Cerebras IPO: 48.8 billion-dollar valuation, whether the "Inveda Challenger" is a bubble or a new king

Original by Black, Deep tide TechFlow

13 MAY PRICING, 14 MAY OPENING DEAL, NASDAQ CODE CBRS。

THIS IS THE LARGEST IPO IN THE WORLD SINCE 2026. THE GROUP, MORGAN STANLEY, CITI CITI, BARCLAYS AND SILVER, RECEIVED 20 TIMES MORE OVER-SUBSCRIPTIONS DURING THE ROAD STAGE, BRINGING THE ORIGINAL OFFER PRICE FROM $115 TO $150 TO $160, WITH A PROJECTED RAISE OF $4.8 BILLION, WITH A CORRESPONDING ESTIMATE OF $48.8 BILLION。

Just three months ago, Cerebras' second-tier valuation was 23 billion. That is to say, the company's book value more than doubled in the last part of the road before IPO。

The story's "sale point" has been repeated 10,000 times: British Challenger, Crystal Chip, 21 times faster than B200, and OpenAI signed a $1 billion start-up, up to $20 billion in calculator contracts. This is a perfect "AI Challenger" script, technical narratives, geographies, star clients, huge orders, each of which is precisely stuck on the main line of "AI infrastructure" in 2026。

BUT READING THE S-1 FILE FROM ONE PAGE TO THE NEXT REVEALS SOMETHING STRANGE: ALL PUBLIC REPORTS ARE ABOUT THE SAME STORY, WHILE THE BOOK IS ABOUT THE OTHER。

Triple paradox

Open the book, and Cerebras presents a "triple paradox"。

First: Technically, Alpha, financially, accounting magic。

The equity book disclosed that $510 million was received in 2025, an increase of 76 per cent over the same period, and the GAAP net profit was $237.88 million. It sounds very beautiful, as a fast-growing and profitable AI hardware company that is almost mythical in the current valuation environment. CoreWeave was still losing in March of this year when IPO, and Cerebras handed over 47% net interest。

However, the $237.8 million "net profit" represented $363.3 million from a one-time, non-cash accounting adjustment resulting from paper gains arising from the cancellation of long-term contractual liabilities associated with G42. Taking this off, adding an equity incentive of $49.8 million, the real net loss in 2025 was $75.7 million, which is 247 per cent worse than the 21.8 million in 2024。

That is to say, the market is seeing an IPO gold boy with a "profitable + 76% growth" and a stock call book with a "failable and fast-growing company with a growing loss." The difference between the two versions is which one the market would like to believe。

Second, the surface was removed from G42 and actually replaced with OpenAI's loop nest。

The story of the first IPO failure in Cerebras in 2024 was not complicated: G42, a customer from an Emirates background, contributed 85 per cent of the collection in the first half of the year, CFIUS filed a case and the company was forced to withdraw its application。

A year and a half later, the list of clients appears to be diversified, adding OpenAI, AWS, these heavyweight clients. But in May 2026, S-1, 2025, the client structure is this:

• MBZUAI (MOHAMMED BEN ZAYED UNIVERSITY OF ARTIFICIAL INTELLIGENCE): 62%
• G42: 24%
• Total: 86 per cent

G42 JUST GAVE THE WEIGHT TO MBZUAI, ALSO LOCATED IN THE UAE AND ASSOCIATED WITH G42. MBZUAI SINGLE CLIENT ACCOUNTS FOR 77.9 PER CENT OF ACCOUNTS RECEIVABLE。

And OpenAI, the so-called Redemption Line, is itself an embedded structure. The contract was valued at over $20 billion, and OpenAI committed itself to purchasing 750 MW. But the same document revealed several things: OpenAI provided a $1 billion loan to Cerebras; OpenAI obtained Crebras 33 million shares of near-free equity recognition; and OpenAI's Master Relationship Agreement contained exclusionary clauses that limited Cerebras' sales to some of the named competitors。

In other words, OpenAI is also a customer, lender, upcoming shareholder and, to a certain extent, strategic controlling party in Cerebras. An anonymous analyst said a terrible thing about an analysis in Mediam:WHEN THE PROCEEDS ARE CIRCULAR, THE VALUATION IS CIRCULAR, THE IPO IS INTENDED TO BE CASHED BY THOSE WHO MAKE THEM, IT IS NOT A MARKET, IT IS FINANCIAL ENGINEERING。

The wording may be too sharp, but at the factual level it is difficult to counter it。

Third, it's the "challenger" on the surface of Inverda, and it's the "narrowband patch."。

This is most easily overlooked by the market。

Cerebras' technology is really hard. WSE-3 is 4 trillion transistors, 900,000 AI cores, 44GB film SRAM, making the whole circle into a chip, bypassing the crosschip communication bottlenecks that all GPU clusters have to face. Independent Artificial Analysis benchmark tests show that running Llama 4 Maverick (400 billion parameters), CS-3 output 2,500+ tokens per second, British flagship DGX B200 approximately 1,000 tokens, Groq and Sambanova are 549 and 794 respectively。

The numbers don't lieCerebras has an intergenerational advantage in deduced this particular scene。

The key word is "adjection." Cerebras has made it clear in his own book that it is best at latacy-sensive inference workloads, and that it does not challenge the capacity or intent of Britain for large model training and generic computing. CUDA ecology has accumulated for almost 20 years from 2007 to date, with model training tool chains, developers ' communities, third-party banks, all of which are still in the moat of England。

More importantly, the market is not standing still. The Vera Rubin structure, published at GTC 2026, is 336 billion transistors, with performance numbers five times greater than Blackwell; AMD MI 400 has tracked down 32 billion transistors; Google TPU v6, Amazon Trainium 3, Microsoft Maia 2, and super-scale manufacturers are doing self-study chips. In Weeda 2025 fiscal year R& D invested more than $18 billion, and last December spent $20 billion on the acquisition of AI ' s initial Groq assets, and in March invested another $4 billion in two optical technology companies。

So more precisely, Cerebras is not replacing Weeda, it's stealing a different position in the narrow band of Inveida. This is a real business, but the $48.8 billion valuation corresponds to $510 million in sales, meaning 95 times the market rate。

Andrew Feldman's third sale

In addition to the numbers, the soul figure of the company needs to be mentioned。

Andrew Feldman, an undervalued serial entrepreneur in Silicon Valley. He wasn't the founder of the technocratic style, and he didn't come out of the ivory towers. He graduated from Stanford Business School, became Vice-President of the Market of Riverstone Networks (IPO 2001) and Vice-President of the Force10 Networks (which sold $800 million to Dale in 2011)。

Together with Gary Lauterbach, he created SeaMicro in 2007 as an "energy efficiency server" to build a small core of low-capacity processors into clusters against the then-mainstream core high-capacity servers. The idea was very forward, but the market was too early. 2012 AMD bought SeaMicro for $334 million and Feldman left after two years of AMD VP。

And then he did Cerebras。

Put Feldman together to see one interesting thing: he's not a chip designer, he's an alternative better. SeaMicro was betting on the small core to defeat the big core, and the AMD was buying it to make its own server CPU platform using its Freedom Fabric connectivity technology, but it didn't work, and the SeaMicro brand disappeared silently. Cerebras is betting on Big Chip versus Little Chip, exactly the opposite of SeaMicro。

In a sense, Feldman did the same thing, finding the mainstream-neglected, seemingly impossible paths, placing heavy bets, and pushing them to the market with great sales power. SeaMicro was able to squeeze Force10's sales team. And in his hand, AMD is looking at his sales network; Cerebras, the most important thing he's doing right this time is to get G42 done, get a product in 2024 and 80% of the hardware company from a single Middle Eastern customer, and eventually sign a $20 billion contract。

The note to the story is: Feldman is a product sales type CEO, not a technology foresight type CEO. He's good at selling a "sounding crazy" product to a client willing to pay a premium for the difference, this is his Alpha。

It is important to understand this as it directly determines the value of Cerebras investments。

SO, IS CBRS WORTH THE VOTE

Put the triple paradox together, the answer is much more complicated than buying or not buying。

If the target is to eat the IPO's first-day fire, 20 times oversubscribe, AI hardware, the hottest track, and lack a pure Britain alternative, the CBRS is probably high on the first day. This is an event-driven short-term transaction that does not require much in-depth judgement。

But if we're going to make investment judgements about long-term holding, there are three things we have to think about:

First, Cerebras is worthless 95 times the market rate

CoreWeave, March of this year IPO, market sales are about 15 times higher. Nvidia's current marketing rate is about 25 times higher. A $510 million collection in 2025, client concentration 86 per cent, and real-employer losses are valued at 95 times the market rate, which is equivalent to the market ' s demand that it earn between $3 billion and $4 billion over the next three to four years and achieve sustained profitability。

Can this happen? The key is to see if OpenAI's 20 billion contracts are on schedule, as disclosed in the equity book, and in 2026 and 2027, about 15% of the remaining performations, or about $3.5 billion. If we follow this pace, Cerebras will be able to reach 2 billion plus in 2027, and the market rate will be expected to fall into a reasonable range. But any delay at any point in time, any one of the OpenAI strategic adjustments, any one of the new clients lost, will make this valuation instantaneous。

Second, how wide is the Corebras moat

The WSE-3 structural advantage is real, but how long will it last? In Weida, Vera Rubin, AMD MI 400, Google TPU v6 are pushing. The generational replacement cycle for the chip industry is 18-24 months. Once Cerebras takes a slow shot, the technological advantage will be recovered. Its R & D expenditure as a percentage of revenue is already low, but the absolute amount is still a quantitative gap relative to a few giants。

The deeper question is: is this line of crystal-circle chips a mainstream path that will be widely adopted, or is it a "special forces" that will always live in a niche scene? The answer to this question is not certain. The optimistic answer is that when the burden of reasoning is increased from 30 percent today to 70 percent in the future, the niche of Cerebras becomes the main battlefield. The pessimistic answer is that if Ingweida can do Rubin's reasoning, niche will always be niche。

Third, governance structures and geo-risks

The equity book reveals two things that are easy to ignore but important:

First, Cerebras uses the Class A/Class B two-tiered equity structure, whereby the IPO has 99.2 per cent voting power. Even though the founding team will hold only 5% of the circulation stock, they control the company. This means that outside small shareholders have little say in corporate governance。

Second, the company disclosed two "significant internal control deficiencies" (metanational controls in international control over financial reporting). As a growth company, it can exempt SOX 404 (b) auditors from certification for five years after IPO. It's a red light, not a big red light, but worth remembering。

Geographically, the CFIUS has cleared G42 ' s voting rights this time, but export controls (CS-2, CS-3, CS-4 export licences to the United Arab Emirates) remain a long-term variable. Trump ' s policy direction towards exports of Middle East AI chips has not been fully stable so far, and any policy swing will rekindle the tail risk of CBRS。

Conclusions

CBRS THIS IPOAS AN EVENT, WAS THE MOST INTERESTING AI HARDWARE CAPITAL EVENT IN 2026It defines the valuation anchor in the secondary market for AI infrastructure, and its performance is channelled to the pricing of all relevant targets。

It's a classic "high odds, high uncertainty" bet for a long timeAnd the bet is on the macrophrase "The Inference of King" + "Cerebras can borrow Openai to run a narrowband monopoly" + "The market is willing to continue paying 95 times the market premium for AI hardware." The three conditions are set at the same time and the returns will be enormous; any collapse will be tragic。

For institutional investors, the idea of siloing is usually not pursued on the first day, waiting for the tri-quarterly reports, such as the progress of key customers, and the digestion of valuations. For individual investors, it is a small tail asset in the AI hardware configuration, yes; it is a faith ticket for all-in, read the triple paradox above。

IT'S ANOTHER ASPECT OF THE STORY THAT IS MORE OF A CONCERN THAN WHETHER THE CBRS IS GOING UP TOMORROW:When a company with 86 per cent of its profits from two associated entities in the United Arab Emirates and its real operations are still in loss, it can be priced at $48.8 billion by the market. This in itself is telling everyone about the extent of capital madness on the track, which has reached a position。