Anthropic–Akamai inference deal
Where it stands: Anthropic has signed an $11.6 billion, seven-year deal with Akamai for CPU-based distributed inference, after an earlier $1.8 billion deal in May, and the total could reach about $20 billion; Akamai issued Anthropic a warrant for roughly 5% of its common stock. The bet is that inference is constrained by memory bandwidth rather than GPUs, so work spreads into conventional data-centre space instead of giant GPU campuses. Watch for which Claude models actually run on CPU and whether other AI labs answer with non-GPU inference deals.
Latest
- 2026-10-17 (~2:30 AM ET, nichochar RT of Notion's Sarah Sachs — x.com/nichochar): the CPU-inference argument gets a named second voice — Notion argues "a lot of things should be served on CPU that are served on GPU," citing "determinism in executed code"; nichochar frames it as "CPUs underrated. GPUs overrated." No workload-disclosure detail from either company in the reporting.
- 2026-09-28 opened from the ai digest (TechCrunch, TechTarget, TNW, techtimes). Second Akamai–Anthropic deal after May's $1.8B (Bloomberg).
Background
- The thread: Anthropic's compute strategy diverging from the GPU-consensus — a $11.6B, seven-year deal with Akamai (announced Sep 24 via GlobeNewswire; entered our feed four days late — a business-wire gap) for CPU-based distributed inference on Akamai Cloud, potentially +$9B to ~$20B total, with a warrant for ~5% of Akamai common stock issued to Anthropic (a first for Akamai). Frame: AI inference as memory-bandwidth-constrained, not GPU-bound — inference distributed into conventional co-location facilities instead of gigawatt campuses.
- Watch for: the WSJ/tier-1 corroboration of the warrant terms, what workloads actually run on CPU (which Claude models/tiers), whether OpenAI/Google answer with non-GPU inference deals, and the capex-vs-revenue math against the IPO frame.