Anthropic's $11.6bn seven-year compute deal goes to Akamai

Anthropic has signed a seven-year, $11.6bn commitment with Akamai for distributed AI infrastructure and software, aimed at supporting its fast-growing CPU workloads. The deal, announced by the legacy internet infrastructure firm, says more about where AI costs actually sit than any model launch this year.

$11.6bn, and it buys CPUs

Under the announcement, Akamai supplies distributed AI infrastructure and software from its cloud business to support Anthropic's CPU workloads. The contract leaves room to expand by up to $9bn more, for a potential total commitment of roughly $20bn.

Spread over seven years, $11.6bn works out to about $1.66bn a year. The word worth underlining is CPU.

For three years, the hardware story around AI has been almost synonymous with GPUs. This time, Anthropic has locked in general-purpose compute and scheduling software spread across multiple locations — the layer that does the daily work for real users once a model is trained. What a leading model company is paying heavily for is not "training smarter" but "running cheaper and more reliably".

Consider a second data point. On September 13, Zhipu announced a placement of new shares alongside a zero-coupon convertible bond of Rmb20.14bn maturing in 2027, raising about HK$39.3bn in total, with roughly 60% directed at next-generation foundation models, a fully self-trained system and compute infrastructure. One company is in California, the other in Beijing, and the money is moving in the same direction.

Training is the down payment, inference is the monthly bill

Model training is a one-off expense. Inference is a bill that grows month by month, with call volume. The former can be covered by a funding round; the latter has to be carried by cheap, stable compute.

By fixing seven years of cost in advance, Anthropic is conceding something: what decides an AI company's survival in the coming years is unit cost per call, not leaderboard scores.

Product moves point the same way. Fortune reported on September 24, without official confirmation, that OpenAI plans to preview a cybersecurity-specific model, GPT-6 Cyber, in the coming weeks, and that its September 29 DevDay will bring more than ten other product launches. The more products ship, the steeper the consumption curve on the inference side.

That is the arithmetic behind $11.6bn: demand is scaling, so supply has to be locked in early.

"This has nothing to do with small firms" is the laziest reading

A popular argument holds that this is a long-term procurement between two American companies, that the real contest is between Google Cloud and Amazon Cloud, and that it has nothing to do with small and medium-sized businesses or go-global teams in Singapore, Kuala Lumpur or Dubai.

We disagree.

Start upstream. Long contracts lock in upstream costs, and retail prices grow out of upstream costs. That is not a moral judgment; it is accounting order. Anthropic would rather sacrifice flexibility for predictability, which, passed through to API pricing, means prices will be steadier but adjustments will arrive more abruptly.

Demand is already scaling. On September 24, at the Chimelong Spaceship park in Hengqin, Zhuhai, more than 300 AgiBot robots were spread across the grounds giving directions, explaining exhibits and playing table tennis with visitors; the same day, AgiBot's 20,000th general-purpose embodied intelligence robot rolled off the line (Global Times, September 25).

On September 22, Kevin Kelly said in a speech in China that within five years 99% of AI will not even touch humans, and that Silicon Valley engineers around him stopped writing code long ago, managing six to ten AI agents at once (Yicai / Entrepreneur). When everyone on a team has several agents reporting to them, compute shifts from the R&D budget to operating expenses.

Then there is money and interest rates. Deloitte's review of the first three quarters, published in September, noted that strong global business demand for AI pushed up the scale of listings tied to the AI and technology value chain; in the same report, Chinese companies' US listing activity had all but stalled. On September 24, the 30-year US Treasury yield rose to its highest since June 2004 (Sina AI hourly digest, September 25). At that cost of funding, a company able to sign a seven-year contract has effectively bought itself insurance against volatility.

Do not forget the pressure from the other direction. Politico reported on September 24 that the White House has asked OpenAI and Anthropic to hold back new models from the UK AI Safety Institute until the US completes its review. Long contracts upstream, reviews downstream: model companies are being squeezed from both ends at once, and that cost will eventually show up in prices and availability.

The judgment

The AI narrative of the past two years was an arms race in model capability. The next two will be a race in procurement and contracts. Whoever holds longer, cheaper compute contracts closer to users can push unit prices lower.

For small and medium-sized businesses and go-global teams, this is bad news — you are not at the negotiating table — and good news — you can benefit on price. The precondition is not betting your budget on the wrong architecture, and not assuming today's prices will still be there next year.