Written by: Xiao Hei, Deep Tide TechFlow

Pricing on May 13, trading opens on May 14, NASDAQ code CBRS.

This is the largest IPO globally so far in 2026. The underwriting syndicate includes Morgan Stanley, Citigroup, Barclays, and UBS, which locked in 20 times oversubscription during the roadshow phase, raising the issue price from the initial $115-125 all the way to $150-160, with expected fundraising of $4.8 billion, corresponding to a valuation of $48.8 billion.

Just three months ago, Cerebras's secondary valuation was still at $23 billion. In other words, in the final stretch before the IPO, the company's book value more than doubled.

The story's "selling points" have been recounted a thousand times: NVIDIA's challenger, wafer-scale chips, inference speed 21 times faster than the B200, and a $1 billion starting contract with OpenAI, with a potential maximum of $20 billion in computing power. This is a perfect "AI challenger" script, with technical narratives, geopolitical narratives, star clients, and massive orders, each part precisely fitting into the main line of AI infrastructure in 2026.

However, reading the S-1 filing page by page reveals something strange: all public reports tell the same story, but the prospectus tells another.

Triple Paradox

Breaking down the prospectus point by point, Cerebras presents a target composed of a "triple paradox."

First Layer: Technically true alpha, financially accounting magic.

The prospectus discloses: $510 million revenue in 2025, a 76% year-on-year increase, with a GAAP net profit of $237.8 million. This sounds very impressive; a rapidly growing and already profitable AI hardware company in the current valuation environment is almost a "mythical" target. CoreWeave was still making losses during its IPO in March this year, while Cerebras delivered a net profit margin of 47%.

However, of the $237.8 million "net profit," $363.3 million comes from a one-time, non-cash accounting adjustment related to G42's forward contract liability extinguishment, producing paper gains. Excluding this item and adding back $49.8 million in equity incentives, the real non-GAAP net loss for 2025 is $75.7 million, which worsened by 247% compared to the $21.8 million loss in 2024.

In other words, the market sees a "profitable + 76% growth" IPO golden child, while the prospectus reveals a "rapidly growing company with expanding losses." Both versions are not incorrect; the difference lies in which one the market is willing to believe.

Second Layer: On the surface, escaped G42, but essentially replaced with OpenAI's circular nesting.

The story of Cerebras's first IPO failure in 2024 is not complicated: G42, a UAE-based client, contributed 85% of revenue in the first half, leading CFIUS to initiate a review, forcing the company to withdraw its application.

A year and a half later, they are back, and the customer list seems diversified, adding heavyweight clients like OpenAI and AWS. But looking at the May 2026 S-1, the customer structure for 2025 is as follows:

• MBZUAI (Mohammed bin Zayed University of Artificial Intelligence): 62%
• G42: 24%
• Combined total: 86%

G42 just shifted its "weight" to the same UAE-based MBZUAI, which is affiliated with G42. The single customer MBZUAI accounts for 77.9% of accounts receivable.

And the so-called "redemption line" of OpenAI itself is a nested structure. The contract is worth over $20 billion, with OpenAI committing to purchase 750 megawatts of computing power. However, the same document also discloses several things: OpenAI provided Cerebras with a $1 billion loan; OpenAI obtained nearly free warrants for 33 million shares of Cerebras; and OpenAI's Master Relationship Agreement includes exclusivity clauses that restrict Cerebras from selling to certain "named competitors."

In other words, OpenAI is simultaneously a customer, lender, soon-to-be shareholder, and to some extent, a strategic controller of Cerebras. An anonymous analyst commented on an analysis on Medium, saying a harsh line: When revenues are circular, valuations are circular, and IPOs are meant for the people generating these revenues to cash out, then this is not the market; it is financial engineering.

The wording may be too sharp, but at the factual level, this statement is hard to refute.

Third Layer: On the surface, an "NVIDIA challenger," but essentially an "NVIDIA bandwidth filler."

This aspect is the easiest for the market to overlook.

Cerebras's technology is indeed solid. The WSE-3 has 40 trillion transistors, 900,000 AI cores, 44GB on-chip SRAM, making the entire wafer into a single chip, bypassing the inter-chip communication bottlenecks faced by all GPU clusters. Independent Artificial Analysis benchmark tests show that running Llama 4 Maverick (400 billion parameters), CS-3 outputs 2500+ tokens per second per user, compared to NVIDIA's flagship DGX B200 at about 1000 tokens, and Groq and SambaNova at 549 and 794 respectively.

Numbers don't lie; Cerebras has a generational advantage over GPU in this specific inference scenario.

The keyword is "inference." Cerebras itself makes it very clear in its prospectus that its forte is latency-sensitive inference workloads and it has no intention or capability to challenge NVIDIA in large model training and general computing. The CUDA ecosystem has accumulated nearly 20 years since 2007, with toolchains for model training, developer communities, and third-party libraries, all still residing within NVIDIA’s moat.

More crucially, the market is not standing still. NVIDIA's Vera Rubin architecture released at GTC 2026 has 336 billion transistors, claiming performance five times better than Blackwell; AMD MI400 has reached 320 billion transistors; Google TPU v6, Amazon Trainium 3, Microsoft Maia 2, major vendors are all developing their own chips. NVIDIA's R&D investments exceeded $18 billion in fiscal 2025, and last December it spent $20 billion acquiring assets of AI inference startup Groq, followed by another $4 billion investment in two photonics technology companies in March.

So a more accurate statement is: Cerebras is not trying to replace NVIDIA; it is competing for a piece of differentiated territory in NVIDIA's "inference" bandwidth. This is a real business, but the valuation of $48.8 billion corresponds to $510 million in revenue, implying a price-to-sales ratio of 95 times.

Andrew Feldman's Third "Product Sell"

Beyond the numbers, it's important to discuss the soul of this company.

Andrew Feldman is an undervalued "serial entrepreneur" in Silicon Valley. He is not a technical genius founder, nor is he someone who has come out of an ivory tower. He graduated from Stanford Business School, served as Vice President of Marketing at Riverstone Networks (which IPO'd in 2001), and was Vice President of Product at Force10 Networks (which was sold to Dell for $800 million in 2011).

In 2007, he co-founded SeaMicro with Gary Lauterbach, creating "energy-efficient servers," piling a bunch of small-core low-power processors into clusters to compete against the mainstream large-core high-power servers of that time. This idea was very avant-garde, but the market was too early. In 2012, AMD bought SeaMicro for $334 million, and Feldman left after two years as VP at AMD.

Then he founded Cerebras.

Looking at Feldman's trajectory altogether reveals something interesting; he is not a "chip designer," he is an "alternative bettor on compute infrastructure." SeaMicro bet that "small cores would defeat large cores"; it was only half wrong—AMD bought it to use its Freedom Fabric interconnect technology for its server CPU platform, but that path didn't work out, and the SeaMicro brand quietly vanished later. Cerebras bets on "large chips defeating small chips," completely opposite to SeaMicro's proposition.

In a sense, Feldman is doing the same thing: identifying those paths in computing architecture that the mainstream neglects and seem "impossible," placing heavy bets, and then using strong sales capabilities to push it to market. At SeaMicro, he could manage the sales team from Force10, and AMD saw value in his sales network; with Cerebras, the most important thing he did right this time is securing G42, allowing a hardware company where 80% of 2024's revenue came from a single Middle Eastern client to ultimately sign a $20 billion contract with OpenAI.

The footnote of this story is: Feldman is a product sales-oriented CEO, not a technology vision-oriented CEO. He excels at selling a product that "sounds crazy" to customers willing to pay a premium for differentiation; that is his alpha.

Understanding this is crucial as it directly determines the judgment of Cerebras's investment value.

So, is CBRS worth investing in?

Looking at the above three-layer paradox together, the answer is actually much more complex than "buy" or "not buy."

If the goal is to capitalize on the explosive trading on the first day of the IPO, with 20 times oversubscription, AI hardware being the hottest track, and lacking pure NVIDIA alternative public targets, CBRS will likely surge on the first day. This is event-driven short-term trading, requiring little depth of judgment.

But if one intends to make a "long-term holding" investment judgment, three key points must be considered:

First, is Cerebras worth a 95 times price-to-sales ratio?

CoreWeave's IPO this March was priced at about a 15 times price-to-sales ratio. NVIDIA's current price-to-sales ratio is around 25 times. A company with $510 million in revenue for 2025, a customer concentration of 86%, and still operating at a real loss is priced at a 95 times price-to-sales ratio, equivalent to the market expecting it to achieve $3 billion to $4 billion in revenue in the next three to four years, with sustained profitability.

Can this happen? The key depends on whether OpenAI's $20 billion contract can be realized as scheduled. According to the prospectus, in 2026 and 2027, about 15% of the remaining performance obligations will be confirmed, which amounts to about $3.5 billion. Following this pace, Cerebras could reach over $2 billion in revenue in 2027, potentially bringing the price-to-sales ratio into a reasonable range. However, any delays at any point in time, OpenAI strategic adjustments, or the loss of any new customers could instantly undermine this valuation.

Second, how wide is Cerebras’s moat?

The architectural advantages of the WSE-3 are real, but how long will these advantages last? NVIDIA's Vera Rubin, AMD's MI400, and Google’s TPU v6 are all pushing forward. The generational replacement cycle in the chip industry is 18-24 months. If Cerebras lags behind, its technological advantage will be caught up. Its R&D expenditure as a proportion of revenue is already not low, but the absolute amount still represents a significant gap compared to several giants.

A deeper question is: will the wafer-scale chip route be widely adopted as a mainstream path, or will it remain a "special forces" only viable in niche scenarios? This question has no definite answer. The optimistic response is: as the share of inference workloads in AI computing rises from today’s 30% to over 70% in the future, Cerebras's niche will become a main battlefield. The pessimistic response is: as long as NVIDIA improves the inference performance of Rubin, the niche will always remain just a niche.

Third, governance structure and geopolitical risks

The prospectus discloses two easily overlooked but important facts:

First, Cerebras adopts a dual-class share structure (Class A/Class B), where insiders will hold 99.2% of the voting rights after the IPO. Even if the founding team only retains 5% of the circulating shares, they still control the company. This means external minority shareholders have virtually no say in corporate governance.

Second, the company disclosed the existence of two "material weaknesses in internal control over financial reporting." As an emerging growth company, it can waive SOX 404(b) auditor attestations within five years post-IPO. This is a red flag, not a big red flag, but worth noting.

Geopolitically, CFIUS has cleared the voting rights issue with G42 this time, but export controls (CS-2, CS-3, CS-4 shipment permits to the UAE) remain a long-term variable. The Trump administration's policy direction on Middle Eastern AI chip exports has not stabilized completely, and any policy fluctuations could reignite the tail risks for CBRS.

Conclusion

CBRS's IPO is, as an event, the most noteworthy AI hardware capital event of 2026, defining the valuation anchor for AI infrastructure in the secondary market, and its performance will transmit to the pricing of all related targets.

As a long-term holding, it represents a typical "high odds, high uncertainty" bet, betting on the macro narrative of "inference as king" + the micro execution of "Cerebras can leverage OpenAI to create a narrow bandwidth monopoly" + the valuation assumption that "the market is willing to continue paying a 95 times price-to-sales premium for AI hardware." All three conditions need to be met simultaneously for returns to be very significant; any single collapse will lead to a severe drawdown.

For institutional investors, the usual position-building strategy is to avoid chasing on the first day, wait for the third quarter report, key customer developments, and valuation digestion. For individual investors, treating it as a small tail asset in AI hardware allocation is acceptable; treating it as an all-in belief ticket, please read the above triple paradox again.

What is more noteworthy than whether CBRS will surge on its opening day is another layer of meaning of this event: when a company that derives 86% of its revenue from two related entities in the UAE, with real operations still in the red, can be valued by the market at $48.8 billion, this itself tells everyone how crazily capitalized the AI infrastructure track has become and what position it has reached.