Cerebras (CBRS): Inside the AI Chip Maker Taking On Nvidia

Cerebras Systems went from a company that had to withdraw its own IPO filing over a national-security review to the largest tech IPO of 2026, and it now trades around a $43 billion market cap. The chip is genuinely fast. Whether the business underneath it can stand on its own is a different question.

Most coverage of Cerebras is either a benchmark slide claiming it beats Nvidia, or a stock-price headline about its wild swings since going public. What gets skipped is how the business actually works: what the wafer-scale chip does differently, who really pays for it, and why a 103% revenue growth quarter still wasn't enough to get the stock back near its debut highs. That's what this piece is for.

Company Snapshot

Market cap per stockanalysis.com, based on a $181.18 close on September 14, 2026 and 237.56 million shares outstanding.

What Cerebras Actually Does

Cerebras builds the Wafer-Scale Engine (WSE), a single chip cut from an entire silicon wafer rather than the small dies Nvidia, AMD, and every other GPU maker stitch together into clusters. The current version, the WSE-3 powering the CS-3 system, covers roughly 57 times the surface area of a flagship GPU, packing far more compute cores and on-chip memory onto one piece of silicon. The pitch is that eliminating the networking overhead between thousands of small chips makes both training and inference dramatically faster for large models — Cerebras has publicly benchmarked Llama 3.1 70B inference at over 1,800 tokens per second, and in September 2026 OpenAI's GPT-5.6 Sol model launched running on Cerebras hardware at 750 tokens per second in production, the fastest disclosed inference speed for a frontier model. I covered that launch in detail in GPT-5.6 on Cerebras: 750 Tokens/Second.

Cerebras sells that hardware three ways: outright system sales to sovereign and enterprise customers, bundled supercomputing packages, and usage-based inference and training capacity rented through its own cloud — the same model competitors like Groq and SambaNova run. Government and sovereign-AI customers (the U.S. Department of Energy, Mayo Clinic, and Abu Dhabi's G42 ecosystem) have historically made up an outsized share of revenue, a concentration profile explored below.

How Cerebras Makes Money

Historically, roughly three-quarters of Cerebras' revenue came from direct hardware sales, with the rest split between cloud inference and services. That mix is inverting fast: in Q2 2026, cloud and services revenue jumped 287% year over year to $127.7 million while hardware revenue actually grew a slower 17% to $82.1 million, according to the Q2 2026 earnings release. Customers increasingly rent CS-3 capacity through Cerebras' own inference cloud rather than buying systems outright — a shift toward recurring, higher-margin revenue that mirrors what happened to cloud computing generally a decade ago.

The unit economics are improving alongside that mix shift: core gross margin hit 40.6% in Q2 2026, up roughly 940 basis points year over year, even though the company is still posting a core operating loss. For the full revenue breakdown by segment, see how Cerebras actually makes money.

Funding History

Cerebras raised roughly $2.8 billion in private venture funding across eleven rounds between 2016 and early 2026, then added $5.55 billion in its May 2026 IPO — putting total capital raised at more than $8.4 billion. The valuation trajectory compressed sharply in its final two years as a private company.

Sources: DataCenterDynamics (Series H), Cerebras press releases (Series F and G), and the Motley Fool on the May 2026 IPO.

Product Portfolio

Cerebras ships three product layers, all built around the same wafer-scale silicon.

The Diversification Bet: Partnering With AMD and AWS Instead of Just Competing

The most interesting strategic move Cerebras made in 2026 wasn't a chip launch — it was agreeing to plug its hardware into two companies it nominally competes with. On July 23, 2026, AMD and Cerebras announced a joint inference architecture that combines AMD's Helios rackscale GPU systems for the compute-heavy “prefill” stage of inference with Cerebras' wafer-scale chip for the latency-sensitive “decode” stage, deploying first through Cerebras Cloud in the second half of 2026. Separately, AWS built a similar disaggregated setup pairing its own Trainium chips with Cerebras CS-3 systems, offered through Amazon Bedrock.

This matters more than another benchmark win because it's a hedge against the most common criticism of the Cerebras story: that its growth is really just one customer's (OpenAI's) spending decisions dressed up as a platform business. Getting AMD and AWS to co-design production inference architecture around Cerebras silicon is a validation signal that's harder to dismiss than a speed claim on a slide, and it opens two large distribution channels — Amazon Bedrock's enterprise customer base and AMD's hyperscaler relationships — that don't run through the UAE or OpenAI relationships driving most of today's revenue.

Revenue and Key Metrics

Revenue growth has accelerated, not decelerated, since going public: $78.7 million in 2024, $509.99 million in 2025 (+76%), and a guided $880-890 million for 2026 — a trajectory that only makes sense if the $25.4 billion backlog keeps converting on schedule. The company is still not profitable; Q2 2026 core operating margin was negative 16%, though that's roughly a 2,600 basis-point improvement from a year earlier.

The customer concentration behind that growth is the number that should worry anyone underwriting the stock. G42 and Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) — disclosed as related parties in Cerebras' own S-1 — together accounted for roughly 86% of 2025 revenue. I go deeper on that risk, and the math behind the backlog, in our full revenue breakdown.

Competitive Landscape

Cerebras' most direct competitor is Groq, another chip startup built to challenge Nvidia on inference speed. The two have landed in very different places: Cerebras took the public-markets route and stayed independent, while Groq lost its founder and much of its technical team to a non-exclusive Nvidia licensing deal worth up to $20 billion, which cut its valuation roughly in half.

Groq figures reflect its August 2026 $350M round; not a perfectly like-for-like comparison since Cerebras is public and Groq is private.

Against Nvidia, the comparison is closer to a rounding error — Cerebras' entire market cap is a small fraction of Nvidia's. Cerebras' wafer-scale architecture wins on raw inference throughput for specific large-model workloads, but Nvidia's CUDA software moat, hyperscaler relationships built over more than a decade, and manufacturing scale remain advantages no single challenger has matched.

SambaNova, valued at $11 billion after a July 2026 Series F, rounds out the field of independent challengers still standing after a brutal two years for alternative AI chip architectures.

Leadership Team

Bull Case / Bear Case

IPO Outlook: What's Next After Going Public

Cerebras already cleared the hardest part — it priced its IPO at $185 a share on May 14, 2026, the largest tech listing of the year, after a CFIUS national-security review over G42's ownership stake had forced it to withdraw its original 2024 filing and spend more than a year restructuring the relationship. The open question now isn't whether it goes public, but whether the stock stabilizes.

Cerebras' lockup isn't a single 180-day cliff; it's staged across several tranches tied to earnings dates, with an aggregate of roughly 171.1 million shares eligible for release over the lockup period. The latest tranche, 14.6 million shares, unlocked on September 16, 2026, under the schedule set out in the company's prospectus, per reporting on the release schedule. The final tranche stays locked until the earlier of Cerebras' Q3 2026 earnings report or the 180-day outside date, which sets up its next earnings print as the key near-term catalyst for both fresh financial disclosure and a fresh round of potential insider selling.

For now, Wall Street sentiment is holding up better than the share price: Morgan Stanley reiterated an overweight rating and raised its price target on September 4, 2026, citing AI inference demand and the new AMD and AWS partnerships, and the average 12-month analyst price target sits well above where the stock trades today. No secondary offering or additional debt raise has been announced as of this writing.

The Bottom Line

Cerebras is a genuinely differentiated chip company that proved it could go public despite one of the messiest pre-IPO ownership stories of the year. The revenue growth is real, the gross-margin trend is improving, and the AMD and AWS partnerships are a credible sign the business is trying to diversify beyond its two core concentration risks, G42/MBZUAI and OpenAI. But a $43 billion market cap on a company that still runs an operating loss, still gets roughly 86% of last year's revenue from two related UAE entities, and still leans on a single OpenAI contract for most of its backlog is a bet on execution over the next several quarters, not a business that has already proven it can stand on a broad customer base.

All financial figures are based on Cerebras' SEC filings, earnings releases, and credible media reports as of September 2026. Headcount figures are third-party estimates; Cerebras does not publicly disclose exact employee counts.