February 2026 · Secondary at ~$26B

Cerebras is the only independent, production-grade AI chip company at scale. They've built genuinely differentiated silicon (Wafer-Scale Engine, a single chip the size of a dinner plate with 4 trillion transistors), and they're selling it to the most important customers in AI.

OpenAI just committed ~$10B+ through 2028, already live in production. NVIDIA absorbed Groq for $20B in December, leaving Cerebras as the only credible independent alternative.

The investment case rests on three pillars:

The core innovation is the Wafer-Scale Engine (WSE-3), built on TSMC's 5nm process. Standard manufacturing cuts a silicon wafer into hundreds of small chips. Cerebras uses the entire wafer as one processor — 46,000 mm², 4 trillion transistors, 900,000 cores, 44GB of on-chip SRAM.

In LLM inference, the binding constraint is memory bandwidth, not raw compute. By putting memory directly next to the cores, Cerebras eliminates the interconnect bottleneck that limits GPU-based inference.

The benchmarks back this up. Cerebras claims 10x+ faster training time-to-solution and 10x+ faster output generation versus GPU alternatives. Cognition's SWE-1.5 coding agent runs on Cerebras at 950 tok/s — 6x faster than Haiku 4.5, 13x faster than Sonnet 4.5. Training a 175B parameter model on 4,000 GPUs requires ~20,000 lines of distribution code; Cerebras does it in 565 lines in one day.

Inference is becoming the bottleneck, and Cerebras built hardware optimized for this specific type of workload.

In January 2026, OpenAI announced a multi-year agreement to deploy 750MW of Cerebras systems for inference through 2028, worth over $10 billion. On February 12th, OpenAI launched GPT-5.3-Codex-Spark running on Cerebras hardware — the "first milestone" in the partnership.

The deal is structured as capacity-as-a-service: Cerebras builds or leases data centers, OpenAI pays for compute. This is recurring, infrastructure-grade revenue. OpenAI evaluated Cerebras as early as 2017, then ran a gpt-oss open-weight collaboration; culminating in a term sheet a few months ago. Sam Altman is personally an investor in Cerebras.

Beyond revenue, this addresses the company's biggest investor concern. The S-1 disclosed G42 accounted for 87% of H1 2024 revenue under a $1.43B purchase commitment, and G42's ties to Chinese tech triggered a CFIUS review that delayed the IPO over a year. The OpenAI contract structurally diversifies the customer base in one stroke. Other named customers now include Cognition, Hugging Face, IBM, GlaxoSmithKline, Mayo Clinic, TotalEnergies, and the major U.S. national labs (Argonne, Lawrence Livermore, Sandia). Cerebras also recently launched an 8 exaflops supercomputer in India.

An important caveat: NVIDIA commands 30-35x forward multiples because its gross margins are ~70-75%, protected by CUDA. Cerebras's margins are improving fast (11.7% → 41.1% in two years) but the capacity-as-a-service model for OpenAI is capital-intensive — building 750MW of data centers requires enormous upfront CapEx. Public markets will likely apply a hardware/infrastructure multiple, not a software one. If gross margins trend toward 45-50% (AMD's AI chip margins are ~47%) and recurring cloud revenue grows as a share of the mix, the story improves. But the margin profile and capital intensity are definitely factors to consider.

Five companies, multiple exits, 25+ years in hardware.

On December 24, 2025, NVIDIA acqui-hired Groq for $20B. Jonathan Ross (Groq founder, creator of Google's TPU) and ~90% of Groq's engineers joined NVIDIA. Bernstein's Stacy Rasgon noted the deal structure was designed to "keep the fiction of competition alive."

The independent AI chip landscape has consolidated:

NVIDIA's data center revenue alone was ~$115B in FY2025. The AI accelerator TAM is projected to reach ~$250B by 2030. Even low-single-digit market share for Cerebras implies a business many multiples of its current size.

This shifts the business model toward something more capital-intensive than a pure chip company and will weigh on near-term margins and free cash flow — though the emergence of several data center companies will offer paths to offload this from the balance sheet (e.g., Fluidstack is handling a good chunk of Anthropic <> Google TPU buildout.)

Cerebras at $26B is a bet on

Key investors: Benchmark, Tiger Global, Fidelity, Atreides, Altimeter, AMD, Coatue, Foundation Capital, Alpha Wave, 1789 Capital.