Cerebras Cloud vs Hardware: Two Businesses Inside One Company, One Up 265% and One Down 29%

Why split it ​

Cerebras reports two sets of numbers, GAAP and Core (non-GAAP). Look only at total revenue and Q3 growth of 55.3% reads as steady acceleration. Split it by line and two curves move in opposite directions.

That structural shift is not unique to Cerebras — moving from selling machines to selling compute is an industry-wide direction — but Cerebras is the first to make the shift this complete on its statements: cloud went from under 30% to nearly 70% in a single year.

The table: two quarters, four lines ​

Difference one: the growth rates reversed cleanly ​

Hardware went from $96.79 million to $68.88 million, a decline of $27.91 million; cloud went from $38.92 million to $141.93 million, an increase of $103.0 million. Net, total revenue rose $75.10 million.

This is not "hardware flat, cloud filling in the rest." Hardware declined in absolute terms while cloud supplied all of the increase and covered the hardware gap as well. On the Core basis hardware fell only 11.2%, indicating part of the decline is warrant amortization accounting rather than unit sales.

Difference two: the margin structure inverted ​

This is the most easily missed and most important item of the quarter.

Last year the structure was "high-margin hardware, low-margin cloud": hardware at 50.4%, cloud at 16.2%. This year it is "low-margin hardware, high-margin cloud": hardware at 11.0%, cloud at 24.3%. The two crossed, and blended gross margin fell from 40.6% to 20.0%.

On the Core basis the inversion is clearer: hardware gross margin 31.0% (down from 50.5%), cloud gross margin 46.9% (up from 16.5%). Cloud is now 15.9 percentage points more profitable than hardware, compared with a 34-point gap in the opposite direction a year ago.

Why the inversion? The company does not explain directly, but two threads point the direction:

• Cloud builds capacity and then sells tokens; unit costs fall as utilization rises, so margins expand with scale.
• Hardware is in a capacity ramp (manufacturing capacity exiting 2026 is more than 10x the start of the year), so fixed manufacturing costs are not yet diluted.

Both are characteristics of a ramp phase, not an end state.

Difference three: revenue recognition timing is entirely different ​

Hardware is recognized on delivery, one-time. Cloud is recognized on usage or over a contract term, spread out.

Two consequences follow:

First, quarterly volatility is amplified. The company itself warns that "revenue mix can vary significantly from quarter to quarter depending on the timing of large cloud-capacity additions and hardware shipments." That explains hardware dropping from $96.79 million to $68.88 million — a delivery schedule effect, not vanished demand.

Second, predictability improves. With cloud at 67.3% of revenue, the revenue base carries a larger stock component, and forward indicators such as RPO of $2.61 billion matter more.

Difference four: capex is the financial-model consequence ​

A hardware seller does not build data centers; a compute seller does.

That shows up directly in the cash flow statement: nine-month purchases of property and equipment were $1.046 billion against $301 million a year earlier, a 3.5x increase. Net property and equipment rose from $437.4 million to $1.508 billion over the same period.

In other words, cloud growth is not free: it buys a higher gross-margin ceiling at the cost of a heavier asset base. The underlying issue is the same one CoreWeave and Nebius face — Cerebras just operates at an order of magnitude smaller scale.

Difference five: the answer is in the guidance ​

Q4 guidance provides the test: Core revenue of $310-330 million, Core gross margin of 47-49%, Core operating margin of -1% to +1%.

Gross margin rising from 41% in Q3 to 47-49% implies two things at once: a higher blended margin as cloud utilization improves, and easing cost pressure from the hardware manufacturing ramp.

If Q4 lands in the upper half of guidance, the structural inversion is not a one-off but a new baseline.

Conclusion: not a good-or-bad question, but a timing one ​

Three judgments:

First, the cloud shift is directionally settled. Cloud at 67.3% of revenue, Core cloud gross margin at 46.9%, Q4 Core gross margin guided at 47-49% — three numbers pointing to the same conclusion: this company's profit path now rests on cloud.

Second, hardware does not disappear, but its role changes. Hardware revenue is still $68.9 million GAAP ($86.0 million Core), a real business line and the entry point for customers into the Cerebras ecosystem. Its job is customer acquisition, not growth.

Third, the profit inflection is a Q4 question. The -1% to +1% Core operating margin guide is the single number most worth watching. It determines whether "can cloud conversion produce profit" gets an answer inside 2026.

What cannot be compared directly ​

• GAAP and Core bases cannot be mixed: GAAP blended gross margin 20.0%, Core 41.0%, same quarter;
• "Hardware gross margin from 50.4% to 11.0%" is GAAP; the Core basis is 50.5% to 31.0%, the difference coming from warrant amortization and stock-based compensation;
• Revenue mix swings widely by quarter on delivery timing, so a single quarter's share cannot be extrapolated linearly;
• Customer concentration is not disclosed; hardware and cloud customer bases may overlap heavily and cannot be assessed separately;
• This article makes no valuation judgment on Cerebras and no forecast for its next quarter.

Sources ​

• Cerebras press release: Cerebras Systems Delivers Record Revenue in Third Quarter 2026, Raises Full-Year Outlook (November 12, 2026)
• Related on this site: Cerebras Q3 2026 financial data, Cerebras WSE vs GPU inference speed comparison