Transcription
It's 9:30 in the morning, May 14, 2026. The opening bell, NASDAQ. The ticker CBRS, $185. Last night's IPO price.
Within seconds, the first trade prints $350. By mid-morning, shares hit $386. The Financial Times reported the market valuation briefly approaching $70 billion. Baron said the fully diluted figure crossed $100 billion intraday in the hours after the open. On a Thursday morning, the public markets had stopped treating Cerebras like a chip company and started treating it like AI scarcity itself.
Then Market Watch ran a headline, "Cerebras' stock looks like a risky bet on AI hype." The company had $510 million in annual revenue. It was still losing money on operations and 86% of its sales came from two entities in Abu Dhabi. So, what exactly had the market just priced?
Those numbers aren't the story. They're the entry point because Cerebras doesn't sit at the center of a single problem. It sits where three separate pressures converged at once. The scramble to own AI infrastructure before the market decides what it's worth. The geopolitics of which countries and companies get to build it and a public market that for one Thursday morning chose to stop asking ordinary questions. That combination is rare and it's why what happened on May 14 matters beyond a single stock's debut.
What the market ran that day was in effect a referendum, a live public test of whether the AI boom has moved from credible growth into something more precarious. What some analysts are already calling peak hype financialization. That is the question this video is built around.
The story that produced that valuation took seven years to build. It started in August 2019 when Cerebras announced something the semiconductor industry had spent decades saying was not viable. A single chip the size of a dinner plate. One trillion transistors. We scale computing had been dismissed as impractical. Too many defects. Too much heat. No path to yield at commercial scale. Cerebrus built it anyway.
Andrew Feldman, the company's co-founder and CEO, framed the approach in one line. Every architectural decision was made to optimize performance for AI work. In April 2021, the WSE2 arrived, 2.5 trillion transistors built on TSMC's 7 nanometer process with more than double the performance of its predecessor. The first chip had not been a stunt. There was a road map.
By November 2022, the story had expanded. Cerebras unveiled Andromeda, a 13.5 million core AI supercomputer delivering more than one exoflop of compute. The company had stopped selling processes. It was selling AI infrastructure. What the Andromeda announcement did not mention, and what the 2024s, one would later confirm, was that the same period included workforce reductions, layoffs in Q3 2022, and then again in Q1 2023, which brought R&D costs down. Behind the ascending headline, the company was managing burn the way startups do. That detail did not travel with the myth.
In July 2023, the scale shifted again. Cerebras and G42, the Abu Dhabi technology group, announced the Condor Galaxy network, nine interconnected supercomputers targeting 36 exoflops of AI compute in total with the first system already delivering four. Ibrahabi had become structurally central to the growth story. The press stopped calling Cerebras a hardware vendor and started calling it a state-scale AI infrastructure company.
March 2024 brought the WSE 34 trillion transistors, 900,000 AI cores, a chip the company's own materials described as 58 times larger than the largest available GPU. Private investors repriced upward at every beat, over $4 billion in the series F, $8.1 billion in September 2025, approximately $23 billion in February 2026.
Then on January 14th, 2026, Open AAI announced it was partnering with Sirius to deploy 750 megawatt of ultra low latency compute rolling out in phases through 2028. Sachinati from OpenAI stated it plainly. Sirius adds a dedicated low latency inference solution to our platform. Feldman's framing was characteristically expansive. Just as broadband transformed the internet, realtime inference will transform AI.
The IPO range opened at $115. Demand pushed it to $125, then $150 to $160. On the night of May 13, it priced at $185. The myth had been built carefully, milestone by milestone, number by number.
What it was built on is a different story. Start with the revenue. The numbers were real. $24.6 million in 2022, $78.7 million in 2023, $290.3 million in 2024, $510 million in 2025. 4 years of documented growth in a hardware category most incumbents had written off. But the composition of that revenue told a different story than the headline did.
In 2023, a single customer, G42, the Abu Dhabi Technology Group, contributed $65.1 million, 83% of total revenue for the year. In the first half of 2024, that share rose further to 87%. That same year, Feldman announced a major funding round by saying, "This new funding allows us to extend our global leadership to new regions, democratizing AI. 83% from one customer in the Gulf is not a democratized customer base. It is a single relationship dressed in platform language." And behind the ascending revenue line, a detail that did not travel with the myth, the company had executed workforce reductions in the same period. layoffs in Q3 2022 and Q1 2023, managing burn the way startups do while announcing supercomputers.
By the time the 2026 filing arrived, G42 was no longer the dominant revenue name. Coverage treated this as evidence that Cerebrus had diversified. Pitchbook's S1 breakdown offered a more precise read. In 2025, the largest single revenue source was MBZ UAI, the Muhammad bin Zed University of Artificial Intelligence. 62% of annual revenue. D42 contributed another 24% combined 86% of 2025 revenue still tied to two entities both in Abu Dhabi. Pitchbook summarized it directly. Cerebras remained plenty reliant on the United Arab Emirates. The label had changed, the structure had not, and the company entered the IPO carrying $24.6 $6 billion in remaining performance obligations, much of it linked to open AI with only about 15% expected to convert to recognized revenue in each of 2026 and 2027. Gigantic future commitments, still narrow present reality.
Then came the profit story, and this is where the numbers require careful reading. In 2026, multiple outlets reported that Cerebras had turned profitable in 2025. Some coverage cited net income of $87.9 million. Other credible summaries cited $237.8 million. Both figures reference the same filing. The divergence is itself the story. FT and Baronss pointed to a different line. An operating loss of roughly $146 million. The company's core business measured by what it earned selling products and services versus what it spent running them was still in the red.
The gap between those two pictures sat in a single balance sheet entry. Pitchbook identified it. Cerebris removed a G42 related liability and recorded a paper gain of $363 million non-cash, non-operating. That accounting treatment made 2025 look substantially stronger than the underlying business performance warranted. Tom's Hardware stated it plainly. Most of its reported profit that year was due to a $363 million accounting gain, not core business operations. The same Abu Dhabi relationship that had nearly killed the first IPO contributed the paper gain that made the second IPO look like a profitability story. The filing disclosed this. What traveled through markets was the headline number. What sat in the footnotes was the operating loss.
Which brings us to why there were two IPO attempts in the first place. In September 2024, the S1 went public. Within days, Reuters reported the road show would likely be postponed. The reason was not weak demand. It was not the operating losses. It was CPHAS, the Committee on Foreign Investment in the United States, reviewing G42's planned 335 million minority stake in Cerebras. This is the moment that inverts the standard Silicon Valley story. The bottleneck was not the product, not the customers, not the numbers. It was geopolitical trust. US national security institutions looking at the cap table and the customer map and concluding the arrangement required extended examination.
The review ran for over a year. By March 2025, TechCrunch reported the listing was still unresolved. Not a temporary delay, still unresolved. By October 2025, Sirius formally withdrew the first filing. By December, Reuters reported G42 was no longer listed as an investor in the renewed version. The reason Reuters noted was not publicly explained. What was clear was that MBZUI, a different Abu Dhabi institution, remained the company's largest revenue source on the way into the IPO. The cap table had been cleaned. The dependent had not.
The last piece involves what? Going public actually meant for the investors who bought in. Barons reported that after the IPO, class B shareholders, insiders, and legacy holders would control approximately 99% of voting power. Public buyers received class A shares. Economic exposure, yes. governance leverage effectively none and the lockup structure deviated from the standard pattern rather than a clean six-month window before any insider selling. Barons flagged that a large volume of eligible stock was structured to become sale eligible in stages beginning earlier than investors typically assume.
Consider the capital stack that surrounded this. In eight months before the offering, a $1.1 billion series G, a $1 billion series H led by Tiger Global, which nearly tripled the private valuation in months, and an $850 million revolving credit facility. Techrunch reported that Benchmark separately raised a dedicated $225 million special vehicle just to increase its position after the series H. Then the IPO raised $5.55 billion. after the underwriters exercised their full option, $6.38 billion.
For the investors who had been in since the series F at a $4 billion valuation, this structure worked exactly as intended. They were selling into a market that had repriced the company to nearly $70 billion in a single morning. For the public buyer at $350 on May 14, the final entry point in a 5-year ladder built by some of the most sophisticated capital allocators in the world. The terms of participation were materially different.
Each of these four things was disclosed. None of them was hidden. The question the market chose not to ask, at least not until the afternoon, was whether disclosed risk and priced risk were the same thing. All four of those things were in the filing, disclosed, documented, accessible to anyone who read past the revenue chart.
The harder question, the one the filings don't answer, is about the people who built this story, who made the decisions at each step. And what those decisions reveal about how the AI era actually operates when the cameras aren't on the chip. The achievements are real and documented. Argon National Laboratory researchers using Cerebra's hardware won the Gordon Bell special prize. The Department of Energies in NL found the CS2 nearly 500 times faster than its own dual supercomput on a specific modeling workload. Three generations of wafer scale processes, a genuine technical lineage.
And yet Cerebrus' owns one the prospectus the company submitted to the SEC contained a disclosure that most of the investors buying $350 on May 14 likely did not reach. The filing stated that the SEC had previously alleged improper conduct related to revenue recognition at Riverstone Networks where Feldman had held a senior role. In December 2007, Feldman plead guilty to one count of circumventing the accounting controls of an issuer and later settled civil claims with the SEC. These events predated Cerebras by nearly a decade. They are not evidence of misconduct at Cerebras.
But the same one also disclosed something present tense, material weaknesses in the company's internal controls over financial reporting. Not historical, current. Barren made a specific editorial decision to pair those two disclosures. the historical and the present day and flag them together as a risk amplifier. Two separate documents, same filing, one from 2007, one from 2026, both disclosed, neither widely circulated in the coverage that drove CBRS to $350.
The OpenAI announcement on January 14, 2026, was the single event that made Siri's legible to mainstream investors. 750 megawatt of low latency compute phases through 2028. a partnership with the most important AI company in the world. What the announcement did not address was the ownership structure underneath it. Cerebras' own investor page listed Sam Alman, Greg Brockman, and Ilia Sutsker among the company's backers. All three were at various points among the most senior figures at OpenAI. Reuters later reported from court filings that Olman held a personal stake in Cereabbras, valued at $3.2 million. Olman stated he used standard recusal procedures for companies in which he invested personally and he rejected the characterization that this created an improper arrangement. Lawyers for Elon Musk's legal team in separate litigation alleged Altman had an obvious conflict, a claim Altman disputed.
None of that is established wrongdoing. Recusal procedures exist precisely for situations like this. What it does mean is that the January 2026 announcement, the event that added billions to Cerebris's implied value and helped price the year's largest US IPO, was a commercial agreement between organizations whose senior leadership overlapped on both sides of the transaction. That is a fact. What weight the market assigned to that fact on May 14 is a different question.
By December 2025, Reuters reported that G42 was no longer listed as an investor in Siri Brass's renewed filing. The headline was clean. Cphus review resolved. Ownership clarified. Path cleared. Reuters also noted in the same article that it could not determine why G42 had been removed. That sentence traveled almost nowhere in the coverage that followed.
The actual revenue table in the 2026 filing told a more complete story. G42, the entity no longer on the investor list, was still generating 24% of Cerebras' annual revenue as a commercial customer. MBZUAI, a different Abu Dhabi linked institution, was generating 62%. US build revenue had moved in the opposite direction, down year-over-year. Even as total revenue expanded, the regulatory optics had improved, the underlying dependency had not. What changed was the label on the largest counterparty, not the geography of the income statement, which leaves the question of who knew all of this and kept going anyway.
By September 2025, Cerebrris had already lived through a failed IPO attempt, a year-long cifhious review, a formal filing withdrawal, disclosed operating losses, a customer concentration problem, and a CEO with a public legal history. The facts were not obscure. They were in Reuters, TechCrunch, Bloomberg, and the company's own filings. That month, Benchmark, already a major holder, did not reduce its position. It raised a dedicated $225 million special vehicle to increase it. Tiger Global led a $1 billion series H in February 2026 that nearly tripled the private valuation in months. An $850 million revolving credit facility closed in April. The IPO raised $5.55 billion. After the green shoe, $6.38 billion.
These were not passive outcomes. They were decisions made by some of the most informed capital allocators in the world with full access to every disclosed risk at every step of the ladder. The public buyer at $350 on the morning of May 14 was the last person onto that ladder. They had the same filings. They had less context, less history, and near zero voting power in the company they were buying. Whether that asymmetry was the market functioning correctly or the market revealing something about itself is the question this story leaves open.
The asymmetry between who knew what and when and at what price is the market functioning. It is also depending on where you stood on May 14, the market revealing something about itself. Go back to that Thursday morning 9:30 NASDAQ opens. CBRS prints at $350, climbed to $386.34, closes at $311.0. 07 by the following day, $279.72. FD Alphaville had run a title during the first IPO attempt in 2024 that aged without losing its edge. Cerebra Systems needs a dose of initial public skepticism. Market Watch on May 14 itself, Cerebras' stock looks like a risky bet on AI hype.
Both of those assessments are accurate. And a third thing is accurate. Alongside them, Cerebras built a genuinely differentiated technology. Researchers used this hardware to produce scientific results that would not have been practical otherwise. The revenue growth was real. The open air partnership is a real commercial commitment. The wafer scale engine is a real engineering achievement. The critique is not that the company built nothing. The critique is more precise than that.
The critique is that the public market priced's genuine achievements as if all the surrounding risks, the customer concentration, the operating losses, the governance structure, the geopolitical residue, the backlog that converts at roughly 15% per year had already been resolved. They had not been resolved. They had been disclosed. That is the specific character of peak hype financialization. Not fraud, not fabrication, but genuine innovation used as permission to stop asking ordinary questions. The chip is real. The questions it stopped are also real.
What comes next for Cerebrus will be determined by a different kind of test than the one that ran on May 14. The first post IPO earnings call will ask how much of the $ 24.6 billion backlog has converted to recognized revenue and how much remains a promise on a slide. The answer to that question will say more about what this company actually is than any single trading day could. Whether May 14, 2026 looks in retrospect like the opening of a new chapter in AI infrastructure or the day we can point to and say that was the ceiling depends entirely on what the numbers look like when the story has to stand without the hype. We don't know yet. But we have for the first time a public document that lets us watch in real