1. Top AI brass depart

Two labs, ten days. On 5 August Demis Hassabis stepped aside as DeepMind CEO and Jeff Dean left after 27 years, taking three of Google’s most senior researchers to start a lab of their own. Then OpenAI lost Brad Lightcap, its chief operating officer until April, and replaced its revenue chief two days later. Nobody outside those rooms knows why. Four private decisions can land in the same August fortnight without adding up to anything.

  • The people who sold you the roadmap at Google and OpenAI are not the ones who will deliver it, and at OpenAI this is happening on the way to a stock market listing.
  • They are not retiring, they are starting labs, and Alphabet has put money into the one that took its own people. More AI suppliers are coming, not fewer.

2. Zuckerberg declares a price war

Meta gave away a capable AI model that runs on an ordinary computer, and Zuckerberg published 6,500 words arguing Washington should clear the way for more of them. The logic is commercial, not philosophical: OpenAI and Anthropic sell intelligence, Meta sells advertising, so free models cost Meta little and take revenue off both. Nvidia gave one away the next day, and Google halved the price of its workhorse. Only DeepSeek went the other way, quadrupling its prices from 16 August.

  • Everyday AI work is heading towards free. Anyone whose plan depends on charging for it, your suppliers included, is being undercut by a company that does not need the money.

3. Anthropic goes for the record

Anthropic’s investors told the FT they expect it to list in October at $2 trillion or more, which would be the biggest stock market debut in history, on revenue they believe will reach $100 to 120 billion a year by December. None of it is fixed yet. In the same fortnight the company committed $9.1 billion to a Texas power site for the next 20 years, and said it will watermark what Claude writes. OpenAI passed $40 billion of annual revenue but may not go public until next year.

  • Anthropic going first, and going this big, sets the price for every AI company you own, back or compete with.
  • Twenty years of power is a utility commitment, not a software one. These are becoming infrastructure businesses, with infrastructure risk, just as public investors are asked to buy in.

4. Wall Street buys the shovels

Nvidia signed up six of the world’s biggest money managers, Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to raise more than $500 billion for AI data centres, at what Nvidia calls attractive rates for its own customers. Big Tech does not need this and builds from cash flow. The specialist AI cloud operators do, and could not borrow so cheaply alone, so Nvidia may cover up to 25% of any loss if the chips end up worth less than lenders assumed. None of it is signed yet.

  • The guarantee pays out only when chip demand has fallen, the moment Nvidia can least afford it. The insurer and the risk are the same company.
  • So the next wave of AI capacity gets built by weaker borrowers, backed by their own supplier, with money drawn from insurance and pension pools. Watch them, not Big Tech, for the first sign of stress.

5. The cyber line gets crossed

Software security changed hands this week. Z.ai’s GLM-5.3, released 14 August, now beats Anthropic’s and OpenAI’s best models at finding flaws in code, having already turned up 2,436 of them in open-source projects, and Z.ai says it will publish the weights within about two weeks, so anyone will be able to run it. OpenAI, by contrast, put its own cyber model behind vetted access. Days earlier researchers documented a near-autonomous AI attack on Taiwan’s government, including a nuclear safety agency.

  • Those are Z.ai’s own scores, for the version it intends to release, with the most sensitive capabilities potentially held back. What any lab or state keeps in-house appears on no leaderboard.
  • Washington’s pre-release safety review is voluntary, unpublished, and only now being widened to cover open models. Attackers will not wait for it.

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.