Anthropic's $11.6B Akamai Bet: AI Agents Run on CPUs

Anthropic has signed the largest contract in Akamai's history: $11.6 billion over seven years for cloud infrastructure. The striking part is what the deal is not about. It is not a bet on the GPU accelerators that dominate AI headlines. It is a bet on the humble CPU, and it quietly signals how the compute economics of AI agents are shifting underneath the industry's feet.

What happened

On Thursday, Anthropic committed $11.6 billion over seven years to Akamai's cloud platform, more than six times the $1.8 billion deal the two companies struck in May, according to TechCrunch. The multi-year agreement will support Anthropic's accelerating CPU workload demands by leveraging Akamai Cloud's distributed AI infrastructure, Akamai said in its official announcement.

The commitment is not ironclad. According to Akamai's securities filing, it depends on the company meeting certain delivery and service-availability requirements, and either party can end the agreement under certain conditions. The money will not flow immediately, either. Akamai told investors it expects $150 million to $300 million in revenue in 2027, starting in the second half, ramping to an annualized pace of roughly $1.7 billion by the end of 2028. To build the capacity, Akamai plans to spend about $5.5 billion and is adding roughly $1.7 billion to this year's capital expenditure to lock in components such as memory in advance.

Why CPUs now

The most telling detail is what the commitment does not buy: GPUs. Anthropic is not paying Akamai for the accelerators that train frontier models or serve their inference. It is paying for general-purpose CPUs, the workhorse chips that execute code, browse the web, and move data around.

That is a tell about how agents actually spend compute. A deployed AI agent does not spend most of its time in a single forward pass through a large model. It spends it on errand work: calling tools, querying APIs, scraping pages, running snippets of code, and shuttling results back and forth. Every one of those steps is a CPU-bound task that scales with the number of actions an agent takes, not with the size of the model behind it.

This is the quiet economics of agentic AI. A single tool call costs almost nothing in compute terms, but an agent that runs for an hour might make hundreds of them. Multiply that across millions of deployed agents, and the general-purpose compute layer becomes the real bill. Nvidia's GPUs remain the scarce resource for training and heavyweight inference, but the marginal cost of running an agent is increasingly dominated by the commodity compute wrapped around the model.

The warrant structure

The deal also inverts a familiar financial pattern. In the circular deals that have defined the AI buildout, suppliers - chipmakers and cloud providers - invest directly in the AI labs that buy their products. Amazon, Google, Microsoft, and AMD have all put money into Anthropic while simultaneously selling it capacity. Anthropic CEO Dario Amodei has said the company does not participate in these arrangements at the same scale as some of its peers.

This time the direction flips. Akamai is granting Anthropic a warrant - the right to buy nonvoting preferred stock convertible into about 7.7 million common shares, up to roughly 5% of the company, at $111.33 a share. About 2% vests once Anthropic makes its first payment, and each additional $3 billion it commits unlocks roughly another 1%, so the agreement could expand by as much as $9 billion to about $20 billion in total. It is the first time Akamai has attached a warrant to a cloud deal. AMD used a similar structure with OpenAI last year, tying equity to chip-purchase milestones.

The message is hard to miss: compute demand has grown valuable enough that a supplier will hand a customer a slice of itself to win the business, rather than the other way around.

Builder impact

For AI builders, the deal is a leading indicator with three practical takeaways. First, model a two-tier compute architecture. The frontier model is only one component of an agent; the surrounding orchestration, tool calls, and data movement carry their own, separately billable footprint. Teams that budget only for GPU inference will systematically underestimate what a production agent actually costs to run.

Second, treat CPU capacity as a first-class infrastructure decision rather than an afterthought. The shift toward edge-distributed, CPU-based serving that Akamai specializes in matters most for latency-sensitive agents that need to execute close to the user or close to the data they manipulate.

Third, read the warrant as a signal about bargaining power. When a compute supplier pays in equity for the right to serve an AI lab, it confirms that demand for general-purpose inference is beginning to outstrip supply. Builders should expect CPU pricing and capacity to tighten, and should architect for portability across clouds instead of locking into a single accelerator vendor's stack.

What to watch

Anthropic has not disclosed exactly what it will use Akamai's CPUs for, and Akamai's revenue from the deal will not begin in earnest until the second half of 2027. The open question is whether this marks the start of a broader repricing of the unglamorous compute layer, and whether other labs follow with CPU-heavy infrastructure bets of their own. If they do, the GPU-first mental model that has dominated AI infrastructure planning for two years will need a serious rewrite.