Something shifted in the semiconductor world during the third week of April 2026. A company that had once been dismissed as a fringe science experiment filed its paperwork with the Securities and Exchange Commission, set its ticker symbol to "CBRS," and quietly announced that the era of specialized AI hardware had grown up enough to face the public markets. Cerebras Systems, the Sunnyvale startup that spent nine years building processors roughly the size of dinner plates, is preparing for a mid May Nasdaq debut. And this time the filing carries weight that its 2024 attempt simply did not.
The previous IPO bid collapsed under regulatory scrutiny tied to an Abu Dhabi investor. The withdrawal stung. Yet in the intervening months, Cerebras did something remarkable. It rewrote its customer list, landed the kind of contract that makes financial analysts reach for their reading glasses twice, and returned with a story that no longer reads like a plucky underdog narrative. It reads like a credible second front in the AI hardware war.
Why the 2026 Filing Looks Nothing Like the 2024 Attempt
Back in 2024, Cerebras had a problem that every risk officer spotted immediately. A single customer, Microsoft backed G42, accounted for 87 percent of revenue during the first half of that year. Concentration that severe makes institutional investors uneasy. When the Committee on Foreign Investment in the United States launched a review of G42's minority stake over concerns about potential technology transfer to China, the filing stalled and ultimately withdrew in October 2025. G42 eventually restructured its holding into non voting shares, clearing the regulatory cloud.
The new S 1, submitted on April 17, 2026, tells a different story. Revenue climbed 76 percent year over year, reaching $510 million in 2025 compared to $290 million the year before. Hardware sales jumped from $212 million to $358 million. Cloud services added another $152 million. G42 now accounts for 24 percent of revenue rather than a dominant majority. Mohamed bin Zayed University of Artificial Intelligence actually represents a larger share at 62 percent of 2025 revenue, which admittedly introduces its own concentration issue, though the customer base is rapidly diversifying with American hyperscale players entering the picture.
Cerebras posted GAAP net income of $237.8 million for 2025, a dramatic swing from the $481 million loss logged in 2024. The caveat is that much of that profit stemmed from a favorable accounting treatment of forward contracts rather than core operating strength. On a non GAAP basis, the company still ran a $75.7 million loss. Research and development consumed 48 percent of revenue, which is either alarming or admirable depending on whether a reader views it as reckless spending or appropriate reinvestment in a capital intensive industry.
The OpenAI Deal That Changed the Entire Valuation Conversation
Everything about the filing orbits around a single commercial gravity well. In January 2026, Cerebras and OpenAI signed a multiyear partnership worth more than $20 billion. The structure calls for Cerebras to supply 750 megawatts of compute capacity, delivered in 250 megawatt tranches during 2026, 2027, and 2028. OpenAI retains the right to purchase an additional 1.25 gigawatts of capacity through 2030.
That headline figure dwarfs the company's current annual revenue by a factor of roughly forty. The remaining performance obligations sitting on the balance sheet now total $24.6 billion, a backlog figure that would be almost inconceivable for a private chip company just eighteen months ago. Cerebras expects to recognize 15 percent of that obligation during 2026 and 2027 combined.
The relationship goes deeper than a simple supply contract. OpenAI extended a $1 billion loan to Cerebras at 6 percent annual interest, which the chipmaker can repay either in cash or by delivering products and services. In December 2025, Cerebras issued OpenAI warrants to purchase up to 33.4 million shares of non voting Class N stock. If Cerebras maintains a $40 billion average valuation for one month after listing, an additional tranche of shares vests for OpenAI. Roughly a 10 percent stake is on the table, structured to reward both parties for commercial success.
Andrew Feldman, the Cerebras chief executive, did not mince words in recent interviews. He told reporters plainly that his company took the fast inference business at OpenAI away from Nvidia, and that Nvidia obviously did not want to lose it. Coming from anyone else, the statement would sound like bravado. Coming from a company whose customer list now also includes Amazon Web Services, Platforms, GSK, the Mayo Clinic, the U.S. Department of Energy, and the Department of Defense, it sounds like a statement of fact.
The Wafer Scale Engine That Makes All of This Technically Possible
None of these commercial wins would mean anything without the underlying hardware actually working. Cerebras bet its entire existence on a counterintuitive idea nearly a decade ago. Rather than slice a silicon wafer into hundreds of small chips and then stitch them back together with cables and circuit boards, the company decided to leave the wafer intact and use the whole thing as a single processor.
The third generation Wafer Scale Engine, or WSE-3, is the result of that gamble. Fabricated on TSMC's 5 nanometer process, the chip measures 46,225 square millimeters, roughly 57 times larger than Nvidia's H100. It packs 4 trillion transistors, 900,000 AI optimized cores, and 44 gigabytes of on chip static RAM delivering 21 petabytes per second of memory bandwidth. That last number is 7,000 times higher than the H100's off chip HBM bandwidth. The chip delivers 125 petaflops of peak AI compute in a single package.
Why does wafer scale integration matter so much for AI? The answer lies in what modern language models actually spend their time doing. Token generation bottlenecks on memory bandwidth rather than raw compute. Each token requires loading model weights from memory into compute cores. On GPU based systems, most memory sits outside the chip in high bandwidth memory modules, and the processor waits idle while parameters travel across the package. On a wafer scale design, the memory lives on the same silicon as the compute, inches away rather than across a chip boundary.
The performance gap shows up clearly in benchmarks. Independent testing has shown the CS-3 system running Llama 4 Maverick at more than 2,500 tokens per second per user, compared to roughly 1,000 tokens per second on Nvidia's flagship DGX B200. Inference workloads on open source models run 10 to 70 times faster than comparable GPU deployments, depending on configuration.
The Structural Risks That Still Shadow the Bull Case
No IPO prospectus reveals only good news. The Cerebras filing includes some items that prospective investors should weigh carefully:
- Customer concentration remains meaningful, with MBZUAI representing 62 percent of 2025 revenue and G42 another 24 percent, leaving limited diversification despite the OpenAI and AWS wins
- The company disclosed material weaknesses in internal financial controls, a flag that typically requires multiple quarters of remediation to fully address
- Operating losses persist despite the headline net income figure, with a $145.9 million loss from operations in 2025 masked by favorable accounting adjustments
- OpenAI retains termination rights if Cerebras fails to deliver compute capacity on schedule or if service quality falls below contractual thresholds
- Nvidia still controls roughly 80 percent of the AI accelerator market, and its ecosystem advantages in software, tooling, and developer familiarity are substantial
- The $20 billion OpenAI contract creates enormous dependence, with the filing itself describing the relationship as "a substantial portion of our projected revenues over the next several years"
There are also second order risks worth considering. AI demand cycles can reverse quickly. Inference workload patterns could shift toward architectures better suited to GPUs. Competitors including Advanced Micro Devices, Groq, and various custom silicon programs at hyperscale cloud providers are all pushing their own specialized chips. Cerebras has a genuine technical lead on certain inference workloads, but holding that lead against deep pocketed rivals requires continuous execution.
What the IPO Signals About the Broader Specialized Hardware Market
Step back from the specifics of Cerebras itself and the filing carries a larger message. Until recently, the conventional wisdom held that Nvidia would dominate AI compute for the foreseeable future because its CUDA software ecosystem, supply chain relationships, and customer familiarity formed a moat that no startup could cross. That narrative is quietly bending.
Several signals now point toward a more plural future. Amazon Web Services, historically a careful partner to Nvidia, has committed to deploying Cerebras hardware in its own data centers. Oracle has mentioned Cerebras alongside other suppliers on recent earnings calls. OpenAI, the company whose ChatGPT practically single handedly drove the AI hardware boom, is backing a direct Nvidia competitor with billions of dollars and equity.
The underlying reason is almost mechanical. AI workloads have bifurcated. Training frontier models remains overwhelmingly GPU dominated territory. But inference, the act of actually running a trained model to answer user queries, has different bottlenecks and different economics. Speed, latency, and cost per token matter enormously at inference scale. Wafer scale architectures and other specialized designs can attack those metrics in ways that general purpose GPUs find harder to match.
If Cerebras prices successfully at its reported $22 to $25 billion target valuation, the IPO will set an important precedent. Retail and institutional investors, who have been starved for quality AI infrastructure plays outside Nvidia, will suddenly have a direct option. Other wafer scale and specialized silicon companies will watch closely, measuring their own funding paths against the reception. And the old assumption that one company would own the AI hardware future will look, in retrospect, like the kind of consensus that usually breaks right before it collapses.
Whether Cerebras ultimately thrives or stumbles in public markets, its arrival marks something genuinely new. The market has matured enough, and the demand has grown deep enough, that specialized AI hardware is no longer a speculative bet. It is a category. And categories, once established, tend to attract the kind of capital and talent that turn ambitious startups into lasting institutions.
Our Telegram Channel
Our YouTube Channel