Cerebras: The Speed Trap

Its wafer-scale chips deliver the fastest inference on earth. Its business model delivers almost none of that value to shareholders.

If you bought Cerebras Systems because you believe in the thesis — wafer-scale chips powering the fastest AI inference on earth, validated by a massive OpenAI commitment — you're right about the technology and completely wrong about who gets paid for it.

That's the paradox at the heart of this stock. Cerebras has shipped the most compelling answer to the "fast tokens" question. Its WSE-3 chip hits roughly 2,500 to 3,000 tokens per second — 5 to 20 times faster than GPU clouds. On the Q1 earnings call, CEO Andrew Feldman demoed a trillion-parameter model: Cerebras finished in 21 seconds; the leading GPU took 4 minutes and 37 seconds. The architecture is real, and it's extraordinary.

But the company that makes the world's fastest chip is booking most of its revenue at cost-plus-3% margin — and the stock has fallen 22% in three weeks.

How OpenAI is using Cerebras speed to attack Claude

The coding market is the most valuable prize in AI right now. Anthropic's Claude Code has been winning it — developers report Claude generates better code, and Anthropic says its success rate on open-ended coding problems jumped to 76%, a 50-point improvement in six months. The WSJ reported in June that Anthropic is pressuring OpenAI precisely because its strongest growth comes from developer workflows where users generate huge token volumes daily.

OpenAI's counterpunch is pure speed — and it runs on Cerebras silicon.

In February, OpenAI launched GPT-5.3-Codex-Spark, the first product of the Cerebras partnership. It runs at roughly 1,000 to 1,200 tokens per second — about 15 times faster than earlier Codex versions and 5 to 10 times faster than Claude on comparable hardware. On Cerebras's own benchmarks, the company claims 3,000 tokens per second versus 650 on leading GPU clouds for comparable models. The pitch is visceral: you type a prompt; you get software. "Developers can translate ideas into working software in seconds," the IPO prospectus reads. "Creating software at the speed of thought."

But the speed comes with a tradeoff. GPT-5.3-Codex-Spark scores 16 points lower on SWE-Bench Pro than the full GPT-5.3-Codex. It's a distilled model — roughly a 120B-parameter derivative of a much larger system. SemiAnalysis, which published the deepest technical analysis of Cerebras at 10,000 words, noted that Cerebras's chips are "only economically capable of serving relatively small models today." In real-world usage, OpenAI customers report roughly 100 tokens per second — not the 1,000 advertised.

OpenAI consumed essentially all available Cerebras capacity to launch Codex-Spark, freezing out every other potential customer. That means the only enterprise customer actually paying for Cerebras compute at scale right now is OpenAI — and that revenue comes through the cloud delivery model.

The margin trap

Here's how the deal actually works. OpenAI's Master Relationship Agreement gives two delivery options: buy and deploy Cerebras hardware in its own data centers, or receive compute through Cerebras Cloud. For the first 250MW, OpenAI chose the cloud route.

Cerebras books pass-through data center costs as revenue, adds a 3% markup, and that's the gross profit. The revenue number looks enormous; the margin is near-zero. And OpenAI is simultaneously developing its own inference chip ("Jalapeño") with Broadcom — a direct potential replacement.

The numbers that broke the stock

Cerebras reported Q1 on June 23. Revenue of $193.4 million beat estimates and rose 94% year-over-year. The stock fell 20% the next day. Here's why: Q2 revenue guidance of approximately $194 million was essentially flat sequentially. Core gross margin: 36% to 38%, down from 47%. Full-year core operating margins: negative 28% to 32%. The hypergrowth narrative broke the moment the market read past the year-over-year headline.

The lockup cascade

Pre-IPO investors paid a weighted average of $5.07 per share. On June 25, 30.2 million shares unlocked — an 87.5% increase in free float. Eclipse Capital, a Series A investor in at $0.85 per share, has already converted 13.5 million Class B shares to tradable Class A.

The full schedule runs through November: another 36.4 million shares at Q2 earnings plus two days, then staggered biweekly releases through October, with the final 21.5% of investor holdings released after Q3 earnings. Total: more than 210 million shares — roughly $38 billion in potential selling at current prices — held by investors sitting on 30x to 200x gains.

What to watch

The bull case rests on inference demand exploding and Cerebras maintaining an architectural moat. The $24.6 billion backlog is real. The AWS partnership — deploying Cerebras systems as the first third-party chip in AWS data centers — could diversify beyond OpenAI.

The bear case is that the margin structure isn't fixable in a quarter or two, customer concentration remains existential, and the unlock cascade has barely started at 200x forward earnings with negative book value.

Q2 earnings, expected in August, is the litmus test. If cloud margins stabilize above 35% and the hardware revenue decline the CFO warned about proves temporary, the bear case weakens. If not, the unlock sellers have a $5 cost basis, 200x their money, and nothing but time.

More research at bargo.ai/research.

Sources: SemiAnalysis "Cerebras — Faster Tokens Please" (May 13, 2026) · Cerebras Q1 2026 earnings call transcript · CBRS S-1/A prospectus · @phithetasigma bear case analysis · @PatrickMoorhead Q1 analysis · WSJ "OpenAI considers deep price cuts" (June 11, 2026) · Cerebras.ai "Why the AI Race Shifted to Speed" · InfoQ "Codex-Spark Achieves Ultra-Fast Coding Speeds" · Turing College "Codex 5.3 vs. Codex Spark: Speed vs. Intelligence"