Cerebras Priced The Biggest IPO Of 2026 Into A Market Huang Locked Up

What The Cerebras Numbers Actually Say

The S-1 update on May 11 raised the marketed range from $115-125 per share to $150-160 and expanded the offering to 30 million shares. Demand outstripped supply by more than 20 times before the roadshow finished. The final pricing came in at $185, well above even the raised band. Shares were indicated to open on the Nasdaq this morning near $336, nearly double the IPO price.

Cerebras sells the Wafer-Scale Engine 3, a single 5nm AI accelerator with 4 trillion transistors and 900,000 cores fabricated on one continuous wafer. Most chips carve a wafer into hundreds of separate dies. Cerebras leaves the wafer intact. The product is a single piece of silicon roughly the size of a dinner plate, with 44 gigabytes of on-chip SRAM running at 21 petabytes per second of memory bandwidth.

The architecture is real. For inference workloads where memory access dominates and the model fits on a single wafer, the chip is meaningfully faster than what Nvidia ships in commodity form. On Llama 4 Maverick inference, Cerebras hits roughly 2,500 tokens per second per user against about 1,000 for an Nvidia DGX B200 system. That is a real advantage on a real workload.

It is not the workload that defines the inference economy.

The Groq Acquisition Recontextualizes Everything

Five months ago, Jensen Huang took the inference question off the table. Nvidia acquired Groq for $20 billion. Groq was the company Wall Street had been pointing to since 2024 as the proof that inference might fragment away from Nvidia. The company was founded by Jonathan Ross, a former Google scientist instrumental in the development of the Tensor Processing Unit. Groq’s Language Processing Units used on-chip SRAM rather than high-bandwidth memory, achieving inference speeds three to thirteen times faster than GPUs at roughly one-third the energy cost.

Huang did not buy Groq because Nvidia needed inference help. Nvidia GPUs have been running the bulk of production inference workloads at the hyperscalers since 2023. Huang bought Groq because the inference economy is large enough that letting any specialist build escape velocity in it was strategically unacceptable.

The Cerebras S-1 still names Groq among its competitors. That document was drafted before the acquisition closed. By the time Cerebras started its roadshow, Groq was a Nvidia business unit and the inference-specialist category had a new owner.

The history of Nvidia under Huang is a history of moves like this. CUDA was Nvidia’s software lock-in built over a decade when no one was paying attention. NVLink with Fusion now integrates third-party silicon, including Intel and AMD CPUs, into Nvidia clusters. Nvidia uses its scale at Taiwan Semiconductor to consume most of the available leading-edge manufacturing capacity through 2027. Groq closed the inference gap in the ecosystem. The pattern is consistent and observable.

What Cerebras Actually Sells

Wafer-scale architecture is a category of compute, not a replacement for GPUs. The economic case for a CS-3 system is narrow and specific. The customer is a hyperscaler or research lab that wants extreme inference latency on a model small enough to fit on a single wafer, where the workload runs continuously enough to justify the higher acquisition cost. OpenAI is reportedly one such customer. So are several sovereign AI buyers in the Middle East.

Those are real customers. They are also a small percentage of total inference compute purchased in 2026, which by any reasonable estimate runs into the hundreds of billions of dollars annually.

Cerebras does not displace CUDA. Every meaningful large model in production today, including the models Cerebras runs in its own cloud demonstrations, was trained against CUDA. The cost of porting model code and training pipelines off CUDA onto a specialist accelerator is not zero, and the engineering organizations that have absorbed that cost remain rare. The Cerebras buyer is not the marginal Nvidia buyer. The Cerebras buyer is the customer who has decided their workload is unusual enough to justify a second hardware ecosystem alongside Nvidia, not instead of it.

The S-1 risk factors are honest about this. Cerebras lists customer concentration as a primary risk. A small number of large customers, including OpenAI, account for the bulk of revenue. That is not the financial profile of a company displacing the incumbent. It is the profile of a specialist serving the largest buyers in a market the incumbent does not need to defend.

The Signal Is Market Size

The structural insight from this IPO is not that Nvidia faces a new competitor. The structural insight is that the inference economy has grown large enough that, after the dominant supplier has already paid $20 billion to absorb the only credible inference-specific startup, a second specialist can come public at $56 billion.

That is not a story about Nvidia losing share. That is a story about how big the addressable market for AI inference has become. The compute build for inference has multi-year lead times, the capacity is already spoken for through 2027, and the buyers are still underweight. A $56 billion IPO for a wafer-scale specialist is more empirical evidence that the supply constraint is the binding feature of this cycle, not the demand picture.

Jensen Huang has been articulating the supply constraint thesis publicly since 2024. The evidence is in supplier behavior, not in stock prices. Nvidia has reserved most of TSMC’s leading-edge capacity through 2027. Hyperscalers are committing tens of billions of dollars annually to data center capacity that does not yet exist. The bookings tell the story regardless of what the IPO market does in any given week.

What This Tells Long-Term Investors

The reusable lesson from every chip cycle since the introduction of x86 is that ecosystem-locked compute holds the workloads its owners want to hold. ARM broke mobile only because the power envelope of mobile was structurally incompatible with x86 silicon. Nothing similar is true of AI inference today. The model code, the training pipelines, the orchestration layer, and the inference serving stacks are all written against CUDA, and the dominant supplier owns the leading inference-specific startup to emerge from the 2024 architectural cycle.

The companies positioned to capture the bulk of the inference economy share three characteristics. They own ecosystem-locked compute, they have multi-year manufacturing capacity reservations at the leading-edge foundries, and they sit somewhere on the supply chain that delivers AI compute to end customers.

The names that fit that description today: Nvidia, which now owns the CUDA software stack, the NVLink interconnect, the bulk of TSMC’s leading-edge capacity through 2027, and the inference architecture it acquired in Groq. Broadcom, which designs custom AI ASICs for the hyperscalers building their own Nvidia-alternative silicon, and which captures the networking layer that connects every AI cluster. Arm Holdings, whose CPU architecture sits inside Nvidia’s Vera Rubin platform and ships direct to hyperscalers running their own AI infrastructure. Astera Labs, which sells the connectivity components that go into essentially every GPU rack built in 2026.

Cerebras is a real company with a real architecture and a real product. The IPO confirms that the inference market is large enough to support specialists at substantial valuations. That is useful information. It is not the same information as “Nvidia is losing inference,” and the readers who confuse the two will run the same play that has been wrong on Nvidia since 2015.

The signal worth tracking is how much inference capacity gets bought, by whom, and against which software stack. That picture has not changed because Cerebras came public. It has changed because the buyer behind the dominant stack already moved to lock the category up last December.