Cerebras Systems filed an S-1 registration statement with the Securities and Exchange Commission on Friday, April 17. [1] The filing is the company's second attempt to go public. It withdrew its first in October 2025 amid a Committee on Foreign Investment in the United States review tied to investments from the UAE-based G42. The new filing arrives after G42 clearance was secured and after OpenAI — through two separate commitments disclosed Thursday by The Information and confirmed in the S-1 — expanded its compute deal with Cerebras to more than $20 billion over three years. [2]

Cerebras reported 2025 revenue of $510 million, up 76 percent year over year. It reported 2025 GAAP net income of $237.8 million — $87.9 million of which is attributable to common shareholders after $149.9 million is allocated to preferred stock — compared with a net loss of $481.6 million the prior year. [3] The company is targeting a listing valuation of approximately $35 billion — a 43 percent premium to its $23.1 billion Series G valuation in September 2025 — and plans to raise more than $3 billion in the offering. The listing window is the second quarter of 2026. Morgan Stanley, Citigroup, Barclays and UBS are leading the underwriting; Mizuho and TD Cowen are also involved.

The paper's April 18 feature on Anthropic's two-track week noted that the AI infrastructure race was moving from private markets into the language of public disclosure. The Cerebras filing is the first document that lets the paper name, in a public-markets instrument, a specific dollar value on OpenAI's compute commitments. Twenty billion dollars over three years, warrants for up to a ten-percent stake in the chipmaker if spending reaches $30 billion, an additional $1 billion that OpenAI has fronted to Cerebras as a working-capital deposit for data-center construction. [4] The customer is also the counterparty. The counterparty is also an equity holder. The equity holder is also the main competitor Cerebras's chips are supposed to free the AI industry from.

What the concentration looks like

In the first half of 2024, G42 accounted for 87 percent of Cerebras's revenue, a share that triggered the CFIUS review that delayed the first IPO attempt. [5] By 2025, the G42 share had reduced — the S-1 discloses the specific percentage, but it is well below 2024's level — and OpenAI had become the largest customer. The OpenAI commitment of $20 billion over three years works out to an annual run-rate of roughly $6.7 billion. Against $510 million of trailing annual revenue, the OpenAI commitment alone is approximately thirteen times the company's current annual revenue base. If the customer base continues to concentrate as the OpenAI deal ramps, Cerebras at the public-market listing is a company whose business is, in economic substance, the execution of a single customer's compute needs.

This is not unique. Nvidia, which Cerebras positions as its competitive target, has the reverse problem of concentration: its customer base is broad, its revenue is diversified, its price-setting power is the envy of the hardware industry. Nvidia's public-market valuation reflects that diversification. Cerebras is offering public-market investors the opposite: a company whose revenue ramp is tied to a single private counterparty whose own valuation is, in turn, tied to its ability to continue acquiring compute at scale. The dependency is bilateral. So is the risk.

What OpenAI gets

OpenAI receives warrants for a minority stake that scales with spending. At $20 billion, the warrants represent a minority position. At the full $30 billion of potential spending, the warrants represent up to ten percent of Cerebras. OpenAI becomes, by the public offering, a meaningful equity holder in the company it is also the anchor customer of, and the company it is also an investor in — Sam Altman is an early Cerebras investor, a point the S-1 discloses. [6] The relationship is multidimensional in a way that public-markets disclosure will have to enumerate but cannot fully resolve.

The $1 billion working-capital deposit toward data-center construction is the most interesting disclosure. Cerebras has booked it as a balance-sheet asset — a deposit receivable — rather than as current revenue. That accounting treatment matters: it keeps the $1 billion from appearing in the 2025 revenue line, which preserves the clarity of the 76 percent revenue growth figure, but it tells public-market investors that Cerebras's operating cash flow in 2025 was augmented by a $1 billion capital advance from its largest customer. The capital advance is a subsidy with teeth. OpenAI helped build the data centers that will then run the inference workloads that OpenAI is contracted to purchase. The vertical integration is extensive. The disclosure reveals how extensive.

What Nvidia should read

Cerebras's wafer-scale engine chips are genuinely different from Nvidia's discrete GPU architecture. The WSE is a single silicon wafer that integrates the equivalent of hundreds of thousands of cores. For specific inference workloads — particularly large-context-window inference at low latency — WSE has documented speed and power-efficiency advantages over H100 and B200 clusters. OpenAI's decision to commit $20 billion of its compute budget to Cerebras is, in one reading, an operational hedge against Nvidia's pricing power.

Nvidia has responded by lowering its own inference-optimized chip prices (the Jensen keynote at GTC 2026 emphasized this) and by securing multi-year compute commitments from hyperscalers. The public-market arrival of a credible challenger, however small in absolute terms, changes the negotiating room Nvidia has with its own largest customers. Cerebras at a $35 billion valuation is roughly one percent of Nvidia's market capitalization. One percent is not a threat. One percent is a pricing comparable that large buyers can cite at contract renewal.

What CFIUS might still do

The G42 review delayed the first IPO. Clearance was secured. Whether CFIUS re-examines the G42 tie once the company is public — there is no procedural reason it would, but there is political pressure from several Senate offices that would welcome a review — is an open question. UAE investment exposure in U.S. AI infrastructure is not yet a settled policy area, and the G42 relationship, even reduced, remains on the cap table. Any CFIUS action in the next six months would be the most consequential risk factor not yet priced into the range.

The window for execution is narrow. If the IPO prices in May or June at the target valuation, Cerebras arrives in public markets with OpenAI's compute commitment as its central demonstration of product-market fit. If OpenAI's own business trajectory holds, the Cerebras story is bulletproof for at least eighteen months. If OpenAI's private valuation — currently above $300 billion — experiences any downward adjustment before Cerebras's 2027 earnings are on record, the concentration risk Cerebras is carrying becomes visible in its own stock price before the company can diversify the customer base.

A $20 billion contract is a miracle if your anchor customer is durable. It is a knife held at your own head if the customer is not. The S-1 is the document where that distinction stops being private and starts being priced.

-- THEO KAPLAN, San Francisco