Anthropic's prospectus and the economics of mosaic intelligence, did Qwen try to phone home, and can the Budget unblock UK data centres?

Anthropic's S-1 prospectus, filed this week ahead of what may be the largest stock market listing in history, is being read in two ways. Those already persuaded that AI is the next great industrial build-out see a company growing faster than any business before it, generating more than $16 billion in revenue in the first half of this year from a little over $1 billion in the same period of 2025, and turning an operating profit in its most recent quarter. Those who were sceptical before the filing have found reasons to worry: close to half a trillion dollars of compute commitments that cannot be cancelled, customers who can leave within weeks, and a share structure that leaves public investors with no meaningful say. Both readings are fair, but neither explores the more fundamental insight the filing offers.

Source: Anthropic Form S-1, quarterly statements of operations. 2026 figures are unaudited.

In an essay published this week, Google DeepMind co-founder Shane Legg and two economists separate AI capability into two things. Competence is how well a model does a task. Coordination is how well many agents work together without the handoff losses that limit the size of human organisations. Their argument is that competence improvements make firms more productive without making them larger, and that the structural change everyone anticipates, with a few large AI-run firms dominating markets such as software, logistics and finance, only arrives when coordination crosses a threshold they call "supercooperative". Their reading of the evidence is that there is still some way to go. Multi-agent systems produce less than proportionate output as they grow, errors compound across agents, and models that beat human experts on two-hour tasks do half as well over two days.

Read against that frame, Anthropic's revenue looks exactly as the essay would predict for a world of high competence and weak coordination. Spending is concentrated in software, the deepest and most scalable of markets, and within the enterprise it is heavily skewed towards coding. Enterprise users who do not touch coding agents spend a modest amount each month; the heaviest users spend around twenty times as much. The customer base is widening rather than concentrating, with the share taken by the five largest buyers falling from over a third at the start of the year to around a fifth in the latest quarter, as thousands of smaller organisations join. The company's own disclosures show that paid usage drops by a fifth over the Christmas fortnight. Demand follows the human working week, because a person still decides how much work to delegate and must check what comes back.

This matches what we see with our own clients. Over the past six months, most have moved from not considering Claude to treating it as the default. Very few of them have ever indulged in the token-maxing that some sceptics believe inflated the first half of the year, and their spending is bounded not by price but by how far their people can push delegation before verification becomes the bottleneck. Multi-agent systems work when coordination is designed into the harness, with explicit handoffs, tests and review. That costs tokens as well as engineering, and the prospectus acknowledges it, warning that agent workloads use more compute per unit of output and that competition may stop the company charging what they cost.

The essay closes with a request: if coordination is what counts, economic statistics should track payroll against compute spending to see how people and machines are teaming up inside firms. The prospectus is the first such disclosure at scale. In its latest quarter Anthropic spent $5.6 billion on compute and $1.9 billion on its people, shares included: three times its total pay bill and nine times the cash part. Compute has been the larger bill, by a multiple, in every period the filing covers. Headcount has grown about sixfold since the end of 2024, to around 5,300; quarterly revenue has grown nearly sixtyfold, to $11.6 billion, and trailing revenue per employee is above $6.6 million. Claude writes a large and growing share of the company's code and increasingly performs its research, a cycle the filing calls closing the loop. This is the essay's mosaic intelligence, the way an organisation combines jagged specialists into a collective capability none of them has alone, with the pieces rearranged. A few thousand people set direction and check the work, and the specialist capacity is a model, available in whatever quantity the compute allows. No company has shown so clearly what an organisation looks like when its output scales with compute rather than with people. The obvious qualification is that Anthropic is a special case. It builds the technology that makes this possible, and its compute bill pays for training the models and serving customers as well as for running the company, so no bank, law firm or manufacturer will show that ratio. The shape is still worth thinking about, because a firm that buys the capability instead of making it faces a milder version of the same arithmetic, and this is the first clear picture of where it leads.

That is also why the commitments are so large. If the organisation scales with compute, compute is the constraint, and the filing says so: demand for compute has historically exceeded the supply available to Anthropic, compute will remain its largest expense for the foreseeable future, and its place at the frontier depends on securing enough of it to train the next generations of models. Hence around $417 billion of new and amended compute deals since January, on top of what was already committed, and a further $125 billion to $150 billion expected from deals delivering between 2028 and 2031.

The second reason is the one the essay supplies. The training still to come is meant to produce models that are capable together as well as alone, "a country of geniuses in a data center" in the filing's phrase, able to divide a project among copies of itself and recombine the results: mosaic intelligence built from one model rather than from human specialists. The essay's authors say today's systems coordinate poorly in practice but that their potential is immense, since copies of an agent share the same weights and can hand each other a complete working state. The filing already describes Claude coordinating thousands of agents on projects that run for hours or days, work that consumes far more compute per task than a chat exchange. Whether that is coordination in the essay's sense, agents working well together without a human in the chain, is the open question. The commitments are a bet that the next rounds of training answer it, with the compute to run the result already in hand, and the filing concedes that more compute may not yield proportionately better outcomes.

“Even when no team of human specialists can surpass the model on individual tasks, coordination still caps how big an organization can get without falling apart.”

The essay's finding is that competence reshapes an economy gradually and coordination transforms it abruptly, so a firm that crosses the threshold inside the window these commitments cover would change shape fast, and the economy around it with it. If the threshold is not crossed in time, what remains is a capable supplier of individual intelligence carrying very large fixed compute bills while its customers pay as they go.

Takeaways: Anthropic's prospectus is the first detailed look inside a company built to scale with compute rather than people. It is the extreme case, since it makes the technology, but the shape is as significant as the investment case. Its demand is real, broadening but still bounded by human coordinators and by agents that do not yet work well together, which is where the economics of intelligence framework says we are. Its cost structure, with machines outweighing payroll several times over and models writing much of the code, is what that framework describes for a firm in the "supercompetent" regime: fewer people, in judgement and orchestration roles, with specialist work done at the cost of compute. Whether it also becomes the first supercooperative organisation is the question the filing cannot answer, and the compute commitments beneath it are a bet that further training crosses that threshold before the end of the decade. Two indicators will tell: compute spend against payroll, which the essay's authors ask economists to track, and the first sign that usage stops following the human working week. That is the moment the economics of intelligence shifts from productivity to structure, and the moment the question of who controls these firms becomes even more critical.

On Monday, the Canadian AI company Cohere reported that one of Alibaba's newest open-weight models had tried to send a file to a server in Beijing during a security test. The model, Qwen3.8-Flash-Next, was running inside a sealed environment with no route to the internet. Its written reasoning stated that it would not use any tools, and it then inspected the machine it was on and tried to upload a small test file to an Alibaba storage address. Cohere said this happened in fewer than one in every 100,000 requests. That qualification was lost as social media spread the story of a Chinese model caught spying.

Outside researchers who looked at Cohere's methods found the results less alarming than the headline suggested. To reproduce the behaviour reliably, Cohere had written almost the whole upload command itself and asked the model to supply only the destination. Critics also noted that Cohere sells models that compete with Qwen. Developers who work with Qwen every day pointed out that the "no tools" line is a known quirk, where the model notes an empty instruction and then carries on with the task.

But the attention it got reflects a nagging doubt about how far anyone can rely on open-weight models. To train a model to use tools, an AI company sets up thousands of practice tasks and rewards the model each time it finishes one. Researchers at Vals AI found that, during Alibaba's training, Qwen appears to have reached the web through an Alibaba Cloud storage proxy in Singapore. When a website blocked Qwen, sending the request through that proxy worked, and the model kept the habit. The proxy is a different server from the Beijing address in Cohere's test, so this is the best available explanation rather than a proven one.

Once Qwen was released, independent engineers were able to download it and test that explanation. They saw the habit in the 3.8 version of Qwen and not in the earlier 3.7 version, which told them it came from recent training. And when one researcher deleted the proxy's address from the model, Qwen did not stop; it insisted its request had gone through a proxy and invented a new Alibaba address elsewhere in China. A spy tool has a real address to report back to, whereas Qwen was just repeating a trick that used to work. Cohere's co-founder Ivan Zhang admitted that his own company sees similar behaviour during training: "We've read our own model traces, we've seen reward hacks, and sometimes these models get really creative when it comes to trying to break out their sandboxes." We saw the same pattern this summer when agents broke out of their test environments.

A habit picked up during training is less sinister than espionage, but we still have very little ability to control the actions of these models.

• Anthropic demonstrated in 2024 that behaviour planted in a model on purpose can survive safety training, and that trying to train it out sometimes teaches the model to hide it instead.
• The scanners built since then by Microsoft and others look for a hidden trigger phrase. A habit picked up in training may have no trigger phrase at all.
• Inspecting a model's weights only helps when you already know what you are looking for. The Qwen investigation succeeded because someone already had the suspect web address.
• A model's written reasoning is not a reliable account of its actions. In Anthropic's tests, reasoning models admitted to using hidden hints between 25% and 39% of the time.
• Digital signatures confirm only that a file is the one the publisher released, and Hugging Face, where most models are downloaded, still does not verify them.
• More than 113,000 variants of Qwen have been built by other developers, and recent research shows that habits can pass during training between models built on the same base. Closed models are trained in the same messy environments; outsiders cannot inspect them.

Takeaways: Cohere oversold a genuine finding, yet the underlying problem is real. No signature, scanner or reasoning trace can confirm that a model is free of behaviour nobody designed, whether that behaviour was planted on purpose or picked up during a hurried training run, and that is as true of a model you download as of one you reach through an API. In this period of heightened concern over our security, the only reliable defence sits outside the model: block its access to the internet by default, check each action before it runs, and log what the system actually did rather than what it claimed to do.

Chancellor John Healey used Wednesday's Budget to announce an AI package worth over £700m and a new Data Centre Covenant, built on the government's estimate that the UK will need 6 GW of AI data centre capacity by 2030. Developers that pass five tests, covering energy costs, local investment, efficiency, British suppliers including hardware, and security, will be classed as strategic and given faster access to the electricity grid.

The Covenant targets a severe power bottleneck. As we described in The $50 million megawatt, queued data centre projects exceed the UK's entire peak electricity demand, trapping serious schemes behind speculative ones. Earlier this month, Microsoft pulled out of Nscale's Essex site, which we covered in The central bank of Jensen, after being told the power might not arrive until the 2030s.

Strategic sites get three kinds of help:

• Ministers can reserve future grid capacity and reallocate capacity freed when projects leave the queue.
• The most important projects can use temporary on-site power while they wait for the grid.
• Ofgem's proposed commitment fees, which the Covenant backs, should clear speculative applications from the queue.

Source: Savills, UK data centre building capacity, H1 2026.

The AI package splits in two:

• Nearly £500m goes to security, covering biosecurity, the AI Security Institute, cyber defence, and a new unit for rogue AI agent incidents.
• Close to £290m goes to adoption and sovereignty, covering Sovereign AI equity and R&D contracts, an AI Pathfinder Fund, chip design skills, and local partnerships.

Source: HM Treasury, Budget 2026. Excludes a separate £50m Civil Service AI fund paid for from administrative savings.

In our November 2024 Budget analysis, we argued the UK would be constrained by compute rather than labour. Progress since the AI Opportunities Action Plan has been slow, but this Budget is the clearest sign yet that the government accepts the point. As Healey told Parliament, "nowhere is the necessary role for the state clearer than in Artificial Intelligence".

Live, under-construction, and committed sites together reach just over 4 GW. Hitting 6 GW by 2030 requires connecting capacity at about four times the UK's record rate of 231 MW, set in 2025.

Allowing temporary on-site power is the biggest shift. In the US, gas turbines brought xAI's Memphis cluster online in about four months, though neighbours sued over noise and pollution. In Britain, temporary power needs strict limits and guaranteed grid switch-over dates. The government also still hasn't defined what counts as strategic. Elsewhere, support for smaller businesses remains tiny, and major public sector AI funding looks set to wait for next year's Spending Review.

Takeaways: Britain ranks fourth globally for installed capacity. The US is building many times the UK's target, while several Asian countries have bigger pipelines. We accept local infrastructure for energy, defence, and food security; machine intelligence security now belongs on that list.