Lynn Doan’s Bloomberg report puts Anthropic’s $11.6 billion agreement with Akamai in the context of the lab’s growing compute commitments. That is the right entry point, but the headline can make a future service schedule sound like present-day scale. The public documents tell a more conditional story. Akamai is promising dedicated cloud capacity; Anthropic is committing to pay for it subject to delivery and service-availability requirements. To support that service, Akamai expects to spend about $5.5 billion on capacity before the contract reaches its full revenue run rate.
My reading is that the deal is both a meaningful demand signal and a delivery test. It shows that Anthropic wants a large block of general-purpose compute from Akamai. It does not tell us what tasks will run there, how much capacity is already installed, or whether the build will arrive on schedule. For infrastructure buyers, those distinctions are more useful than treating $11.6 billion as cash already earned.
A commitment with a delivery gate
Akamai announced the expanded agreement on September 24, 2026. Its release describes a seven-year contractual commitment for Anthropic to use Akamai Cloud’s distributed infrastructure and software as its CPU workloads grow. The companies may expand the relationship by up to another $9 billion, but that additional business would require future purchases on mutually agreed terms. It is not part of today’s committed amount.
Akamai’s Form 8-K adds conditions that change how the number should be read. The $11.6 billion comes from two project plans under a master services agreement signed on May 5. Each plan has its own seven-year initial term, starting on its respective service start date. The filing says the commitment is subject to delivery and service-availability requirements, as well as termination rights. Anthropic may terminate a project plan after a material outage if the agreement’s conditions are met; the contract also includes remedies for uncured breaches and certain corporate events.
That makes “seven years” an operating term, not a clock that began when the press release appeared. It also means the public commitment is not unconditional in the way a headline figure can imply. Akamai has to supply the agreed capacity and service; the SEC filing does not say that Anthropic owes the entire sum regardless of delivery or availability. The full master agreement and its detailed schedules were not included in the filing, so the precise service thresholds and remedies remain unknown.
This is a real commercial commitment, not merely a statement of interest. But the amount alone does not establish that all the compute is running or that all the revenue is secured against every operational outcome. The contract places delivery and availability at the center of the story.
The supplier spends before the service ramps
Akamai’s September 24 investor presentation makes the timing more concrete. The company estimates about $5.5 billion in capital spending tied to the commitment, with most of the outlay before the projected full revenue rate. Akamai says the first phase includes supply-chain purchases, including memory, in late 2026. It expects the service to begin in the second half of 2027 and ramp through 2028.
| Timing | Akamai’s estimated capex | Akamai’s estimated service revenue |
|---|---|---|
| Q4 2026 | About $1.7 billion | None |
| 2027 | About $3.1 billion | $150 million–$300 million for the year |
| 2028 | About $700 million | Ramping toward full rate by year-end |
| After full ramp | No additional build capex in this schedule | About $1.7 billion annualized |
These are Akamai’s estimates, not audited outcomes. The company says it expects to reach the full contracted revenue rate by the end of 2028, then recognize roughly $1.7 billion a year for the remaining term. The initial service start is projected for late in the second quarter of 2027, with $150 million to $300 million of revenue for that year. The schedule therefore contains a substantial gap between buying and deploying equipment and collecting the full service revenue. Akamai reported $99 million of Cloud Infrastructure Services revenue in the quarter ended June 30, 2026, up 39% year over year. At the projected $1.7 billion annual rate, this one agreement would amount to about 4.3 times a simple annualization of that quarter’s CIS revenue ($396 million). That is a scale comparison between one future contract estimate and one quarter of an existing segment; it is not a like-for-like forecast, and other customer commitments and segment growth will change the baseline.
That gap is where the operational risk lives. Akamai must secure components, arrange data-center capacity, install systems, and meet the project plans’ requirements before the service can produce its projected run rate. The investor deck says the 2026 spending includes pre-purchasing critical components such as memory. It does not publish component quantities, supplier names, site locations, or a detailed equipment delivery calendar.
Akamai also gave a power estimate for its wider cloud-infrastructure contract portfolio. It expects the Anthropic commitment, together with other large customer commitments totaling about $14.4 billion in contract value, to require 95–105 megawatts once deployed and to support about $2.2 billion in annual revenue at full ramp. That is a portfolio-level estimate. It cannot be assigned to Anthropic alone, and it does not tell us how much power the new project plans will draw at any single site.
The company’s schedule is evidence of a large planned build, not proof of the final economics. Subtracting $5.5 billion in projected capex from $11.6 billion in contract value would leave out colocation, power, operating costs, financing, equipment life, and the warrant’s accounting treatment. Akamai says the warrant’s fair value reduces recognized revenue over the contract. The public materials do not give enough detail to calculate the agreement’s profit or return on capital.
CPU capacity is not a workload map
Akamai describes the agreement as supporting Anthropic’s growing CPU workload. A CPU is a general-purpose processor, and Reuters and Bloomberg both identify CPU capacity as the defining feature of the deal. That makes the agreement notable in a market where public infrastructure announcements often foreground specialized accelerators.
But the public announcement stops at that level. It does not name the CPU supplier or model, give server counts, specify where the systems will sit, or explain which Anthropic services will use them. It does not say whether the capacity is for application logic, data movement, agent tools, API services, or another task. Those are examples of work that can involve CPUs in computing systems generally; they are not disclosed uses in this contract.
That distinction matters. The deal supports a narrow conclusion: Anthropic has committed to a major CPU-focused cloud purchase, subject to delivery and availability terms. It does not show that CPUs are replacing accelerators for model training or inference, nor does it establish a new industry-wide CPU shortage. Without workload details or utilization data, the agreement cannot tell readers how much CPU capacity one AI service needs or what performance Anthropic expects to achieve.
The shape of the purchase still raises a useful infrastructure question. A frontier AI service includes more than the processors that perform neural-network calculations. It also runs conventional software and service layers around those calculations. The contract suggests Anthropic is securing a substantial block of that general-purpose capacity from Akamai. It leaves unanswered whether the reason is price, network design, geographic placement, supply availability, or some combination. Akamai describes its network as distributed and spanning thousands of points of presence, but neither company has said which features drove Anthropic’s decision.
Akamai has also described a separate four-year, $200 million agreement with an unnamed U.S. technology company for a multi-thousand NVIDIA Blackwell GPU cluster. That disclosure shows Akamai sells accelerator capacity as well as CPU-focused services, but the customer was not named and nothing links that cluster to Anthropic. It cannot establish whether this agreement complements another Akamai deployment or how Anthropic divides work across processors.
The restrained reading is also the most useful one for planners: treat the CPU designation as a clue about the kind of capacity being bought, not as a blueprint for Anthropic’s architecture.
A warrant aligns future business, with a cost
The agreement also includes an equity warrant. Akamai issued it on September 18 alongside the project-plan expansion. The SEC filing describes a right to buy up to 387,051 non-voting convertible preferred shares, initially equivalent to about 7.7 million common shares, at an exercise price equivalent to $111.33 per common share. The warrant is not the same as Anthropic already owning a stake in Akamai.
Vesting is tied to commercial activity. Akamai’s release says a portion representing about 2% of its common stock is expected to vest with the $11.6 billion commitment. Further tranches, representing up to about 3%, would depend on as much as $9 billion in additional purchases. The filing describes the first tranche as tied to the first payment under Project Plan 3 and the remaining tranches to successive additional commitments, subject to stated conditions. The additional purchases still require mutual agreement.
This structure links a potential customer benefit to more business between the companies. It also affects how the headline commitment translates into reported revenue: Akamai’s investor materials say the fair value of the warrant is deducted from revenue over the contract’s life. That accounting treatment makes the contract total and recognized revenue different measures. It does not, by itself, tell us the agreement’s cash profit or whether the warrant will be exercised.
The warrant is one more reason to read the deal as a negotiated infrastructure arrangement rather than a simple cloud subscription. Yet the public evidence does not show that Anthropic received the equity right in exchange for taking on Akamai’s build risk, or that the arrangement replaces other compute contracts. The documents establish the commercial terms they disclose; they do not reveal the full bargaining history.
What would count as delivery?
Akamai’s $11.6 billion agreement is consequential because the company is planning a multi-year capacity build around a named customer. The investor schedule is unusually useful: it states when the provider expects to spend, when service revenue should begin, and when the full run rate could arrive. The SEC filing is equally important because it makes delivery and availability part of the commitment rather than details to assume away.
The limits are just as concrete. The $9 billion expansion remains prospective. Akamai’s power estimate covers multiple customers. The specific CPUs, sites, workload mix, and actual utilization are undisclosed. The full master services agreement and detailed availability provisions were not in the 8-K. The independent coverage confirms the announcement, but it does not independently validate future deployment or economics.
The main turn in this story is from buying compute to operating it. Demand matters because no provider builds this much capacity without a customer willing to contract for it. Delivery matters because Akamai plans to spend billions before the revenue reaches its expected rate. My view is that the headline is a credible signal of demand, while the schedule and availability clauses are the test of whether that demand becomes useful infrastructure.
Readers should watch two things in later filings: whether Akamai’s reported spending and service revenue track the schedule, and whether the company reports the service start and availability milestones needed for the project plans to proceed. Those disclosures will say more than the possible $20 billion total. Until then, the agreement is real as a commitment and unfinished as a service.