The short answer
The largest contract in Akamai’s history is a bet that AI agents will need huge amounts of ordinary compute. Anthropic has spent the year locking in GPU and accelerator capacity. This deal is different: it buys distributed CPU capacity, the kind that runs tool calls, code execution, browsing and orchestration around a model rather than the model itself.
For teams building on LLMs, the lesson is practical. An agent’s bill and its bottlenecks are not only tokens. Sandboxes, workers, queues and API calls add CPU load that grows with every step an agent takes. Plan that layer the way you plan any production system, as part of your cloud and DevOps architecture, not as an afterthought to the model choice.
What did Anthropic and Akamai agree?
Akamai announced the agreement after the market closed on Thursday, September 24, and filed the release with the SEC. Anthropic commits $11.6 billion over seven years to Akamai’s cloud services to support growth of CPU workloads at scale. CEO Tom Leighton said Anthropic chose Akamai’s capabilities “for building and operating AI infrastructure at scale.” Akamai shares jumped by double digits on Friday before giving back part of the gain.
The structure is unusual. Anthropic gets a warrant for up to about 5% of Akamai, and more of it vests the more Anthropic buys. That ties the supplier’s upside to the customer’s spending and gives Akamai the confidence to pre-buy components and memory now, in a market where both are scarce. The Decoder counts this as one of a series of compute deals Anthropic has signed in under a year, alongside arrangements with Amazon, Google, Microsoft and AMD for accelerators and cloud capacity.
Why does an AI lab need this many CPUs?
Training and inference run on GPUs and other accelerators. Agents add a second layer of work. Each task can mean dozens of steps: calling APIs, running generated code in a sandbox, fetching and parsing pages, reading files, checking results and deciding what to do next. That is classic general-purpose computing. Bloomberg noted that networks of CPUs are generally better for the coordinated tasks agentic AI requires.
Akamai brings a widely distributed footprint built for content delivery and edge computing. Placing agent execution close to users and to the systems agents call can cut latency on each step, and small delays multiply across a long agent run. Neither company has published technical details of how the capacity will be used, so treat specific architectures as expectations rather than announced facts.
What it means for US & EU software teams
First, agent costs are not only token costs. Many business cases for agents are built on per-token pricing alone. In production, sandboxes, browser sessions, background workers and retries can add a large share of the bill, and they scale with the number of steps, not the number of users. Instrument CPU time and memory per agent task from the first pilot, so the unit economics you present are real.
Second, capacity is still being fought over. A lab that already has deals with the largest clouds is still signing multi-year contracts for more. Tight supply shows up downstream as rate limits, regional gaps and changing prices. If your product depends on Anthropic’s Claude models or any single provider, design for throttling: queues, backoff, fallback models and graceful degradation.
Third, keep the execution layer portable. The model provider may be fixed by quality or contract, but the code that runs tools, sandboxes and orchestration can live on any cloud or at the edge. Containerized workers and infrastructure as code make it possible to move that layer for cost, latency or data residency, which matters for EU workloads under GDPR and, for financial entities, DORA’s rules on third-party concentration risk.
What to do now
- Measure the non-model side. Track CPU, memory and wall-clock time per agent step alongside tokens.
- Set limits per agent. Cap concurrency, steps per task and spend per run so one runaway loop cannot flood your workers.
- Isolate execution. Run generated code and browsing in short-lived sandboxes with tight network and file access.
- Plan for throttling. Add queues, retries with backoff and a fallback model path for when a provider limits you.
- Keep it portable. Package workers as containers and define infrastructure as code so you can move regions or clouds without a rewrite.
Frequently asked questions
What did Anthropic and Akamai announce?
On September 24, 2026, Akamai disclosed a seven-year agreement under which Anthropic commits $11.6 billion to Akamai cloud services, focused on CPU workloads at scale. The deal can grow by about $9 billion to roughly $20 billion. Akamai also issued Anthropic a warrant for up to about 5% of its common stock at $111.33 per share.
Why would an AI lab buy CPUs instead of GPUs?
GPUs run model training and inference, but agentic AI also creates a lot of general-purpose work around the model: running tools and code, sandboxes, browsing, orchestration, data handling and API calls. That work runs on CPUs. Bloomberg reported that networks of CPUs are generally better suited to the coordinated computing tasks agents require.
When does the Akamai capacity come online?
Akamai said the deal has no impact on its 2026 revenue guidance. It expects about $5.5 billion of capital spending tied to the commitment and is adding about $1.7 billion to 2026 capex to buy components and memory in advance. TechCrunch reported that revenue from the deal is expected to begin in the second half of 2027.
Does this change how we should use the Claude API?
Not directly or immediately. The deal adds capacity behind Anthropic’s services; it does not change API terms, pricing or data residency today. It is a signal to plan for: agent workloads are compute-heavy on both GPU and CPU, and providers are still securing capacity, so build for rate limits, fallback models and cost controls.
What should teams building AI agents plan for?
Budget and size the non-model side of agents, not only tokens. Measure CPU time for tool execution, sandboxes and orchestration, keep that layer portable across clouds, set per-agent cost and concurrency limits, and design fallbacks for when a model provider throttles or degrades.
Sources
Akamai Technologies — Form 8-K, Exhibit 99.1 press release (SEC)
Bloomberg — Anthropic Strikes $12 Billion Deal With Akamai for AI Computing
TechCrunch — Anthropic to pay Akamai $11.6 billion over seven years in cloud deal
CNBC — Shares of Akamai surge after deal with Anthropic
The Decoder — Anthropic signs $11.6 billion cloud deal with Akamai