Six Things We've Learned from Cerebras' New S-1

What the prospectus filed last week says about the company’s material weaknesses, OpenAI’s circular deal and the $250 billion valuation potential. 

When we last wrote about Cerebras, the company had confidentially refiled for IPO and the debate was theoretical: would public markets reward a capital-intensive AI infrastructure company on strategic positioning alone, as longtime backer Coatue argued, or would “price matter” once a real-time pricing engine got involved?

On April 17, Cerebras moved the conversation forward. The publicly filed S-1 gives investors the first clean look at 2025 financials, the shape of the OpenAI deal, and how the company plans to frame its story ahead of its Nasdaq listing. As the roadshow process begins, Cerebras is said to target a $35B valuation while raising over $3B from the offering. Six things from the filing changed how we’re reading it.

1. Revenue jumped, but the customer concentration only rotated.

Top-line growth was the number everyone wanted to see. Revenue rose 75%, from $290 million in 2024 to $510 million in 2025. That inflection was driven by the launch of the CS-3 system in 2024 and the company's expansion into hosted inference.

But the filing also reveals that customer concentration did not resolve, so much as move within the same ecosystem. G42, which accounted for 85% of 2024 revenue, dropped to 24% in 2025. In its place, Mohamed bin Zayed University of Artificial Intelligence (MBZUAI, a related party to G42) stepped up to 62%. Combined, the two UAE-linked counterparties still delivered 86% of 2025 revenue. MBZUAI alone made up 78% of year-end accounts receivable.

Future diversification now hinges almost entirely on the ramp of OpenAI and AWS as customers/partners, neither of which contributed meaningfully to 2025 results.

2. The GAAP profit is largely accounting noise.

The filing produced an eye-catching headline: Cerebras swung from a $482 million net loss in 2024 to $238 million of GAAP net income in 2025. Several outlets led with the flip to profitability.

The income statement tells a more complicated story. Nearly all of the year-over-year swing is a single non-cash line: a $363 million gain in 2025 from the extinguishment and change in fair value of forward contract liabilities, against a $401 million loss on the same line in 2024. That's roughly $765 million of mark-to-market movement tied to how preferred stock and funding commitments get re-measured at each period end.

Strip it out and the picture inverts. Non-GAAP net loss widened from $22 million in 2024 to $76 million in 2025. Non-GAAP operating loss more than doubled, from $43 million to $96 million. The underlying business is clearly in investment mode as it scales inference infrastructure while spending a whopping 48% of revenue ($243 million) on R&D. Capex (purchase of PP&E) ballooned to $383 million last year, versus $23 million in 2024, reflecting the build out of capacity. That doesn't include what could be billions of future operating lease commitments from a Canadian data center and AWS deals. That level of spending adds context to the recent funding rounds, including $1 billion equity raise and $1 billion in revolving credit from earlier this year, along with the expected ~$3 billion target to be raised from the IPO. 

3. The OpenAI deal is bigger than $20B, but could be slower to recognize.

Cerebras confirmed the OpenAI Master Relationship Agreement, signed December 2025, is valued at “more than $20 billion” for 750 megawatts of AI compute capacity. Three details in the S-1 reshape how that number should be read.

First, the warrant. OpenAI received a warrant for 33.4 million shares of Class N (non-voting) stock at a nominal $0.00001 strike. It only fully vests if OpenAI exercises expansion options scaling the commitment to 2 gigawatts of compute, nearly triple the headline deployment.

Second, the revenue ramp. Cerebras expects to recognize approximately 15% of the OpenAI contract value over the first 24 months ending December 2027 (i.e. ~$1.5 billion / yr at the $20 billion headline mark), with 43% between months 25 and 48 (i.e. ~$4-4.5 billion / yr), and the balance afterward. As expected, it looks like the vast majority of sales won't hit the revenue line until 2028 and beyond.

Third, the financing. OpenAI lent Cerebras $1 billion at 6% interest under a secured promissory note, repayable over three years following delivery of the first 250MW. Interestingly, accrued interest can be satisfied in compute, hardware, or services instead of cash.

The result is a complex, tangled deal. OpenAI is now simultaneously Cerebras's largest contracted customer, its largest creditor, and its largest warrant holder. Concentration along any one of those dimensions would be a risk factor on its own. The economic alignment runs deep, but so does the dependency. And it's a clear reminder that the company was worth just ~$8 billion before the landmark OpenAI deal was announced over the winter, catapulting valuation >$20B.

4. Founder voting lock and a new performance-equity package.

Cerebras will list as CBRS on the Nasdaq Global Select Market under a three-class share structure: Class A (public) carries one vote per share; Class B (founders and pre-IPO investors) carries 20 votes and converts 1-for-1; Class N (customers including OpenAI and G42) is non-voting. Combined with concentrated pre-IPO ownership, the structure means Cerebras likely qualifies as a Nasdaq “controlled company,” exempting it from requirements like a majority-independent board and independent compensation and nominating committees.

One related item from the compensation disclosures is worth surfacing alongside the governance discussion. In February 2026, with Compensia's input, the compensation committee granted Founder/CEO Andrew Feldman and CTO Sean Lie a three-part founder equity package: "make-whole" RSUs explicitly sized to bring founder equity ownership to the top 10% of peer companies, annual refresh RSUs, and — most significantly — performance-RSUs that vest against trailing 90-day market-capitalization hurdles of $75 billion, $150 billion, and $250 billion. Feldman's PRSU grant is 5.7 million shares; Lie's is 3.3 million. The hurdles require roughly 3.3x, 6.5x, and 10.9x the implied $23 billion private mark. Whatever else they signal about internal ambition, they set a useful benchmark for the kind of upside public investors are being asked to imagine at pricing.

The dollar magnitude is easy to lose sight of in the multipliers. The combined 9.0 million PRSUs represent roughly 3.5% of Cerebras on a fully diluted basis. At the $75 billion first hurdle, the 3.0 million shares in tranche one are worth approximately $870 million combined — roughly $550 million to Feldman and $320 million to Lie. At the $150 billion hurdle, cumulative vested value lifts to roughly $3.5 billion ($2.2 billion Feldman, $1.3 billion Lie). At the $250 billion full-vest, the package is worth approximately $8.7 billion, split $5.5 billion to Feldman and $3.2 billion to Lie. The structure echoes Elon Musk's 2018 Tesla CEO performance award (since voided by the Delaware Chancery Court) and Vlad Tenev's 2021 Robinhood grant, both of which used tiered market-capitalization hurdles to align pay with long-dated outcomes.

5. Internal controls: four consecutive years of unresolved weaknesses.

Cerebras discloses material weaknesses in its internal control over financial reporting for both fiscal 2025 and fiscal 2024. The company groups the deficiencies into two buckets. The first is inadequate or missing accounting resources and expertise affecting revenue recognition, inventory management and costing, data center asset accounting, and equity administration. The second is a failure to maintain adequate IT general controls, including ineffective segregation of duties. Remediation, including new hires, formalized processes, new controls etc. is in flight, but the filing notes the weaknesses will not be considered remediated until new controls operate for a sufficient period and are tested as effective.

A material weakness disclosure at IPO is, on its own, an uncomfortably normal thing to see at this stage of the cycle. KPMG's 2024 IPO material weakness study found roughly 43% of all U.S. IPOs and around 55% of TMT issuers disclosed at least one material weakness in their registration filings, with "lack of resources and expertise" the single most cited root cause. Cerebras fits the pattern. Two features sharpen it here: CEO Andrew Feldman's 2007 guilty plea to circumventing accounting controls at Riverstone Networks, which contemporary coverage of the 2024 filing already flagged; and the overlap of the disclosed control gaps with the exact accounts — revenue recognition, inventory, data center assets — where Cerebras is scaling fastest under its largest contracts. Those are the areas a post-IPO auditor will look at first.

Worth noting for continuity: the September 2024 S-1 disclosed substantially the same material weaknesses, for fiscal 2022 and fiscal 2023, using the same two-bucket framing (accounting resources and expertise; IT general controls and segregation of duties). Two changes sharpen the 2026 version. First, the accounting-resources bucket expands beyond "inventory" to "inventory management and costing" — a signal that costing methodology itself is now part of the weakness. Second, "data center assets accounting" is added for the first time, directly coinciding with the period in which Cerebras began scaling data center capacity for OpenAI and AWS. That is four consecutive fiscal years of unresolved material weakness, and the scope has broadened rather than narrowed even as remediation has been underway since the first filing.

6. The pricing ask, grounded in forward EBITDA.

Despite the growth in revenue, unit economics are moving the wrong way on the trailing view. Gross margin slipped from 42% in 2024 to 39% in 2025, and management guides that margins will remain pressured in early 2026. Margin pressure is expected as a function of customer warrants recorded as contra-revenue (Cerebras issued G42 warrants at a $0.01 strike in both December 2025 and April 2026), pass-through data-center cost arrangements, and elevated lease and depreciation as the OpenAI and AWS footprints deploy. The question for underwriting the IPO is whether those margin headwinds reverse fast enough for the forward EBITDA math to pencil out.

Cerebras's February 2026 Series H priced the company at approximately $23 billion. Secondary prints through early April 2026 have pushed the mark closer to $30 billion — a ~30% step-up in roughly two months, concentrated in anticipatory private-market activity ahead of a mooted IPO rather than any disclosed round. A public offering around this $30 billion would be roughly 59x trailing sales against 2025 revenue of $510 million, which sits near the top of the fabless and broader semiconductor distribution shown in the table below. 

Source: AlphaSense / Canalyst semiconductor peer comparison as of April 17, 2026; Cerebras implied multiple based on Series H valuation reported by Bloomberg (Feb 2026) and 2025 revenue disclosed in S-1 filed April 17, 2026. EV in $ millions except Cerebras shown in $ billions.

Trailing sales is useful as sanity check, but it is not the frame the public market will apply to the stock once it lists. Given the company’s expected high growth, the frame that matters instead could well be forward valuation multiples like EV / EBITDA. Capital-light Arm, Astera Labs Broadcom, Credo, Marvell, NVIDIA, and AMD collectively trade in a roughly 15–20x FY+2 EV/adj EBITDA range at current prices. Capital-intensive neclouds trade lower, into mid-teens. Anchoring Cerebras to a blend of that multiple range at its floated ~$35 billion enterprise value implies 2028E adjusted EBITDA of approximately $2.0 billion. That is the level the underwriting has to build toward.

Where the company is today sets the starting point. In 2025, Cerebras recognized $510 million of revenue, posted a non-GAAP operating loss of $96.1 million, and depreciated $34.5 million. Adjusted EBITDA — non-GAAP operating loss plus D&A — is roughly negative $62 million, or about -12% margin. In 2024, adjusted EBITDA was approximately negative $31 million, also about -11% margin. 

If you allow them three years of scaling and anchor on 2028, maybe it gets EBITDA margin to up to roughly 33%, minimum among peers. That is roughly 50 percentage points of EBITDA margin expansion from the current -12% in three years, which is aggressive but not unprecedented. NVIDIA moved from the mid-20s to the mid-60s in roughly three years post-AI inflection. Astera Labs went from breakeven to 30%+ in two. The path requires both hardware gross-margin expansion and scaled operating leverage as the OpenAI and AWS footprints deploy.

Estimating 2028E revenue depends on two components. First, a non-OpenAI base of $510 million in 2025, growing somewhere between 50% and 100% per year as G42 and MBZUAI mature and other hyperscale and enterprise customers ramp. Second, OpenAI revenue. The MRA recognition schedule disclosed in the S-1 could put OpenAI contribution north of $4 billion in 2028. At ~75% baseline CAGR plus $3-4 billion of OpenAI, 2028E revenue lands in a band of roughly $4 billion (low) to $8 billion (high), with a midpoint closer to $6 billion.

The summary version: an IPO at or above $30 billion is a forward bet that two things hold together. First, that OpenAI recognizes ~$4 billion per year of revenue by 2028. Second, that the underlying business rapidly settles at a Marvell- or Astera-like EBITDA margin after the current investment period moderates. 

One supporting signal on private-market demand lives in the compensation notes. Cerebras recognized $31 million of stock-based compensation expense in 2024 and $14 million in 2025 specifically attributable to secondary market transactions where current and former employees sold common stock to existing equity holders at prices above the company's determined fair value. In this case, the excess over fair value is booked as SBC under ASC 718. The December 2025 tender cleared $78 million of Class B stock at the Series G price of $36.23, including $1 million of stock apiece from CFO Robert Komin and COO Dhiraj Mallick. Combined with the company's express waiver of its right of first refusal on secondary sales since January 2023, the pattern suggests  a company supportive of its employees taking some exposure off the table as a hot secondary market began bidding up valuations, particularly into 2H25. 

What we’re watching at pricing

Pull back and Cerebras is one of the cleaner AI-infrastructure bets public markets will see this cycle. On one side are the fabless semiconductors — Nvidia, Broadcom, AMD, Marvell, Astera — where valuation ranges are established and AI exposure is diluted through broader portfolios and cyclical end markets. On the other are the neoclouds — CoreWeave and Nebius in public, with Lambda, Crusoe, and others queuing behind them — a younger cohort built on commodity GPU fleets where gross-margin and customer-concentration economics are still being stress-tested. 

Cerebras is structurally both: silicon IP owner, cloud operator and, as of March 2026, co-designed hyperscaler deployment inside AWS. Crucially, essentially every dollar of revenue comes from AI compute. There is no horizontal software platform, legacy CPU line, or mixed end-market to untangle.

For a company whose marked valuation in the past year has moved from ~$8 billion to over $25 billion in recent secondary prints, the S-1 is the most honest disclosure public investors have had. The next milestones, including price ranges and firmed up raise targets, will help us judge just how much of that private-market story the public market investors are prepared to buy.

Care to get more insights and access to our research? Fill out the form below.

RESEARCH DISCLOSURE This article is for informational and educational purposes only. It represents independent thematic analysis prepared by PrePublic Equity Partners ("PEP") and is intended to discuss industry trends and company dynamics in the private markets. This content does not constitute investment advice, a recommendation, or a solicitation to buy or sell any security or investment product. PEP is not a registered investment adviser or broker-dealer. Any offer or solicitation relating to securities will be made only through definitive offering documents to eligible investors. PEP and its affiliates may hold financial interests in companies discussed herein and reserve the right to trade such positions at any time without notice. Private market investing involves significant risk, including illiquidity and potential loss of principal. All data is sourced from publicly available information and has not been independently verified.

IMPORTANT DISCLOSURE This content is published by PrePublic Equity Partners ("PEP") for informational and educational purposes only. It does not constitute an offer to sell, or solicitation of an offer to buy, any security. No such offer or solicitation is made except by means of a confidential Private Placement Memorandum or other definitive offering documents delivered to eligible investors only. PEP is not a registered investment adviser with the SEC or any state securities regulator. Nothing in this article should be construed as personalized investment, financial, legal, or tax advice. All views are the opinions of the author as of the date of publication and are subject to change without notice. Private market and pre-IPO investing involves a high degree of risk, including illiquidity, potential total loss of principal, and reliance on unverified private company data. Past analytical observations are not indicative of future results. PEP and its affiliates, officers, or employees may hold financial interests in companies discussed in this article. PEP reserves the right to buy or sell such positions at any time without notice. PEP does not receive compensation from issuers mentioned in its research. PEP is an independently operated subsidiary of Alumni Ventures, LLC.