Cerebras Systems: The Record on the Company That Bet Against Nvidia

Picture the engineer who, in 2015, looked at the AI chip business and decided the whole industry was building it wrong. Everyone else was wiring thousands of small chips together. Andrew Feldman's team wanted to carve a single computer out of an entire silicon wafer — one chip the size of a dinner plate. It was the kind of idea that gets you laughed out of a fab. Eleven years later, that bet went public and, for one day in May 2026, the market acted like it was the smartest idea in computing. This is the record on the company behind the chip — the founders, the money, the one customer that nearly sank it, and what actually happened at the IPO.

Records over spin: every number below is attributed. Where a claim is a company benchmark or a forward-looking bet rather than an audited fact, it is marked.

The founders and the founding bet

Cerebras Systems was founded in 2015. Its CEO and public face is Andrew Feldman, who previously co-founded the server startup SeaMicro and sold it to AMD in 2012. The co-founding team is widely reported to include Gary Lauterbach, Michael James, Sean Lie, and Jean-Philippe Fricker — chip architects who had worked together at SeaMicro.

The founding bet was contrarian: instead of stitching many small GPUs together, build the largest single chip physically possible. The result is the Wafer-Scale Engine (WSE). According to Cerebras, the current WSE-3 packs 900,000 cores, 44GB of on-chip memory, and 4 trillion transistors on a 5nm TSMC process — a chip of roughly 46,225 mm², versus about 814 mm² for Nvidia's H100 (Cerebras specs, via Introl and HeyGoTrade reporting). The company quotes 125 petaflops of peak compute for the WSE-3, which ships inside its CS-3 system.

The bet against Nvidia

Cerebras's pitch is speed, especially for inference — the moment a model actually answers you. The company claims it ran Llama 4 Maverick (a ~400-billion-parameter model) at 2,500 tokens per second per user, which it says is more than double Nvidia's DGX B200 Blackwell system on the same model (Cerebras benchmark). It has also claimed up to 21x faster inference than a B200 on certain Llama 3 70B reasoning workloads at lower total cost (Cerebras benchmark).

Read those as vendor benchmarks, not independent results — they're the seller's stopwatch on the seller's track. The structural claim underneath is more durable and worth weighing on its own: because the whole model can live on one giant chip's memory, Cerebras avoids the chip-to-chip data shuffling that slows GPU clusters. That is the company's real argument against Nvidia. Whether it scales economically is exactly what the IPO put to the market.

The money

Cerebras raised steadily as a private company. The relevant recent rounds, per the company and reporting:

• Series G — $1.1 billion at an $8.1 billion valuation, October 2025 (Cerebras press release).
• Series H — about $1 billion, valuation ~$23 billion, February 2026 (reporting via NextPlatform and TECHi).

On revenue, the company reported $510 million in 2025, up 76% year-over-year (Cerebras S-1, via TechCrunch and Futurum). That's real growth — but the next number is the one investors circled.

The G42 problem — and where the concentration went

Cerebras has a customer-concentration problem so large it delayed the company's first attempt to go public. In 2024, roughly 85% of revenue came from G42, the Abu Dhabi-based AI firm (reporting via NextPlatform). G42 is also an investor, and its stake triggered a long review by the Committee on Foreign Investment in the United States (CFIUS) that stalled the 2024 IPO filing.

Here's the honest twist the spin tends to bury: the concentration didn't go away — it moved. For 2025, G42 accounted for about 24% of revenue, but Mohamed bin Zayed University of Artificial Intelligence accounted for about 62% — meaning those two Abu Dhabi-linked customers together were roughly 86% of the $510 million (reporting via fwdstart/mostlymetrics on the S-1). The risk shifted from one company to one region. That is a material fact, not a footnote.

The IPO: what actually happened

Cerebras went public on the Nasdaq under the ticker CBRS on May 14, 2026 (TechCrunch, CNBC). The record:

• Priced at $185 per share — well above the initial $115–$125 range, which had been raised to $150–$160 (TechCrunch).
• Sold about 30 million shares, raising ~$5.5 billion (TechCrunch).
• Opened around $385, up roughly 108%, then closed near $311, implying about a $66 billion valuation on day one (TechCrunch).
• The stock slid the following day (CNBC, Yahoo Finance).

TechCrunch and The Motley Fool called it the biggest U.S. tech IPO since Snowflake in 2020 — the first blockbuster listing of 2026's IPO season.

What's verified vs. what's still a bet

Verified: the founders, the funding rounds, the $510M 2025 revenue, the customer concentration, and the IPO mechanics — all attributed above to the S-1 and major outlets. Still a bet: the performance claims (vendor benchmarks), and the larger wager that wafer-scale silicon can pull durable, diversified business away from Nvidia. A 108% first-day pop is the market's hope, not its proof. The record to watch next is the 2026 revenue mix — specifically whether Cerebras can sell to customers who aren't in Abu Dhabi.

Image: Cerebras Systems logo, via Wikimedia Commons (file: Cerebras_logo.svg), released under CC0 (public domain).

Sources: TechCrunch · CNBC · Yahoo Finance · The Motley Fool · Cerebras S-1 (SEC) · Cerebras Series G press release · The Next Platform · fwdstart · Futurum Group · Introl