OpenAI has published a document arguing that the value of AI infrastructure has very little to do with how much of it you own. For a company running one of the largest capital programmes in corporate history, that is an odd thing to put in writing. It makes more sense once you notice who else is reading. Britain has put £2 billion behind a twentyfold increase in national compute capacity by 2030. If OpenAI’s argument holds, most of that money is buying the part of the problem that matters least.
What the abundance argument actually claims
The post, published on 31 July under the title “Building abundant intelligence”, opens with a line that does a lot of work: “AI infrastructure is not valuable because it is large. It is valuable because of what it makes possible.” From there it describes a self-reinforcing commercial loop. Cheaper intelligence makes more work worth doing. More work done creates more value, which produces revenue and demand signal, which funds the next round of research and capacity.
The evidence offered is mostly pricing. On 30 July, OpenAI cut the price of its smallest GPT-5.6 tier, Luna, by 80 per cent, and its mid-tier, Terra, by 20 per cent. That leaves Luna at $0.20 in and $1.20 out per million tokens, against $2 and $12 for Terra. We covered the cuts when they landed, and the reading then was that buyers had forced them. The post makes the opposite case: that they are what efficiency gains look like when they reach the price list.
Both can be true. What is more interesting is the yardstick the post proposes to replace tokens with. Customers, it argues, want their support issue closed or their contract reviewed, so the honest measure is “the cost of a successful outcome, including the time, retries, oversight, and errors required to get there.” That is a genuinely useful idea and most UK finance teams are not yet measuring it.
Strategic Reality: A model that costs four times as much per token and finishes the job first time can be the cheaper option. If your AI cost reporting is denominated in tokens or seats rather than completed tasks, you cannot see which of your workloads is on the wrong tier.
| The numbers OpenAI puts on the record | What they establish |
|---|---|
| 80% and 20% | Price cuts to GPT-5.6 Luna and Terra, effective 30 July 2026 |
| $0.20 / $1.20 per million tokens | Luna’s input and output pricing after the cut |
| 13.3% to 38.3% on ARC-AGI-3 | Score improvement from context and reasoning changes, with the model itself unchanged |
| One sixth of the output tokens | Consumed in reaching that higher score |
| 20% | Reduction in end-to-end serving costs from software optimisation work |
| More than 15% | Gain in token-generation efficiency from improved speculative decoding |
| 99.8% | Share of OpenAI’s weekly output tokens now produced by agentic work through Codex |
| More than 1 billion users and 2 million businesses | The demand base the capital programme is justified against |
The gains are in the system, not the silicon
The ARC-AGI-3 line is the one worth stopping on. OpenAI reports that changes to how the model retains reasoning and manages context took GPT-5.6 Sol from 13.3 per cent to 38.3 per cent on a public task set, almost tripling the score whilst consuming a sixth of the output tokens. Its own summary of what happened: “The model did not change. The surrounding system did.”
Put that beside the other efficiency claims and a pattern appears. A 20 per cent cut in serving costs came from optimising production software. A 15 per cent gain in token-generation efficiency came from better speculative decoding. None of it came from more hardware. The post is explicit that delivering value takes more than additional data centres, and that the surrounding system decides how much of each unit of compute becomes useful work.
If that is where the near-term cost curve actually bends, it has an uncomfortable implication for anyone whose AI strategy is a procurement plan. The scarce asset is not the accelerator. It is the operating capability wrapped around it: routing, context management, evaluation, the product design that removes steps from a task.
Critical Context: Every efficiency gain OpenAI reports in this post is a systems gain. Britain’s compute programme is almost entirely a hardware and buildings programme. Those are not the same investment, and only one of them shows up in a ribbon-cutting.
Look at what the UK has actually bought. The government’s own one-year progress review records five AI Growth Zones designated across Britain, £2 billion for twentyfold compute expansion by 2030, up to £250 million earmarked for AI Research Resource cloud capacity, £750 million for the next national supercomputer at Edinburgh, and up to £500 million in the Sovereign AI Unit. The Growth Zones are credited with generating £28.2 billion of investment and over 15,000 jobs, with each zone receiving £5 million in targeted local funding.
That is a serious programme by British standards. It is also, line by line, a programme for building and powering things. There is no comparable line for the serving stack, the evaluation infrastructure, or the engineering capability that determines whether a publicly funded supercomputer produces useful work per pound or merely produces heat. We made a version of this argument when the first big infrastructure commitments landed, in our analysis of how the UK was redefining AI infrastructure strategy. The gap has widened since.
So is it an infrastructure argument or a lobbying position?
Both, and the fact that it can be both is the thing to understand. The economics described are real and the engineering results are specific enough to be checked. It is also a document with three audiences at once: customers, who are told to move workloads down a tier; investors, who are told the capital programme is disciplined; and governments, who are told that the hard part is coordination rather than ownership.
The passage aimed at that third audience is easy to miss. Describing its approach to capacity, the post notes that it “does not require owning every asset or building every component ourselves” and that the company can “own, partner, or buy depending on what best serves the customer and makes the most economic sense.” Read as corporate strategy, that is prudent. Read as a signal to a host government negotiating an investment announcement, it is a statement about optionality — and optionality is exercised in both directions.
Britain has already seen it exercised. OpenAI withdrew from the North Tyneside site attached to Stargate UK, and freedom of information documents later showed that neither OpenAI nor its partner Nscale held recorded meetings with local authorities before the project was unveiled, with £20 billion of the headline investment notional. We covered the questions raised over that site in July.
The sharpest line in the post is the one the UK should turn around and apply to itself: “The objective is not to build the most infrastructure. It is to deploy the right capacity, at the right time, against credible demand.” OpenAI says it tests capital against growth in usage, enterprise contracts, API demand, utilisation and revenue. Britain designated Growth Zones on the basis of a 500MW build target and a route to power that, on the evidence of the first announcements, was not always verified. A Capgemini survey of 600 senior energy executives found 77 per cent worried demand will outstrip supply. Globally, the Uptime Institute expects roughly half of 250 large datacentre projects announced since 2021 to be cancelled or delayed.
Reality Check: OpenAI applies a demand-evidence test to its own capital. The UK has not applied one to the investment it announces on OpenAI’s behalf. That asymmetry, not the size of the cheque, is the governance problem.
Abundance is a price argument. Sovereignty is an access argument
Here is where the two conversations get conflated, and the conflation is expensive.
Abundance, as the post defines it, is about cost and capability: “intelligence that keeps getting more capable, more affordable, and more valuable to the people who use it.” Nothing in that definition addresses whether the intelligence remains available to you. Those are orthogonal properties. A supply that is cheap, plentiful and revocable is not a strategic asset. It is a dependency with attractive unit economics.
Britain got a demonstration of the difference in June, when Washington ordered Anthropic to stop foreign nationals using its frontier models at roughly 90 minutes’ notice, reopening the sovereignty debate in the middle of London’s densest AI district. The Commons Science, Innovation and Technology Committee reported on 7 July that the government has “no coherent strategic framework” for sovereign technology capability, that the UK “may not be able to count on its allies” for critical technology, and that it risks “substituting activity for strategy”. Committee chair Dame Chi Onwurah put it directly: the government “needs a realistic plan to achieve sovereign capabilities in critical areas or risk having its access cut off at the whim of its partners.”
Falling prices do nothing about any of that. If anything they make it harder to argue for, because the cheaper the commercial option looks, the more expensive a domestic alternative looks by comparison. That is the structural reason the abundance frame functions as lobbying whether or not anyone intends it to. An argument that intelligence is becoming plentiful and affordable is, in a policy setting, an argument that building your own is a poor use of public money.
| Stakeholder | What follows from this |
|---|---|
| UK government | Compute capacity and compute sovereignty are separate purchases; the £2bn programme buys the first and mostly assumes the second |
| Enterprise buyers | Switch cost, not unit price, is the number that determines your exposure when access terms change |
| Public sector procurement | Contracts should specify continuity of access and exit terms, not only price per token or per seat |
| UK AI start-ups | Systems and serving efficiency is the layer where a small team can still be competitive; frontier training is not |
| Boards and audit committees | A supplier concentration risk denominated in pounds understates a risk that is actually about jurisdiction |
Hidden Cost: The cheaper a provider’s tokens get, the more work you move onto it, and the more expensive leaving becomes. Falling unit prices quietly increase switching cost. That is the mechanism by which an abundance of intelligence turns into a concentration of dependency.
For scale, the Financial Times reported in April that Alphabet, Amazon, Microsoft and Meta plan roughly $725 billion (£548 billion) of AI capital expenditure in 2026 alone. Britain’s entire compute commitment through to 2030 is under half a per cent of one year of that spending by four American firms. This is not a gap that can be closed. It is a reason to stop treating capacity as the variable that decides the outcome.
What to do about it
The useful response is not to abandon the compute programme. It is to stop treating hardware as the whole of the strategy, and to buy the things the abundance argument itself identifies as decisive.
- Fund the operating layer, not only the estate. If a 25-point benchmark improvement is available from context management and reasoning changes with the model held constant, then serving-stack engineering is a national capability question. Public compute procurement should carry efficiency and utilisation targets, not only FLOPs and megawatts.
- Write access continuity into contracts. The sovereignty risk that materialised in June was contractual and jurisdictional, not technical. Ask what happens to your workload if an export decision lands on a Friday afternoon, and get the answer in the agreement.
- Measure cost per completed task. Take the one genuinely portable idea in the post. Instrument a real workflow end to end, including retries and human correction, then re-benchmark it against the cheaper tiers. Price cuts do not reach your budget on their own.
- Price your exit before you need it. Run one workload on an alternative provider, including an open-weight option, so that the cost of moving is a figure you know rather than one you discover under pressure.
Where you start depends on where you are. If AI is still in pilots, the cost-per-task instrumentation is the cheap move and it will change which tier you buy. If you have production workloads on a single US provider, the migration estimate is the priority, because that number converts an abstract sovereignty worry into a board decision. If you are a public body, the access-continuity clause is urgent, because it is the one thing procurement can fix without waiting for a policy.
Take Action: Pick your largest AI workload this quarter and produce two numbers for it — the fully loaded cost of one completed task, and the cost of moving it to another provider. Most organisations have neither, and both are decision-grade.
Four things this reading understates
-
The efficiency gains may not generalise. OpenAI’s serving improvements were achieved on its own stack, at its own scale, with its own models. A national facility running heterogeneous research workloads has far less room for the same optimisations. Mitigation: treat the systems argument as directional rather than as a transferable engineering benchmark, and require utilisation reporting from publicly funded compute.
-
Cheaper tokens have not produced cheaper bills. Reasoning models consume far more tokens per task than their predecessors, and usage-based billing means spend scales with how hard the model decides to think. A headline price cut can coincide with a rising invoice. Mitigation: budget against completed tasks and set token ceilings per workflow.
-
Sovereignty has more than one failure mode. Access restriction is the one Britain has just experienced, but price discrimination, deprecation of a model you depend on, and terms-of-service changes all produce the same operational result. Mitigation: test your continuity plan against model deprecation, not only against export controls.
-
The UK compute programme has objectives beyond frontier competitiveness. Free public compute for researchers and SMEs has a research and regional development rationale that does not depend on matching hyperscaler capex, and judging it purely against a sovereignty test is unfair to it. Mitigation: be explicit about which objective each pound is serving, because the current framing lets one programme claim credit for both.
The strategic takeaway
OpenAI’s argument is better than the reflex response to it. Intelligence is getting cheaper, the gains are increasingly coming from the system rather than the chip, and measuring the cost of a finished outcome is a genuine improvement on counting tokens. British organisations should take all of that seriously.
What the argument does not do — and was never going to do — is address the question Britain actually faces. The post closes on the goal of “more useful intelligence within reach”, and reach in that sentence means affordability. The Commons committee is asking about reach in the other sense: whether the intelligence stays available when someone else’s government decides otherwise. A strategy that answers the first question is not a strategy for the second, and a compute programme sized for neither will satisfy nobody.
Three things worth holding onto. Capacity and access are different purchases, so stop letting one budget line claim to deliver both. The competitive layer is the operating stack, which is the one place a mid-sized country and a mid-sized company can still make ground. And when a supplier publishes an argument for why you do not need to build what it sells, the argument can be entirely correct and still be worth reading twice.
- Instrument one workload for cost per completed task, including retries and human correction
- Produce a migration cost estimate for your largest single-provider workload
- Add an access-continuity and exit clause to your next AI contract renewal
- Re-benchmark existing workloads against the cheaper GPT-5.6 tiers rather than assuming the saving arrives
Britain can buy compute. Whether it can keep it working, and keep it switched on, are separate problems that the current strategy has folded into one.
Source: OpenAI, “Building abundant intelligence”, 31 July 2026. UK compute figures from the government’s one-year review of the AI Opportunities Action Plan. Sovereignty findings: Commons science committee report on science diplomacy, 7 July 2026. Hyperscaler capital-expenditure figures reported by the Financial Times in April 2026; datacentre cancellation estimates from the Uptime Institute and the energy executive survey from Capgemini, both as reported by The Guardian.
Analysis by Resultsense — making sense of AI in the UK. For strategic guidance on AI procurement and compute strategy, get in touch.