Anthropic just made the most interesting bet in AI infrastructure — and it's not about GPUs.
The Summary
- Akamai signed a $11.6B, 7-year deal to provide CPU-based computing power to Anthropic, sending AKAM shares soaring
- The deal is entirely CPU-focused, breaking from the GPU-obsessed AI infrastructure playbook every other frontier lab is following
- Signal: Anthropic sees a compute arbitrage opportunity that competitors are blind to, or they're building inference infrastructure at scale while others burn cash on training
The Signal
Anthropic just committed $11.6 billion over seven years to Akamai for CPU compute. Not GPUs. CPUs. The chips your laptop runs on, not the $40,000 Nvidia H100s that OpenAI and Google are stockpiling like nuclear warheads.
This is either brilliant or catastrophic, and the answer tells you everything about where AI economics are headed. Akamai's entire deal with Anthropic is CPU-based, which means one of two things: Anthropic has cracked inference efficiency in ways that don't require the GPU gold rush, or they're betting that the future of AI workloads splits cleanly between training (GPU-heavy) and serving (CPU-viable at scale).
"Analysts say the deal should boost Akamai's growth and its profitability profile looks attractive, given the contract is entirely focused on CPUs."
Here's why this matters. Training frontier models is a GPU bonfire. You burn billions to create the model. But inference, the actual serving of responses to users, is where the money gets made or lost at scale. If you're serving millions or billions of queries, GPU costs eat your margins alive. CPU inference is orders of magnitude cheaper per query, but historically too slow for complex models.
Anthropic is signaling they've solved the latency problem, or they're willing to trade some response speed for unit economics that actually work. Either way, it's a direct challenge to the "more GPUs equals better AI" theology. Consider the implications:
- Akamai gets a revenue stream larger than its entire 2025 annual revenue, paid over seven years
- The profitability profile is "attractive" because CPUs don't have Nvidia's margin embedded in every chip
- Anthropic gets compute costs that could be 5-10x lower than GPU-based inference at scale
This is the kind of infrastructure bet that either makes Anthropic the most profitable AI lab or leaves them serving slower responses than competitors. The $11.6B commitment suggests they're confident it's the former. They're not renting this capacity month-to-month. They're locking in seven years of CPU compute, which means they see their inference workload scaling massively and they want cost certainty.
The Akamai angle is equally telling. They're a content delivery network that's been trying to break into cloud compute for years, getting steamrolled by AWS, Azure, and Google Cloud. Now they're landing the biggest AI compute contract announced this year, and it's for the one thing the hyperscalers aren't competing on aggressively: CPU-based AI inference at massive scale.
The Implication
Watch how other AI labs respond. If Anthropic's CPU inference bet works, you'll see a wave of "inference optimization" announcements from competitors who suddenly discover they too can run on cheaper hardware. If it doesn't, Anthropic will be back in the GPU market within 18 months, and this will be studied as a $11.6B lesson in why you don't fight the hardware curve.
For Akamai, this is existential validation. They just became an AI infrastructure player overnight. For the rest of us, it's a signal that the AI infrastructure stack is still being written, and the assumptions everyone's making about what hardware you need might be wrong.