Original author: Xiao Hei, Deep Tide TechFlow
Pricing on May 13, trading opens on May 14, NASDAQ ticker CBRS.
This is the largest IPO globally so far in 2026. The underwriting syndicate includes Morgan Stanley, Citigroup, Barclays, and UBS, and they managed to pull in 20x oversubscription during the roadshow, pushing the issue price from an initial range of $115-125 all the way up to $150-160, aiming to raise $4.8 billion, which corresponds to a valuation of $48.8 billion.
Just three months ago, Cerebras' secondary valuation was at $23 billion. This means that in the final stretch before the IPO, the company's book value more than doubled.
The story's 'selling point' has been retold a thousand times: Nvidia's challenger, wafer-level chips, inference speed 21 times faster than B200, and a $1 billion to $20 billion computing power contract signed with OpenAI. This is a perfect 'AI challenger' script, with tech narratives, geopolitical narratives, star clients, and huge orders, each part fitting seamlessly into the 2026 AI infrastructure main line.
However, if you read through the S-1 document page by page, you'll notice something odd: all public reports tell one story, while the prospectus tells another.
Triple paradox.
Breaking down the prospectus reveals that Cerebras is structured by a 'triple paradox.'
First layer: technically real Alpha, but financially it’s accounting magic.
The prospectus reveals: $510 million in revenue for 2025, a 76% year-on-year growth, and a GAAP net profit of $237.8 million. It sounds very attractive for a rapidly growing and profitable AI hardware company, considered a 'mythical' asset in the current valuation environment. CoreWeave was still losing money when it IPO'd in March this year, while Cerebras directly delivered a 47% net margin.
However, this $237.8 million 'net profit' includes $363.3 million from a one-time, non-cash accounting adjustment related to the extinguishment of forward contract liabilities with G42, creating paper gains. Excluding this and adding back $49.8 million in equity incentives, the real non-GAAP net loss for 2025 is $75.7 million, worsening by 247% compared to the $21.8 million loss in 2024.
In other words, the market sees a 'profitable + 76% growth' IPO golden boy, while the prospectus discloses a 'rapidly growing company with worsening losses.' Both versions aren’t wrong; the difference lies in which one the market is willing to believe.
Second layer: it appears to have shed G42, but has actually replaced it with the nested cycle of OpenAI.
The story of Cerebras' first IPO failure in 2024 isn’t complicated: G42, the UAE-based client, contributed 85% of revenue in the first half, prompting CFIUS to launch an investigation and forcing the company to withdraw its application.
A year and a half later, the client list seems to have diversified, adding heavyweight clients like OpenAI and AWS. However, if you look at the customer structure in the S-1 from May 2026, it looks like this:
MBZUAI (Mohammed bin Zayed University of Artificial Intelligence): 62%
G42: 24%
Combined: 86%
G42 merely passed its 'weight' to MBZUAI, also based in the UAE and an affiliated entity. MBZUAI accounts for 77.9% of receivables.
The so-called 'salvation line' with OpenAI is itself a nested structure. This contract is valued at over $20 billion, with OpenAI committing to purchase 750 megawatts of computing power. But the same document reveals several things: OpenAI provided Cerebras with a $1 billion loan; OpenAI obtained nearly free warrants for 33 million shares of Cerebras; and the Master Relationship Agreement with OpenAI includes exclusivity clauses restricting Cerebras from selling to certain 'named competitors.'
In other words, OpenAI is simultaneously a customer, lender, upcoming shareholder, and a strategic influencer of Cerebras. An anonymous analyst once said something harsh in an analysis on Medium: when the revenue is cyclical, the valuation is cyclical, and the IPO is for cashing out those generating the revenue, that’s not a market, it’s financial engineering.
The wording might be too sharp, but on the facts, this statement is hard to refute.
The third layer: on the surface, a 'challenger' to Nvidia; in essence, a 'narrowband filler' for Nvidia.
This point is the easiest for the market to overlook.
Cerebras' tech is indeed solid. The WSE-3 has 40 trillion transistors, 900,000 AI cores, and 44GB on-chip SRAM, turning an entire wafer into a single chip, bypassing all the inter-chip communication bottlenecks that GPU clusters face. Independent Artificial Analysis benchmark tests show it running Llama 4 Maverick (400 billion parameters) at over 2500 tokens per second per user, while Nvidia's flagship DGX B200 outputs around 1000 tokens, and Groq and SambaNova are at 549 and 794 respectively.
Numbers don’t lie; Cerebras has a generational advantage over GPUs in this specific inference scenario.
The keyword is 'inference.' Cerebras' own prospectus makes it clear that its strength lies in latency-sensitive inference workloads, and it has no intention or capability to challenge Nvidia in large model training or 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 within Nvidia's moat.
More importantly, the market hasn’t been stagnant. Nvidia's Vera Rubin architecture released in GTC 2026 boasts 336 billion transistors, claiming performance to be five times better than Blackwell; AMD's MI400 has already caught up with 320 billion transistors; and major players like Google TPU v6, Amazon Trainium 3, and Microsoft Maia 2 are all developing self-designed chips. Nvidia is investing over $18 billion in R&D for FY 2025 and spent $20 billion last December acquiring the assets of AI inference startup Groq, plus another $4 billion on two photonics technology companies in March.
So a more accurate statement is: Cerebras is not looking to replace Nvidia; it’s trying to carve out a differentiated niche within Nvidia's 'inference' narrowband. This is a real business, but the $48.8 billion valuation corresponds to $510 million in revenue, implying a 95x price-to-sales ratio.
Andrew Feldman's third 'sell product.'
Beyond the numbers, we need to talk about the company's soul figure.
Andrew Feldman is an undervalued 'serial entrepreneur' in Silicon Valley. He’s not a tech genius founder nor someone who walked out of an ivory tower; he graduated from Stanford's business school and served as VP of marketing at Riverstone Networks (which went public in 2001) and product VP 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' by stacking a bunch of low-power processors to compete against the mainstream high-power servers of the 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 created Cerebras.
When looking at Feldman's path as a whole, an interesting observation arises: he’s not a 'chip designer'; he’s a 'contrarian bet on compute infrastructure.' SeaMicro bet on 'small cores beating big cores,' half wrong, as AMD intended to use its Freedom Fabric interconnect technology for its server CPU platform, but that path didn’t pan out, leading to SeaMicro's brand quietly fading away. Cerebras bets on 'big chips beating small chips,' which is the exact opposite of SeaMicro's proposition.
In a sense, Feldman is doing the same thing: identifying paths in computing architecture that are overlooked by the mainstream and seem 'impossible,' placing heavy bets, and then using strong sales skills to push them to market. At SeaMicro, he could leverage Force10's sales team, and AMD valued his sales network; for Cerebras, the most important thing he did right this time was securing a $20 billion contract with OpenAI, enabling a hardware company with 80% of its 2024 revenue coming from a single Middle Eastern client.
The footnote of this story is: Feldman is a product sales CEO, not a visionary tech CEO. He excels at selling a 'crazy-sounding' product to customers willing to pay a premium for differentiation, and that’s his alpha.
Understanding this is crucial, as it directly impacts the assessment of Cerebras' investment value.
So, is CBRS worth investing in?
When you stack the above three paradoxes together, the answer is much more complex than simply 'buy' or 'not buy.'
If the goal is to ride the IPO's first-day hype, with 20x oversubscription, AI hardware being the hottest sector, and lacking pure Nvidia alternatives, CBRS will likely spike on day one. This is event-driven short-term trading, requiring little deep judgment.
But if you're making a 'long-term hold' investment decision, there are three things you must clarify first:
First off, is Cerebras really worth a 95x price-to-sales ratio?
CoreWeave went public in March this year at a 15x price-to-sales ratio. Nvidia's current price-to-sales ratio is around 25x. A company with projected revenues of $510 million in 2025, an 86% customer concentration, and still operating at a loss is priced at a 95x price-to-sales ratio, implying the market expects it to generate $3-4 billion in revenue over the next three to four years and achieve sustained profitability.
Can this work? The key depends on whether OpenAI's $20 billion contract can be fulfilled on schedule. According to the prospectus, 15% of the remaining performance obligations are expected to be confirmed in 2026 and 2027, roughly $3.5 billion. If this pace holds, Cerebras could hit $2 billion+ in revenue by 2027, with its price-to-sales ratio potentially squeezed into a reasonable range. But any delay at any point, any strategic shift from OpenAI, or any new customer loss could make this valuation crumble instantly.
Second, how wide is Cerebras' moat?
The architectural advantages of WSE-3 are real, but how long will this advantage last? Nvidia's Vera Rubin, AMD MI400, and Google TPU v6 are all pushing ahead. The chip industry's generational replacement cycle is 18-24 months. If Cerebras falls behind, its technical advantage will be bridged. Its R&D spending as a proportion of revenue is already significant, but in absolute terms, it still lags behind the giants by an order of magnitude.
The deeper question is: is the wafer-level chip path going to be a widely adopted mainstream route, or just a 'special forces' niche? There’s no definite answer to this. The optimistic answer is: as inference workloads rise from 30% of total AI compute today to over 70% in the future, Cerebras' niche may become the main battlefield. The pessimistic answer is: as long as Nvidia ramps up Rubin's inference performance, the niche will forever remain a niche.
Third, governance structure and geopolitical risks.
The prospectus revealed two easily overlooked but important things:
First, Cerebras employs a dual-class share structure (Class A/Class B), with insiders holding 99.2% of the voting power post-IPO. Even if the founding team only holds 5% of the circulating shares in the future, they still control the company. This means that external minority shareholders have virtually no say in corporate governance.
Second, the company disclosed two 'significant internal control deficiencies' (material weaknesses in internal control over financial reporting). As an emerging growth company, it can waive SOX 404(b) auditor proof for five years post-IPO. This is a red flag, not a big one, but worth noting.
Geopolitically, CFIUS has cleared the voting rights issue with G42, but export controls (shipping permits for CS-2, CS-3, CS-4 to the UAE) remain a long-term variable. The Trump administration's policy direction on AI chip exports to the Middle East has yet to stabilize completely; any policy swings could reignite tail risks for CBRS.
Conclusion.
This CBRS IPO is, as an event, the most noteworthy AI hardware capital event of 2026; it defines the valuation anchor for AI infrastructure in the secondary market, and its performance will ripple through the pricing of all related assets.
As a long-term hold, it's a typical 'high-risk, high-reward' bet, wagering on the macro narrative of 'inference is king' + the micro execution of 'Cerebras leveraging OpenAI to create a narrowband monopoly' + the valuation assumption that 'the market is willing to pay a 95x price-to-sales premium for AI hardware.' All three conditions need to hold for returns to be massive; any one of them collapsing could lead to severe drawdowns.
For institutional investors, the strategy is usually to wait for the first day to settle, watch the quarterly reports, monitor key customer progress, and wait for valuation digestion. For individual investors, treating it as a small tail asset within an AI hardware allocation is fine; but if you're considering it as an all-in faith bet, please revisit the above three paradoxes.
What’s more noteworthy than whether CBRS will skyrocket tomorrow is the deeper implication: when a company that derives 86% of its revenue from two related entities in the UAE and is still losing money can be valued at $48.8 billion, it tells everyone just how crazy the capital scene in the AI infrastructure space has gotten.