stand the pessimists. They point to the hundreds of billions being invested in chips, data centers and power supply for AI companies that barely make a profit. In their view, expectations, investments and stock valuations are being inflated faster than actual revenues.

Both positions probably contain part of the truth. AI can become a foundational technology while, at the same time, too much money flows into the wrong companies, overpriced infrastructure and unsustainable business models. The success of the technology does not guarantee that every investor, chipmaker, model developer or data center operator earns an adequate return on it.

The most important question is therefore not just:

Is AI real?

The more relevant question is:

Can the future revenues from AI justify today's investments and valuations?

That distinction is essential for understanding the AI bubble.

The central distinction

The technology can win while a large part of the capital loses.

Why it matters: Railways, telecom networks and the internet all created lasting infrastructure while destroying badly timed capital along the way. AI may follow the same pattern.

The fear, concretely: what if the subscribers churn?

The bubble fear has a very concrete shape. The business model of the big AI labs rests on two pillars: monthly subscriptions from consumers and professionals (twenty to two hundred dollars a month) and API revenue from businesses. Against that stand investments of an entirely different order: according to the International Energy Agency, the largest tech companies spent more than 400 billion dollars on data centers in 2025 alone, with roughly another 75 percent growth expected for 2026.

The doom scenario is easy to retell. A recession arrives, or simply subscription fatigue. Consumers are the first to cancel the AI subscription they mostly used for emails and recipes. Businesses discover that a cheap open-weight model does ninety percent of their tasks just as well. The labs' revenue growth stalls while the depreciation on chips and the energy bills keep running. Financing dries up, the weakest players collapse, and the dominoes — chip suppliers, data center builders, utilities with AI contracts — fall with them.

That fear is amplified by the circular money flows in the sector: chipmakers investing in AI labs that use the money to buy chips, cloud providers issuing revenue guarantees to customers who rent capacity with borrowed money. Such constructions are not fraud by definition — but they do make the sector more sensitive to a confidence shock, because one falling link hits several balance sheets at once.

Two nuances belong to this story.

First: the paying consumer is not the whole market. The bulk of expected AI revenue sits in business processes — coding agents, customer service, document processing, research. That demand is less cyclical than a consumer subscription, precisely because AI replaces labor costs there: in a recession, automation becomes more attractive, not less.

Second: that subscribers will eventually churn at current prices is almost certain — and that need not be a disaster. As we will see later, the price of a fixed level of intelligence is falling by roughly a factor of ten per year. The question is not whether AI at twenty dollars a month is sustainable, but whether the labs can move fast enough toward a world in which the same service costs two dollars and the volume is a hundred times larger.

To weigh whether that will succeed, we first need to understand who is driving this price war — and why. That story begins in China.

Why China changed the conversation

For a long time it seemed the United States held an almost untouchable lead. American companies had the best chips, the largest cloud platforms, the best-capitalized AI labs and the strongest models.

The rapid rise of Chinese models such as DeepSeek, Qwen and Kimi changed that picture. According to the Stanford AI Index 2026, the performance gap between the best American and Chinese models has all but disappeared; since 2025 they have traded places several times. The United States still produces most of the internationally leading frontier models, while China is strong in publications, patents, industrial robotics and open models.

That does not automatically mean China has won the AI race — because there is no single AI race. Several contests are running at once:

• who builds the most intelligent frontier model;
• who delivers AI at the lowest price;
• who has the best chips;
• who can build the most energy and data center capacity;
• who develops the best robots and industrial applications;
• who builds the strongest ecosystem of developers and companies;
• who can deploy AI into the real economy fastest.

China has become strong above all in efficiency, open-weight models and rapidly turning research into competitive products. The United States retains major advantages in advanced chips, global cloud infrastructure, investment capital and much of the frontier research. The competition has become far more even — but it has not been decided.

What is model distillation?

One of the most important topics in this debate is model distillation.

In distillation, a strong and often large model is used as a teacher for a smaller or cheaper model. The strong model generates answers, solutions, code and reasoning; those outputs then serve as training data for the student model. OpenAI even offers distillation as a legitimate technique within its own platform. It is not a dark trick but a normal machine-learning method.

It becomes controversial when a company collects the answers of a competing closed model at scale to train its own competitor with them. OpenAI's terms of service prohibit using outputs to develop competing models.

In February 2026, Anthropic reported that it had discovered large-scale distillation campaigns which, according to the company, were linked to Chinese AI developers — using large numbers of accounts and millions of interactions to harvest Claude's capabilities. This is Anthropic's own analysis and accusation and should be read as such, but it illustrates how valuable access to a frontier model is: a competitor does not have to discover every skill itself, it can harvest the knowledge, reasoning patterns and post-training already inside the teacher model. That accelerates development considerably.

Are Chinese models mostly copies?

The claim that Chinese models merely copy American ones is too simple.

According to its technical report, DeepSeek-V3 was pre-trained on 14.8 trillion tokens, with a Mixture-of-Experts architecture of 671 billion parameters of which only 37 billion are active per token. DeepSeek reported roughly 2.8 million H800 GPU-hours for the full training run and developed its own optimizations along the way: Multi-head Latent Attention, DeepSeekMoE, efficient load balancing, multi-token prediction, hardware-software co-design and reinforcement learning for reasoning.

DeepSeek-R1 also showed that reasoning ability can improve sharply through reinforcement learning, even without human examples for every intermediate step. Distillation does play an important role — especially in transferring reasoning to smaller models.

The correct conclusion is therefore:

Chinese companies use distillation as a powerful accelerator, but combine it with their own pre-training, their own architectures, reinforcement learning, engineering and large domestic compute clusters.

Without external distillation they would probably move more slowly. But they would not suddenly stop developing new models.

Can OpenAI and Anthropic block distillation?

They can make it harder — but probably not impossible.

AI companies can demand stricter customer identification, set request limits, detect coordinated accounts and suspicious query patterns, block intermediaries and API resellers, watermark outputs, tighten legal terms and shield very detailed reasoning traces.

The fundamental problem: a distillation request often looks exactly like a normal request. "Solve this math problem" or "write and test this code" are perfectly ordinary tasks — but collected systematically and at enormous scale, the answers form a training dataset. An attacker can spread requests across thousands of accounts, countries and points in time, and paraphrase or rewrite answers before using them as training data.

The realistic goal is therefore not to make distillation impossible, but to make it more expensive, slower, riskier, more detectable and legally harder. That can matter strategically: without easy access to the latest behavior of Claude or GPT, it takes more time and compute to develop comparable capabilities independently — especially for new skills such as agentic behavior, coding and complex reasoning. It will slow China down, not switch it off permanently.

Why China can keep going without American models

Chinese AI companies now have alternative routes. They can distill from their own largest models, combine outputs from multiple Chinese and open-weight models, and automatically verify answers for math and programming: code can be tested, a proof can be checked, an agent can be judged on whether it actually completed a task. Models can generate millions of attempts and use only the successful ones for new training.

Beyond that, much of the crucial AI research is simply public. Architectures, reinforcement-learning methods, datasets and evaluations are shared in papers and open-source projects. Blocking Claude and GPT cuts off valuable information — not the world's AI research.

China's most important constraints thus sit less and less in the algorithms and more and more in: access to the latest chips, fast interconnects, high-bandwidth memory, scale of compute, reliable software infrastructure, energy supply and manufacturing capacity for advanced semiconductors.

The battle for AI is therefore as much a battle for factories, energy, networks and materials as a battle for algorithms.

Does China win by making America foot the bill?

Everything above converges in one popular reading: China is already winning this contest — not by being better, but by being smarter with other people's money. The reasoning: American labs pay the billions for frontier research and gigantic training runs. Chinese companies watch along via public papers and (partly) via distillation, build comparable models for a fraction of the cost and then give them away almost for free as open-weight models. In doing so they demolish the margins with which the Americans were supposed to earn back their investments. America carries the cost of innovation; China harvests the commoditization.

There is something to it. DeepSeek, Qwen and Kimi have demonstrably put price pressure on the entire market, and an open-weight model you can run yourself is the ultimate bargaining chip against a closed API. For Chinese players this is also rational strategy: if you cannot win the frontier because of American chip restrictions, you can still make it worthless.

For companies that expect to charge high prices for model access for a long time, this is a serious risk. If an open-weight competitor delivers comparable quality for a fraction of the price, the margins of closed providers come under pressure. That can lead to commoditization: the model itself becomes a cheap commodity, while value shifts to applications, business data, distribution, cloud infrastructure, energy, robots, integration with real business processes, and trust and safety.

But as a winning strategy the story is too neat. Three caveats:

• China pays too. Chinese labs run the same price war on their home market, earn almost nothing on their models and struggle with expensive, scarce compute. Whoever pushes the price to zero pushes their own revenue to zero as well.
• Free models grow the American market too. Cheap intelligence drives usage — and much of that usage still runs on American chips, American clouds and American energy infrastructure. A company that could not deploy an expensive AI agent profitably may well be able to at 90 percent lower cost.
• The frontier stays valuable. As long as the best model runs six to twelve months ahead and that genuinely matters for agents, coding and research, a price premium remains for the leader. Commoditization mostly hits the mid-market.

Chinese efficiency thus causes two effects at once: lower prices and margins per AI task, and far more AI tasks in a much larger total market. Which effect dominates differs per company and per market.

The honest conclusion: China has structurally lowered the margins of the American AI industry and thereby sharpened the bubble question — but it has not won the race by doing so. It has made the race more expensive for everyone.

What is recursive self-improvement?

A next step in AI development is recursive self-improvement (RSI). The term conjures the image of an AI that autonomously builds a more intelligent version of itself, which then builds an even better successor — a self-accelerating intelligence explosion. That full scenario has not been publicly demonstrated. Milder forms already exist.

Level one: improving an answer

A model drafts an answer, checks it and tries again: reflection, self-critique, best-of-N, majority voting, self-consistency, running and repairing code. The model temporarily gets better at one task, but its weights do not change.

Level two: automated training loops

An AI system can generate new training problems, produce multiple solutions, have them checked automatically, select the successful ones, train a new model on them — and repeat. Chinese and American labs already use parts of this. DeepSeek-R1 developed reasoning through reinforcement learning. Google DeepMind built AlphaEvolve, which can propose, test and improve new algorithms; some improvements touched processes that were also relevant to the models behind AlphaEvolve itself.

This is an improvement loop — but humans still set the goal, the environment, the evaluation, the budget, and which changes actually get applied.

Level three: full recursive self-improvement

Under full RSI, an AI would autonomously determine its own limitations, which architecture works better, which data and experiments are needed, how its successor should be trained — and how to improve that entire process again. Anthropic describes this as a plausible future development, not as something already achieved.

The current situation sits between advanced automation and limited self-improvement: AI already helps with AI research, but does not yet demonstrably control the full development chain on its own.

Can Chinese companies use recursive self-improvement?

Yes. There is no fundamental technical reason why only American companies could build such loops. Chinese labs have strong foundation models of their own, reinforcement learning, synthetic data generation, model-as-judge systems, code and math verifiers, open scientific knowledge and large clusters.

What is probably still missing — in China and in the United States alike — is a fully reliable AI researcher that can independently lead the development of a successor frontier model. The core problem is verification: how does the system know a proposed improvement is genuinely better? For math and code that is sometimes easy; for general intelligence, reliability, creativity and long-term planning, judgment is much harder. And a system that judges itself can learn to fool the evaluation instead of genuinely improving — reward hacking.

RSI is therefore not just a matter of more intelligence. It requires reliable evaluations, strong safety boundaries, sufficient compute and systems that do not keep amplifying their own errors.

Will self-improvement drastically lower the cost of frontier models?

Probably yes — for a fixed level of performance. AI can help researchers select better training data, design more efficient architectures, write better GPU kernels, catch training errors earlier and run fewer hopeless experiments. A model that costs billions today may be reproducible for a fraction of that in the future.

But that does not mean AI companies will spend less. Whoever can reach the same model quality for one billion instead of ten has two options: save nine billion — or spend ten billion anyway on a far more powerful model. Frontier labs almost always choose the second.

More efficient AI does not automatically mean less AI investment. More efficient AI pushes the frontier further out.

Chips get faster and models get more efficient

The compute needed for one token or one task will almost certainly keep falling — through Mixture-of-Experts, sparse attention, quantization, distillation, speculative decoding, caching, specialized small models, better chips, faster memory and hardware-software co-design.

NVIDIA claims its Rubin platform can deliver up to ten times lower inference cost per token than Blackwell for certain workloads. Those are vendor claims — but they illustrate the pace. At the same time, not every task needs the largest model: a router sends easy questions to a small model and only the hard ones to the frontier.

How fast this is going is now well measurable. In March 2023, GPT-4 cost about 60 dollars per million output tokens. By mid-2026, several providers deliver that same intelligence class for roughly 40 cents — a decline of over 99 percent in just over three years, or roughly a factor of ten per year. If the trend continues even at half that pace, "GPT-4 level" will cost mere pennies in 2028.

The price of GPT-4-level intelligence is collapsing

API price per million output tokens for GPT-4-class models, in USD. Logarithmic scale. Dashed line: forecast at roughly 10x decline per year.

The price per unit of intelligence is collapsing. But note what the chart does not say: it says nothing about total spending. Total spending is rising — because usage is growing even faster than the price is falling.

Less energy per task does not mean less total energy

A common fallacy is that more efficient chips and models automatically lead to lower total power consumption. Suppose an AI task costs one euro today and is therefore run one million times. If the same task later costs one cent, usage can increase a hundredfold or a thousandfold. AI then no longer answers the occasional question but is continuously active: as a programmer, research agent, customer service worker, robot controller — a digital team that works day and night.

This is the classic Jevons effect: when a technology becomes cheaper and more efficient, total consumption can rise.

Energy per AI taskTotal AI energy useTotal in breakthrough scenario

The Jevons effect: down per task, up in total

Illustrative index (2024 = 100). Dashed line: breakthrough scenario in which recursive self-improvement plus the hardware roadmap cut costs roughly 100x and push consumption toward the physical ceiling.

The dashed line deserves explanation, because this is exactly what a breakthrough would do. NVIDIA's announced roadmap is already an annual cadence — Blackwell, Vera Rubin (2026), Rubin Ultra (2027), Feynman (2028) — with claims of up to ten times lower inference costs per generation. Stack another software factor of ten from recursive self-improvement on top, and the price of intelligence falls not ten times per year, but a hundred times in two to three years. Demand for intelligence has so far proven extremely price-elastic: every price drop led to an even bigger explosion of usage. At a hundred times cheaper, things become viable that are unthinkable today — a permanent agent per employee, per device, per business process. The constraint then shifts from demand to supply: power grids, transformers, permits and memory production set the ceiling, and the price of intelligence converges to the price of energy. Note: an efficiency breakthrough does not resolve the bubble question — it relocates it. From "is there enough demand?" to "who owns the energy and the memory?". Bearish for whoever sells expensive model access, bullish for whoever owns infrastructure.

The International Energy Agency's numbers confirm the picture. In 2025, data center electricity consumption rose about 17 percent — considerably faster than total global electricity demand. The IEA expects consumption to roughly double from 485 TWh in 2025 to some 950 TWh in 2030, about three percent of global power demand. AI-focused data centers grow fastest within that: their consumption triples toward 2030.

Data centers on their way to a doubling

Global data center electricity consumption, in terawatt-hours (TWh). Translucent bars: IEA base case forecast.

This looks contradictory but is logical: every individual computation gets more efficient, and the number of computations grows even faster.

The back-of-the-envelope: the scientist versus the agent

To feel what hundreds of millions of research agents mean energetically, it helps to run a Fermi estimate that is rarely made: what does a human researcher actually cost in energy? Not just the brain — the whole chain. The car to work. The heated office and the lab with fume hoods and freezers at minus 80 degrees. The flight to the conference. The coffee — including cultivation, roasting, transport, the machine, and the van with a driver that comes to refill the machine.

There are roughly ten million full-time researchers worldwide. Per researcher, per year, just the work-related part: about 6,000 kWh for office and lab, about 2,000 kWh commuting, roughly 1,500 kWh in conference travel, and a small 300 kWh for the coffee and canteen chain. Together on the order of 10,000 kWh per year — for about 1,700 working hours. Times ten million researchers: about 100 TWh per year, roughly a fifth of total global data center consumption in 2025. Humanity already spends a data-center-sized amount of energy facilitating thinking — largely on concrete, asphalt and cappuccino.

Office + labCommutingFlightsCoffee chainSurrounding organizationAgent, 24/7

What does a year of thinking cost?

Work-related energy per year, in kWh. The scientist delivers about 1,700 working hours (less than half of them actual research); the agent 8,760. Order-of-magnitude estimate.

The comparison that matters is per hour of thinking. The scientist: 10,000 kWh divided by 1,700 hours is roughly 6 kWh of overhead per working hour. The agent: 0.07 to 0.7 kWh per working hour. Per hour of thinking, an agent is roughly ten to a hundred times more energy-efficient than the logistical overhead of a human researcher alone — never mind that the agent never sleeps, never sits in traffic and needs no conference in San Diego. The coffee deserves special mention: one scientist's annual coffee chain can keep a light agent thinking non-stop for about five months. Lab robots do not change the picture much: an automated lab station draws 1 to 5 kilowatts — per hour comparable to a human's lab share, but it runs 168 hours a week instead of 35.

And this bill is still far too kind to the human, because a scientist does not work alone. There is an assistant, a manager, a department head; there are teaching duties, committees, review cycles and grant applications to prove oneself to institute and university. Around the institute stands an entire apparatus: administration, IT, finance, HR, facilities — and all those people drink coffee themselves, devour sandwiches and pizzas, and commute back and forth every day. Every extra organizational layer brings its own cappuccino chain; the curve is, so to speak, exponential. This is not guesswork: on research grants, universities routinely charge 50 to 60 percent overhead for precisely these indirect costs, and support staff often outnumber academic staff. Add sick leave (about 5 percent) and the restaurant lunch, and the fully loaded bill lands closer to 20,000 to 30,000 kWh per scientist per year. More important still: with all the teaching, meetings and grant writing, a researcher often spends less than half their time on actual research. Per hour of genuine research, the overhead climbs to 30 to 40 kWh — and the agent becomes not ten to a hundred, but fifty to five hundred times more efficient.

Then comes the sting, because Jevons strikes here too. Three hundred million light agents cost about 180 TWh per year — that fits almost exactly into the growth room the IEA projects for AI data centers toward 2030; the forecasts effectively imply room for precisely such an agent army. But three hundred million heavy agents cost about 1,800 TWh — more than the entire projected data center consumption of 2035. Per unit of research, thinking gets ten to a hundred times cheaper, but because we deploy thirty times more of it, the total still rises — until the power grid draws the line. One caveat belongs here: the scientists do not disappear, so their commuting energy is only truly "saved" if behavior changes. For now, agent consumption simply comes on top.

The underrated wildcard: what if AI solves its own energy problem?

All the forecasts above share one silent assumption: that energy stays scarce and expensive, and must be delivered through the existing grid. But if research accelerates on every front — and that is exactly what hundreds of millions of research agents would do — then energy generation itself is just another research problem.

The history of hardware efficiency shows how structurally we underestimate this kind of gain. A color CRT television from the 1970s drew 150 to 200 watts for a 25-inch picture; a modern LED TV shows four times the screen area for less power — per square centimeter of picture an improvement of roughly a factor of ten. The 60-watt incandescent bulb became an 8-watt LED. And for computation itself the trend is even more extreme: under Koomey's law, the number of computations per kilowatt-hour doubled every one and a half to two years for decades — an improvement of many billions of times since the 1950s. Nobody staring at a humming CRT in 1970 predicted we would one day charge a pocket supercomputer with a phone charger.

On the generation side, the same movement has begun. The hyperscalers have become the de facto biggest drivers of new nuclear power: Microsoft had a mothballed reactor at Three Mile Island restarted, Google and Amazon invested in small modular reactors (SMRs) — small, serially producible nuclear plants that fit next to a data center or industrial site — and Microsoft already signed a purchase agreement for fusion energy with Helion. Increasingly, capacity is built "behind the meter": generation right next to the data center, so the power does not have to travel through the overloaded grid and past old-fashioned transmission towers. And AI is already accelerating this research itself: DeepMind used reinforcement learning to stabilize fusion plasma and discovered hundreds of thousands of new crystal structures with GNoME for batteries and superconductors, among other things.

A mini-reactor per household, by the way, is almost certainly not coming — shielding, security and regulation make it unrealistic. But the decentralized idea behind it is already happening by another route: solar panels plus home and neighborhood batteries are step by step making households less dependent on those same transmission towers, and an SMR per data center campus is a serious scenario for the 2030s.

For this research piece, this means a possible feedback loop that could tilt the entire energy story: AI needs energy, AI accelerates energy research, cheaper energy enables more AI. If that loop gets going, the supply limit in the Jevons chart is not a wall but a moving ceiling. Two sober caveats guard the optimism. First, the pace: energy projects run in decades, AI demand grows by the quarter — the squeeze of the next five to ten years is not solved by this wildcard. Second, Jevons itself: the LED made light dirt-cheap, after which we started illuminating everything. Cheap energy will not tame AI consumption but unleash it. Only then the problem is no longer scarcity — and a bubble bursting on expensive energy is a different story from a bubble bursting on missing demand.

Is today's data center buildout a bad investment, then?

Not necessarily. A data center is not just GPUs: it is also land, buildings, grid connections, transformers, substations, fiber, cooling, security, backup power and permits. The chips can be replaced after a few years; the building, the power connection and the infrastructure last much longer. If new chips get ten times more efficient, the same data center can do ten times more AI work.

There is a real financial risk, though. Companies can overpay for grid connections, build in the wrong locations, depreciate hardware too slowly, or sign contracts based on overly optimistic demand expectations. So the AI bubble need not mean all data centers become worthless — it can mean that certain data centers, chips and companies were financed at too high a price.

Data centers in space

One of the most futuristic ideas is building AI data centers in space. The concept has attractive sides: in a suitable orbit, solar panels produce energy continuously, no neighbor objects to a new substation, no drinking water is needed for evaporative cooling, and satellite data can be processed on the spot. ESA is researching technologies for large reusable space structures that this would require.

Still, the technical problems are enormous.

Cooling in space

Space is cold, but a vacuum does not cool a computer: there is no air to carry heat away. Heat must ultimately be radiated away as infrared, and that requires large radiators. A 2026 technical analysis calculated that a representative one-megawatt orbital data center needs about 5,640 square meters of solar panels and 2,500 square meters of radiator surface — tens of kilograms of infrastructure per kilowatt of compute. The study's conclusion: launch and hardware costs must fall much further before space compute can compete economically. It is a model-based preprint, not a commercial installation, but it shows the scale of the problem.

Communication

Frontier training requires extremely fast links between large numbers of chips. On Earth these run through networks with enormous bandwidth and minimal latency; between satellites, distances are larger and links are more limited. For local processing of Earth-observation data that is no problem — for training one gigantic model across thousands of satellites it is.

Maintenance and radiation

Chips in space endure radiation, extreme temperature swings and debris, and repairs are exorbitantly expensive. Space data centers will therefore first serve satellite imagery, scientific space data, communication, defense and local inference. Large-scale frontier training in space first requires breakthroughs in cheap launches, autonomous assembly, radiation-hardened chips and optical communication. Not impossible — but for now no replacement for the clusters on Earth.

What about quantum computers?

Whenever a future breakthrough is speculated about, the word quantum soon appears. The state of play in mid-2026: the first genuine logical qubits exist — Quantinuum demonstrated about 94, QuEra 96, and Google's Willow chip showed that errors genuinely decrease as error correction scales up. But production-ready, fault-tolerant quantum computers are expected only somewhere between 2030 and 2035, and scaling to the millions of physical qubits needed for useful applications is still years away.

More important for this story: a quantum computer is not a faster GPU. Neural networks run on enormous amounts of relatively simple linear algebra over gigantic datasets — precisely the kind of work quantum computers are bad at, if only because loading large amounts of classical data into qubits is a fundamental bottleneck. There will be no "quantum ChatGPT" that suddenly decimates training or inference costs.

Where quantum can touch the AI race is indirect: simulation of materials and chemistry (better batteries, chips and cooling for the data centers of the 2030s), specific optimization problems, and cryptography — the post-quantum migration of all security is already underway. For today's bubble debate, the honest conclusion: quantum is not a deus ex machina that lowers AI costs this decade. It is rather the logical candidate for the next hype cycle, once this one cools down.

And Europe? Poor Mistral — or is it?

In this whole story about America and China, a continent has been missing. That is no coincidence: in the AI race of 2026, Europe is mostly a spectator in the contest for capital.

The numbers are sobering. According to the Stanford AI Index, Europe attracted about 21 billion dollars in private AI investment in 2025. The United States: 286 billion. Even China, which leans heavily on state capital rather than venture capital, sits at 12 billion in private investment — the same order as all of Europe.

The capital gap

Private AI investment in 2025, in billions of dollars. China additionally leans heavily on state capital that stays out of this picture.

Europe's one serious frontier player is France's Mistral AI. "Poor Mistral" is the reflex — and compared to the American war chests, that is fair: Mistral has raised less capital in total than OpenAI burns in an average quarter. But the picture deserves nuance. In September 2025, Mistral raised 1.7 billion euros at an 11.7 billion valuation, with chip-machine maker ASML as its striking lead investor (11 percent of the shares). By mid-2026 the company is in talks about a new round of some 3 billion at roughly a 20 billion euro valuation, revenue grew to more than 400 million dollars annualized, and it targets a billion euros in revenue by the end of 2026.

The alliance with ASML is strategically more interesting than the amount. ASML is the only company in the world that makes the EUV lithography machines with which all advanced AI chips are produced — the real chokepoint of the entire global AI supply chain sits in Veldhoven, the Netherlands. Europe does hold a system-critical trump card; it just never built the layer above it.

Why not? The explanations are familiar: fragmented capital markets, pension funds that barely touch venture capital, energy prices several times those of Texas or Sichuan, and regulation that steers toward risks rather than opportunities. The AI Act — whose heaviest obligations take effect in August 2026 — gives Europe a lead in trust and compliance, but not a single extra data center.

Brussels is trying to turn the tide with the AI Continent plan: the InvestAI program is meant to mobilize 200 billion euros, of which 20 billion for four to five AI gigafactories, each with more than 100,000 advanced AI processors. The interest was there — 76 proposals for 60 sites in 16 member states — but execution is faltering: in mid-2026, Bloomberg reported delays and funding problems that are scaring off potential partners. The pattern is familiar: ambitious targets, slow procurement, and meanwhile European companies simply rent their compute from American hyperscalers.

Yet Europe's position in the bubble scenario is paradoxically favorable. Precisely because Europe barely joined the investment wave, it has little to lose if the bubble bursts. And in the commoditized world China is forcing — in which strong models are cheap or free — value shifts to exactly the things Europe is strong in: industrial application, business data, machine building, and the chip machines themselves. Europe does not have to win the frontier race to profit from cheap intelligence; it mainly has to make sure it can run and apply that intelligence sovereignly.

The realistic European strategy is therefore not "build our own OpenAI", but: embrace Mistral and the open-weight world as suppliers, actually finish the gigafactories for sovereign capacity, treasure ASML as a geopolitical trump card, and push AI into industry faster than anyone else. So Mistral is not poor — it is just playing a different game, with different stakes.

Japan and South Korea: the quiet winners of every scenario

Two countries are strikingly absent from the story about America, China and Europe — and that is not weakness but strategy. Japan and South Korea are not competing in the frontier model race. They have nested themselves in the layer beneath it: the picks and shovels of the gold rush.

South Korea may be the second most important chokepoint in the world, after ASML and TSMC. Samsung and SK Hynix together supply roughly 80 percent of all high-bandwidth memory — the stacked fast memory inside virtually every AI accelerator; SK Hynix alone held a 58 percent market share in early 2026. No Korean memory, no NVIDIA GPU, no American data center, no Chinese cluster. On top of that, in mid-2026 Seoul announced a national investment wave of hundreds of billions of dollars in chips, data centers and robotics led by Samsung, NVIDIA deepened its alliances with SK Hynix, SK Telecom (which is building a gigawatt-scale AI cloud, planned live in 2027), Naver, LG and Hyundai, and Samsung and SK Hynix signed a letter of intent with OpenAI for the eventual supply of 900,000 DRAM wafers per month to the Stargate project. Korea has its own models too — Naver's HyperCLOVA, LG's Exaone — but they play regionally, not at the frontier.

Japan plays the capital and integration card. SoftBank is one of the largest external financiers of OpenAI and the Stargate project; a consortium around SoftBank, Sony, Honda and NEC is working on a Japanese foundation model; and chip project Rapidus is aiming for its own 2-nanometer inference chips around 2028 — though for training, Japan remains dependent on NVIDIA, TSMC and Korean memory for the foreseeable future. Japan's real trump card sits further down the chain: its robotics industry (Fanuc, Yaskawa) is where AI enters the physical economy — exactly the contest that, according to this piece, ultimately matters most.

The strategic logic is that of a casino: do not gamble, run the table. Korea earns in every scenario — whoever wins the model race buys Korean memory. If the bubble bursts, HBM demand takes a temporary hit, but the near-monopoly remains. For Europe this is the most painful mirror in this whole story: Korea and Japan both deliberately chose the infrastructure layer and built a national strategy worth hundreds of billions around it. Europe holds exactly such a position with ASML — but never built the strategy around it.

When agents do the work, who collects the tax?

There is another layer beneath the European problem, and it is fiscal. Suppose the agent economy truly breaks through: companies and individuals build agents that earn serious money, and those agents almost all run on the models and data centers of American — or Chinese — owners. Who collects the tax on all that work?

The answer is more uncomfortable than it seems, because the European welfare state is built on one assumption: that work is done by people in employment. In most European countries, more than half of all tax revenue comes directly or indirectly from labor — payroll tax, income tax, social contributions. Employment taxes alone account for about 24 percent of tax revenue on average across OECD countries, with peaks of 35 percent in Germany and 37 percent in France.

The European treasury leans on payslips

Employment taxes (payroll taxes and social contributions) as a share of total tax revenue.

Now follow the money when an agent takes over a job. The salary — on which Europe levied payroll tax and contributions — disappears. In its place comes a subscription or API invoice to a company in San Francisco. The profit on that agent lands with the owner of the model, and corporate tax is generally levied where the company is domiciled — the United States — or wherever it has parked its intellectual property for tax purposes. What Europe keeps: the VAT on the service (levied where the customer sits) and the corporate tax of the European company that uses the agent. But the wage slice — by far the largest slice — evaporates. In that sense the intuition is right: if Europe sleeps while the work runs through American or Chinese agents, the tax base shifts structurally toward the owners' countries.

Three caveats keep this from becoming too smooth a story.

One: Washington earns almost nothing from it for now. Corporate tax requires profit, and there is barely any. OpenAI lost tens of billions in 2025 and, according to reported projections, does not expect to break even until after 2029–2030; Anthropic is the first frontier lab on course for operating profit (during 2026). Moreover, American companies can carry losses forward almost indefinitely, so even the first profitable years remain largely tax-free. The great tax stream toward the US is future music — and requires, additionally, that the bubble is not a bubble. The doom scenario "subscribers churn" and the scenario "the US gets rich on agent taxes" cannot both be true at once.

Two: "the US" is not automatically the winner — the shareholders are. Big tech is historically excellent at shifting profits to low-tax jurisdictions. A sizable share of agent profits may ultimately land in Ireland, Singapore or the Cayman Islands rather than with the American treasury.

Three: Europe is instrumentally weak here — and that is partly its own choice. The international answer to precisely this problem already existed: Pillar One of the OECD tax deal would partially assign taxing rights to the countries where the customers sit instead of where the company is domiciled. That process has stranded: the US withdrew from the OECD Global Tax Deal by executive order in January 2025, and Washington threatens retaliation against countries levying their own digital services taxes. About half of European OECD countries have introduced or announced such a DST, but it taxes revenue (typically 2 to 5 percent), not profit — a band-aid, not a solution. Until this changes, Europe has hardly any fiscal handle on value that is consumed within its borders but booked outside them.

There is also a counter-voice: think tanks such as ITIF argue that earlier automation waves did not destroy the labor share or the tax base either — work shifted, new jobs emerged, and the gloomy predictions did not come true. That can happen again. But even in that milder scenario the direction stands: the larger the share of work running through agents with foreign owners, the more the fiscal pie shifts from European payslips to profits elsewhere.

Is the welfare state in danger, then?

In the extreme scenario: yes — and the squeeze is double. The IMF estimates that some 60 percent of jobs in advanced economies are exposed to AI. If a substantial share of those actually shifts to agents, the state's revenue side erodes twice at once: the payroll tax disappears with the salary, and people without wages consume less, so the VAT base shrinks too. The two biggest pillars of the European treasury sink simultaneously.

At the same time, the spending side rises. Healthcare and education remain labor-intensive and become relatively ever more expensive as the rest of the economy automates — the classic Baumol effect — with an aging population on top. A shrinking base under a growing welfare state: that is the structural squeeze.

Still, "in danger" is not the same as "lost". The core problem is not scarcity but collection: if agents truly make productivity explode, there is more wealth to distribute than ever — it just lands in places European tax authorities cannot reach. The conceivable answers already exist on paper: shifting the center of gravity from labor to capital and profit taxation, levies on compute or data center consumption, and the Norway model — the state takes stakes in AI infrastructure itself, so that dividends rather than payroll taxes carry the welfare state. The real danger sits in the transition phase: the base can erode faster than new instruments arrive. And the one international instrument designed for this has — see above — just stranded.

For Europe, this makes the stakes of this entire part more concrete. Sovereign AI is not just a matter of pride or privacy — it is ultimately a matter of tax base. Whoever owns the agents collects the profit; whoever merely rents them pays the invoice and loses the payroll tax.

Where are we: AI, AGI, ASI — and what if most things get solved?

Everything in this piece ultimately hangs on one question: how far along are we, and where is this going? Definitions first. AI is what we have now: systems that do specific tasks at or above human level. AGI (artificial general intelligence) is the point where a system does virtually all economically relevant thinking work at least as well as a competent human. ASI (artificial superintelligence) is the stage beyond: better than the best humans at virtually everything, including improving itself — the completion of the recursive self-improvement from Part II.

The state of play in mid-2026

In narrow domains, humans have already been passed: math olympiads, competitive programming, protein folding. The most interesting yardstick on the road to AGI is the task horizon tracked by research group METR: how long can a task be before an AI only completes it half the time? That horizon doubled every seven months from 2019 to 2024, and every four months since 2024; the best models now sit around five hours of autonomous expert work. Extend the line, and within a few years systems will handle tasks that take humans weeks. What is still missing for AGI is well known: reliability over long horizons, continual learning and memory, and the physical world. Predictions vary accordingly: experts put the median around 2030 for systems that autonomously complete 80 percent of eight-hour software tasks, superforecasters say 2028, the general public 2037 — and for full ASI most serious estimates point to the 2030s. One thing is near certain: AGI will not be a day but a gradient; we will keep arguing about whether we have arrived while the economy is already being transformed by it.

What it means for humanity

The economic consequences are the through-line of this piece: the labor and tax story, the welfare state, the question of who the owners are. On top of that comes the scientific acceleration — the "decade of progress in a year" that Anthropic CEO Dario Amodei, among others, writes about: hundreds of millions of research agents on diseases, materials and energy. And beneath it lie the risks that grow with the capability: alignment (does a system smarter than its makers still do what its makers intended?), concentration of power (the tax chapter times a thousand: whoever owns ASI owns the economy) and misuse. If you take the optimistic and the grim scenario seriously at the same time, you are in the right place in this piece.

Will ASI solve the bottlenecks — and what if most things then get solved?

The expectation that an ASI would first attack its own bottlenecks is entirely logical: energy, chips, memory and algorithms are precisely the problems limiting its further growth, and it is recursive self-improvement at the system level. The wildcard from the energy part then stops being a gamble and becomes policy. There is just a bootstrap problem inside it: the ASI that solves the energy problem first needs the energy to exist. Hence today's construction frenzy.

But then the intriguing follow-up hypothesis: suppose the infrastructure is built, ASI is running, and much has been solved — does the need for compute then collapse? There is something to be said for it. Research has diminishing returns once the big questions are answered, and for most everyday tasks "smart enough" eventually really is enough — nobody needs a supergenius to manage a calendar. The historical parallel exists too: per-capita electricity consumption plateaued in rich countries after decades of growth, and after the railway construction boom, track was never laid at that pace again.

Yet there are three strong counterforces. First: every answered question breeds new ones — science has never turned out to be "finished", and an ASI that solves cancer starts on aging, and after that on things we do not yet have words for. Second, demand shifts from thinking to doing: even a world where most things have been invented keeps consuming compute to run — agents, robots, simulations, personalization, entertainment. Third, there is Jevons itself: cheap, abundant intelligence has never yet chosen to go unused. The most likely shape is therefore not a collapse in demand but an S-curve: explosive growth, then a plateau at a high level — like electricity. For the bubble question, that is an important nuance: on an S-curve, not everyone loses, but above all whoever built last and most expensively. The capacity delivered just before the plateau has the worst business case. Timing, not technology, then determines who the loser is.

Three scenarios — and what the charts say

Lay the three lines from this piece side by side — token prices falling about 10x per year, energy consumption doubling toward 2030, and a capital gap in which the US invests 13 times more than Europe — and the scenarios draw themselves.

The forecast that speaks from the data itself: the price of intelligence keeps collapsing (the best-documented trend in the entire sector), total consumption and total investment keep rising (Jevons), and the shock — if it comes — will be financial, not technological. The question is not whether capital gets destroyed, but how much, whose, and whether the infrastructure remains useful afterwards. With the railways and the fiber network, the answer was the same every time: much capital destroyed, infrastructure remained. And should demand saturate after the ASI buildout, the curve ends not in a crash but in an S-bend — in which the biggest loser is whoever built last and most expensively.

Conclusion

The AI bubble cannot be judged by looking only at the cost of one Chinese model or the latest efficiency gain of a chip.

Chinese AI companies use distillation, but also build their own models, architectures and training systems. If OpenAI and Anthropic get better at blocking distillation, that slows China down — it does not stop China. The strategy of letting America carry the heavy costs while harvesting the commoditization is real, but it is a strategy that destroys margins on both sides of the Pacific.

Recursive self-improvement already exists in limited form: AI helps with data generation, code, algorithm design and training processes. Fully autonomous self-improvement has not been convincingly demonstrated.

Chips and models will get more efficient, so costs per token and per task fall — by roughly a factor of ten per year. But cheaper AI means more AI: efficiency saves cost per task and simultaneously enlarges total demand for compute and energy. Though that demand has a wildcard: if AI-accelerated research makes energy generation itself cheaper — the way the CRT became an LED screen and the incandescent bulb an LED lamp — then the supply limit is not a wall but a moving ceiling. Today's infrastructure buildout is therefore not automatically wasted money; the real risk is that certain companies build too early, too expensively or in the wrong place — or lose their subscribers before the costs are recouped.

Europe stands largely on the sidelines of this race in capital terms, but holds a chokepoint in ASML and, as a heavy consumer of newly cheap intelligence, can actually profit from the price war others are fighting — provided it finally builds its sovereign compute instead of announcing it. And more is at stake than competitiveness: a welfare state that runs on labor taxes for more than half its revenue cannot afford an agent economy in which the profits — and thus the tax base — structurally land with American or Chinese owners. Japan and South Korea, meanwhile, show how it can be done differently: whoever occupies the infrastructure layer earns in every scenario — including a bursting bubble.

AI can become one of the most important technologies of this century and simultaneously cause a financial correction among companies that paid too much for compute, market share or future expectations.

The biggest battle is ultimately not just about who owns the smartest model. The winner holds the best combination of intelligent models, efficient chips, cheap energy, scalable data centers, strong researchers, automated AI development, robots and practical applications, access to capital — and the ability to use AI in the real economy.

The AI race is therefore not a short sprint between ChatGPT and DeepSeek. It is the construction of an entire new industrial infrastructure — and we are probably still only at the beginning.

Sources and figures

• Stanford HAI, AI Index Report 2026 — US-China performance gap, private AI investment by region.
• IEA, Energy and AI and Key Questions on Energy and AI — data center consumption 415 to 485 TWh (2024–2025), about 950 TWh (2030), about 1,200 TWh (2035); big-tech capex above 400 billion dollars (2025), 75 percent growth expected (2026).
• TokenCost, AI Price Index and BenchLM, LLM API Pricing History — GPT-4-level price decline from 60 dollars (2023) to about 0.40 dollars (2026) per million output tokens.
• Mistral AI, Series C announcement (1.7 billion euros, ASML lead); Tech Funding News on the 20 billion euro round; MLQ on 400 million dollars-plus ARR.
• European Commission, AI Continent and AI Gigafactories; Bloomberg on the delays (June 2026).
• Deloitte/Forbes, How AI Will Reshape Employment Taxes and Public Revenues — employment taxes about 24 percent OECD, Germany 35, France 37; The Economy, When Consumption Fades — VAT erosion under AI job losses.
• Congressional Research Service, Pillar 1 and Digital Services Taxes; Tax Foundation, Digital Services Taxes in Europe 2026 — US withdrawal from the OECD deal (January 2025) and the stranded Pillar One process.
• Ed Zitron, OpenAI financials 2025; Forbes, Anthropic and OpenAI: opposite paths to profitability; counter-voice: ITIF, AI Is Not Going to Destroy the Tax Base.
• DCD, Samsung and SK Hynix scale up HBM production (about 80 percent of the world market); TechTimes on the South Korean investment wave; Igor's Lab on the NVIDIA-Korea alliances; Tech-Insider on the Japanese foundation model consortium.
• The Quantum Insider, logical qubits and error correction; PostQuantum, Quantum Machine Learning: State of the Field — fault-tolerant systems expected 2030–2035.
• AI Futures Project, Timelines Update — METR task horizon doubling every four months since 2024; Forecasting Research Institute, expert and superforecaster timelines.
• Scientist-versus-agent estimate: own Fermi calculation based on UNESCO-order researcher counts (about 10 million FTE), typical energy intensities for offices, laboratories and transport, and reported inference energy per token; order-of-magnitude only.
• Energy innovation: public announcements 2023–2025 — Microsoft/Constellation (Three Mile Island reactor restart), Google/Kairos Power and Amazon/X-energy (SMRs), Microsoft/Helion (fusion offtake agreement); DeepMind research on fusion plasma control and GNoME materials discovery; Koomey's law (computations per kWh).
• DeepSeek-V3/R1 technical reports; NVIDIA Rubin performance claims; ESA research on space structures; preprint analysis of orbital data centers (2026).

---

At Sentinel Alpha we build privacy-first AI agents that run on your own infrastructure — because in every scenario above, owning your intelligence beats renting it.