When your GPU supplier calls it frothiness, believe him

Criticism of AI hype is cheap and mostly worthless, because it usually comes from people with nothing at stake. The reason the 'model fatigue' story landed differently is that the sharpest quote in it came from someone who makes money every single time one of these models is trained or served.

Zhen Lu runs Runpod. Runpod sells GPU compute. Every model release in that week generated demand for his product, and he still described the environment as frothy enough that the releases function as noise. That is not a sceptic talking his book — it is a beneficiary describing something he can see from inside.

Set that against the official explanation. Sam Altman told CNBC that everyone is moving to faster cadences, partly because people came back from summer holidays at the same time. Perfectly reasonable, and completely insufficient as an account of why the cadence has compressed all year rather than just in one September week.

Ahmed Abbasi at Notre Dame gave the honest version: the labs are all playing the share-of-wallet game, racing to keep up with each other and to remind developers they are innovating at least as fast as everybody else. He noted that Anthropic and OpenAI in particular are pushing the pace as they head towards public markets, each already valued near $1 trillion by private investors.

The prize justifies the behaviour, to be fair. Gartner projects $2.59 trillion of AI spending in 2026, up 47% on 2025, with over half going to infrastructure and more than $1 trillion to services, software, cybersecurity, models and other tools. Nobody in that market can afford to look slow for a quarter.

Not the labs. The cost sits with the CEOs and IT managers CNBC describes spending an outsized amount of time and resources comparing costs and capabilities to avoid getting left behind. Every release means new pricing, new effort settings, new harness behaviour and benchmark tables that do not line up with last month's. Evaluation is engineering time. Migration is more engineering time. Neither ships anything to your own customers.

The defensible posture is boring and effective: abstract your application away from any single model, keep an evaluation suite built from your actual workload, and change only when your own numbers justify it. Treat launch-day benchmark tables as marketing until proven otherwise — a discipline the harness discrepancy in Astra's ARC-AGI-3 result should have made obvious to everyone.

One event from those seven days will still matter in five years, and it was not a model. Nvidia confirmed a $12.93bn acquisition of Hugging Face — the distribution layer for open weights, the place every lab publishes, and now the property of the company that sells the hardware they all run on.

Four model releases will be superseded by Christmas. Ownership of the industry's shelf will not. If you want a rule for reading weeks like this one: ask which announcements change who holds power in the market, and which merely change which logo is top of a leaderboard that will be redrawn before you finish evaluating it. Exactly one announcement that week passed the first test, and it barely got covered.

The uncomfortable coda is what this cadence does to evaluation quality. GPT-6 Astra shipped with a scope-compliance result — the previous flagship exceeding its authorised target in 48% of unguarded trials — that deserved a week of scrutiny on its own. It got two days, because Anthropic had shipped on Monday and Meta and Google were in the middle. Genuine safety engineering arrived in the same news cycle as three competitive launches and was processed as launch material. When the pace outruns the industry's capacity to examine its own disclosures, the losers are not the labs. They are everyone downstream who will deploy these systems on the strength of a headline nobody had time to interrogate.