Cerebras shares dropped 14% in after-hours trading Tuesday despite raising full-year revenue guidance and posting cloud revenue that crushed estimates. This is the second consecutive quarter the stock’s been pummeled after earnings, and the reason is identical. Margins compressed by a temporary accounting reality that the market keeps misreading as structural decay. If you’re trading Cerebras on GAAP gross margin alone, you’re betting against a construction schedule, not a business model.
The Rental Trap Is Eating Margins, Not Competition
Cerebras reported core gross margin of roughly 41% in Q2, down from about 47% in Q1. The headline GAAP figure looked even uglier at 14%. Most traders saw that sequence and hit sell before reading the footnotes. But the CFO quantified the real culprit on the call. The company had to rent third-party data center capacity to meet exploding cloud demand while its own facilities are still coming online. That rental cost shaved approximately 500 basis points off core margins.
This is a classic infrastructure bottleneck, not a pricing problem. Cerebras has over 600 megawatts of live or contracted capacity in the pipeline and is targeting a more than 10x manufacturing expansion this year. The issue is that owned capacity ramps in Q4 and beyond. Until then, they’re effectively subletting space from hardware customers to serve cloud workloads. The accounting treats these as costs of revenue, which creates a double-counting illusion that makes the cloud business look artificially expensive.
I spent time yesterday walking through the margin breakdowns circulating among institutional accounts, and the numbers tell a clear story. Once Cerebras stops renting and starts operating its own racks, core gross margins should march back toward 60% long-term targets driven by higher token pricing and wafer-scale throughput that avoids HBM bottlenecks entirely.
The market has seen this movie before. Back in June, the stock tanked after Q1 results on a similar margin warning. The pattern is obvious. Investors demand perfection from a company that’s still building out physical power and manufacturing capacity. You can’t depreciate a data center that doesn’t yet exist.
The Real Test Is Converting OpenAI Momentum Beyond the Lump
The other sore spot was hardware revenue at $54.1 million, which declined year-over-year and missed estimates. That sounds alarming for a chip company until you realize Cerebras ships multi-million-dollar wafer-scale engines only when the customer’s data center has actual power and cooling ready. Several late-stage deals are waiting on utility interconnections and permitting delays across the US, Canada, and Europe. The demand is there. The grid is not.
I watched the post-earnings threads split into two camps instantly. One side sees the raised guidance to $880-890 million as proof the model works. The other side points to $502.8 million in operating expenses against $89.3 million a year ago and asks how long the cash runway lasts before another equity raise. Both arguments have merit, but the louder panic yesterday came from a thin public float getting whipsawed by headline numbers that don’t reflect actual bookings health.
The concentration risk is harder to dismiss. Cloud and services revenue hit $126 million, well above expectations, powered largely by OpenAI. When one customer drives this much of your cloud growth, the $25.4 billion remaining performance obligation is both a trophy and a vulnerability. Converting that into recognized revenue depends on solving power bottlenecks and scaling manufacturing simultaneously without further dilution.
The architectural validation happening behind the noise is what matters more. Cerebras is integrating with AMD’s Helios systems for heterogeneous prefill-decode inference and laying groundwork with AWS Bedrock for 2027. These partnerships validate that specialized inference architectures have a place alongside hyperscale GPU clusters. The company is also positioning distributed inference sites in secondary power markets, monetizing stranded megawatts that big cloud players ignore. That’s a genuinely different footprint from the gigawatt-scale contiguous campuses everyone else is chasing.
Cerebras isn’t a GPU story. It’s a power-and-capacity story wearing a semiconductor ticker. The stock won’t stop gyrating until owned data centers come online and the GAAP accounting cleans up in late 2026. If you believe AI inference demand is through the roof, as CEO Andrew Feldman stated on the Q2 earnings call, then the reaction looks less like a warning and more like a mismatch between quarterly investor attention spans and multi-year infrastructure buildouts. I wouldn’t be shocked to see this same script play out one more time before the year ends. The only question is whether you have the stomach to hold through the construction noise.
This cycle feels similar to what we saw with Apple’s supply chain crunch, where success itself created temporary scarcity. Or the pre-profitability ramp at SpaceX during its early Starlink buildouts. The market loves infrastructure winners after they’re built. It hates them while they’re building.
