Anthropic's $11.6B Akamai deal is a CPU bet

Every compute headline this year has been about GPUs. This one is not. Anthropic committed $11.6 billion over seven years to Akamai for distributed cloud capacity built around CPUs — the ordinary chips that run code, hit APIs and serve web requests. If you operate agents, the Anthropic Akamai deal is a pricing signal about the half of your bill that nobody benchmarks.

What actually happened

Per Akamai's 8-K filing and TechCrunch, the structure is unusual in both directions:

• $11.6 billion over seven years, with an option to expand by another $9 billion — roughly $20 billion at the ceiling.
• The master services agreement dates to May 5, 2026; the two project plans behind the $11.6 billion were signed September 18, 2026.
• Akamai issued Anthropic a warrant for non-voting convertible preferred stock equal to up to 7.7 million common shares — about 5% of Akamai — at $111.33 per share. Roughly 2% vests with the initial commitment; the remaining ~3% only if Anthropic spends the additional $9 billion.
• Akamai is raising 2026 capital expenditure by about $1.7 billion to lock in components including memory, with roughly $5.5 billion in total capex tied to the commitment.
• Akamai guides to $150–300 million of revenue in 2027, starting in the second half, reaching an annual pace near $1.7 billion by the end of 2028.

Note the direction of the equity. Nvidia invests in its customers. Here the supplier handed the customer a path to a fifth of its own stock — payment for a commitment large enough to require rebuilding the supplier's balance sheet around it.

Why CPU capacity matters for your business

Agents are mostly not inference. A single agent run is one or two model calls wrapped in dozens of ordinary operations: fetching pages, parsing documents, running code in a sandbox, hitting your CRM, retrying a flaky endpoint, writing to Postgres. That work runs on CPUs. When the company with the clearest view of agent workloads signs an eleven-figure CPU contract, it is telling you which line grows fastest as agents move from demo to production.

Your own bill has the same shape and you probably are not measuring it. Most teams track tokens per run because the dashboard makes it easy. Almost nobody tracks CPU-seconds, egress and storage per run. Instrument one workflow end to end — model spend, compute, bandwidth, database — and you will usually find the token line is not the one that surprised you at scale.

Distributed beats centralized for the boring half. Akamai's asset is edge presence, not a hyperscale campus. Agent work is latency-sensitive and chatty: hundreds of small calls where round-trip time, not FLOPs, sets the wall-clock. Put the orchestration close to the data it touches. That is a design choice you can make this quarter without a seven-year contract.

Seven-year commitments at your scale are a mistake. The lesson from the deal is the workload shape, not the term. Keep your orchestration layer portable — containers, no proprietary runtime primitives — so the CPU half of your stack can move when someone prices it better. It will happen faster than seven years.

Sources: Akamai 8-K exhibit (SEC), TechCrunch.