Cerebras IPO raises $5.5B; signs 750MW OpenAI compute deal

Full analysis

Cerebras, the maker of chips built on a whole silicon wafer instead of cut into smaller pieces, raised $5.5 billion in a May 2026 IPO and signed OpenAI to a deal for 750 megawatts of its compute running from 2026 through 2028. For a business reader who buys AI by the token, the question is simple: does this crack NVIDIA's grip on the price you pay to run models, or is it one big customer propping up one big bet?

This is a reversible thing to watch and an irreversible thing to bet on. You can route a workload to a new inference provider next quarter and route it back the quarter after. But if you rearchitect your pipeline around Cerebras's tooling, that is expensive to undo. Nothing here sets a deadline for you. The dates that matter belong to Cerebras: 600 megawatts live or contracted by end of 2027, first European data center by end of 2027, manufacturing up more than tenfold during 2026.

The Skeptic. A $5.5 billion IPO where your biggest customer is also your anchor tenant should make anyone uneasy. The OpenAI deal is the bull case and the single largest risk in the same sentence. If OpenAI leans harder into its own silicon or Google's TPUs, Cerebras loses its marquee name and the story deflates. Wafer-scale has been "about to scale" since 2019. A tenfold manufacturing jump from a company that has never published its yield rates is the kind of number that impresses until you ask what the starting point was. And the European data center is a slide until it is drawing power.

The Compute Pragmatist. 750 megawatts is a real number, roughly the draw of a very large data center, not a demo. The claim to stress-test is the tenfold manufacturing scale. Wafer-scale yield does not improve in a straight line, and the bottleneck is not Cerebras's ambition, it is TSMC allocation and keeping defect rates down at volume. If CS-4 is on an advanced node, Cerebras is fighting for the same wafer starts as everyone else in a foundry environment that is still tight. The 600-megawatt figure by end of 2027 also assumes power and cooling that normally take 18 to 24 months to permit and build. That timeline is aggressive.

The Researcher. 750 megawatts of committed capacity moves wafer-scale out of the curiosity bin. The design does something real: it collapses the memory-bandwidth bottleneck that slows conventional chip-stitched designs on long-context work, where the model has to hold a lot of text in play at once. The open question is yield science. Tolerating defects across an entire wafer is an unsolved materials problem, and Cerebras's redundancy tricks have not been checked by outsiders at this production volume. Hold at tenfold scale and this is a genuine break from the NVIDIA pattern. Miss, and the IPO was the exit.

The Enterprise Buyer. If you buy inference, this is a second source, and a second source is worth something when NVIDIA capacity is scarce and priced like it. But you do not sign a multiyear commitment to a vendor whose survival rides on one customer's roadmap. You would want pricing on a per-token basis you can compare directly to an H100 cluster, a clear exit if Cerebras slips its buildout, and honesty about whether your workload even fits their access pattern. For most buyers the right move is a pilot on a narrow, long-context job, not a migration.

The deepest disagreement between the Researcher and the Skeptic rests on one unpublished fact: the yield rate at volume. The Researcher sees an architecture that genuinely beats stitched-together chips on a specific job. The Skeptic sees a company that has never shown the number that would prove it can build at scale. The Compute Pragmatist sides with the Skeptic on physics, foundry allocation and power permitting both argue the 2027 timeline slips. The Enterprise Buyer does not care who wins the argument as long as there is a credible second supplier to pit against NVIDIA on price.

What this hinges on: can Cerebras actually deliver the contracted megawatts on schedule, and does the OpenAI relationship stay a customer rather than quietly becoming a dependency that OpenAI engineers around. The council leans skeptical on the timeline and neutral on the architecture. Before anyone reorganizes a pipeline around this, run a real workload, long-context inference, side by side against your current cluster, measure tokens per dollar and tail latency, and get a contract that lets you walk if the buildout stalls.

Prediction: Cerebras will have less than 450 megawatts of data center capacity actually live and serving traffic by December 31, 2027, falling short of the 600-megawatt "live or under contract" figure it cited at TechCrunch Disrupt in September 2026.

Confidence: Medium. Power permitting and wafer yield both argue the buildout slips.

Why: The 600-megawatt figure blends capacity that is already running with capacity that is only contracted, which lets a shortfall hide behind a signed order rather than a working facility. Large data center power and cooling typically take 18 to 24 months to permit and energize, and a first European site promised by end of 2027 is starting that clock late. On top of that, wafer-scale yield does not scale in a straight line, and the tenfold manufacturing claim rests on defect rates Cerebras has never published. For both the construction and the chips to land on the stated schedule, two hard timelines have to go right at once, which is the less likely outcome. The opposite case, that everything energizes on time, requires power procurement and yield to both outperform their usual behavior.

Revisit by 2027-12-31: We're right if Cerebras's own reporting or credible third-party tracking shows under 450 megawatts live and serving traffic by year-end 2027. We're wrong if 450 megawatts or more is actually operational and running workloads by then.