It Is CPUs, Not GPUs — the Line Most Likely to Get Skipped
Amid a steady stream of GPU news, this $11.6 billion buys general-purpose CPUs. Akamai did not say what Anthropic will use them for, so what follows is inference rather than fact: **a great deal of an agent workload is not matrix multiplication.** Running tools, executing code, parsing web pages, calling APIs, managing sandboxes, pre- and post-processing around retrieval — that is ordinary computing, and its share rises as the number of agent steps in a single task grows. On September 24 this site covered Claude using roughly 950 agents over 21 hours and 210 million tokens to run a screen; in work like that, model inference is only one segment, and the orchestration and execution around it all runs on CPUs. If that inference holds, the deal reveals something more interesting than its size: **a frontier lab is signing a separate seven-year contract for the non-inference part of agent work.** Which implies that part has grown large enough to warrant planning capacity for it specifically, rather than buying it incidentally on existing cloud. It also explains why Akamai. Its strength was never GPU clusters; it is globally distributed edge nodes and conventional compute capacity — a better shape for an agent execution layer that needs network proximity, high concurrency and light individual calls than a centralized GPU campus is.
The Warrant Turns a Supply Contract Into Half an Equity Deal
The genuinely new part of this deal is the structure. Akamai issued Anthropic a warrant for nonvoting preferred stock convertible into 7.7 million shares, about 5% of shares outstanding, at $111.33; roughly 2% vests on the first payment, with the rest unlocking about 1% per additional $3 billion committed. It is the first time Akamai has attached a warrant to a cloud deal. Two things to see in that design. **First, it turns the customer into a stakeholder.** The more Anthropic spends, the more Akamai stock it gets; the higher Akamai's stock, the more that warrant is worth. The relationship shifts partly from buyer-pushing-down-seller-pushing-up to both sides pointing the same direction. **Second, the $111.33 strike is an anchor.** Akamai traded as high as $129.60 in late trading after the announcement — meaning the warrant was already in the money on the day the news broke. That is not coincidence but the ordinary effect of this structure: announcing a large deal lifts the stock, and the warrant's value rises with it. Note the double edge: for Akamai's existing shareholders, up to 5% of potential dilution is a real cost, exchanged for a seven-year revenue commitment that could extend to roughly $20 billion. Whether that trade is good depends on the margin on that $5.5 billion of build-out, and on whether the contract runs its full seven years.
$150–300 Million Next Year, Against $11.6 Billion
The two numbers most worth reading side by side: an $11.6 billion contractual commitment, and **$150 million to $300 million of expected revenue next year.** There is no contradiction — service only begins in the back half of next year, and revenue on a seven-year contract naturally sits later. But the comparison confirms the pattern this site has written about all month: **the headline number in AI infrastructure is total contracted value, and realization is spread across years.** Nscale on September 9, Crusoe's $140 billion against 1 gigawatt on September 20, SoftBank's staged contributions on September 22, Oracle's force majeure notice on a 2028 campus because of one pipeline on September 25 — today's story is a relatively healthy sample of the same pattern: **it at least put next year's revenue range in the announcement**, rather than only the total. Also worth recording is the shape of the up-front spend: about $5.5 billion to build the capacity, plus roughly $1.7 billion added to this year's capex to buy components such as memory in advance. That last detail matters — in the current component supply environment, being willing to lock in memory early says the company considers the delivery window more important than the price.
"Not Ironclad" Is Akamai's Own Language in the Filing
Finally the boundary, and this time Akamai stated it itself in a securities filing: the commitment depends on Akamai meeting certain delivery and service-availability requirements, and either company can end the agreement under certain conditions. So the accurate reading of "$11.6 billion over seven years" is: **a commitment with performance conditions, not a check already cleared.** That is not skepticism — every long-term supply contract carries such clauses — but a reminder to price in the conditionality before using the number in any valuation math. On the industry picture, this says one more thing: Anthropic's compute sourcing is diversifying. This site covered its six-year, $45 billion capacity contract with Nscale on September 9, and roughly $517 billion of compute contracts signed over eleven months on September 17; today's is seven years and $11.6 billion for CPUs. **A single lab signing multi-year contracts with multiple suppliers at once is both locking capacity and avoiding dependence on any one of them.**
via: Akamai's press release, TechCrunch, Business Standard, Investing.com