Cerebras nearly doubled its revenue and guided 2026 sales above Wall Street’s estimates. Its stock still fell about 10%. The reason was a margin squeeze, caused not by a shortage of chips but by a shortage of buildings.
Cerebras just learned the price of going public into a mania. On Tuesday the AI chipmaker posted its first results since a blockbuster May listing. Revenue almost doubled. The full-year forecast came in ahead of analysts. And the shares dropped roughly 10% anyway.
The numbers were good. First-quarter revenue hit $193.4m, up 92% on a year earlier, CNBC reported. The net loss narrowed to $14m from $23.9m. Both lines beat what Wall Street had penciled in. Cerebras even guided full-year revenue to between $855m and $865m, comfortably above the $824.8m analysts expected.
So why the sell-off? Because the story investors cared about was not demand. It was margins. And the cause of the margin problem is the most telling part of all.
A grand irony #
Cerebras warned that its core gross margin will fall to between 36% and 38% this quarter, down sharply from 46.5% in the first. That is a steep drop for a company sold to investors as a high-margin chip story. The reason is almost poetic.
“It’s a grand irony that after all this technology that we’ve invented, and Nvidia’s invented, buildings are the limiting factor,” chief executive Andrew Feldman said. Cerebras cannot get enough data-centre space to house its chips.
So it is renting back some of its own systems from a customer and building out its own capacity at speed. Those costs will shave 10 to 15 points off margins this year, finance chief Bob Komin told analysts.
This is the bottleneck almost no one priced in. The constraint on AI is shifting from silicon to concrete and power. Even the company with the world’s largest chip cannot find enough plugged-in floor space to run them. US utilities now plan to spend $1.4 trillion by 2030 to feed this demand, and the wait for power is exactly what is eating Cerebras’ margins.
Why Wall Street was so unforgiving #
A beat that gets punished tells you something about the mood. Cerebras went public in May at $185 a share. The stock opened at $350 and closed its first day above $311. It has since slid hard, ending Tuesday at $226.72, down about 28% from its peak.
The listing itself was historic, raising more than $6bn. That made it the biggest semiconductor IPO ever, and the largest US tech debut since Uber in 2019.
The timing did not help. Cerebras reported into a brutal day for chip stocks. The Philadelphia Semiconductor Index fell 7.9%, a rout one analyst called a “chip-wreck”, with Micron leading the broader sell-off. When sentiment turns on the whole sector, a single good quarter struggles to swim against it.
There is a deeper expectations trap, too. Investors have grown used to Nvidia and a handful of peers blowing past estimates every quarter. That has set a brutal bar. For a newly public Nvidia challenger, merely beating is not enough. Anything short of spectacular reads as a disappointment.
The business under the share price #
Strip out the market noise and Cerebras is still a striking company. It builds the Wafer Scale Engine, a single chip the size of a dinner plate, and sells it on one promise: the fastest AI inference in the world.
Inference is the work of running models for users, rather than training them. It is the niche Cerebras chose, and its pitch is to partner with everyone except Nvidia.
The customer list backs the ambition.
In January, Cerebras signed a deal with OpenAI to supply 750 megawatts of inference capacity, worth more than $20bn at full expansion, plus a $1bn working-capital loan. It is also pairing with Amazon Web Services on a split approach, where AWS chips handle one half of a query and Cerebras runs the fast part.
Group 42 and an Abu Dhabi AI university round out a concentrated roster of big names. The mix is shifting, too. Cloud and services revenue, the part that runs models for customers, jumped 178% on the year, a sign demand is moving toward exactly the inference work Cerebras is built for.
The product story is real, too. Cerebras co-launched Codex-Spark, a coding model tuned for near-instant responses, which it says runs at more than 1,000 tokens per second. Speed is the whole sell. Fast AI, Feldman likes to argue, is simply more useful than slow AI, and worth paying for.
The case for caution #
The risks sit in plain sight. Cerebras leans heavily on a few customers, OpenAI above all, so any wobble there would land hard. Its IPO ranked among the largest in chip history, which raised the bar it must now clear quarter after quarter. And the capacity build that is denting margins will keep doing so while it races to catch demand.
The competitive picture is crowded as well. Cerebras is one of many firms chasing the inference market, the part of AI where rivals see the best opening against Nvidia. Speed advantages can narrow. Customers can build their own silicon, as OpenAI just showed by unveiling a home-grown inference chip of its own.
None of that makes the bet wrong. Revenue is nearly doubling, the backlog is real, and the demand for fast inference is not in doubt. The doubt is about timing and cost. Cerebras is one of a wave of AI names, alongside SpaceX, OpenAI and Anthropic, whose valuations now assume years of flawless execution.
So the open question is not whether people want what Cerebras sells. They plainly do. It is whether the company can grow into its valuation while its single biggest constraint is not its technology but the buildings and power it takes to run it.
For now, the AI chip race has a strange new chokepoint, and even the fastest chip in the world is stuck waiting for the lights to come on.
Get the TNW newsletter #
Get the most important tech news in your inbox each week.