On May 13, Cerebras Systems (NASDAQ: CBRS) became one of the most anticipated public listings of 2026, pricing at $125–$135 per share before opening at $350 and closing at $311 — a 68% first-day surge that valued the company at roughly $26.6 billion.
It was the kind of debut that makes headlines. And for anyone watching the AI chip space, the question is unavoidable: does Cerebras threaten Nvidia's (NASDAQ: NVDA) dominance?
The Cerebras Thesis: Wafer-Scale Is Real
Unlike Nvidia, which builds traditional GPUs and stitches them together in massive clusters, Cerebras takes a radically different approach. Their Wafer Scale Engine 3 (WSE-3) is a single, enormous silicon wafer — the size of a dinner plate — that eliminates the need to communicate between separate chips.
This gives Cerebras a genuine advantage in AI inference — the task of actually running a trained AI model. While Nvidia's H100s and B200s split inference work across 8 GPUs, Cerebras can hold an entire model on one massive chip, dramatically reducing latency.
The numbers back this up. Cerebras reported $510 million in revenue for 2025 with a 47% net margin — impressive profitability for a pre-IPO AI hardware company. And their deal with OpenAI — a reported $20 billion compute agreement covering 750MW of capacity — signals that the biggest names in AI see value in their architecture.
The Nvidia Counterargument: Blackwell Changes Everything
Nvidia isn't sitting still. Their Blackwell B200 architecture, now in full production, has dramatically closed the inference performance gap. Blackwell's second-generation Transformer Engine and FP4 precision deliver inference throughput that's 4x the H100 — and Nvidia's CUDA ecosystem means any AI startup can deploy Blackwell without rewriting their stack.
This is Cerebras's real vulnerability. Their hardware is impressive, but it requires customers to adopt a non-standard programming model. Nvidia's moat — CUDA — is 15 years of developer tooling, libraries, and workflows that everything in AI already runs on.
As one analyst put it: "Nvidia doesn't have to be the fastest at inference. It just has to be good enough — and it already is."
Where Cerebras Wins: The OpenAI Deal
The $20B OpenAI deal is the single most important data point for Cerebras bulls. OpenAI — the company that effectively created the modern AI market — chose Cerebras for a massive compute buildout. That 750MW deal is not trivial. It's a bet that Cerebras's wafer-scale architecture delivers something Nvidia's clusters can't: lower latency per inference at scale.
The key question: Is the OpenAI deal a proof point that Cerebras has a unique moat, or is it just a hedge while Nvidia's supply remains constrained?
What Investors Should Watch
- Revenue concentration: If OpenAI represents 40%+ of Cerebras's revenue, that's a single-customer risk that would terrify most institutional investors.
- Nvidia's inference pricing: If Nvidia aggressively prices Blackwell for inference workloads, Cerebras's value proposition narrows.
- Lockup expiration: Cerebras's lockup period will end around November 2026. Insider selling could pressure the stock regardless of fundamentals.
- The CUDA flywheel: Every AI startup already runs on Nvidia. Cerebras needs to make switching worth the engineering cost.
The Verdict
Cerebras is a real company with real technology and real revenue. The WSE-3 is genuinely impressive hardware, and the OpenAI deal gives them credibility that no other Nvidia challenger has.
But calling the Cerebras IPO a "threat" to Nvidia overstates the case. Nvidia's market cap is roughly 100x Cerebras's valuation. Blackwell is real. CUDA is a fortress. And Nvidia's supply chain — without which most of the AI industry can't function — is unmatched.
The better framing: Cerebras makes the AI chip market a two-horse race in inference, with Nvidia leading training uncontested. For investors, CBRS offers exposure to an AI pure-play with asymmetric upside. NVDA offers the certainty of a platform that the entire industry is built on.
— The Signal





Discussion
Comments powered by Giscus — sign in with GitHub to join the discussion.