Cerebras Systems · one lens
Hamilton Helmer on Cerebras Systems
Wafer-scale silicon reads as a genuine cornered resource — CUDA's switching costs are the wall it still has to clear.
7 Powers sets five tests for a genuine Cornered Resource: idiosyncratic, non-arbitraged, ongoing, non-replicable, sufficient (7 Powers). Cerebras’ wafer-scale engine — one large piece of silicon per chip, instead of the small dies most chipmakers cut from a wafer and wire together — clears the idiosyncratic and non-replicable tests cleanly; no competitor ships the same architecture at comparable yield. AWS agreeing in March 2026 to deploy the Wafer-Scale Engine for inference alongside its own Trainium chips is a real vote from a hyperscaler with every incentive to build an alternative instead. That’s the Benefit half of Power, and it looks genuine.
The non-arbitraged test is where the case complicates. Abu Dhabi’s G42 supplied 87% of Cerebras’ first-half-2024 revenue while separately committing $335 million to Cerebras stock — a customer capturing equity in the resource it depends on, close to the structure 7 Powers uses to disqualify a coveted asset from counting as Power: preferential access purchased at a price that captures the rents is a transaction, not a Barrier (7 Powers). OpenAI’s compute relationship with Cerebras has followed a similar shape, growing from roughly $10 billion to more than $20 billion with OpenAI reportedly taking equity alongside it.
The larger Barrier Cerebras still has to clear isn’t Nvidia’s chips — it’s Nvidia’s CUDA software stack, the Switching Costs power built from years of developer tooling and trained models that make leaving costly in ways no single benchmark erases (7 Powers). A cornered chip design is a real advantage. Whether it converts into share depends on whether customers judge the migration off CUDA worth the wafer-scale gain — and so far, Mistral and Mayo Clinic aside, Cerebras’ largest relationships run through customers who are also its shareholders.
The seven Powers, one by one
| Power | Status | Why |
|---|---|---|
| Scale Economies | absent | not argued in the record |
| Network Economies | absent | not argued in the record |
| Counter Positioning | absent | not argued in the record |
| Switching Costs | absent | the record frames CUDA switching costs as Nvidia's barrier facing Cerebras, not a power Cerebras itself holds |
| Branding | absent | not argued in the record |
| Cornered Resource | contested | wafer-scale silicon clears idiosyncratic and non-replicable tests, and AWS's WSE deployment is a real vote — but G42 (87% of 1H2024 revenue) and OpenAI both hold equity in Cerebras, close to the 'price captures the rents' disqualifier on non-arbitraged |
| Process Power | absent | not argued in the record |
An editorial application of Hamilton Helmer's published framework — not their actual view, affiliation, or advice.
Where this lens comes from
Hamilton Helmer's framework is set out at the lens page, drawn from 7 Powers. This page applies it to Cerebras Systems and nothing else — the company's full record is in the dossier.