Cerebras Raised $6.4 Billion in the Biggest Chip IPO Ever. Its Margins Look Like a Cloud Provider's. That's a $50 Billion Problem.

Sixty-eight percent. That is the share of Cerebras Systems' revenue it spent on capital expenditures last quarter, a figure buried in the company's first public earnings release on June 23 and one that tells you more about its future than the headline revenue beat or the $20 billion OpenAI contract that dominated coverage. NVIDIA, the company Cerebras is most often compared to and whose valuation multiples the market has generously applied, spent about 8% of revenue on capex in the same period. AMD spent roughly 5%, and Cerebras spent 68%.

That single ratio crystallizes the central tension of the biggest semiconductor IPO in history: Cerebras built a genuinely differentiated chip, raised $6.4 billion selling it to public markets in May, and is now being valued at $49.8 billion on $860 million in guided annual revenue, a 58× multiple that assumes the fat margins and capital-light economics of a semiconductor designer. But the company's own financials reveal something closer to a vertically integrated data center operator, a business model that historically commands multiples of 10 to 15× revenue and margins twenty to thirty points below the semiconductor standard.

Which one is Cerebras? The answer matters to the tune of about $40 billion in market capitalization.

What the Numbers Actually Say

Start with the Q1 income statement. Revenue hit $193.4 million, up 94% year over year and 13% sequentially, beating Wall Street's $181 million consensus by a comfortable margin. Hardware revenue of $110.6 million grew 59% while cloud and services revenue nearly tripled to $82.8 million, reflecting Cerebras's aggressive push to sell compute rather than boxes. GAAP gross margin came in at 45%; the company's preferred "core" metric, which strips out stock-based compensation, pass-through data center costs, and customer warrant amortization, printed at 47%. GAAP net loss narrowed to $14 million from $23.9 million a year ago. Operating cash flow turned positive for the first time: $12.3 million.

Good numbers in isolation, but now look at the guidance Cerebras gave for Q2: core gross margin of 36 to 38%, a collapse of nine to eleven percentage points from Q1's 47%. For the full year, management guided 38 to 41% core gross margins alongside core operating margins of negative 28 to 32%. CFO Bob Komin attributed the margin compression to "increased data center deployment costs and operational scaling challenges," which is corporate speak for a business that is rapidly shifting from selling chips at semiconductor margins to renting compute at infrastructure margins.

Compare that trajectory with the peer group investors are implicitly benchmarking Cerebras against:

Company	Gross Margin	Capex / Revenue	Revenue Multiple	Business Model
NVIDIA (Q1 FY27)	75.0%	~8%	~15×	Fabless chip designer
AMD (Q1 2026)	~50%	~5%	~8×	Fabless chip designer
TSMC (Q1 2026)	~55%	~45%	~10×	Contract foundry
Equinix (Q1 2026)	~47%	~15%	~10×	Data center REIT
Cerebras (Q1 2026)	45% (→36-41%)	68%	58×	Chip maker + cloud

Cerebras's capex intensity exceeds every semiconductor company in the table, exceeds TSMC (which operates some of the most capital-intensive fabrication facilities on Earth), and even exceeds the implied capex burden of major cloud providers. Its guided margins sit in the zone historically occupied by infrastructure operators, not chip designers. Yet the market has awarded it the richest revenue multiple of the group by a factor of four.

Revenue per Megawatt: Where the Math Gets Uncomfortable

The OpenAI deal is the fulcrum. Reuters reported in April that OpenAI agreed to pay more than $20 billion over three years for 750 megawatts of Cerebras inference compute, a massive commitment that has since been confirmed in Cerebras's press release though not itemized at the deal level in SEC filings. Run the division: $20 billion across 750 MW over three years works out to roughly $8.9 million per megawatt per year.

Now run the payback arithmetic. Industry benchmarks from JLL and McKinsey peg data center construction costs at $20 million to $40 million per megawatt of IT capacity, depending on Tier level, cooling technology, and power infrastructure. At Cerebras's guided core gross margin of 38%, each megawatt generates about $3.4 million in annual gross profit. Dividing the low-end build cost by that figure yields a payback period of 5.9 years on the capital invested to service a contract that lasts three. At the high end, payback stretches to 11.8 years.

Cerebras is not building all 750 MW at once, and some infrastructure will be leased rather than owned, which alters the math in nuanced ways. But the directional finding is clear: on a pure gross-profit basis, the data center capex required to fulfill the OpenAI deal is unlikely to pay for itself within the deal's own timeframe, meaning Cerebras needs either renewals, additional customers on the same infrastructure, or margin expansion that defies its own forward guidance to make the unit economics work.

Your Customer Is Also Your Banker

Buried in Cerebras's balance sheet is a line item labeled "Loan from customer," split between $621.3 million in current liabilities and $361.6 million in non-current liabilities, totaling $982.9 million. This is the $1 billion working capital loan OpenAI extended in January to help fund data center development. Alongside it sits $516 million in "customer warrants," an asset representing equity instruments issued to OpenAI that are amortized as contra-revenue over the life of the relationship.

Reuters reported, citing The Information, that OpenAI's cumulative spending could reach $30 billion, in which case the warrants could represent up to 10% of Cerebras. At the current $49.8 billion market capitalization, that is roughly $5 billion in equity on $30 billion in spending, an effective 17% rebate that never shows up in the headline revenue figure but shows up precisely where it matters: in the gap between reported revenue and economic value received.

Patrick Moorhead, chief analyst at Moor Insights & Strategy, summarized the structural problem in an X post following earnings: "Concentration did not go away; it rotated. The answer is a $20B-plus, 750MW commitment to one customer, OpenAI, who also lent Cerebras $1B and gets paid partly in warrants. Trading G42 risk for OpenAI risk could be looked at as a bigger single point of failure, not diversification, and most of the $24.6B backlog is that one contract."

In 2025, UAE-affiliated customers including G42 and Mohamed bin Zayed University of Artificial Intelligence accounted for roughly 86% of Cerebras's revenue, according to the company's IPO prospectus. The OpenAI partnership was supposed to diversify that base. Instead, it appears to have rotated concentration from one dominant customer to another, while adding a layer of financial entanglement that did not previously exist: your largest customer is now simultaneously your largest creditor, a future equity holder, and the source of most of your backlog.

Strongest Counterargument

Cerebras's Wafer Scale Engine is not marketing theater. The WSE-3 uses the entire silicon wafer as a single processor, eliminating the inter-chip communication latency that constrains GPU clusters, and independent benchmarks from Artificial Analysis have measured it delivering nearly 1,000 tokens per second on Moonshot AI's Kimi K2.6, a trillion-parameter model, a throughput figure no GPU-based system can match today for inference workloads at that parameter count. AWS validated this by choosing Cerebras for the decode phase of a disaggregated inference architecture, pairing it with AWS's own Trainium 3 chips for prefill, a partnership that effectively makes Amazon a co-architect of Cerebras's compute model rather than a competitor.

Margin compression may also be transitory. In Q1, cloud and services revenue grew 178% year over year at a 49% gross margin, well above the 41% hardware margin and trending higher, because once data center infrastructure is built and depreciated, incremental compute revenue falls almost entirely to gross profit. If cloud revenue grows from 43% of the mix today to 70% or more, blended margins could recover into the mid-40s or higher without any pricing improvement whatsoever, purely through mix shift. The heavy capex of 2026 and 2027 may be planting the infrastructure that sustains a structurally higher-margin business by 2028. Early AWS was capital-intensive and margin-thin too, and it became the most profitable division in Amazon's history.

Limitations

Several caveats constrain this analysis. OpenAI deal terms are drawn from Reuters and The Information reporting, not from Cerebras's own SEC disclosures, which describe the partnership in general terms without itemizing per-megawatt pricing, warrant conversion schedules, or annual minimum commitments. Our data center build-cost range of $20 million to $40 million per megawatt is an industry-wide average; Cerebras's actual costs could be lower if the company is leasing rather than building, or higher if custom cooling infrastructure is required for wafer-scale processors. We do not know what percentage of Cerebras's FY2026 revenue is attributable to OpenAI specifically. And the 10% warrant figure comes from a single publication's unnamed sources; the actual dilution could be materially different.

What You Can Do

If you hold CBRS stock or are considering a position: Watch the cloud/services mix ratio in Q2 and Q3 earnings more than top-line revenue. If cloud gross margins hold above 50% and cloud revenue crosses 55% of total, the margin recovery thesis gains real ground. If hardware remains dominant and gross margins compress below 36%, the infrastructure identity wins and the stock needs to re-rate toward infrastructure multiples, which implies significant downside from $168.

If you work in AI infrastructure procurement: Cerebras's AWS partnership and the disaggregated inference architecture it enables are worth testing now, because the speed advantage on decode-heavy workloads is real and measurable and the pricing model (cloud-based, per-token) removes the commitment risk of buying hardware. Request benchmark access through inference.cerebras.ai and run your own latency tests against GPU-based alternatives on identical model families before locking into multi-year GPU commitments.

If you follow semiconductor valuations broadly: Cerebras is the clearest test case yet for whether vertically integrated AI chip-to-cloud companies deserve semiconductor multiples or infrastructure multiples. The answer will set pricing expectations for every startup in the space, from Groq to SambaNova to d-Matrix, all of which face the same strategic choice between selling silicon and selling compute.

The Bottom Line

Cerebras built something genuinely remarkable. A single chip that fills an entire 300-millimeter silicon wafer is a feat of engineering that should not be dismissed because the business model surrounding it is complicated. But engineering achievement and business model clarity are different things, and right now the market is paying a semiconductor premium for a company whose financial signature increasingly resembles a cloud infrastructure provider with one very large customer who is also its banker. The next four quarters will determine whether margin compression is a growing pain or a permanent identity. The $40 billion gap between chip-company pricing and infrastructure-company pricing is where that question lives, and Q1's 68% capex-to-revenue ratio just made it a lot harder to avoid.