On Friday, April 17, Andrew Feldman filed a document in Washington that his company had filed once before.
The first time was September 2024. That filing, for an initial public offering of the wafer-scale chipmaker Cerebras Systems, was shelved after federal regulators spent more than a year reviewing the company's ties to G42, a United Arab Emirates technology conglomerate that at one point represented 87% of Cerebras's revenue. In October 2025, Cerebras formally withdrew it.
This second S-1 is different. The customer concentration that stalled Cerebras in 2024 is still there. It is just a different customer.
The Scoop That Broke While the Filing Went Up
The same day Cerebras filed, The Information reported that OpenAI had doubled its commitment to buy Cerebras chips. The new order replaces a January agreement that the two companies had announced three months earlier. The financial relationship now runs in four tranches.
| Item | Amount |
|---|---|
| Chip purchase commitment over 3 years | More than $20 billion |
| Previous January commitment (now superseded) | $10 billion |
| OpenAI loan to Cerebras for data center construction | $1 billion |
| OpenAI warrants | 33.4 million shares (up to 10% stake) |
| Warrants vest when OpenAI buys | 2 gigawatts of capacity by 2030 |
Cerebras is targeting a valuation near $25 billion, with a price range of 22 to 25 billion dollars in the preliminary terms. Morgan Stanley is leading the deal. The expected ticker on Nasdaq is CBRS, and pricing is scheduled for mid-May.
The numbers in the filing look like a story of vindication.
| Metric | 2024 | 2025 |
|---|---|---|
| Revenue | $290.3 million | $510 million |
| Net income (loss) | ($481.6 million) | $87.9 million |
That is a 76% revenue jump and the first profit in Cerebras's eleven-year history. Nearly all of it came from two customers.
Why OpenAI Needs Wafer-Scale Chips
To understand why OpenAI committed this much money to a company that barely cleared $87 million in profit, you have to understand what Cerebras sells and why Nvidia does not sell it.
Every large language model has two jobs. The first is training, where the model learns from data. The second is inference, where it generates a response. For most of the generative AI boom, training has consumed the attention, the talent, and the compute budgets. Cerebras is a bet that inference is about to eclipse training in total spend.
The WSE-3, Cerebras's third-generation wafer-scale engine, is a single piece of silicon 58 times larger than Nvidia's B200 GPU. It contains 4 trillion transistors and 900,000 AI cores, delivering 125 petaflops of compute. Its signature advantage is not raw FLOPs. It is memory bandwidth of 27 petabytes per second, roughly 200 times what Nvidia's NVLink interconnect provides between H100 GPUs.
Memory bandwidth is the bottleneck that determines how fast a model generates the next token. On a WSE-3, OpenAI's open-weight model gpt-oss-120B runs at up to 3,000 tokens per second. On Nvidia hardware, the same model generates responses in the low hundreds. A reasoning query that stalls on an Nvidia cluster completes in about a second on Cerebras.
For teams building agentic products, where a model might chain dozens of reasoning steps before producing an answer, that latency difference is the product. Anthropic's Claude Code, Cursor, and OpenAI's own Codex all suffer when responses stall mid-thought. Cerebras sells the speed that makes real-time agents feel alive.
The 2016 Coffee That Became a $20 Billion Contract
Sam Altman was one of the first people Andrew Feldman pitched.
In 2016, when OpenAI was a year-old nonprofit with eleven employees and Cerebras existed only on slide decks, the two met in San Francisco. Altman wrote one of the first checks into Cerebras. Over the next nine years, the companies talked repeatedly about a partnership. Nothing shipped.
The first production deal finally closed in January 2026: 750 megawatts of inference capacity through 2028 for more than $10 billion. Feldman, on stage at the announcement, described the category shift in a single line:
"Just as broadband transformed the internet, real-time inference will transform AI, enabling entirely new ways to build and interact with AI models." — Andrew Feldman, CEO of Cerebras Systems, January 14, 2026
Altman framed the deal carefully. It was "incremental" to Nvidia, he said, adding "we plan to increase our Nvidia purchasing over time."
Three months later, the commitment doubled. The new S-1 discloses that OpenAI's order now covers 750 megawatts plus an additional option for 1.25 gigawatts through 2030. Warrants vest when OpenAI purchases 2 gigawatts of total capacity. If OpenAI exercises them, it will own roughly one dollar of Cerebras equity for every ten it spends on Cerebras chips.
It is a structure unusual in semiconductor procurement. It is not unusual in Sam Altman's playbook. OpenAI has cut similar deals with CoreWeave, AMD, and Nvidia itself. LDS covered the fresh CoreWeave commitment earlier this month in CoreWeave Signed Meta for $21 Billion. The Next Morning, Anthropic Called.
The Customer That Became a Dependency
Read the risk factors in the S-1 and the story gets more complicated.
The filing discloses that the OpenAI contract "represents a substantial portion of our projected revenues over the next several years." In 2024, a single customer (G42) accounted for 87 percent of Cerebras's revenue. That concentration is what regulators spent fourteen months investigating before Cerebras withdrew its first IPO.
If OpenAI delays an order, re-routes inference traffic to Nvidia or Google TPUs, or chooses a competing custom-silicon partner, Cerebras's revenue forecast collapses. The loan from OpenAI makes the dependency structural. Cerebras is building data centers with capital borrowed from its largest customer to serve that same customer's workloads.
The hedge works in the other direction too. In September 2025, OpenAI signed a non-binding letter of intent with Nvidia worth up to $100 billion. It has a separate supply agreement with AMD.
The Cerebras deal is one leg of a three-way bet on who will supply the inference compute OpenAI needs to hit $650 billion in cumulative infrastructure spending over the next five years.
Nvidia still dominates. Its data center revenue was $51.2 billion last quarter, roughly 100 times what Cerebras earned in all of 2025.
The Other Side
Not everyone who reads the S-1 sees a breakout story.
The most cited skeptical view is structural. Nvidia's data center gross margins sit above 70 percent. Cerebras does not break out a comparable number in the filing.
The gap between $510 million in revenue and 87.9 million dollars in net income suggests the margin picture is still thin. A chip company with soft margins and one large customer is not the same kind of bet as Nvidia.
The other consistent critique is software. Most production AI code is optimized for CUDA, Nvidia's programming model, which has roughly fifteen years of developer mindshare behind it. Cerebras's own developer documentation acknowledges that porting a model to the WSE-3 requires translating CUDA kernels to its Software Language for Linear Algebra (SLA), a process that remains non-trivial. For a buyer who wants to move workloads between vendors, that lock-in cuts both ways.
Cerebras has answers to both. On yield, it points to a published technical note describing how redundancy and dynamic routing across 970,000 physical cores let it activate 900,000 usable cores per wafer. On software, it argues that inference serving does not require the full CUDA stack, only the narrow subset that runs hot during decode.
Feldman addressed none of this directly in his X post the day of the S-1 filing. Instead he wrote that gpt-oss-120B "runs fastest on Cerebras," and left the reader to decide whether speed is the point.
How It Unfolded
What It Means for Practitioners
For ML engineers, the interesting part is not who wins the inference war. It is that the war is finally happening.
Through all of 2024 and most of 2025, the only realistic answer to "where do I run my model?" was Nvidia or a cloud GPU rental that sat on Nvidia. AMD Instinct GPUs existed but were rare outside of hyperscalers. Google TPUs were walled inside Google Cloud. Custom silicon from startups was a rounding error.
The Cerebras S-1 is evidence that inference-specific silicon is now a category with its own pricing, its own procurement cycles, and its own public comparables. It also names the other entrants. Groq, SambaNova, and Tenstorrent are cited directly. AWS Trainium and Google TPU v7 are referenced as indirect competition. The filing reads like a market map.
For anyone deploying production models, the implication is tactical. A reasoning model served at 3,000 tokens per second on Cerebras enables UX patterns that simply cannot run on a 200-tokens-per-second Nvidia stack. Latency budgets change. Chain-of-thought becomes interactive. Agents stop feeling like batch jobs.
The broader trend is clear from other recent LDS coverage, including Anthropic Tripled Its Revenue in 90 Days. Then It Signed for 3.5 Gigawatts. and SpaceX Filed for the Largest IPO in History. The Money Isn't for Rockets. The capital funding inference capacity is now measured in hundreds of billions, and the customers writing those checks increasingly want equity in the suppliers.
The Bottom Line
Cerebras spent eleven years building a chip that is 58 times the size of a standard GPU. The $20 billion that will keep it public belongs to the customer it spent nine years trying to close.
If OpenAI executes the full order, it will have committed more to Cerebras than Cerebras is worth as a public company on day one. If it walks away, Cerebras will be the largest IPO failure in chip history. The S-1 does not tell you which is more likely. It tells you the company is comfortable with the coin flip.
Cerebras is not going public on the strength of its chip. It is going public on the strength of a single customer's ten-year bet. That customer is Sam Altman, who signed the first check in 2016 and the last one three months ago. The rest of the market will now decide whether OpenAI's conviction is enough.
Sources
- The Manila Times (Reuters): OpenAI to spend more than $20 billion on Cerebras chips, receive stake (April 18, 2026)
- Benzinga: OpenAI's $20 Billion Deal Sets Up Second Most Likely IPO After SpaceX (April 17, 2026)
- SiliconANGLE: AI chip developer Cerebras Systems files to go public amid rapid revenue growth (April 17, 2026)
- CNBC: AI chipmaker Cerebras files to go public after scrapping IPO plans last year (April 17, 2026)
- TechCrunch: OpenAI signs deal, worth $10B, for compute from Cerebras (January 14, 2026)
- Andrew Feldman on X: In 2016, Sam and I first met...
- Techi: Cerebras IPO (CBRS): Valuation, Timeline & How to Invest 2026 (April 18, 2026)
- IEEE Spectrum: Cerebras WSE-3: Third Generation Superchip for AI
- Tech Insider: Cerebras IPO — $510M Revenue, $10B OpenAI Deal, $23B Valuation (April 17, 2026)