Akamai rose more than 17% after hours on an Anthropic commitment to spend $11.6B over seven years on its cloud services. Spread evenly, that is about $1.66B a year, somewhere near two-fifths of what Akamai books in revenue annually, landing on a company the market has priced for years as a slow-growing CDN and security vendor. The move is proportionate to the revenue. It is less proportionate to the terms. Anthropic also gets an option to take a stake of up to 5% of Akamai, so part of the price of winning the contract is paid by Akamai’s own shareholders, in their own equity, to the customer.
This is the second time in six weeks the structure has shown up in a large AI supply agreement. Marvell issued Google a warrant for roughly 7% of its shares in August, vesting against Google’s custom-products revenue. The logic is identical. The vendor wins a customer big enough to re-rate the company, and the customer asks for a claim on that re-rating as a condition of signing. Customer-acquisition cost moves off the income statement and onto the cap table, where it doesn’t dent reported margins and doesn’t appear in an after-hours print. What decides the cost is the strike. A warrant struck below where AKAM trades after the announcement is dilution from day one. One struck well above it is a performance giveback that only costs holders if the deal delivers what the 17% already assumes.
The other side of the tape came from Oracle, which sent Blue Owl a force majeure notice reserving the right to delay payments on the 2.45GW Project Jupiter data center in New Mexico if the site fails to launch in 2028. A force majeure notice filed two years ahead of the date it concerns is unusual. It works as a pre-positioned exit. Oracle is telling its financier that the payment stream underwriting the campus depends on a delivery date nobody in the chain fully controls, whether the holdup is grid interconnection, turbines, transformers or permits. Blue Owl is private credit. The tail risk of the largest AI campuses is being assigned, clause by clause, to lenders who mark quarterly at best and whose investors can’t exit on a Thursday evening the way AKAM buyers can enter.
Read together, the two items describe the same bargaining position from opposite ends. Suppliers are paying customers in equity to secure demand. Tenants are writing contractual escape hatches against capacity they committed to. In both cases risk flows away from whoever controls compute demand, and in September 2026 that party is the model lab and the hyperscale tenant, with the equipment vendor and the lender absorbing what they hand off.
Scale makes the allocation matter more than the aggregate. Brookings now projects $10.3T of US AI spending between 2025 and 2032, averaging 3.6% of GDP a year, the largest single-industry build-out measured against the size of the economy. A figure that large gets spent even if individual projects slip, because demand keeps getting signed. The live question for public equity is who carries the delay, and Oracle has just answered it for one 2.45GW site.
The demand itself is real, and it is getting financed at both ends of the quality spectrum. DeepSeek’s annualized revenue has reached $1B, up from under $500M a few months ago, and it wants to close a roughly $7.5B raise by late October, about 7.5 times run-rate revenue. Firmus, the Australian data center operator, expects a $77M after-tax loss for the first half of FY27 while preparing to raise $5B in an IPO. Capital is reaching the model company with doubling revenue and the infrastructure operator with losses on similar terms. That happens when capacity is scarce enough that borrower quality stops being the filter.
Private marks are moving faster than any public multiple can. TypeSafe announced a $40M seed last week at a reported $200M valuation and is now in talks at $10B or more, a 50x repricing in roughly seven days. DensityAI, started by former Tesla Dojo engineers, is raising hundreds of millions at $10B. OpenEvidence raised $250M at $15B, up from $12B in January, and may be open to a sale. Island raised $400M at $6.4B. None of these rounds publish their preference stacks, ratchets or side terms, and the Akamai warrant is a fair reminder of how much of an AI deal’s economics lives in terms the headline leaves out.
One data point cuts against the idea that the build-out is broad. ASML’s revenue share from Europe fell to 0% in Q1 and Q2 2026, from 1% in 2025 and 5% in 2024, and the company is now asking EU authorities to create demand for European chips. On the same day, Carney, Macron and other Western leaders pushed for a global AI supervisory body. The spending is concentrated in the US and Asia, and Europe’s share of the most important tool order book in the industry has gone to zero. For ASML that changes nothing in the next four quarters of revenue. It does change how durable its political standing at home is.
Akamai’s filing will carry the strike, the vesting conditions and whether exercisable shares scale with what Anthropic actually spends. If exercise tracks spend, as Marvell’s schedule tracks Google’s revenue, the dilution schedule is the revenue schedule, and the after-hours move priced only one of them.
The strike price is the number that matters.