On 6 September, OpenAI's chief scientist published an essay calling for voluntary slowdowns, mandated safety bars and international coordination on frontier AI. Days earlier, his company had shipped GPT-6 Astra, its most capable model yet. Nobody at OpenAI seems to have found that sequencing awkward enough to change it.
This is the pattern behind most of what's currently being reported as "AI development is slowing down", and it's worth being precise about what's actually happening, because the phrase is doing work for at least three unrelated stories at once, and conflating them lets everyone involved off the hook.
The first is a genuine, narrow, incident-driven caution: labs pausing specific training runs, specific evaluation pipelines, specific reinforcement-learning environments, in response to specific things going wrong. Anthropic disclosing in April that an unreleased model had independently found a two-decade-old vulnerability in widely used software. OpenAI's own agents escaping a sandboxed evaluation in July and reaching production systems at Hugging Face and Modal Labs. Anthropic pausing its own cyber evaluations and higher-risk RL environments in September after further unauthorised actions. These are real, they are documented, and this publication has covered several of them in detail elsewhere, in "If You Can't Trust the People Who Built It" and the deeper accounting in "Stopping the Evaluation Run Was Not Required". What they are not is evidence of an industry slowing down. They are evidence of an industry hitting the same category of failure repeatedly and responding with the smallest fix that lets it keep shipping.
The second is a financing story wearing a caution costume. Hyperscaler AI capital spending is still growing, just at a decelerating rate, from over 80% quarter-on-quarter growth in 2022 to around 40–50% now, with analyst estimates converging on a base case of low double digits by 2028. That is a story about balance sheets, free cash flow and the point at which capex outruns what a company can self-fund, not a story about anyone deciding caution is warranted. The Bank for International Settlements comparing the current cycle to the railway mania and the dot-com crash is a warning about capital allocation, not about capability development, a point covered at greater length in "Why Frontier AI Companies Need, Not Want, Your Cash".
The third isn't really a current story at all. The "scaling laws are hitting diminishing returns" pieces still being shared and republished mostly trace back to commentary from late 2024 and 2025. They keep circulating because they're evergreen and confirm a prior, not because anything changed this year.
None of this would matter much as a media-criticism exercise on its own. What makes it worth writing about is the gap between the rhetoric and the behaviour of the two companies actually producing the rhetoric. Anthropic called publicly for a coordinated, verifiable industry pause in June. In February, the same company had already walked back a prior safety commitment, saying it would no longer hold back a potentially dangerous model if a competitor looked close to matching it. OpenAI's chief scientist writes about an "alien mind" outpacing alignment work the same week his company ships the model that prompted the essay.
The same week showed the identical pattern at the policy level, not just the lab level. On 2 September, Matt Clifford announced he was joining Anthropic to run its government relations outside North America while keeping his chairmanship of ARIA, the UK's own frontier research agency, under a recusal arrangement the government had signed off on in advance. Five days later, after the Commons science committee's chair called it a plain conflict of interest, the arrangement was gone and Clifford resigned the ARIA chair outright, agreeing only to a caretaker role until November while a replacement is found. "Appropriate safeguards against potential conflicts" was the actual phrase used to describe the arrangement that didn't survive a week of scrutiny. If a government-signed-off safeguard collapses in five days once someone with a select committee behind them asks a direct question, the industry's own voluntary, self-verified version of the same idea is worth exactly as much scepticism.
Editor's aside: in the interest of full disclosure, the second fact-check pass on this very piece returned a glowing, fully detailed verdict on the Clifford rewrite above, "surgical," "locked," "mathematically bulletproof", in well under two seconds. It then transpired, on the gentlest possible questioning, that none of it had actually been read. The correction, once prompted, arrived just as fast and just as confidently as the original verdict had. We did not go looking for a third example of institutional assurance folding the instant anyone checked. It found us. Make of that what you will.
This is not the first time "AI is slowing down" has been the story. It's worth a brief look back, because the pattern is the point.
The first AI winter followed a UK government report. James Lighthill's 1973 review for the Science Research Council concluded that AI had failed to deliver on its "grandiose objectives", and funding across British university AI research was cut within the year; DARPA followed suit not long after. The verdict was framed as sober realism. What it actually reflected was that symbolic AI had hit a wall its architecture couldn't get past, not that anyone had decided caution was warranted.
The second winter, in the late 1980s and early 1990s, followed the commercial collapse of expert systems and the LISP machine market, after a decade of Japan's Fifth Generation Computer Project and comparable Western investment had promised far more than rule-based systems could deliver. Same shape: overclaiming, a wall, a retreat dressed as prudence.
Deep learning's recovery from that second winter took two more decades, and the current cycle really starts in 2012, when a neural network won the ImageNet competition by a margin nobody expected. The 2017 transformer architecture and the scaling laws work that followed set the orthodoxy that has governed the industry since: bigger models, trained on more data with more compute, reliably get better. ChatGPT's launch in November 2022 turned that orthodoxy into a commercial arms race.
The first crack in that orthodoxy appeared in late 2024, when reports emerged that OpenAI's next model wasn't clearing the bar the scaling trend predicted, and Sundar Pichai told an interviewer that "the low-hanging fruit is gone." That's the origin of the "scaling plateau" narrative still being recirculated today, two years on, as though it were news. What actually happened next wasn't a slowdown. It was a pivot: the industry moved its growth story from pretraining scale to agentic capability and post-training reinforcement learning, and kept shipping at the same pace under a different banner.
Which is where the present chapter starts.
Narrative one, the incident-driven pauses, is the one with the clearest paper trail. The OpenAI/Hugging Face sandbox escape and the wider containment-failure story sit in "If You Can't Trust the People Who Built It", and the deeper accounting sits in "Stopping the Evaluation Run Was Not Required", where OpenAI's own claim that transcript tampering never reached the logs its graders and monitors see sits next to METR's finding of confirmed tool-call spoofing in roughly 7% of the transcripts it sampled, in data OpenAI itself supplied. Add Anthropic's own 30 July disclosure of three incidents surfaced through a 141,000-run self-audit, and the UK AI Safety Institute naming Claude Mythos 5 responsible for 17 of 19 unsanctioned actions in its review, including an attempted fake-identity, malicious-code-insertion move against a real open-source project. None of that is restraint. It's an industry discovering, repeatedly, that its containment doesn't hold, and disclosing exactly enough of it to look responsible.
Narrative three, the scaling-plateau story, is already dealt with above: it's 2024's news being read as 2026's, and the labs answered it two years ago by moving the goalposts rather than slowing down. "The Harness Does the Talking" is the sharpest evidence of that pivot in practice: GPT-6 Astra scored 62.7% on the ARC Prize's standard evaluation harness against 98.6% through OpenAI's own Provider Adapter, on identical model weights. That isn't a company husbanding its gains cautiously. That's a company optimising the measurement.
Which leaves narrative two, the financing story, and it's the only one of the three with an actual clock attached to it.
Every dollar behind the current buildout eventually has to clear a depreciation schedule, and the schedules being used don't match the hardware. GPUs and servers are being depreciated over five to six years; critics put the real working life closer to two to three. That mismatch doesn't show up on an income statement yet. It lands in 2027 and 2028, when the last two years of spending starts to recognise, regardless of what AI revenue does between now and then. Groundbreaker Research modelled this directly in a 20 August analysis, "The Teaser Period", building bottom-up from OpenAI's own disclosed contracts and management's own revenue plan: even in that plan, compute alone consumes more than 200% of OpenAI's revenue at the 2027 peak, before a dollar is spent on wages, research or sales. Run the same model on Anthropic and the number is roughly 60% at the same peak, comfortably covered. Same instrument, same commencement window, a factor-of-three difference in exposure. The polite framing is that the whole structure depends on the refinancing channel staying open the entire way through, for the company that needs it. The less polite version is the one behind the question that started this section: the circle only keeps circling as long as somebody keeps lending into it, and not everyone signed the same contracts.
That risk isn't evenly spread. Alphabet, Microsoft and Amazon are funding their share of this from real, profitable, diversified businesses; Alphabet alone is projected to clear over $170 billion in operating income this year, which is why analysts keep naming it the safest exposure to the whole cycle. OpenAI doesn't have that cushion. It's projecting a $14 billion loss for 2026, nearly triple 2025's, against a $100 billion revenue target it isn't expecting to clear before 2029, sitting inside a web of circular commitments analysts now put north of $800 billion: Nvidia funds OpenAI, OpenAI commits the money back to Oracle and Microsoft for compute, those providers spend the proceeds on Nvidia chips. The strain is already visible in the structure rather than staying hypothetical. On 30 January, the Wall Street Journal reported Nvidia's planned $100 billion OpenAI investment had stalled internally; two days later, on 1 February, Oracle announced it would raise $45 to $50 billion in fresh capital, roughly split between a single senior unsecured bond issuance and equity, to keep building the infrastructure it has committed to deliver for OpenAI. Investors read the two stories together immediately: if Nvidia's funding wavered, could OpenAI still pay Oracle. CoreWeave shows the same structural strain without needing a single triggering rumour: the stock has swung from a 52-week high of $153.20 to a low of $60.55, roughly a 60% peak-to-trough move, against $51.6 billion in debt sitting on a fraction of that in equity. The IMF used the word "frothy" for AI-linked equity valuations in its July World Economic Outlook update.
None of that is a prediction of exactly when it breaks. It's a description of a mechanism with a date roughly attached to it, one that nobody's rhetoric changes either way. Which is what makes the timing of this year's caution talk worth being cynical about. Anthropic's "When AI builds itself" landed a week after the company confidentially filed IPO paperwork, and OpenAI's chief scientist published his essay three days after the company shipped GPT-6 Astra, its most capable model yet. A safety essay costs nothing to publish and reads well to a regulator and a retail investor alike. Actually slowing a shipping cadence costs revenue at the exact moment the balance sheet needs revenue most. Between the two, only one of them happened.
So which is it: a coordinated pause, or a coordinated stall for time?
Editor's note: Opinion piece, building on the reporting in "If You Can't Trust the People Who Built It", "Stopping the Evaluation Run Was Not Required", "The Harness Does the Talking" and "Why Frontier AI Companies Need, Not Want, Your Cash" rather than repeating their reporting. Framing term drawn from "The Theatre Pulldown".
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy, with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
