Summary

An analysis of OpenAI research on how enterprises actually use its products, and the single most useful adoption entry in the knowledge base. The finding is a divergence: the gap between the heaviest and the average corporate users more than tripled in six months. The episode also carries an unusually candid Sam Altman admission about being wrong on timelines, and a second one about his own behaviour that is more revealing than the research.

• The headline number: the gap between frontier firms and typical firms, measured in output tokens per active user, went from 2.6x in January to 8.3x by the end of June. Through all of 2025 it had been roughly 2x. Frontier firms now use 17 times as many tokens as eighteen months ago; average firms about twice as many.
• OpenAI defines frontier firms as the top 10% of usage in a month by output tokens per active user, with typical firms between the 45th and 55th percentile. The measure is volume of work done, not number of people using it.
• The agentic flippening, traced month by month in enterprise output tokens. Before October it was essentially all ChatGPT. February: 87% chat, 13% agentic. March: 73/27. Late April: agentic passes half at 53%. By June, where the data ends: 36% chat, 64% agentic.
• What separates the two groups is not spending but tooling. At typical firms 9% of weekly active users use plugins and 3% use skills. At frontier firms it is 21% and 19%. At OpenAI itself, 95% and 93% — so even the frontier firms are early.
• The fastest growth in agentic use is now OUTSIDE software and engineering. OpenAI's explanation is that software moved first because codebases give agents clear context and tests make output verifiable, while general knowledge work has limited context, is hard to specify and lacks criteria for checking the result.
• The host disputes the credit: agentic use grew among general knowledge workers because users worked out the patterns themselves, not because OpenAI improved the models for them.
• The pattern he offers is a ladder — generation (draft an email, a formula), then synthesis (combine disparate sources), then execution (act inside existing systems), then maintenance (keep a system running over time). Chat work sits on the bottom rungs; agentic work climbs.
• Legal work makes the shift concrete. In chat, 57% of the work is writing and 20.5% knowledge retrieval, with system operation at 0.2%. In agentic, writing falls to 16.2%, retrieval to 8.3%, while system operations rises to 17.7%, workflow automation to 7.7% — and coding, meaning actually building applications, to 32.9%, by people who are not software engineers.
• Sam Altman on having been wrong: "I thought when we got to GPT-4... that very quickly after that there was going to be much more disruption... I think I was wrong about a few things, but one in terms of the speed. The economy just has so much inertia... we've all been too ambitious on timelines." The host's summary: who needs a pause AI movement when you have corporations.
• The more interesting admission is about himself. Altman says he has had Codex for months and still clicks between messaging apps, still scrolls email, still keeps a to-do list the old way. "By revealed preference, I have a better way to do it now and I still do it the old way... we build intellectual mind muscle memory." The barrier at the top of the industry is habit, not capability.
• From the headlines, on how much AI staff actually use: a voluntary internal Microsoft spreadsheet showed monthly AI spend ranging from $1 to $28,000 per employee, with medians of roughly $150-500 in seven of eight departments and $975 in core AI. Only 350 of Microsoft's 223,000 employees reported a figure, so this is indicative rather than representative — and there was no correlation found between token spend and pay or promotion.
• Nvidia is being described as the central bank of compute: $42.3bn invested in private companies as of March, plus recent moves on Poolside, Mercor and Perplexity. The argument offered against calling this circular is that Nvidia cannot reinvest in itself — it owns no fabs, so its growth is capped by suppliers it does not control, and investing outward is the only way to grow the demand side.
• Taiwanese prosecutors charged nine people over smuggling Blackwell 300 systems to China, including a manager in Nvidia's distribution business and two people at Supermicro. Of 130 servers ordered, 74 reached buyers in China and 56 were stopped — fewer than 10,000 chips, which the host notes is real but nowhere near enough for a frontier training cluster.

Why it matters

This inverts the story the knowledge base has been collecting about AI and work. The risk these numbers describe is not mass unemployment; it is divergence. Firms that learned to use agents are pulling away from those that did not at a rate that tripled in six months, and the separating factor is mundane — skills and plugins, which cost nothing and which most companies simply have not adopted.

Altman conceding he was wrong about the speed of disruption is worth recording precisely because the knowledge base holds so many predictions in the other direction, including his own. His reason — institutional inertia — is the same force that makes the archive of confident timelines in this database worth keeping.

The most quotable thing here is his admission that he has the tool, knows it is better, and still works the old way. For any reader of the Handbook wondering why their organisation has not changed despite everyone agreeing it should, that is the answer, from the person with the least excuse.

usNZ0fWbTok-transcript.txt