Cerebras CEO: Look at all our orders…

Investors: You mean “sales” right?

Cerebras CEO: (pretends the Zoom froze) 

Cerebras shares got sent to the AI dungeon (-16%) after investors read the fine print underneath the “booming” cloud business and discovered that fewer customers are buying its enormous AI machines.

Sure, revenue still jumped 74% year over year to $180.1 million. But analysts expected $194 million, and Cerebras’ hardware business picked a terrible quarter to start moving backward.

Turns out taking orders and actually making sales are two very different things. Case in point: hardware revenue dropped 23% to $54.1 million as customers struggled to take delivery of the company’s enormous wafer-scale processors.

If you’re unfamiliar with Cerebras, the company’s entire pitch is that normal computer chips are for peasants.

Instead of slicing a 300-millimeter silicon wafer into hundreds of smaller processors like the industry has done for decades, Cerebras keeps nearly the whole thing intact and turns it into one BIGLY AI engine.

Its latest “Wafer-Scale Engine” crams trillions of transistors into a processor the size of a dinner plate, allowing AI models to haul ass without repeatedly passing data back and forth across a warehouse full of separate chips.

The downside is that installing one requires considerably more effort than sliding it into the office Dell and restarting Windows. Hence why so many orders… not so many verified sales.

CEO Andrew Feldman said that mismatch will keep hardware sales “lumpy” because data-center space isn’t always ready when the company finishes building its systems. 

Meaning the company created a chip so large and powerful that some buyers apparently need to renovate the house datacenter before bringing it home. (Not exactly an ideal situation, amirite?).

The good news is that Cerebras found a way to monetize customers who want the horsepower but don’t have room in the garage. 

Cloud and services revenue nearly quadrupled to $126 million as companies rented access to its processors rather than installing the ginormous systems themselves. 

That may eventually become the more important business anyway.

Rather than convincing every customer to buy, house, and maintain its specialized hardware, Cerebras can install the machines in its own data centers and sell computing power by the token.

It’s essentially the difference between selling people commercial pizza ovens and opening Domino’s (or if you hate their greasy crust, Papa John’s).

Cerebras believes customers will pay premium prices for inference speed, which determines how quickly a trained AI model responds after receiving your prompt. 

Speaking of prompts, the company recently landed a YUGE compute agreement with Scam Altman (read: OpenAI) and is preparing to run GPT-5.6 Sol at speeds of up to 750 tokens per second. Cerebras has also struck agreements involving Amazon and AMD as it attempts to loosen Nvidia’s stranglehold on AI infrastructure.

But as we all know, feeding the insatiable AI beast isn’t cheap.

Cerebras posted a $450.5 million net loss for the quarter, compared with a $309.5 million profit a year earlier. Although about $377 million of the ass-kicking came from stock-based compensation, because historically speaking… ringing the Nasdaq bell also triggers an executive-stock pinata. 

So just to reiterate, Cerebras didn’t dump nearly half a billion dollars into the office shredder. It just celebrated going public by making it rain stock on employees while shareholders waited outside holding the dilution bag.

Meanwhile, GAAP gross margin fell to a malnourished 14%, although the adjusted core figure looked much healthier at 41%. Management blamed some of that pressure on renting additional computing capacity while its own data-center footprint catches up with demand.

But have no fear, Cerebras expects the financial wedgie to loosen as new capacity comes online and component costs fall. 

For the third quarter, management guided core revenue to between $214 million and $216 million, just a smidge above expectations.

Then CFO Bob Komin whipped out the kind of growth target normally found inside a gas-station boner-pill advertisement.

Cerebras plans to more than triple revenue in 2027.

How, you ask? 

Well, the company already has $25.4 billion in remaining performance obligations, largely supported by its enormous OpenAI relationship. If Cerebras can install enough capacity, turn that backlog into recognizable revenue, and improve margins along the way, last night’s selloff could eventually look spectacularly stupid.

But those are three very large “ifs.”

Cerebras has proven there are buyers hungry for faster AI computing. Now it needs to prove it can deliver that speed profitably and at enormous scale. Until then, nobody’s buying the earnings-call theater… no matter how many times management screams “$25.4 billion backlog” into the computer microphone. 

At the time of publishing this article, Stocks.News holds positions in Amazon as mentioned in the article.