Akamai’s new agreement with Anthropic draws attention to a part of AI infrastructure that is easy to overlook: the computing capacity needed to keep applications running around the model.
Akamai announced approximately $11.6 billion in commitments over seven years to support Anthropic’s growing CPU workloads. Its SEC filing says the two project plans were signed on September 18 under an existing services agreement. Payment commitments depend on delivery and availability requirements and are subject to termination provisions. Each plan’s seven-year term begins on its respective service start date.
The company also describes a possible further $9 billion expansion. That additional business is not yet a committed purchase. The accompanying warrant covers non-voting preferred shares, with vesting tied to payments and further commitments. It should not be read as Anthropic immediately owning five percent of Akamai’s voting common stock.
What happens around the model
An agent working through a long assignment may use a browser, retrieve documents, call other software, or run code. The model helps decide what to do, while other systems carry out those actions and retain the results. The full assignment therefore depends on more than the speed or quality of a single model response.
Akamai identifies CPU demand as the reason for this agreement. That supports a narrower conclusion than saying all AI should move to the edge: at least one major model developer is committing substantial resources to the surrounding compute infrastructure. The announcement doesn’t establish where every agent workload should run.
For organizations adopting agents, the useful question is what the assignment requires. A job that runs for hours needs a place to preserve its state. An agent that uses business software needs appropriate access. If work fails halfway through, someone needs to understand what completed and whether it is safe to resume. These requirements remain even as models improve.
What smaller organizations should look for
I see an opportunity for infrastructure providers to make that work easier to operate. A smaller organization could benefit from being able to choose an execution environment suited to its needs without building one from scratch. It could also benefit from changing models without having to redesign every connected process.
Those benefits depend on the terms and behavior of the systems offered. More infrastructure spending doesn’t automatically create customer choice. I’d want to know whether an organization can move its data, understand its costs, and replace a component without losing the work already done.
The same scrutiny applies to reliability. A service should make failures visible and give the person responsible a way to intervene. That is part of what makes long-running agents usable beyond a demonstration.
There is plenty of room to be enthusiastic about AI while asking these practical questions. As agents take on longer assignments, the surrounding systems will determine how much of their capability organizations can actually use. Akamai’s agreement is a substantial example of investment in that work; its eventual results will need to be judged as the capacity is delivered.