Cerebras IPO: The Wafer-Scale Chip Challenging Nvidia’s AI Reign

The Chip That Defies Semiconductor Convention

Every chip maker in the world follows the same process. A 300mm silicon wafer enters a fab, hundreds of identical dies are etched onto it, and the wafer is cut apart. Good dies ship; defective ones are discarded.

Cerebras Systems rejected that logic entirely. Instead of cutting a wafer into hundreds of small chips, Cerebras uses the entire wafer as a single processor. The Wafer-Scale Engine 3 (WSE-3), built on TSMC’s 5nm process, packs 4 trillion transistors and 900,000 AI-optimized cores onto 46,225 square millimeters of silicon — roughly the size of a dinner plate. It carries 44 gigabytes of on-chip SRAM and delivers 125 petaFLOPS of compute.

For context, Nvidia’s H100 GPU contains 80 billion transistors across 814 square millimeters. The WSE-3 is 56 times larger by area and holds 50 times more transistors. This is not an incremental improvement — it is a fundamentally different architecture, and investors have decided it is worth $23 billion.

From CFIUS Turbulence to $23 Billion

Cerebras first filed its S-1 for a Nasdaq IPO in September 2024. A Committee on Foreign Investment in the United States (CFIUS) review then examined the company’s relationship with Group 42 (G42), a UAE-based technology conglomerate that was a major customer and investor. CFIUS granted clearance in March 2025, but by October 2025, Cerebras withdrew the IPO because its financial filings had become stale — they no longer reflected the company’s current valuation or cash position.

That same month, Cerebras closed a $1.1 billion Series G at an $8.1 billion valuation, led by Fidelity Management & Research and Atreides Management. Then, in February 2026, the company raised another $1 billion in a Series H round at $23 billion, led by Tiger Global with participation from Benchmark, Fidelity, AMD, Coatue, and others. G42 is no longer listed among Cerebras’ investors in the new filing.

Now Cerebras is targeting a Q2 2026 IPO re-filing on the Nasdaq — entering public markets at a moment when AI hardware companies command extraordinary premiums. CoreWeave, the GPU cloud provider that went public in March 2025, has surged 123% since its IPO, with its market cap reaching approximately $42 billion. Cerebras offers something CoreWeave does not: proprietary chip technology rather than rented Nvidia GPUs.

Why Wafer-Scale Wins at Inference

The AI compute market is undergoing a structural shift. Training a frontier model is a one-time capital expenditure. Inference — running the trained model for every user query, every agentic workflow, every API call — is an ongoing operational cost that scales with adoption. By 2026, inference accounts for roughly 67% of total AI compute spending, up from about 50% in 2025, and is projected to reach 80% or higher by 2028.

Nvidia’s GPU architecture was designed for graphics rendering and adapted for AI training. For inference, particularly sequential token generation in large language models, GPUs face three structural limitations. First, LLM inference is memory-bandwidth-bound: each token generation reads the model’s parameters from memory, and GPUs stall waiting for data. Second, GPUs achieve high utilization only at large batch sizes, but real-time low-latency applications require small batches. Third, models too large for a single GPU must be split across multiple chips, introducing communication overhead.

The WSE-3 addresses all three. Its 44 gigabytes of on-chip SRAM can hold entire models without external memory access. Its 900,000 cores maintain utilization at any batch size. And the single-chip design eliminates multi-chip overhead entirely. Cerebras claims the CS-3 system delivers 21x faster inference than Nvidia’s DGX B200 Blackwell for Llama 3 70B workloads. Independent benchmarks from Artificial Analysis measured 2,522 tokens per second for Llama 4 Maverick on Cerebras versus 1,038 tokens per second on Blackwell — a 2.4x advantage on that specific test. Performance varies by workload, but the directional advantage is consistent.

The CS-3 system consumes 23 kilowatts and requires water-cooled cold plates with micro-fin channels — no standard rack configuration works. This is both a barrier to adoption and a competitive moat: the integrated cooling-compute design is extremely difficult to replicate.

The $10 Billion OpenAI Partnership

Cerebras’ most powerful commercial validation is its reported $10 billion multi-year deal with OpenAI. Under the agreement, OpenAI rents Cerebras compute capacity — 750 megawatts through 2028 — rather than purchasing hardware. Deployment began in early 2026 for latency-sensitive workloads including agentic AI.

For OpenAI, the logic is supply chain diversification. Its inference infrastructure relies almost entirely on Nvidia GPUs, creating single-vendor dependency. Adding Cerebras as a second platform reduces this risk and creates pricing leverage.

For Cerebras, the deal provides revenue visibility that transforms the IPO narrative from “promising technology” to “contracted revenue from the world’s most demanding AI customer.” A $10 billion committed pipeline makes the $23 billion valuation significantly easier for public market investors to underwrite.

A Crowded Challenger Field

Cerebras is not alone in targeting Nvidia’s dominance, but the competitive landscape is shifting fast.

Nvidia acquired Groq for approximately $20 billion in December 2025, absorbing the inference-optimized LPU chip maker into its own ecosystem. What was once an independent challenger is now part of the incumbent.

SambaNova builds reconfigurable dataflow chips and has raised approximately $1.49 billion, including a $350 million Series E in February 2026 led by Vista Equity with Intel. It focuses on enterprise AI deployments.

Tenstorrent, led by chip architect Jim Keller, raised $800 million at a $3.2 billion valuation and has pivoted to an IP licensing model — Samsung, LG, and Hyundai license its RISC-V CPU and Tensix AI cores.

Google TPUs remain the most scaled Nvidia alternative but are available only through Google Cloud, limiting their addressable market. AWS (Trainium/Inferentia), Microsoft (Maia), and Meta are each building custom ASICs for their own workloads.

Among all challengers, Cerebras holds the most radical architectural position, the most dramatic performance claims, and the highest private valuation. Its IPO will serve as a referendum on whether fundamentally different hardware can break Nvidia’s grip.

What Could Derail the Bet

Manufacturing risk. Every wafer must function as a single system — there is no sorting good dies from bad. A defect that exceeds the redundancy budget destroys an entire chip worth over $100,000. Scaling from hundreds to thousands of wafers introduces failure modes that only Cerebras has navigated, and any fab disruption impacts every chip produced.

Customer concentration. If a substantial share of revenue comes from OpenAI, Cerebras’ financial health is tied to that single relationship. Public markets penalize companies with more than 30-40% customer concentration through lower valuation multiples.

Nvidia’s response. Nvidia has a history of defending market share through targeted products, aggressive pricing, and CUDA software ecosystem enhancements. The CUDA moat — millions of developers, two decades of tooling — represents the highest switching cost in AI hardware. Unless Cerebras’ performance advantage is overwhelming and sustained, many organizations will stay with Nvidia.

Sources & Further Reading

• Cerebras Systems Raises $1 Billion Series H — Cerebras
• OpenAI Partners with Cerebras for Inference Compute — OpenAI
• Cerebras CS-3 vs. Nvidia DGX B200 Blackwell Benchmark — Cerebras
• Cerebras WSE-3 Third Generation Wafer-Scale Engine — IEEE Spectrum
• CoreWeave Stock Soars 123% Since IPO — Motley Fool
• Cerebras $10 Billion Inference Deal with OpenAI — Next Platform
• Nvidia Buying Groq for $20 Billion — CNBC
• AI Compute Shift from Training to Inference — Computerworld