Anthropic commits $11.6bn to Akamai's cloud⁠.

Anthropic will spend $11.6bn over seven years on Akamai's cloud infrastructure, Akamai said on Thursday.² That is more than six times the size of a $1.8bn deal between the two companies reported in May.² Akamai says the commitment lets Anthropic draw on its distributed infrastructure to meet "accelerating CPU workload demands".¹

The processors in question are the general purpose kind that run code and browse the web, and demand for them has grown as AI agents take on more tasks.² Akamai did not say what Anthropic will use them for.² The deal is the largest in Akamai's history.²

The commitment is not ironclad.² According to Akamai's securities filing, it depends on Akamai meeting certain delivery and service availability requirements, and either company can end the agreement under certain conditions.² Akamai will see no revenue from it this year, and executives expect $150m to $300m in 2027, reaching an annual pace of about $1.7bn by the end of 2028.² To build the capacity, Akamai expects to spend about $5.5bn.²

The unusual part is the warrant.² Akamai has given Anthropic the right to buy non-voting stock convertible into 7.7 million common shares, up to about 5% of the company, at $111.33 a share.² About 2% is expected to vest once Anthropic makes its first payment, and each further $3bn it commits adds roughly another 1%, so the deal could grow to about $20bn in total.² It is the first time Akamai has attached a warrant to a cloud deal.²

That turns the usual pattern round, in which chipmakers and cloud providers invest in the AI labs that buy their products.² Amazon, Google, Microsoft and AMD have all invested or agreed to invest in Anthropic while selling it chips or cloud capacity.² In May, Anthropic also agreed to use all the compute capacity at the Colossus 1 data centre in Tennessee, roughly 220,000 Nvidia GPUs.¹

What the agents actually run on

Most talk about AI spending is about specialist chips. This deal is a reminder that an agent spends much of its day on ordinary computing: running code, fetching pages, calling other systems. That work runs on general purpose processors, and the firms that build the models are now booking years of them in advance. None of the reporting gives a British angle, a sterling figure or a word from a UK regulator, so nothing here changes what a UK customer pays or signs today.

The part worth your attention is the shape of the deal. The headline number is conditional, either side can walk away, and the supplier has handed its customer a share of its own upside to win the work. When a supplier of yours announces a large AI commitment, ask the same three questions: what is firm, what depends on conditions, and who carries the cost if the demand does not arrive. Capacity bought seven years ahead is a forecast, not a fact.

Also today

• Oxford let OpenAI train its models on Bodleian texts

  The University of Oxford has allowed OpenAI to train its models on historical texts digitised from the Bodleian Library, and staff minutes record worries about the reputational risk of the partnership.³

• The Pentagon can keep Anthropic on its risk list

  A divided federal appeals court in Washington refused to overturn one of the Pentagon's supply chain risk labels on Anthropic, so the department's block on Claude can continue for now.⁴

• Villagers object to Google's datacentre in Andhra Pradesh

  Farmers in Tarluvada, in the Indian state of Andhra Pradesh, say land has been taken back from them for Google's $15bn AI datacentre, its largest planned outside America.⁵

• Meta's adults-only agent comes with a cuddly mascot

  Meta says its Muse agent is for over 18s only, but a children's advocacy group says its soft, rounded mascot is clearly a character that would appeal to very young children.⁶