Inference compute: the stack
AI infrastructure — who actually sells inference compute
Inference compute is no longer one market — it's a three-layer stack. We mapped who actually sells it, and where each player stands commercially.
The shape of it:
→ Host CPUs are suddenly strategic. Agentic AI is CPU-bound, and ARM's response was its first in-house chip in 35 years — co-developed with Meta, production H2 2026. Qualcomm is coming for the same socket using ARM's own instruction set.
→ The accelerator layer is splitting into a speed tier and a cost tier. SRAM architectures (Cerebras, Groq's LPU — now licensed into Nvidia for $20bn) own latency-critical agentic serving. LPDDR architectures (Qualcomm AI200/AI250) win batch, long-context and multi-model hosting on cost per token. Both are attacks on the incumbent GPU + HBM model.
→ Memory is the binding constraint. HBM is effectively sold out through 2026. Three suppliers. That's where pricing power lives.
The commercial stages tell their own story: Nvidia at >$100bn run-rate; Cerebras now public at ~$52bn market cap on $828m of FY26E revenue with a >$10bn OpenAI pipeline; ARM's CPU not producing material revenue until 2028; Graphcore now captive silicon inside SoftBank's Stargate buildout.
Same technology wave, wildly different points on the commercialisation curve — which is exactly where the interesting valuation questions sit.
Full map below. Happy to share the working behind any of the stage calls.