Cerebras Has a Real Moat and a Real Problem: Great Silicon, a Two-Customer Revenue Base
Cerebras is two companies wearing one ticker. One is a genuinely differentiated silicon engineering shop with an architectural edge that the rest of the industry is now scrambling to copy. The other is a revenue base so concentrated that until roughly two months ago it had, functionally, two customers — both in Abu Dhabi. Any honest read of CBRS has to hold both at once, and the second one deserves far more weight than the momentum around the stock suggests.
The Edge Is Real
Start with what works, because it is not nothing. Cerebras builds wafer-scale processors — single chips the size of a dinner plate — carrying 44 GB of on-chip SRAM and delivering roughly 21 petabytes per second of memory bandwidth. For inference, which is bottlenecked by how fast you can move model weights out of memory rather than by raw compute, that translates to claimed speedups of up to 15x over a top-end GPU.
The strategic elegance is what the design avoids. CEO Andrew Feldman’s pitch is that the binding constraints in AI hardware today — HBM supply, TSMC’s CoWoS advanced-packaging queue, leading-edge process capacity — are constraints Cerebras sidesteps entirely. It uses SRAM printed on its own logic wafer, whose supply Feldman describes as effectively infinite, rather than the scarce, expensive HBM that Micron and SK Hynix ration out. When memory prices spike, Cerebras’s cost structure barely notices. That is a durable advantage, and the fact that Nvidia bought Groq to bring a comparable SRAM-centric inference architecture in-house is the strongest possible validation that the approach matters.
The Problem Is Also Real
Now the part the bull case skips. In 2024, Cerebras was effectively a single-customer company: G42, the Abu Dhabi AI group, was 85% of revenue. The IPO story sold declining dependence — and on G42 specifically, that was true, its share falling to 24% in 2025. But the gap did not fill with a diversified book of enterprise business. It filled with a second Abu Dhabi entity. Mohamed bin Zayed University of Artificial Intelligence accounted for 62% of 2025 revenue. Add them together and roughly 86% of 2025 revenue came from two UAE-affiliated, related buyers. The remaining 14% is a fragmented base of smaller government, enterprise, and cloud customers, none large enough to move the needle.
This is the crux of the suspicion, and it is the correct suspicion. A company selling a supposedly general-purpose inference breakthrough should be winning a broad, repeatable book of customers the way Nvidia and AMD do. Cerebras instead sells a handful of large, project-based deployments to state-aligned sovereign buyers whose procurement tracks national AI strategy, not competitive product wins in the open market. MBZUAI’s purchases run on multi-year capital cycles tied to UAE policy — durable, but concentrated in exactly the way that makes a revenue base fragile. The 76% revenue growth from 2024 to 2025 was overwhelmingly UAE-driven, not evidence of commercial breadth.
The OpenAI Substitution Doesn’t Solve It
The counterargument is that OpenAI changes everything: a multi-year agreement announced in January valued at more than $20 billion, anchoring a $24.6 billion backlog. That is real and it is large. But look closely at what it does to the risk profile. It does not diversify the revenue base — it swaps one concentration for another. Instead of depending on two related sovereign entities, Cerebras becomes dependent on a single hyperscale counterparty whose own strategic direction, compute-sourcing decisions, and competitive relationships are in constant flux. A deterioration in the OpenAI relationship would hit Cerebras’s revenue directly and proportionately. Concentration in a $20 billion customer is still concentration.
And crucially, OpenAI contributed $0 to 2025 revenue and only began contributing in February 2026. The AWS relationship — the one piece of genuinely mainstream diversification, putting Cerebras silicon in front of ordinary cloud customers via Bedrock — isn’t expected to show up financially until 2027. So the diversification thesis is a 2027 event being priced into a 2026 stock. Genuine breadth won’t be verifiable until then.
What You’re Actually Paying For
The valuation makes the concentration matter more, not less. After its May debut, Cerebras traded around 90x trailing revenue, and above 50x even on the forward guidance midpoint of ~$860 million. The company posted a $146 million operating loss in 2025 and guided Q2 gross margin down to 36–38% from Q1’s 47% — the margin hit stemming from leasing hardware back to deploy OpenAI capacity, a sign the business model is shifting from selling chips to renting compute, which is more capital-intensive and slower to prove out.
So the buyer at these levels is underwriting three bets simultaneously: that the SRAM architecture stays ahead of Nvidia’s Groq-powered inference push, that the OpenAI backlog converts on schedule, and that AWS plus a real enterprise book materializes to turn a sovereign-dependent supplier into a diversified one. The engineering bet is the one I’d be most comfortable making. The customer bet is the one that justifies your suspicion.
The Bottom Line
Cerebras is a genuinely impressive engineering company wrapped around an uncomfortable customer list, priced for a diversification that hasn’t happened yet. The moat is real: SRAM immunity to the HBM crunch is a structural advantage that grows more valuable the tighter memory supply gets. But a moat around a revenue base that is 86% two related buyers is a moat around a narrow keep. The single most bullish thing Cerebras could report in any coming quarter is not another OpenAI headline — it’s a hard revenue number attached to a US or European enterprise customer nobody in Abu Dhabi controls. Until that line appears, the skepticism isn’t cynicism. It’s just reading the filing.