Entity — Cerebras
Status: active Owner: Finance / Charlie AGT-002 Entity ID: ENTITY-AI-CEREBRAS Visibility: PUBLIC Last verified: 2026-07-10
Snapshot
Cerebras is the wafer-scale challenger: one full 300mm-wafer die (WSE-3, 46,225 mm2, 4T transistors, 44 GB on-wafer SRAM) instead of reticle-sized GPUs, sold as CS-3 systems and, increasingly, as a low-latency inference cloud service. The company has pivoted its go-to-market emphasis from training systems to inference-as-a-service, where SRAM-resident weights give it a structural tokens/s-per-user advantage. Cerebras completed a Nasdaq IPO (ticker CBRS) in Q2 2026 per press reports; exact deal terms vary across secondary sources and should be pulled from the prospectus.
Chip Table
| Chip | Status | Key spec highlights | As-of | Source | Grade |
|---|---|---|---|---|---|
| WSE-3 (CS-3) | shipping | TSMC 5nm; 46,225 mm2; 4T transistors; 900,000 cores; 44 GB on-wafer SRAM; 125 PF peak "AI compute" (precision/sparsity basis not stated on chip page); MemoryX external memory 1.5 TB - 1.2 PB; clusters to 2,048 CS-3 | 2026-07-10 | https://www.cerebras.ai/chip + https://www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine | 🟢 (vendor page; FLOPs precision qualifier missing — flagged) |
| WSE-4 / next gen | rumored | no official disclosure verified this pass — n/a | 2026-07-10 | n/a | 🟠 |
Technical Trajectory
- Architecture: wafer-scale monolithic die trades yield/packaging complexity for on-chip memory bandwidth — weights live in 44 GB SRAM (secondary sources cite ~21 PB/s on-chip bandwidth; not on the official chip page — 🟡), eliminating the HBM round-trip that bounds GPU inference latency.
- Generation cadence: WSE-1 (2019, 16nm) → WSE-2 (2021, 7nm) → WSE-3 (2024, 5nm); ~2-3 year cadence. No WSE-4 disclosure verified as of 2026-07.
- Business pivot: from selling CS systems to operating an inference cloud; press-reported multi-thousand tokens/s per user on large open models (🟡, vendor-adjacent benchmarks — not MLPerf; Cerebras does not submit MLPerf closed-division results, so cross-vendor comparison stays qualitative).
- Customer concentration: pre-IPO filings reportedly showed heavy revenue concentration in G42/MBZUAI-related entities (🟡); a large OpenAI inference capacity agreement was widely reported 2026 with conflicting dollar/MW figures across outlets (🟠 until filing-confirmed).
- IPO: Nasdaq listing (CBRS) completed Q2 2026 per The Register (🟡); valuation/raise figures conflict across secondary sources ($23-27B valuation range reported) — use S-1/424B prospectus numbers only.
Watch Items
- Pull the actual prospectus (SEC EDGAR: S-1/424B for Cerebras Systems) — revenue, customer concentration, OpenAI agreement terms — upgrades most financial facts from 🟡/🟠 to 🟢.
- First post-IPO quarterly report (revenue mix: systems vs inference service).
- WSE-4 announcement signals (Hot Chips 2026, TSMC advanced-node commentary).
- Independent inference benchmarks (e.g. Artificial Analysis) for tokens/s claims; MLPerf participation remains absent.
Claims
- CLAIM-AI-SILICON-001 — Cerebras is the non-GPU merchant challenger data point in the accelerator-mix question (wafer-scale, not hyperscaler ASIC).
- CLAIM-AI-BOTTLENECK-001 — SRAM-only architecture is a natural experiment on whether HBM is the binding constraint; also newly IPO-filings-visible.
Sources
- 🟢 Cerebras chip page (WSE-3) — https://www.cerebras.ai/chip (accessed 2026-07-10)
- 🟢 Cerebras press release, WSE-3 announcement — https://www.cerebras.ai/press-release/cerebras-announces-third-generation-wafer-scale-engine (accessed 2026-07-10)
- 🟡 The Register, Cerebras IPO outcome — https://www.theregister.com/ai-ml/2026/05/15/cerebras-wafer-scale-ai-bet-delivers-blockbuster-ipo/5240821 (accessed 2026-07-10)
- 🟠 Assorted secondary IPO/OpenAI-deal figures (tech-insider.org, nerdleveltech.com, marketwise.com) — conflicting; not relied on for numbers (accessed 2026-07-10)
Changelog
- 2026-07-10: Created; first verification pass. IPO deal terms and OpenAI agreement figures deliberately left unpinned pending prospectus pull.