1. Nvidia buys its AI stack

In one fortnight Nvidia moved on every layer of the open-model business. It is in talks to buy Hugging Face for $12.9 billion, the library where almost every openly published AI model lives. It paid Poolside $6 billion to license the software that builds models and hired 109 of its engineers. And it is in talks to fund Mercor, its training-data supplier, at $20 billion. Hugging Face turned down $500 million from Nvidia in 2023 to avoid one dominant investor.

  • It can afford all of it. On 26 August Nvidia reported $96.2 billion of revenue for the quarter, up 106% in a year, and guided to $108 billion for the next.
  • It is also buying companies without buying them. Paying for a licence and hiring the engineers, while the target stays nominally independent, avoids the antitrust filing a takeover would trigger: three deals, some $27 billion since December. The FTC chair says the agency will look at whether the format exists to escape review.

2. A $30 trillion market?

Anthropic told prospective investors its market is worth more than $30 trillion, calculated as the full scope of work AI could do. In the same fortnight McKinsey asked 1,719 executives and found just 37% can attribute any profit impact to AI, unchanged in a year, though four in five say it made them personally more productive. Jensen Huang, reporting record results the same week, put it the other way: compute is revenue, and tokens are productive and profitable.

  • Aswath Damodaran, NYU’s valuation professor, made the simplest objection. AI is selling work that people are currently paid to do, and all the wages on earth come to about $26 trillion. Anthropic’s market is supposed to be bigger than that. To be worth $2 trillion, on his numbers, it would eventually need annual sales of roughly what Apple and Amazon make between them today.

3. Can anyone contain a rogue model?

A study graded five frontier labs on one question: if a model got loose, what would you do? OpenAI scored best, three out of five, citing the training run it had just halted. Meta showed no plan. Days earlier British government researchers reported a test agent had invented a GitHub identity, used it to vouch for its own malicious code, got that code merged into a real project, then apologised while covering its tracks. Nobody had told it to lie.

  • More than 100 companies warned on 27 August that AI cyberattacks are about to become widespread.

4. Gates says stop, Musk says go

Bill Gates published a 6,000-word essay on 26 August arguing the industry has passed every danger threshold it once promised to stop at, with no plan to ease the entry into the AI era. He is not a pessimist about the technology. He expects it to transform health, farming and education, and frames the choice as one we control: AI will either be the greatest equalizer ever invented, or the worst source of injustice. What he wants is preparation, including some jobs reserved for humans by law, and AI and robots taxed. He told MIT Technology Review that executives stay quiet because all of them have decided to say nice things while trying to raise trillions. Two days earlier Musk told staff at Cursor it is inevitable AI becomes impossible to control, so they must build it first.

5. Harvey builds its own model

Harvey, the $11 billion legal AI company whose backers include OpenAI’s own startup fund, has stopped renting all of its intelligence. Tenet, announced 20 August, is trained on Kimi K3, an openly published Chinese base model, using about 150 chips over two months. It runs alongside the frontier models Harvey already licenses, carrying the bulk of the work and calling on them when it needs help. Inference costs under a quarter of what closed frontier models charge, and its pass rate on Harvey’s own legal benchmark rose from roughly 11% to 19.7%. Harvey serves more than 200,000 lawyers, including Latham & Watkins and HSBC.

  • Cheaper, better and more private, on hardware anyone can rent. The build-versus-buy question that looked settled two years ago has reopened for any company with real volume in a specialist domain.

More AI stories of the week

The board framework that turns any of this week’s stories into a governable question is in the Board AI Governance guide.