Telcos, Akamai’s $13 Billion Anthropic Deal Is a Blueprint

A pattern is emerging in AI infrastructure: the most interesting deals are not coming from hyperscalers. Akamai’s $13B Anthropic Deal: A Blueprint for Telcos is a prime example of this trend. Anthropic has committed $11.6 billion over seven years to Akamai for dedicated CPU capacity and managed support. Added to a $1.8 billion agreement signed in May, the relationship is worth about $13.4 billion, with a further $9 billion possible later. Industry analyst Sebastian Barros argues the deal holds a lesson for telecom operators. There is not a single GPU in the new contract.

Not Colocation, Not a Hyperscaler Deal

Akamai expects to spend about $5.5 billion building the capacity, including $1.7 billion this quarter to secure components such as memory. It will not build one giant campus. Instead, it plans to secure power in blocks of roughly 10 to 30 MW from several colocation providers, fill them with servers it owns and operates, and connect them through its network.

Calling this colocation misses the point. Anthropic is not renting racks or megawatts. It is buying compute, and the colocation providers supply only one input. It is not an AWS-style deal either. Akamai combines network, cloud software, hardware procurement and third-party facilities into one dedicated product. Telcos already own many of these ingredients: buildings, power, fiber, redundant paths and field operations. Operators that have already moved to a cloud-native 5G SA core will recognize the same cloud-operations discipline.

Why the Order of Commitment Matters

The most important detail is the sequence of risk. Anthropic committed in May and expanded in September, and only then did Akamai commit its capex. It did not forecast demand. It had demand under contract.

Much of telecom’s edge strategy ran the other way: deploy sites, launch a platform, then look for applications. The price of Akamai’s certainty is lower economics. It normally targets about 50 cents of annual recurring revenue per dollar of capex, while this contract works out closer to 31 cents. Anthropic also received warrants that could reach about 5% of Akamai. In return, Akamai gets scale and a seven-year anchor customer. By Barros’s estimate, the contract could grow to more than four times the current annualized size of its cloud infrastructure business.

Why CPUs, and Why Now

Agents do more than generate tokens. They plan, call tools, compile code, run tests, query databases and drive browsers, and much of that work runs on conventional CPUs, memory and storage. Intel sees the CPU-to-GPU ratio moving from roughly 1:8 in training toward 1:4 as inference grows, and potentially close to 1:1 in agentic environments. TrendForce projects a similar direction.

Anthropic has not disclosed what will run on Akamai’s servers. But the commercial signal is clear: frontier AI is creating multibillion-dollar demand for conventional compute. That kind of capacity fits in small power blocks spread across many sites, which is what telco footprints already offer.

What This Means for Operators and Vendors

Get the customer before the capex. The operator that signs a workload first can deliver an initial tranche, prove it can operate the infrastructure, and expand only when the customer commits to more.

Sell compute with a delivery date, not “edge computing.” The product is dedicated capacity with an availability commitment and a price, delivered across several facilities that behave as one footprint.

Choose the layer deliberately. Verizon already shows the first layer working: repurposed central offices, a dark-fiber agreement with Google worth well over $1 billion, and a partnership with Vultr. Akamai shows the next layer up, where the operator also sells the compute itself.

Revenue per megawatt is not return per megawatt. Akamai’s large cloud contracts are expected to generate roughly $22 million per MW per year, against $3 to $4 million for colocation. But the compute seller also carries servers, depreciation, procurement and technology risk. A telco selling powered space and fiber may earn an excellent return with far less capital.

Plan for the downside. Akamai is concentrating a large share of future cloud revenue in one customer. CPU and memory shortages can ease, and agents can become more efficient.

Frequently Asked Questions

Is the Akamai and Anthropic deal a colocation deal? No. Akamai leases power from colocation providers, but it assembles, owns and operates the servers and sells the resulting compute to Anthropic.

Why does a CPU-only contract matter? Agentic AI shifts a growing share of work to conventional CPUs, memory and storage. Intel and TrendForce both project CPU-to-GPU ratios moving toward 1:1 to 1:2 for agentic workloads.

What should telcos take from it? Secure the customer first, then let the contract pull the capex. Decide how far up the stack to go: powered real estate and fiber, a partnership with a neocloud, or selling compute directly.

Going Deeper

Understanding how telco infrastructure, edge, and AI-native network architectures connect is covered throughout our 5G training catalog. For operators assessing which layer of the AI stack to target, our 5G consulting services can help structure the decision, and you can contact our team to scope a program.