Cerebras targets a $26 billion IPO valuation: the first real crack in NVIDIA's monopoly

Cerebras Systems is preparing Wall Street's blockbuster IPO for 2026. The Sunnyvale-based AI chip startup confirmed on May 4 its plan to raise $3.5 billion at a targeted valuation of $26.6 billion at the high end of the range. And according to Bloomberg, investor demand is so strong that the price range is expected to be raised before the IPO date set for May 14, 2026 (ticker: CBRS).

The order book is oversubscribed 20x. That's a level of enthusiasm Wall Street hasn't seen since Snowflake's 2020 IPO. And the stakes are massive: Cerebras is the first serious challenger to NVIDIA in five years.

The plate-sized chip

Cerebras's technical argument fits in one object: the WSE-3 (Wafer Scale Engine, third generation). It is an AI chip etched onto an entire silicon wafer — roughly 462 cm². For comparison, an NVIDIA Blackwell B200 chip is about 813 mm². The WSE-3 is therefore dozens of times larger.

Why does it matter? Because in a monolithic chip, all communication between cores happens at the speed of light on silicon — without going through external interconnects (PCIe, NVLink, etc.) that bottleneck classical NVIDIA clusters. The WSE-3 packs:

• 900,000 AI cores on a single chip
• 44 gigabytes of SRAM on-die (the fastest memory in existence)
• Internal memory bandwidth of approximately 21 petabytes per second

For large LLM inference, the impact is spectacular. Cerebras posts response times 10x to 20x faster than an equivalent H100 cluster when serving Llama-3.1 405B or DeepSeek-V3. It is exactly what agentic AI workloads need to maintain sub-50ms latencies.

OpenAI, the customer who changes everything

The detail that tips the IPO from "interesting" to "major" is OpenAI. Sam Altman announced a partnership with Cerebras covering 750 megawatts of low-latency compute. Concretely, OpenAI uses Cerebras to serve part of its ChatGPT and Codex inference.

Why is OpenAI betting on Cerebras when it could simply buy more NVIDIA GPUs? Three reasons:

1. Latency: for real-time flows like voice and agents, wafer-scale shaves hundreds of milliseconds per request.
2. Diversification: OpenAI wants to reduce its NVIDIA dependence. Stargate, Microsoft, and now Cerebras form a strategic patchwork.
3. Cost: Cerebras claims a 30% to 40% lower cost per token versus an equivalent H100 cluster for 100B+ parameter model inference.

The OpenAI-Cerebras contract is the most cited element in the S-1 filings to the SEC. It also justifies the valuation multiple: with $510 million in 2025 revenue, Cerebras trades at roughly 52x revenue. Stratospheric for a hardware maker — except this isn't comparable to Intel or AMD. Cerebras is positioned closer to NVIDIA, which trades at 25x revenue but with far more mature growth.

The IPO math

The jump from $23 billion (February 2026 private round) to $26.6 billion (May IPO) represents a 15% rise in three months. And if the range is indeed raised, the $30 billion mark could be crossed on day one of trading.

Why it matters for the AI race

The Cerebras IPO is a strategic barometer for three reasons.

1. First credible NVIDIA challenger. For five years, NVIDIA has held a quasi-monopoly on AI silicon — over 90% market share. AMD with MI300X, Intel with Gaudi, Google with TPUs, and Amazon with Trainium have all attempted breakouts without moving the lines. Cerebras attacks from a different angle: not by making competing GPUs, but by offering a radically different architecture. If the IPO succeeds, the signal to investors is clear: there is room for something other than NVIDIA GPUs.

2. Inference validated as a distinct market. For a long time, the AI hardware market conflated training and inference. Cerebras positions itself explicitly on inference — the market that will explode as LLMs get embedded everywhere. According to McKinsey estimates, the inference market will be larger than training as soon as 2027.

3. Signal for hardware consolidation. If Cerebras is worth $26 billion publicly, then Groq, SambaNova, Tenstorrent, Etched, MatX, and other inference startups will see their valuations explode. It's the exact scenario that played out after Snowflake's 2020 IPO — a wave of late-stage raises and follow-on listings. Expect to see Groq in the IPO pipeline within 12 months.

The competitive landscape

Cerebras occupies a very specific niche: large models served at high speed. As long as LLMs keep getting bigger and agents need sub-second responses, that niche will keep expanding.

Key takeaways

• Cerebras Systems is preparing its Wall Street IPO for May 14, 2026 under the ticker CBRS
• Targeted raise: $3.5 billion at a $26.6 billion valuation (potentially $30B+ if range is raised)
• Order book oversubscribed 20x, a historic level
• Flagship chip: WSE-3, the largest AI chip ever made (462 cm², 900,000 cores, 44 GB SRAM)
• Strategic customer: OpenAI with a 750 MW low-latency compute contract
• 2025 revenue: $510 million, valuation multiple ~52x
• Positioning: first credible NVIDIA challenger on AI inference

Cerebras is not the next NVIDIA killer — Jensen Huang sleeps fine. But it is the first time since 2018 that an AI hardware player reaches public markets with a coherent thesis, solid revenue, and an anchor customer the size of OpenAI. While Anthropic stacks NVIDIA GPUs via SpaceX, OpenAI bets on diversification. And Wall Street rewards contrarian bets. On May 14, if the Cerebras IPO clears $30 billion in valuation at the closing bell, we will officially enter the second phase of the AI silicon war — the one where NVIDIA no longer owns the narrative.