Good Friday,
OpenAI expands its Cerebras commitment (now past $20B over 3 years)
Cerebras builds AI chips. But they built them in a way nobody else did.
The standard approach (Nvidia, AMD, everyone else): You take a silicon wafer (a thin disc about 12 inches across), cut it into hundreds of small rectangular chips, package each one individually, then wire thousands of them together with cables and switches to form a big AI cluster. A GPU is one of these small chips. A data center rack holds dozens. A training cluster uses tens of thousands.
What Cerebras did: Don't cut the wafer. Keep the entire 12-inch disc as one single chip.
That's it. That's the whole insight. Their flagship product, the WSE-3 (Wafer-Scale Engine 3), is roughly 46,000 square millimeters of silicon. An Nvidia H100 is about 800 square millimeters. So one Cerebras chip is about 56x the physical size of a top-end GPU, packing around 900,000 cores and 44GB of on-chip memory.
Wafer-scale is brilliant for one thing: serving a trained model as fast as possible.
Cerebras had effectively one anchor customer before this (G42), and the deal is specifically for low-latency ChatGPT inference. It sits alongside Samsung HBM4 for Project Titan, AMD accelerators, and Nvidia training clusters. This is multi-silicon diversification, not consolidation.
What the deal does is validate Cerebras as an alternative to the Nvidia GPUs. It gives rivals another door to knock on.
We also notice Broadcom is now the implementation layer for OpenAI's Titan ASIC AND the 3.5GW Anthropic-Google TPU expansion announced April 7. Two of the three US frontier labs are routing custom silicon through the same vendor on the same TSMC N3 process. Watch item: TSMC N3 capacity disclosures and any Broadcom ASIC backlog commentary in next earnings. That's where the real scarcity signal shows up.
Know someone allocating to AI infra? They probably saw the OpenAI headline without the analysis. Forward this email.
Stories That Matter
OpenAI bypasses SK Hynix and Micron to source HBM directly from Samsung, tightening spot-market supply for rivals. Vertical integration becomes the shortage playbook.
Claude Opus 4.7 in GitHub Copilot shifts inference costs to Microsoft's balance sheet. GitHub's per-seat model faces margin pressure if inference spending outpaces subscription revenue.
Regime Snapshot
Compute (CRS): 66, scarcity.
Memory (MRS): 79, Shortage.
The Daily Chain gives you the headline. The terminal shows you which companies are exposed, how the constraint map is shifting, and what the regime history says about timing.
See you tomorrow,
Teng