Cerebras Bets on 'Fast' Inference Computing with OpenAI
The company aims to reshape the economics of inference computing for the era of AI agents, offering low-latency solutions based on its architecture.
Cerebras and OpenAI: How 'Fast Tokens' Became the New Currency of the AI Industry
While Nvidia reports record revenues and the market digests its partnership with Corning, a far more radical shift is brewing on the periphery. Cerebras just went public, and its founder Andrew Feldman articulated a thesis that should make Jensen Huang's eye twitch: "Nvidia didn't want to lose the fast inference business at OpenAI, and we took it." Behind this bravado lies not just a $200+ billion contract, but a fundamental rethinking of what exactly customers are willing to pay for in the age of AI agents.
The Core: What's Really Happening
Cerebras and OpenAI have struck a deal whose structure resembles a vassal oath more than a commercial contract. OpenAI has committed to gradually purchase 750 MW of AI inference capacity by 2028, with the option to expand to 2 GW by 2030. The remaining obligations under the contract amount to $246 billion. At the same time, OpenAI extended Cerebras a $1 billion loan at 6% annual interest and received warrants for 33.4 million shares with an exercise price near zero. If all conditions are met, OpenAI will own approximately 10-12% of Cerebras.
Sam Altman and his team—Greg Brockman, Ilya Sutskever—have personally invested in the company's equity. In essence, OpenAI has transformed from a customer into a quasi-parent structure for the chip manufacturer. No one in the industry has done this: not Microsoft, not Google, not Amazon. They lease capacity or order chips. Altman, however, is building an alternative vertical of "consumer—model—silicon," where the equipment supplier is not just a contractor but part of the empire.
The technical basis of the bet: the WSE-3 chip can deliver up to 2000 tokens per second on a model comparable to GPT-5.3-Codex—15-21 times faster than a comparable Nvidia GPU cluster. At the same time, the cost per token is 32% lower. The key word is "comparable." The model running on Cerebras is not a full GPT-5.3. It is a distilled version with 10 times fewer parameters, trained on the original. Essentially, it's a "fast copy" of the flagship model, and customers—remarkably—are already voting with their wallets.
Timeline and Context
2024: Cerebras files its S-1 for the first time, but the SEC blocks the IPO because 86% of the company's revenue comes from two UAE entities—MBZUAI and G42. CFIUS launches a review due to G42's ties to Huawei and Chinese genomics companies.
2025: Cerebras revenue reaches $510 million with 76% year-over-year growth, but operating losses persist. The $237.8 million profit shown in GAAP reporting is an accounting illusion: a one-time non-cash gain from restructuring liabilities to G42 worth $363.3 million. Without it, the loss is $75.7 million.
January 2026: OpenAI signs a Master Relationship Agreement with Cerebras. The amount exceeds $200 billion, making it the largest contract in AI infrastructure history.
March 2026: AWS announces a partnership with Cerebras. The scheme is "disaggregated inference": AWS Trainium3 handles prefill, Cerebras CS-3 handles decode. This is an architectural revolution, detailed below.
May 14, 2026: Cerebras lists on NASDAQ with 20x oversubscription. The offering price is $185, but on the first day of trading, shares soar to $350 and close at $311. Market capitalization at its peak exceeded $100 billion, dropping to $280 two days later. The market oscillates between euphoria over the "Nvidia killer" and horror at the 130x revenue multiple.
Who Wins and Who Loses
OpenAI wins. The "customer becomes co-owner of the supplier" model is high-level financial engineering. OpenAI gets exclusive access to capacity, a stake in a growing asset, and leverage through the loan and warrants. If Cerebras takes off, OpenAI profits from the capitalization growth. If it fails, OpenAI as a creditor gains control over collateral accounts. A win-win position.
Sam Altman personally wins. His goal is to build trillions of dollars worth of AI infrastructure independent of Nvidia. Cerebras is the second brick after the AMD deal for 6 GW. Altman now has two alternative silicon suppliers, both more dependent on him than he is on them.
The concept of 'fast tokens' wins. SemiAnalysis revealed a staggering fact: 80% of their team's AI budget goes to Anthropic's fast mode Opus 4.6, not the smarter Opus 4.7. Engineers are abandoning the "better" model in favor of the "faster" one. This is a tectonic shift: speed becomes a standalone product for which a premium is paid, not a free add-on to intelligence.
Nvidia loses—but not obviously. Nvidia controls 94.4% of the data center GPU market. Losing part of OpenAI's inference business is not fatal, but it is a symptom. Nvidia sees the threat: that's why it already bought Groq for $20 billion last year—a direct response to Cerebras' architecture. Vera Rubin, Nvidia's next chip, is being designed as a heterogeneous system where different types of computation are distributed across different chips within a single rack. This is an admission: monolithic GPU architecture is no longer optimal.
Investors who bought Cerebras shares at the peak lose. With a 130x revenue multiple, GAAP profit that is 100% a one-time accounting item, and a single customer that is also a creditor, this is not an investment but speculation. The history of tech IPOs is clear: in the first years after listing, newcomers underperform the market by an average of 3.6% annually.
What the Media Isn't Saying
Non-obvious insight: Cerebras' real goal is not to beat Nvidia, but to vertically integrate with OpenAI at the 'model-silicon' level.
What is described as an "inference contract" is actually the creation of a new type of AI company. The traditional model: a chip manufacturer (Nvidia), a cloud provider (AWS), a model developer (OpenAI). Between them are market relationships. Cerebras and OpenAI are building a fundamentally different structure: a unified stack where the model, chip, and data center are designed for each other and linked by cross-ownership.
The model on Cerebras is not GPT-5.3. It is GPT-5.3-Codex-Spark, a distilled version with 10 times fewer parameters. It can only be efficiently run on WSE-3. And WSE-3, in turn, is tailored precisely for such models: 44 GB of SRAM on the chip means larger models simply won't fit. This is a closed loop—or, if you will, a closed ecosystem. OpenAI is creating silicon on which only its own (distilled) models run efficiently, and then sells access to this stack as a service.
If this model proves successful, the industry will split not into "chip manufacturers" and "model developers," but into vertically integrated "AI conglomerates," each owning the full stack. Google is already close with TPU and Gemini. OpenAI is building its vertical through AMD and Cerebras. Who's next—Anthropic? Microsoft?
Second insight: the disaggregated inference architecture launched by AWS is a death sentence for monolithic GPU clusters.
The scheme where prefill (understanding the query) runs on AWS Trainium and decode (generating the response) runs on Cerebras marks the end of an era where one type of chip does everything. Different phases of inference have fundamentally different computational natures: prefill is compute-bound, decode is memory-bound. Optimizing one piece of hardware for both tasks is like making a car that must be both a truck and a sports car: expensive and inefficient.
This means that in 3-5 years, a typical AI data center will consist of several types of specialized chips, not hundreds of thousands of identical GPUs. Nvidia understands this—Vera Rubin is already being designed as a heterogeneous system with different chips for different tasks. But Nvidia has a problem: disaggregated inference means customers can buy chips from different vendors for different phases. And that breaks the CUDA monopoly.
Forecast: 30 Days and 90 Days
30 days (until mid-June 2026):
Cerebras shares will continue their volatile movement. After the correction from $350 to $280, a bounce to $320-330 is possible on news of new contracts, but there is no fundamental support at these levels. Any negative signal—delays in delivery, changes in OpenAI's terms, production issues—will send shares below $200.
Nvidia will report earnings on May 20, and Jensen Huang will almost certainly devote significant time to the topic of inference, trying to show that Blackwell and Vera Rubin are not inferior to Cerebras. Expect announcements about new optical connections (building on the partnership with Corning) and possibly the acquisition of another fast inference startup.
OpenAI will begin negotiations for the next funding round ahead of its IPO. Having its own chip vertical through Cerebras will be a strong argument for a $1 trillion valuation.
90 days (until mid-August 2026):
Cerebras will announce its first major customer outside of OpenAI-AWS-UAE. This is critical: as long as 86% of the company's revenue depends on three counterparties, institutional investors will not seriously enter the stock. If no new customer emerges, Cerebras shares will face a second wave of decline to $200-220.
OpenAI may exercise some warrants or convert the loan into additional Cerebras shares, strengthening control. This would signal to the market that Altman does not consider the current price overvalued and believes in long-term growth.
Nvidia will respond asymmetrically. Instead of a price war, Huang will bet on the software ecosystem: an announcement of CUDA for heterogeneous computing, allowing developers to use specialized chips without abandoning the familiar stack. This will neutralize Cerebras' main advantage—speed—by lowering the barrier to migration to new architectures.
The key signal to watch: if by the end of August any of the hyperscalers (Microsoft, Google) announces support for disaggregated inference involving Cerebras, it will be a point of no return. The industry will acknowledge that the era of monolithic GPU clusters is over, and a race for specialized chips will begin. If AWS remains the only major proponent of this architecture, Nvidia still has 3-5 years of dominance, and Cerebras' bet will remain a niche experiment.
The bet is placed. The next three months will show whether Feldman is right in claiming that the "fast search market" is as large as internet search—or whether fast tokens will remain a premium niche for coding while the rest of the AI world continues to buy GPUs. Either way, one thing is already clear: "fast enough" AI is no longer sufficient.
— Editorial Team
No comments yet.