Akamai Signs $1.8 Billion Seven-Year Cloud Contract With Anthropic

Akamai Inference Cloud, introduced later, uses Nvidia RTX PRO 6000 Blackwell servers and BlueField-3 data processing units. These products target inference rather than training. Training large models requires tightly coupled GPUs in a few centralized facilities.

Inference involves millions of low-latency requests best served from locations near users, an architecture that aligns with a network originally built for content delivery.

The Akamai agreement is the latest in a series of infrastructure moves by Anthropic. The company reported 80-fold year-over-year growth in the first quarter against an internal target of 10 times, pushing its annualized revenue run rate above $30 billion.

It recently secured all available capacity at SpaceX's Colossus 1 data center in Memphis, encompassing more than 220,000 Nvidia GPUs and over 300 megawatts of power. Anthropic also expanded its commitment with Google Cloud in April in a five-year arrangement reported to total roughly $200 billion.

The company maintains additional agreements with AWS for Trainium 2 chips, CoreWeave for GPU access, and direct deals with Nvidia and Broadcom for custom silicon. The Akamai deal provides distributed inference capacity to complement those centralized resources.

Enterprise users accessing Claude models through Anthropic's application programming interface now depend on the AI company's routing decisions across multiple suppliers rather than any single hyperscaler's infrastructure. This fragmentation challenges procurement assumptions built around co-location with model providers.

Akamai's forward-looking statements cite standard risks including the customer's ability to meet purchase obligations and the company's capacity to deploy infrastructure on schedule. Supplier relationships in frontier AI have shifted rapidly, with Anthropic itself altering its mix in the past 90 days.

Akamai's edge network offers advantages for low-latency inference but its overall compute scale remains smaller than that of major hyperscalers by an order of magnitude. The deal positions edge infrastructure as a complement to centralized training capacity rather than a substitute.

The transaction indicates that boards and technology leaders should treat the AI supplier base as broader than the three primary hyperscalers. The development underscores how inference workloads are reshaping perceptions of long-established internet infrastructure layers.

A network built to accelerate video delivery has become paid infrastructure for one of the leading AI assistants, with contract values large enough to drive significant share-price movement.