Cerebras Systems, a Silicon Valley AI chipmaker whose first IPO attempt collapsed under federal security scrutiny eighteen months ago, raised $5.55 billion on Nasdaq on Thursday after pricing 30 million shares at $185 each — more than double its original target range — making it the largest global IPO of 2026 and a test of whether retail and institutional investors will overlook a customer base still overwhelmingly concentrated in the United Arab Emirates. Anyone holding a pension fund, a 401(k) with technology exposure, or shares in Fidelity or Benchmark now has an indirect stake in whether that bet pays off.
Arm and SoftBank Sought to Acquire Cerebras Weeks Before the Listing
The morning before Cerebras priced its shares, Bloomberg reported that Arm Holdings and its parent company SoftBank had made a preliminary approach to acquire the company in the weeks prior. Cerebras rejected it. Representatives for all three companies declined to comment. The rejected offer signals how fiercely the AI chip sector is consolidating: SoftBank closed its $6.5 billion acquisition of Ampere Computing in late November 2025, and Arm launched its own branded data-center processor — the Arm AGI CPU, co-developed with Meta — in March. A Cerebras acquisition would have given SoftBank a shortcut into AI accelerator silicon it cannot yet build organically. Instead, Cerebras chose public markets, and the market responded: the stock opened at $350, touched $386.34 intraday before a circuit breaker triggered, and closed its first session at $311.07. As of Friday it trades near $279.
A Chip the Size of a Dinner Plate — and Why That Matters for AI Speed
Cerebras builds the Wafer-Scale Engine 3, a processor that occupies an entire silicon wafer rather than the thumbnail-sized die produced when a wafer is cut into individual chips. At 46,225 square millimeters — roughly 57 times the area of Nvidia's Blackwell GPU package — the WSE-3 packs 4 trillion transistors, 900,000 compute cores, and 44 gigabytes of on-chip SRAM. The practical result is that a single Cerebras system can run inference on very large language models with latency that a rack of GPUs cannot match, because the weights never have to travel between chips.
Cerebras CFO Bob Komin told CNBC on Thursday: "For our fast inference product, there's so much demand that our biggest challenge is actually trying to supply it. We are adding as much manufacturing and data center capacity as we possibly can, and we're still sold out into 2027." The chip is manufactured at TSMC on a 5-nanometer process node — less advanced than the 2-nanometer node used for the most cutting-edge AI chips, a limitation competitors will likely exploit.
OpenAI's $20 Billion Deal Carries an Unusual Price: Equity, a Loan, and Clawback Rights
The central pillar of Cerebras' bull case is a multi-year Master Relationship Agreement with OpenAI worth more than $20 billion committing to deploy 750 megawatts of Cerebras inference capacity through 2028. The structure is more entangled than a standard vendor contract. OpenAI advanced Cerebras a $1 billion working capital loan in January 2026 at 6% annual interest, secured by warrants to purchase up to 33.4 million shares of Cerebras stock at a near-zero exercise price. At Thursday's opening price, those warrants would be worth approximately $11.7 billion if fully vested.
Morningstar senior equity analyst Brian Colello characterized the arrangement bluntly: "The OpenAI deal came as part of a $20-billion arrangement that took the form of a commitment to buy billions of dollars of Cerebras' chips in exchange for a stake in the company and a $1 billion loan." Colello added that the greatest risk for Cerebras investors is "intense competition in AI inference, especially versus market leader Nvidia and its Groq business unit," and that "customer concentration would be another, especially since OpenAI has made large chip deals and will need to continue to grow to justify such deals."
The S-1 discloses a significant risk buried in the loan terms: if the Master Relationship Agreement is terminated for any reason other than OpenAI's own uncured breach, OpenAI can seize control of the loan funds and demand immediate repayment. Cerebras is also handing OpenAI warrants worth roughly half the gross profit it stands to make on the deal, according to Financial Times calculations. AWS holds a separate warrant package, having agreed in March to become the first major cloud provider to deploy Cerebras hardware inside its own data centers, accessible through Amazon Bedrock.
The UAE Concentration That Blocked the First IPO Has Largely Shifted, Not Shrunk
The national security review that forced Cerebras to withdraw its first IPO filing in late 2024 centered on G42, an Abu Dhabi AI conglomerate under CFIUS investigation for its historical ties to Huawei and the risk of advanced American AI technology reaching China through UAE intermediaries. G42 provided $335 million in investment and accounted for 85% of Cerebras' revenue in 2024.
The amended S-1 filed in May 2026 shows G42's revenue share has fallen to 24% — but the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), a UAE entity that the prospectus identifies as a related party of G42, accounted for 62% of 2025 revenue. Together the two UAE-linked entities represented roughly 86% of revenue. MBZUAI alone accounted for 77.9% of accounts receivable as of December 31, 2025. The S-1 warns explicitly: "The loss of, any substantial reduction in sales to, or the default on payments by, any of our significant customers would harm our business." Geopolitical tensions in the Middle East and potential escalation of U.S. export controls are listed as material risks.
The company's own prospectus acknowledges that G42 and MBZUAI "are considered related parties with respect to each other." For investors, the practical implication is that the revenue base Cerebras presented as diversified in 2025 largely reflects a reallocation between two connected UAE entities — not the addition of independent Western customers. CEO Andrew Feldman acknowledged the issue on CNBC Thursday, describing the business model matter-of-factly: "There's some whales out there, there's some really big customers. That is one of the characteristics of this market."
Reported Profit Rests on a $363 Million One-Time Accounting Gain
Cerebras reported $510 million in 2025 revenue, up 76% from $290 million in 2024, and GAAP net income of $237.8 million — numbers that circulated widely as evidence of profitability at an IPO-stage hardware company. The S-1 tells a narrower story. The swing to profit was driven by a one-time, non-cash gain of $363.3 million from extinguishing a forward-contract liability tied to G42's original investment. Strip out that gain and adjust for stock-based compensation, and Cerebras posted a non-GAAP net loss of $75.7 million — a 247% deterioration year-over-year. The company's GAAP operating loss for 2025 was $145.9 million.
At Thursday's closing price of $311.07, the stock traded at between 130 and 190 times trailing 2025 revenue, depending on the share-count basis used — compared with roughly 26 times for Nvidia and 21 times for AMD, according to CNBC's Jim Cramer, who said: "Really, if you buy this stock up here, you're betting on the idea that Cerebras will have much better growth for many years in the future." Motley Fool analysts described the valuation as "borderline euphoric" and declined to include Cerebras on their list of recommended buys.
The $24.6 billion in remaining performance obligations cited in the prospectus is real, but management estimates only roughly 15% will be recognized across 2026 and 2027 combined — meaning the bulk of that backlog converts to revenue in 2028 and beyond, years during which Nvidia, AMD, Google, Amazon, and Microsoft will all be competing in the same inference market.
Nvidia Paid $20 Billion to Enter Cerebras' Core Market Three Months Before the IPO
Cerebras' primary commercial advantage — low-latency inference using fast SRAM memory rather than the high-bandwidth DRAM used in GPUs — is now directly targeted by its largest competitor. In December 2025, Nvidia paid $20 billion to acquire assets from Groq, a startup whose Language Processing Units use an architecture more similar to Cerebras' than to a traditional GPU. Nvidia announced Groq-based products at GTC in March 2026. Morningstar's Colello identified Nvidia's Groq unit as "likely Cerebras' fiercest rival, since both are attacking AI inference using smaller but faster SRAM memory."
Cerebras also now competes directly with the cloud divisions of Google, Microsoft, Oracle, and Amazon — all of which are building or have already deployed their own custom inference chips. The company's transition from chip seller to cloud operator, while strategically necessary to reach new customers, places it in a market where its largest customers are also its largest competitors.
Founders and Early Investors Converted Years of Risk Into Billions of Dollars on Thursday
The IPO crystallized substantial personal wealth for Cerebras insiders. CEO Andrew Feldman, who co-founded SeaMicro before AMD acquired it in 2012, held a stake worth approximately $1.9 billion at the IPO price. CTO Sean Lie's stake was valued at roughly $1 billion. Greg Brockman, OpenAI's co-founder and president, held approximately 78,000 Cerebras shares — worth $14.4 million at the IPO price — a holding disclosed during Elon Musk's ongoing litigation against OpenAI, in which a 2017 email from Brockman was entered into evidence stating that "exclusive access to Cerebras hardware would give OpenAI an overwhelming hardware advantage over Google." OpenAI CEO Sam Altman held roughly 89,000 shares, valued at approximately $16.5 million at pricing.
Among institutional holders: Fidelity held a position worth approximately $3.8 billion at the IPO price; Benchmark roughly $3.3 billion; Foundation Capital approximately $2.8 billion; Eclipse approximately $2.5 billion. The offering was reported to be approximately 20 times oversubscribed.
CoreWeave's 162% Post-IPO Return Set the Appetite Cerebras Priced Into
The clearest benchmark for Cerebras is CoreWeave, which listed on Nasdaq in March 2025 at a $23 billion fully diluted valuation, raising $1.5 billion. CoreWeave stock has risen 162% since its debut, providing a concrete tailwind for Cerebras' marketing. But the two companies carry different risk profiles: CoreWeave resells Nvidia GPUs as a cloud service and operates at a loss; Cerebras designs its own silicon, carries higher gross margins, and — on a non-GAAP basis — remains loss-making. PitchBook senior analyst Dimitri Zabelin argued that the AI market's shift from training toward inference favors Cerebras: "The AI hardware market rotated from training-cycle dominance toward inference-cycle scaling, where token generation speed and cost per query determine competitive positioning." Cerebras' first earnings report as a public company will be the first real test of whether that argument translates into diversified, recurring revenue.
What Public Investors Are Actually Buying — and What Remains Unproven
Cerebras has never operated cloud infrastructure at hyperscaler scale. The OpenAI deployment ramps across 2026 through 2028. The AWS arrangement, signed as a binding term sheet in March 2026, has not yet converted into deployed revenue. The $2.2 billion Canadian data center lease signed in March — with expected minimum payments over ten years — and the $344 million in existing non-cancelable data center commitments mean Cerebras' capital obligations are substantial and front-loaded even as its revenue is back-loaded.
For ordinary investors considering shares at current prices: the stock now trades at roughly 160 times the company's actual operating-loss-adjusted earnings from 2025, in a sector where Nvidia — which dominates the market Cerebras is trying to enter — trades at 26 times sales. The price embeds an assumption of near-perfect execution on contracts with two customers that are either a UAE-linked entity or a company whose founders hold personal stakes in Cerebras. Every prospective shareholder deserves to know that before buying.
Cerebras trades on Nasdaq under the ticker symbol CBRS.
ⓒ 2026 TECHTIMES.com All rights reserved. Do not reproduce without permission.
