Cerebras
They put an entire silicon wafer on a single chip and called it a processor. The market took eleven years to agree they were right.
Founded in 2015, Cerebras spent years as the AI chip industry's most audacious long bet — a company that had solved a real problem the market wasn't quite ready to pay for. Then inference became the defining compute challenge of the decade, and a decade of proprietary wafer-scale engineering suddenly looked like exactly the right preparation.
By Aaron Rose · Tech Reader Magazine · September 3, 2026
Who They Are
Cerebras Systems was founded in Sunnyvale, California in 2015 by Andrew Feldman, Sean Lie, Gary Lauterbach, Michael James, and Jean-Philippe Fricker — a team with deep roots in high-performance computing and semiconductor design. The company employs around 708 people and trades on Nasdaq under the ticker CBRS, having completed its IPO in May 2026 — the largest US tech listing in nearly five years at the time of pricing.
The company's core product is the Wafer Scale Engine, or WSE — a processor that uses an entire silicon wafer as a single chip rather than cutting the wafer into many smaller dies. The WSE-3, the current generation, packs more than 4 trillion transistors onto a piece of silicon roughly 57 times larger than the largest competing GPU die. That scale gives it vastly more on-chip memory, more processing cores, and higher interconnect bandwidth — all without the latency overhead of coordinating across a multi-chip cluster.
The flagship commercial product is the CS-3 system, a turnkey AI compute environment that packages the WSE with its own software stack, designed to run large AI workloads as a complete unit rather than requiring customers to assemble and integrate components themselves. In August 2026, Cerebras unveiled the CS-4 — doubling inference speed over its predecessor and delivering up to 30 times the throughput of comparable GPU configurations.
$386All-time high share price reached on IPO day, May 14, 2026.CBRS currently trades near $190, with a market cap of approximately $44 billion.
The Problem They Are Solving
The conventional approach to AI compute is to manufacture many small chips and connect them in clusters. That architecture works, and Nvidia has made it the dominant paradigm for both training and inference. But it comes with a structural tax: every time data has to travel between chips, it consumes time and energy. At inference scale — where a model responds to millions of queries per day — that inter-chip communication overhead compounds into real cost and real latency.
Cerebras built around the opposite assumption. Keep everything — compute cores, memory, interconnect — on one very large piece of silicon, and you eliminate the communication overhead entirely. The tradeoff is manufacturing complexity: wafer-scale production requires managing defects across an unusually large surface area, and the yield and packaging challenges took years to solve. That's the decade of work that preceded the IPO.
The pitch to customers is straightforward: fewer systems, less power draw, less cooling infrastructure, and faster per-token response times. The total cost of ownership argument is what attracted OpenAI, AWS, and a growing roster of enterprise customers to the platform — and it's the same argument that makes Cerebras directly relevant to any organization running AI inference at scale.
The company built around the opposite assumption from Nvidia.Keep everything on one very large piece of silicon,and you eliminate the communication overhead entirely.
The IPO Arc and What It Tells You
Cerebras filed its original IPO in September 2024, withdrew it in October 2025 amid a national security review of UAE-based investor G42 — which had accounted for over 80 percent of revenue in 2024 — and refiled in early 2026 with a substantially different customer story. By the time the book closed in May, it was 20 times oversubscribed. The company priced at $185 per share, above the already-raised guidance range, raising $5.55 billion and opening at a fully diluted valuation of $56.4 billion. The stock touched $386 on its first day of trading.
The customer mix had changed materially between the two filings. OpenAI signed a multi-year agreement for 750 megawatts of AI compute capacity — one of the largest AI infrastructure contracts ever publicly disclosed. AWS added a partnership. The G42 concentration risk remained in the prospectus disclosures, but it was no longer the center of gravity of the business. The 2025 revenue figure of $510 million, with a 47 percent net margin, gave public market investors a real financial story to underwrite.
Since the IPO high, CBRS has traded back to the $190 range — a significant correction from the opening-day peak, but one that reflects the volatility expected of a newly public company in a rapidly repricing sector rather than any fundamental deterioration in the business. Analyst consensus as of this writing sits at Strong Buy, with a 12-month price target averaging $291.
Plans and Road Ahead
The CS-4, unveiled at Cerebras's Supernova 2026 event in August, entered general availability in the September quarter. Early access partnerships include OpenAI, AMD, Figma, Cognition, and CrowdStrike — a customer list that spans foundation model labs, enterprise software, autonomous coding, and cybersecurity. The company's roadmap targets 20 times current throughput by 2027, which, if delivered, would extend its performance lead over GPU-based inference systems considerably.
Cerebras is also expanding its geographic footprint, with a new AI data center in Finland announced in late August. That expansion reflects the growing international demand for inference capacity that is not solely dependent on US-based hyperscaler infrastructure — a consideration that is increasingly relevant for European enterprises navigating data sovereignty requirements.
The competitive picture is crowded and getting more so. Fractile, Groq, and Etched are each pursuing inference-optimized architectures with significant venture backing and, in Groq's case, a strategic investment from Nvidia itself. But Cerebras enters that competition as the only inference-focused chip company with a public market track record, commercial revenue, and a product already shipping at scale. The decade of manufacturing work that preceded the IPO is, at this point, a moat that new entrants will take years to replicate.
The AI Compute Problem
Cerebras is the inference chip story that took the longest to tell and arrived at exactly the right moment. The market spent years waiting for inference to become the defining AI compute problem. Cerebras spent those same years solving it. The timing, for once, lined up.
The Inference Race
Cerebras is shipping. Fractile is raising. Next in this series: Unitree, the Chinese robotics company bringing sub-$20,000 humanoids to a factory floor near you — and what that actually means for the industry.