Skip to main content

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)

ItemWSE-3 (CS-3)WSE-4 (CS-4)Improvement
Process NodeTSMC 5nmTSMC 3nm (estimated)+1 generation
Transistor Count4 trillion~5-6 trillion (estimated)1.4×
Wafer Size300mm full300mm fullSame
AI Core Count900,000~1,500,000 (estimated)1.67×
On-chip SRAM44 GB~80 GB (estimated)1.8×
On-chip SRAM Bandwidth21 PB/s~40 PB/s (estimated)1.9×
BF16 Compute125 PFLOPS~200 PFLOPS (estimated)1.6×
FP8 Compute250 PFLOPS (estimated)~400 PFLOPS (estimated)1.6×
TDP~25 kW~30-35 kW (estimated)1.3×
Launch20242027 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

TimeEventDetails
2015FoundedAndrew Feldman founded Cerebras Systems
2019WSE-1 announcedFirst wafer-scale chip (400K cores)
2021WSE-2 announced850K cores, 20nm
2024WSE-3 (CS-3)900K cores, 44GB SRAM, 125 PF BF16
2024-12Saudi G42 investmentG42 invested hundreds of millions (Middle East customer)
2025OpenAI partnershipOpenAI procured Cerebras inference compute
2026-04-17IPO filingSubmitted S-1 for listing
2026-05IPO listedTarget Nasdaq "CBRS", valuation $22-25B
2027WSE-4 expectedFirst 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

MetricWSE-3 (CS-3)WSE-4 (estimated)Rubin R200
Form FactorSingle-chip wafer-scaleSingle-chip wafer-scaleSingle-card multi-die
AI Core Count900,000~1,500,000~10,000 (SM)
SRAM44 GB~80 GB288 GB HBM4
SRAM Bandwidth21 PB/s~40 PB/s22 TB/s HBM4
BF16 Compute (sparse)125 PFLOPS~200 PFLOPS25 PFLOPS
Single-chip BF16 Compute Ratio1× (baseline)
TDP25 kW~30 kW1.8 kW
CoolingLiquid + massive systemLiquid + massive systemLiquid (card-level)
EcosystemPyTorch + JAXSameCUDA

💡 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)

ItemConfiguration
WSE Count1 (single-chip wafer-scale)
Server Dimensions1U + large cooling (similar to CS-3)
MemoryX1.5 TB extended DRAM (external)
SwarmXMulti-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

ItemDetails
VendorCerebras Systems
First Disclosure2027 (expected)
Product Pagehttps://www.cerebras.net/
FoundryTSMC
IPO StatusListed May 2026 (Nasdaq: CBRS)
CustomersOpenAI, G42, Mistral, Meta, and others