The closing door

The most important fact about Anthropic’s newest model is not what it can do. It is who can use it. Mythos is currently being released through a tightly managed preview: twelve named launch partners and somewhere over forty other organisations. The announcement on 7 April pushed it out through a defensive-security consortium called Project Glasswing. The twelve are Amazon Web Services, Apple, Microsoft, Google, Nvidia, JPMorgan Chase, the Linux Foundation, CrowdStrike, Cisco, Broadcom, Palo Alto Networks and Anthropic itself. The other forty have not been named, and probably will not be.

What matters, for the argument that follows, is not the guest list. It is the structure. Frontier AI capability is beginning to enter the world through curated corporate channels, not through the front door of a public API. Cost is one filter on access. Security, in several different senses of the word, is the other. The rationale is defensible. The structure is novel and important, and it deserves much more attention than it has so far had.

Below the top tier, the rest of the market is commercial. Anyone with a corporate card can pay for frontier access. On paper, the prices still look manageable. Claude Opus 4.6 charges $5 per million input tokens and $25 per million output. GPT-5.4 sits below it at $2.50 and $15, with a Pro variant at $30 and $180 for the heaviest reasoning workloads. DeepSeek V3.2, the Chinese open-weight model that has been leading that particular pack, costs 30 cents input and 50 cents output. Put those numbers side by side and the Western frontier runs somewhere between fifteen and twenty times more per input token than the Chinese alternative, and a good deal more on the output side.

Those numbers look bearable until you run them against the workloads this capability is actually meant to handle. A real research or coding task does not sip tokens. It can burn hundreds of thousands on the input and tens of thousands on the output, and that is before any retries. Heavy usage by a research team touches millions of tokens per analyst per day. On Opus 4.6 at that intensity, the bill runs to tens of dollars per analyst per day and well into five figures per analyst per year. Multiply by a team of a hundred and inference alone approaches a million dollars annually. The same work on DeepSeek costs a small fraction of that. For any institution that cannot support the Western pricing, parts of the job simply do not get done.

The arithmetic understates the problem in two ways, and they reinforce each other. The first is that not all of the prices being quoted are market-clearing prices. Sam Altman said last year that OpenAI loses money on its $200-a-month Pro subscription because customers use it more than expected. OpenAI booked $5 billion of losses on $3.7 billion of revenue in 2024. Anthropic is currently providing around $100 million in usage credits to its Project Glasswing partners, which is a direct subsidy on Mythos access. This is not indefinitely sustainable. As the subsidy layer thins out, list prices will climb and the slope will get steeper. Anthropic has already started. On 4 April it blocked Pro and Max subscribers from routing flat-rate plans through third-party agent frameworks, which means the heaviest workloads now have to be paid for separately, by the token. Flat-rate access is becoming harder to sustain just as the most demanding workloads are scaling up.

The second understatement is on the usage side. Headline per-token prices have been dropping for years. Reasoning token consumption per organisation, over the same period, has risen 320 times, on OpenAI’s enterprise data. Both of those things are true simultaneously. The workloads have changed. A simple chat query used to cost a few hundred tokens. An agentic workflow on the same task uses several times as much. Agentic coding workflows, once you count retries and self-correction loops, can eat millions of tokens per task. New tokenisers in recent models also count more aggressively, which quietly inflates the bill for unchanged inputs. None of this is visible on the price list. All of it shows up on the invoice.

What the market ends up with is three tiers, more or less. At the top, hedge funds, hyperscalers, magic-circle law firms and tier-one enterprises will run their agentic workloads on Opus 4.6 or GPT-5.4. For them, set against the value of what comes out, the marginal cost is trivial. In the middle, mid-sized enterprises and well-funded public sector bodies will run on the tier below. At the bottom, smaller firms, civil society organisations and a long list of institutions across the developing world will run whatever they can get on Chinese open-weight models, usually on commodity hardware. Those models typically lag the leading edge by a year or more, and even then GPU capacity is not easy to come by.

Nobody publicly designed this sorting. It is what falls out when frontier agentic work matures from a consumer curiosity into an enterprise category, and pricing starts to reflect what enterprise buyers will bear. The market will describe that as emergent. In practice, it is being reinforced at every layer, by actors whose incentives are all aligned in the same direction. Frontier AI is not being democratised. It is being tiered. The tiers are being drawn at the infrastructure layer, not at the software layer, which is why the access-policy debate keeps landing in the wrong place.

This sorting does more than allocate compute bills. It allocates who gets to work at the top. Give one research team at a leading US university Opus 4.6 across its literature review and experimental design, and over the same period it will produce materially more than an equally talented group in Lagos or Nairobi working with less. A magic-circle law firm running the same capability across discovery and case strategy will pull away from mid-tier competitors, and further still from the small firms that cannot match the throughput. A national security agency with Mythos-class models can find and patch vulnerabilities that an under-resourced agency in a smaller country may only encounter after they have been exploited. The inequality here is concrete. It compounds across research, law, national security and large parts of the wider economy, and the longer it is left unaddressed the harder it becomes to reverse.

None of this self-corrects. The forces underneath are structural, and they are intensifying.

The access gradient is widening

The standard story about how information technology evolves goes like this. Capability starts expensive and concentrated, then gets cheaper and more widely distributed over time. Mainframes gave way to personal computers. Smartphones now carry more compute in a pocket than NASA had in 1969. Frontier AI, the assumption runs, will follow the same arc. The training cost of a 2023-class model can already be replicated on a serious university research budget. On that reading, today’s top tier becomes tomorrow’s commodity.

I would be more comfortable with that argument if it were not being made about a technology that does not behave like the ones that came before it.

The earlier waves distributed outward because the silicon kept getting smaller and cheaper, and because the infrastructure they ran on was modest enough to fit in an office or a pocket. None of that applies at the top of the AI curve. Each new generation of frontier model has tended to demand roughly an order of magnitude more compute than the one before it. That compute has to be run at a power density only a handful of sites in the world can supply. And the regulatory posture around it, inside the labs and at the level of the governments that host them, is tightening rather than loosening. The forces are stacking up rather than cancelling out, and none of the earlier transitions faced all of them at once. The gap between what the leading edge can do and what commodity kit can do is not closing as you go up. The ceiling is rising faster than the floor.

The labs themselves are also gating the top tier more tightly than the old trajectory would predict. Several forces pull in the same direction at once. Commercial logic is one: exclusivity preserves pricing power, and enterprise customers will pay for provenance. Safety is another, because Mythos-class models can do things that should not sit behind a casual API key. Operational capacity is a third, in the plain sense that compute is finite and somebody has to be turned away. Underneath all of that, the governments hosting the labs have their own national-security motives, and increasingly treat this category of model as a dual-use technology, in the same regulatory bracket as advanced semiconductors and precision optics.

All of these pressures push one way. They tighten the ring around the most advanced capability and loosen the one around everything below it. Most of the people who will spend the next decade living on the lower ring have not yet worked out how steep the ground is going to be.

The substrate that makes the gradient durable

Access concentration would be a passing problem if the underlying capability were portable. It is not. Frontier AI runs on physical infrastructure that takes years to build and sits in a very small number of jurisdictions. A frontier training run can consume hundreds of thousands of the most advanced GPUs running continuously for months. The power draw of a single hyperscale AI campus is now pushing towards a gigawatt. Stargate, the joint OpenAI and SoftBank build-out, is sized at $500 billion across facilities whose individual power envelopes begin at 500 megawatts. These are not numbers any normal industrial process has historically reached.

The constraint is physical and it is binding. PJM and other US grid operators cannot connect new capacity fast enough, which is why the hyperscalers have shifted towards behind-the-meter gas generation. xAI secured 1.2 gigawatts of dedicated gas capacity at its Mississippi site to power the Colossus complex. Meta and Amazon are building combined-cycle plants next to multiple sites. US utility power to data centres is forecast to rise from around 62 gigawatts in 2025 to 134 gigawatts by 2030, which is more than doubling in five years, and it is concentrating further in the states where firm power, planning consent, fibre and cooling water can be assembled at speed.

The list of jurisdictions that can actually put all of that together, on a single contiguous site, is very short. Parts of the US Gulf Coast. Parts of the Nordics. The UAE, Saudi Arabia, and, with some effort, Singapore and South Korea. Most of Western Europe does not make that list today. Most of the Global South does not make it either. The hierarchy is durable because the plant beneath it cannot be redistributed by any intervention at the model layer. You can rewrite the rules of API access in an afternoon. You cannot rewrite a 500 megawatt build-out.

This is also why the export-control regime looks the way it does. US controls on advanced accelerators have not actually stopped Chinese labs from training frontier models. Stanford’s 2026 AI Index measured the US–China model performance gap at 2.7 per cent on the Arena Leaderboard. What the controls have done is constrain the operational capacity to deploy those models at scale. Zhipu was, earlier this year, rationing sales of its coding product to around 20 per cent of previous capacity because of server constraints. Chinese labs can now get close to the leading edge on benchmarks. They have not yet matched it as a commercial service for hundreds of millions of users. The gap is an infrastructure gap, not a research gap, and it sits underneath almost every conversation being had about AI sovereignty and so-called democratisation.

What countries are actually doing

Some countries have understood this and are responding to it. Singapore has been investing in public AI infrastructure and in skills, on the visible assumption that the capacity to host and run frontier-adjacent models is a strategic asset rather than a purely commercial one. India’s IndiaAI programme combines public compute investment with explicit autonomy goals. Both have worked out that buying access from the major US and Chinese labs is not a long-term strategy for a country that wants its AI sector to remain answerable to domestic priorities. Saudi Arabia and the UAE have gone further, deploying sovereign capital at a scale other states have not attempted, with more than $50 billion of named AI and data-centre investment across the two. Microsoft’s UAE commitment alone runs to $15.2 billion through 2029. The Gulf bet was that frontier AI would end up being hosted wherever power is cheap and politics are relatively frictionless. So far, that bet has largely paid off, although the proximity of those assets to the energy infrastructure Iran spent six weeks attacking has introduced a resilience question the original investment cases did not model.

China is doing something different, and I think underestimated. Rather than compete only for the most advanced models, it is quietly exporting low-cost AI and cloud capability into countries where the alternative is dependence on a US-controlled stack. Belt and Road digital infrastructure programmes and Huawei cloud build-outs across Africa and South-East Asia are the most visible part of this. The open-weight strategy that DeepSeek and others briefly pursued, before pivoting back towards closed source when the commercial case for openness weakened, fits the same pattern. If you cannot dominate the leading edge, you build the substrate the next billion users will run on, and you are patient about it. The strategy does not require frontier parity to work. What it locks in, instead, is dependency at the infrastructure layer, which is much harder to unwind than a choice of model.

The United States is doing something else again. Its strategy is national champions plus an alliance system. Frontier capability sits in a handful of US-domiciled labs, and access is negotiated bilaterally, through a mix of commercial relationships and government partnerships. The Anthropic–Pentagon dispute that has been winding through the courts since February is, in part, a test of where the line between commercial and state control should sit. Anthropic refused to drop its prohibitions on mass surveillance and autonomous weapons. The Pentagon responded by designating the company a supply-chain risk. The legal outcome is still open. The underlying pattern is not: the United States is treating frontier AI as a strategic asset to be made available to allies on terms it largely sets itself.

Most countries sit in none of these camps. They risk being locked into long-term dependence on infrastructure they neither own nor can materially influence. The UK is one of them. Britain has excellent AI research and a serious talent pipeline. The AI Security Institute did some of the most authoritative public technical work on Mythos within a week of the announcement. That has not translated into a frontier AI position, largely because it does not solve the plant problem. The issue has never been talent. It is power, planning, cost and speed, in roughly that order. Industrial electricity prices in the UK are among the highest in the OECD. Grid-connection queues in the regions where data centres would plausibly site now run into 2035 and beyond, which is something anyone who has actually tried to get connected in the last three years can confirm. On 9 April OpenAI paused its Stargate UK build entirely, citing energy costs and the regulatory environment as reasons it could not commit to long-term infrastructure investment in this country. A week later, on 16 April, the £500 million Sovereign AI Unit launched, chaired by James Wise. It is a sensible instrument for the problems it has been set up to address. It is also roughly thirty times smaller than Microsoft’s commitment to the UAE on its own, and it does not change the megawatt arithmetic that will decide where the next generation of frontier capacity is built. The two announcements arrived in the same week. They told the same story.

What democratisation would actually require

The conversation about democratising frontier AI is still happening, mostly, at the wrong layer. Calls for compulsory licensing of frontier models, open-source mandates, international access frameworks and inclusive multilateral governance are all attempts to reshape the access regime at the software level. They proceed as though the core bottleneck is that labs are unwilling to share, and therefore the policy answer is to make them share. That is not, for the most part, where the bottleneck actually sits. Labs are gating access partly because they prefer to, commercially, but also because compute capacity is genuinely constrained and because governments are leaning on them. Even if every frontier lab adopted the most permissive access regime imaginable tomorrow morning, the underlying capability would still be concentrated in a small number of facilities, owned by a small number of corporates, in an even smaller number of jurisdictions. Democratising the access regime without democratising the plant underneath just moves the bottleneck. It does not remove it.

A serious democratisation agenda has to operate on the infrastructure itself. Public investment in compute, at a scale that can actually compete with hyperscaler build-outs, is the starting point. Grid capacity and planning consent have to be instruments of industrial policy, and so does the price of the power itself. Managing them as regulatory questions at arm’s length is no longer adequate to the problem. Mid-sized democracies that individually lack the scale to host frontier capacity could, in principle, underwrite shared facilities together. Some of these choices also sit uncomfortably with existing climate commitments, because the thermal generation the hyperscalers currently need to run AI workloads is not compatible with most stated net-zero pathways. That trade-off is going to have to be confronted rather than evaded, and it is still not the conversation most policymakers are having in public.

The default trajectory, without intervention of that kind, is that frontier AI ends up as a private utility for the governments and large corporates that can afford to host it, and a rented service for everyone else. The lessor sets the terms. The lessee has very little recourse when prices rise or when permitted uses narrow. In practice that means the cognitive frontier of medicine, law, security, scientific research, finance and engineering will be set, for the next several decades, by the relatively small number of organisations that own the plant underneath. Everyone else works with a lagged, supervised version of the same capability. The gap between frontier-equipped institutions and the rest will not look like the familiar productivity gap from an earlier wave of automation. It will look more like a category gap. Whole classes of work, from drug discovery to long-horizon agentic research, will simply not be available below a certain spending threshold. The countries and institutions below that threshold will fall further behind on every problem where frontier capability is what does the work.

To be clear, this is not a thought experiment. It is the structural endpoint of what is already being built. If the substrate question is not pulled into the democratisation debate in a serious way, that endpoint could arrive by the end of the decade. The door is closing. The question worth asking is who ends up outside it.