Targeting Q2 2026 IPO
CEREBRAS
WSE-3 // Wafer Scale Engine
The most radical approach in semiconductor history: an entire 300mm silicon wafer fabricated as a single processor. Cerebras has bypassed the memory wall entirely, enabling AI models to run with data locality that traditional architectures can only dream of. This is not an incremental improvement—it's a fundamental reimagining of how computers should be built for artificial intelligence.
$8.1B
Valuation (2025)
$1.91B
Total Funding
4T
Transistors
900K
AI Cores
Wafer-Scale Architecture
How Cerebras eliminated the memory wall
SPARSE LINEAR ALGEBRA CORES
900,000 AI-optimized processing elements
900,000
ON-CHIP SRAM
Distributed memory across entire wafer
44 GB
SWARM INTERCONNECT
2D mesh fabric for core-to-core communication
20 PB/s
Technical Specifications
WSE-3 Detailed Breakdown
Die Size
46,225 mm²
The entire 300mm wafer utilized as a single chip—approximately 56x larger than the largest GPU. This eliminates the need to partition models across multiple chips.
Process Node
TSMC 5nm
Leading-edge fabrication enabling 4 trillion transistors. Each core is independently programmable and optimized for sparse linear algebra operations.
Memory Bandwidth
20 PB/s
On-chip memory bandwidth that dwarfs HBM-based systems. Data stays local to compute, eliminating the off-chip memory bottleneck entirely.
Inference Speed
3,000+ tok/s
World-record inference performance on models like Llama 3.2. The architecture excels at both training and inference workloads.
Cluster Scaling
2,048 CS-3
Wafer-Scale Clusters can connect up to 2,048 CS-3 systems for 256 exaflops of computing power for the most demanding AI workloads.
Power Efficiency
High
By keeping data on-chip, Cerebras dramatically reduces energy consumption compared to systems constantly shuttling data to external memory.
Performance Comparison
Cerebras CS-3 vs NVIDIA H100
Company Timeline
From stealth to IPO
2016
Founded by SeaMicro Alumni
Andrew Feldman, Gary Lauterbach, and team from SeaMicro (sold to AMD for $334M) founded Cerebras with a radical vision: make a computer from an entire wafer.
2019
WSE-1 Unveiled
First Wafer Scale Engine announced with 1.2 trillion transistors and 400,000 cores. The largest chip ever built, proving the concept was manufacturable.
2021
WSE-2 & National Lab Partnerships
Second generation with 2.6T transistors. Major partnerships with DOE national labs including Sandia, Livermore, and Los Alamos for scientific computing.
March 2024
WSE-3 Launch
Third generation on TSMC 5nm with 4 trillion transistors and 900,000 cores. CS-3 system can train Llama2-70B in a single day. Named TIME Best Invention 2024.
August 2024
Inference Cloud Launch
Launched world's fastest AI inference service, claiming 10-20x faster than NVIDIA H100 systems. Began competing directly in the inference market.
September 2025
$1.1B Series G at $8.1B Valuation
Massive funding round led by Fidelity and Atreides. Revenue jumped from $6M in Q2 2024 to $70M in Q2 2025. G42 divested stake to clear regulatory hurdles.
Q2 2026
IPO Target
After withdrawing October 2025 IPO filing, Cerebras is now targeting Q2 2026. Analysts predict listing valuation exceeding $15 billion.
Strategic Analysis
Strengths, risks, and market position
✦ Competitive Advantages
- Only company with wafer-scale technology in production—no direct competitor
- Eliminates memory wall entirely with 44GB on-chip SRAM
- Dramatically simplified distributed computing—no complex multi-GPU programming
- Strong government and scientific computing relationships (DOE, Mayo Clinic)
- Revenue growth trajectory: 534% YoY in 2024
- First-mover in offering inference as a service at unprecedented speeds
⚠ Key Risks
- Extremely complex manufacturing—yield management on full wafers is unprecedented
- Heavy customer concentration: G42 was 83% of 2023 revenue
- NVIDIA's CUDA ecosystem creates massive switching costs
- High system costs may limit addressable market
- Regulatory scrutiny of Middle East relationships delayed IPO
- Competition from other well-funded startups intensifying
◎ Market Position
- Carving niche in high-end inference where latency is critical
- Strong position in scientific computing and national security applications
- Partnership with Dell for enterprise go-to-market
- Meta partnership announced for Llama API with 18x faster inference
- Expanding data center footprint: 6+ facilities across US and Europe
- Cloud offering democratizes access without capital expenditure
⟁ Investment Thesis
- AI hardware market projected to reach $400B+ by 2030
- Inference workloads growing faster than training—Cerebras positioned for this shift
- Unique technology creates genuine differentiation vs. GPU alternatives
- IPO in 2026 offers liquidity event for early investors
- Revenue trajectory suggests path to profitability as scale increases
- Strategic value could attract acquisition interest from hyperscalers