Cerebras IPO Priced at $48.8 Billion Valuation, Sparks Debate on AI Hardware Bubble

Article by Xiao He, Shenchao TechFlow

Priced on May 13, trading opens on May 14, NASDAQ ticker CBRS.

This is the largest IPO globally to date in 2026. The underwriting syndicate consists of Morgan Stanley, Citigroup, Barclays, and UBS. During the roadshow, the offering received 20 times oversubscription, pushing the offering price up from the initial range of $115–$125 to $150–$160, with expected proceeds of $4.8 billion and a valuation of $48.8 billion.

Just three months ago, Cerebras's secondary valuation was $23 billion. That means, in the final stretch before its IPO, the company's book value more than doubled.

The story’s “hook” has been repeated ten thousand times: NVIDIA’s challenger, wafer-level chip, 21 times faster inference than the B200, a $1 billion starting contract with OpenAI, and up to $20 billion in computing power deals. It’s a perfect “AI challenger” script—technical narrative, geopolitical narrative, star client, massive orders—each element perfectly aligned with the 2026 AI infrastructure主线.

But as you read through the S-1 filing page by page, something strange emerges: all public reports tell the same story, while the prospectus tells a different one.

Breaking down the prospectus item by item, Cerebras presents a company defined by a "triple paradox."

First layer: Technically genuine alpha, financially accounting magic.

The prospectus discloses: 2025 revenue of $510 million, a 76% year-over-year increase, and a GAAP net profit of $237.8 million. This sounds exceptional—an AI hardware company experiencing rapid growth and already profitable, making it nearly a "mythical" prospect in today’s valuation environment. CoreWeave was still losing money at its IPO in March this year, while Cerebras delivered a net margin of 47%.

However, of this $237.8 million in "net profit," $363.3 million came from a one-time, non-cash accounting adjustment—the paper gain from the extinguishment of the forward contract liability related to G42. Excluding this item and adding back $49.8 million in stock-based compensation, the company’s actual 2025 non-GAAP net loss was $75.7 million, a 247% worsening compared to the $21.8 million loss in 2024.

In other words, the market sees a profitable IPO star with 76% growth, while the prospectus reveals a rapidly growing company with expanding losses. Neither version is wrong—the difference lies in which version the market chooses to believe.

Second layer:表面上摆脱了 G42, but actually replaced it with a nested loop of OpenAI.

The story of Cerebras's failed IPO in 2024 was straightforward: G42, a customer based in the UAE, accounted for 85% of the company's revenue in the first half of the year; CFIUS launched a review, forcing the company to withdraw its application.

After a year and a half, the customer list appears to have diversified, now including heavyweight clients like OpenAI and AWS. But flipping to the S-1 filed in May 2026, the customer structure for 2025 was as follows:

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

G42 merely transferred the "weight" to MBZUAI, which is also based in the UAE and affiliated with G42. A single customer of MBZUAI accounts for 77.9% of its accounts receivable.

And OpenAI’s so-called “lifeline” is itself a nested structure. This contract, worth over $20 billion, commits OpenAI to purchasing 750 megawatts of computing power. But the same document also reveals several other details: OpenAI provided Cerebras with a $1 billion loan; OpenAI received nearly free warrants for 33 million shares of Cerebras; and OpenAI’s Master Relationship Agreement includes exclusivity clauses restricting Cerebras from selling to certain “named competitors.”

In other words, OpenAI is simultaneously a customer, lender, upcoming shareholder, and to some extent, a strategic controller of Cerebras. An anonymous analyst made a sharp comment on a Medium analysis: “When revenue is circular, valuation is circular, and the IPO exists merely to allow those creating the revenue to cash out, this isn’t a market—it’s financial engineering.”

The wording may be somewhat harsh, but factually, this statement is difficult to dispute.

Third layer: On the surface, it's a "challenger" to NVIDIA; in essence, it's a "narrowband complement" to NVIDIA.

This point is most easily overlooked by the market.

Cerebras’s technology is truly impressive. The WSE-3 features 4 trillion transistors, 900,000 AI cores, and 44 GB of on-chip SRAM, integrating an entire wafer into a single chip and bypassing the inter-chip communication bottlenecks that all GPU clusters must confront. Independent benchmarking by Artificial Analysis shows that the CS-3 delivers over 2,500 tokens per second per user running Llama 4 Maverick (400 billion parameters), while NVIDIA’s flagship DGX B200 achieves approximately 1,000 tokens per second, and Groq and SambaNova achieve 549 and 794 tokens per second, respectively.

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

The keyword is "inference." Cerebras clearly states in its own prospectus that it excels at latency-sensitive inference workloads and has neither the capability nor the intention to compete with NVIDIA in large model training or general-purpose computing. The CUDA ecosystem has accumulated nearly two decades of development since 2007, and the toolchains for model training, developer communities, and third-party libraries all remain firmly within NVIDIA’s moat.

More critically, the market is not standing still. NVIDIA’s Vera Rubin architecture, unveiled at GTC 2026, features 336 billion transistors and is claimed to deliver five times the performance of Blackwell; AMD’s MI400 has already reached 320 billion transistors; Google’s TPU v6, Amazon’s Trainium 3, and Microsoft’s Maia 2 demonstrate that hyperscalers are all developing their own custom chips. NVIDIA invested over $18 billion in R&D during fiscal year 2025, acquired the assets of AI inference startup Groq for $20 billion last December, and further invested $4 billion in two photonics technology companies in March.

So a more accurate statement is: Cerebras isn’t trying to replace NVIDIA; instead, it’s carving out a differentiated position within NVIDIA’s narrow “inference” segment. This is a legitimate business, but a $48.8 billion valuation against $510 million in revenue translates to a price-to-sales ratio of 95x.

Beyond the numbers, let’s talk about the visionary behind this company.

Andrew Feldman is an underappreciated serial entrepreneur from Silicon Valley. He is not a technical genius founder nor someone who emerged from an ivory tower; he graduated from Stanford Business School and served as Vice President of Marketing at Riverstone Networks (which went public in 2001) and as Vice President of Product at Force10 Networks (acquired by Dell for $800 million in 2011).

In 2007, he co-founded SeaMicro with Gary Lauterbach to build "energy-efficient servers," clustering many low-power, small-core processors to challenge the dominant high-power, large-core servers of the time. The concept was highly forward-thinking, but the market was too early. In 2012, AMD acquired SeaMicro for $334 million. After serving as a VP at AMD for two years, Feldman left the company.

Then he created Cerebras.

Looking at Feldman’s path as a whole, an interesting pattern emerges: he is not a "chip designer," but rather an unconventional bettor on compute infrastructure. SeaMicro bet that small cores would beat large cores—partially wrong. AMD acquired it hoping to leverage its Freedom Fabric interconnect technology for its own server CPU platform, but that path ultimately failed, and the SeaMicro brand quietly disappeared. Cerebras, by contrast, bet that large chips would beat small chips—exactly the opposite of SeaMicro’s thesis.

In a sense, Feldman does the same thing: he identifies overlooked, seemingly “impossible” paths within computing architectures, makes bold bets, and leverages his exceptional sales acumen to bring them to market. With SeaMicro, he was able to control Force10’s sales team—the very sales network that AMD valued. This time with Cerebras, his most crucial move was securing G42, enabling a product to generate 80% of its 2024 revenue from a single Middle Eastern client, ultimately landing a $20 billion contract with OpenAI.

The footnote to this story is: Feldman is a product-sales CEO, not a technology-visionary CEO. His alpha lies in selling products that sound "crazy" to customers willing to pay a premium for differentiation.

It's important to understand this, as it directly impacts the assessment of Cerebras's investment value.

When viewed together, the three paradoxes above reveal an answer that is far more complex than simply "buy" or "not buy."

If the goal is to capitalize on the first-day surge of an IPO, with 20x oversubscription, in the hottest sector—AI hardware—and with a lack of pure-play NVIDIA alternatives among listed assets, CBRS is highly likely to spike on day one. This is an event-driven short-term trade that doesn’t require deep analysis.

But if you're making an investment decision for "long-term holding," there are three things you must first consider:

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

CoreWeave went public in March this year at a price-to-sales ratio of approximately 15x. Nvidia’s current price-to-sales ratio is around 25x. A company with $510 million in revenue for 2025, 86% customer concentration, and still operating at a loss on a true business level is being priced at a 95x price-to-sales ratio—implying the market expects it to achieve $3 to $4 billion in revenue and generate sustained profitability within the next three to four years.

Will this work out? It all hinges on whether OpenAI’s $2 billion contract can be delivered on schedule. According to the prospectus, approximately 15% of the remaining performance obligations—roughly $3.5 billion—are expected to be recognized in 2026 and 2027. If things proceed at this pace, Cerebras’s revenue in 2027 could reach $2 billion or more, bringing its price-to-sales ratio into a reasonable range. However, any delay at any point, any strategic shift by OpenAI, or any loss of a new customer could instantly undermine this valuation.

Second, how wide is Cerebras's moat?

The architectural advantages of WSE-3 are real, but how long will they last? NVIDIA’s Vera Rubin, AMD’s MI400, and Google’s TPU v6 are all advancing. The semiconductor industry’s generational cycle is 18–24 months. If Cerebras falls even slightly behind, its technological edge will be erased. While its R&D spending as a percentage of revenue is already high, the absolute dollar amount still pales in comparison to that of the major players.

A deeper question is: Will the wafer-scale chip approach become a widely adopted mainstream path, or will it remain forever a “special forces” unit confined to niche scenarios? There is no definitive answer. An optimistic view is that as the share of inference workloads in total AI computing rises from today’s 30% to over 70% in the future, Cerebras’ niche will become the main battlefield. A pessimistic view is that as long as NVIDIA improves Rubin’s inference performance, the niche will remain just a niche.

Third, governance structure and geopolitical risks

The prospectus discloses two easily overlooked but important matters:

First, Cerebras employs a two-class share structure of Class A and Class B shares; after the IPO, insiders hold 99.2% of the voting rights. Even if the founding team retains only 5% of the outstanding shares in the future, they will still control the company. This means external minority shareholders have virtually no say in corporate governance.

Second, the company disclosed two "material weaknesses in internal control over financial reporting." As an emerging growth company, it is exempt from the SOX 404(b) auditor attestation requirement for up to five years after its IPO. This is a warning sign—not a major one—but worth noting.

On the geopolitical front, CFIUS has resolved G42’s voting rights issue, but export controls (licenses for shipments of CS-2, CS-3, and 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 fully stabilize; any policy shift could reignite CBRS’s tail risk.

The CBRS IPO, as an event, is the most significant AI hardware capital event of 2026, establishing the valuation benchmark for AI infrastructure in the secondary market; its performance will influence pricing across all related assets.

As a long-term position, it’s a classic “high payout, high uncertainty” bet, relying on three simultaneous assumptions: the macro narrative that “reasoning is king,” the micro execution that “Cerebras can achieve a narrow monopoly by leveraging OpenAI,” and the valuation premise that “the market will continue to pay a 95x price-to-sales premium for AI hardware.” If all three hold true, the returns could be enormous; if any one fails, the drawdown could be severe.

For institutional investors, the typical strategy is to avoid chasing positions on the first day, instead waiting for the quarterly earnings report, key client developments, and valuation adjustment. For individual investors, treating it as a small tail-position allocation within your AI hardware exposure is acceptable; but treating it as an all-in belief play? Please re-read the threefold paradox above.

More noteworthy than whether CBRS will surge upon tomorrow’s open is another layer of meaning behind this event: the fact that a company generating 86% of its revenue from two related entities in the UAE, while still operating at a loss, can be valued at $48.8 billion, speaks volumes about just how extreme the capital frenzy has become in the AI infrastructure space.