Cerebras WSE-4 (CS-4)
Overview
Cerebras WSE-4 (system codename CS-4) is Cerebras Systems' 4th-generation wafer-scale AI chip (Wafer-Scale Engine, WSE), expected to be announced in 2027 (after Cerebras IPO in May 2026). WSE-4 is a process upgrade + performance improvement over WSE-3, built on a 5nm or 3nm process (depending on TSMC N3 capacity), with AI Core count expected to increase from WSE-3's 900K to ~1.5M.
WSE-4 is Cerebras' first "fully public" product post-IPO, offering detailed SPEC benchmarks, pricing, and ecosystem support for the first time.
Core Specifications (Estimated)
| Item | WSE-3 (CS-3) | WSE-4 (CS-4) | Improvement |
|---|---|---|---|
| Process Node | TSMC 5nm | TSMC 3nm (estimated) | +1 generation |
| Transistor Count | 4 trillion | ~5-6 trillion (estimated) | 1.4× |
| Wafer Size | 300mm full | 300mm full | Same |
| AI Core Count | 900,000 | ~1,500,000 (estimated) | 1.67× |
| On-chip SRAM | 44 GB | ~80 GB (estimated) | 1.8× |
| On-chip SRAM Bandwidth | 21 PB/s | ~40 PB/s (estimated) | 1.9× |
| BF16 Compute | 125 PFLOPS | ~200 PFLOPS (estimated) | 1.6× |
| FP8 Compute | 250 PFLOPS (estimated) | ~400 PFLOPS (estimated) | 1.6× |
| TDP | ~25 kW | ~30-35 kW (estimated) | 1.3× |
| Launch | 2024 | 2027 expected | — |
⚠️ Not officially announced: Above are estimates. Cerebras only has publicly released WSE-3 info. WSE-4 detailed specifications subject to future Cerebras announcements.
Cerebras History and IPO
| Time | Event | Details |
|---|---|---|
| 2015 | Founded | Andrew Feldman founded Cerebras Systems |
| 2019 | WSE-1 announced | First wafer-scale chip (400K cores) |
| 2021 | WSE-2 announced | 850K cores, 20nm |
| 2024 | WSE-3 (CS-3) | 900K cores, 44GB SRAM, 125 PF BF16 |
| 2024-12 | Saudi G42 investment | G42 invested hundreds of millions (Middle East customer) |
| 2025 | OpenAI partnership | OpenAI procured Cerebras inference compute |
| 2026-04-17 | IPO filing | Submitted S-1 for listing |
| 2026-05 | IPO listed | Target Nasdaq "CBRS", valuation $22-25B |
| 2027 | WSE-4 expected | First post-IPO product |
Cerebras IPO Key Data:
- Valuation: $22-25B (based on latest funding round)
- 2025 Revenue: ~$510M (+150% YoY)
- 2025 Net Loss: ~$200M (still unprofitable)
- OpenAI major deal: $10B inference compute long-term contract
- Listing window: May 2026
WSE-4 vs WSE-3 vs NVIDIA Rubin R200
| Metric | WSE-3 (CS-3) | WSE-4 (estimated) | Rubin R200 |
|---|---|---|---|
| Form Factor | Single-chip wafer-scale | Single-chip wafer-scale | Single-card multi-die |
| AI Core Count | 900,000 | ~1,500,000 | ~10,000 (SM) |
| SRAM | 44 GB | ~80 GB | 288 GB HBM4 |
| SRAM Bandwidth | 21 PB/s | ~40 PB/s | 22 TB/s HBM4 |
| BF16 Compute (sparse) | 125 PFLOPS | ~200 PFLOPS | 25 PFLOPS |
| Single-chip BF16 Compute Ratio | 5× | 8× | 1× (baseline) |
| TDP | 25 kW | ~30 kW | 1.8 kW |
| Cooling | Liquid + massive system | Liquid + massive system | Liquid (card-level) |
| Ecosystem | PyTorch + JAX | Same | CUDA |
💡 WSE core advantage: On-chip SRAM + high bandwidth + high compute. Single-chip BF16 compute is 5-8× that of Rubin R200 (per single-chip comparison), but power and physical footprint are major issues (25-30 kW/chip vs 1.8 kW/card).
Cerebras CS-4 System (Estimated)
| Item | Configuration |
|---|---|
| WSE Count | 1 (single-chip wafer-scale) |
| Server Dimensions | 1U + large cooling (similar to CS-3) |
| MemoryX | 1.5 TB extended DRAM (external) |
| SwarmX | Multi-WSE interconnect (optional, up to 192 WSEs) |
| Total BF16 Compute | ~200 PFLOPS (single WSE-4) |
| TDP (single WSE-4) | ~30 kW |
Software Ecosystem
- Cerebras Software Platform (CSoft): Based on PyTorch
- JAX + Cerebras backend: Google integration
- vLLM 0.7+ Cerebras backend (estimated)
- HuggingFace integration
- Triton + Cerebras backend
- OpenAI compatible API (Cerebras Inference)
Use Cases
- ✅ Ultra-large-scale LLM training (WSE-4 single chip can hold 1T+ models)
- ✅ Ultra-low-latency inference (CS-4 Inference)
- ✅ Government/scientific computing (high compute + domesticization options)
- ✅ OpenAI inference customers (WSE-4 expected to support some OpenAI inference workloads)
- ❌ Edge deployment (power/physical footprint)
- ❌ Small/medium model training (not cost-effective)
Vendor Information
| Item | Details |
|---|---|
| Vendor | Cerebras Systems |
| First Disclosure | 2027 (expected) |
| Product Page | https://www.cerebras.net/ |
| Foundry | TSMC |
| IPO Status | Listed May 2026 (Nasdaq: CBRS) |
| Customers | OpenAI, G42, Mistral, Meta, and others |
Related Products
- Cerebras WSE-3 - Current flagship
- Cerebras WSE-2 - Previous generation
- NVIDIA Rubin R200 - Same-generation GPU
- NVIDIA Groq 3 LPX - Inference comparison
- Google TPU 8t - Training ASIC comparison
- Full Comparison Table