Thesis. A genuine architectural bet on inference latency with $24.6B of contracted backlog and OpenAI committing 750MW — but 86% of revenue comes from two customers, gross margin is guided down 10 points, and five times the IPO float unlocks by November.
Reviewed 19 Sep 2026, after Q1 2026 — the company's first public quarter. ⚠️ IPO'd 14 May 2026 at $185, so every conclusion here rests on a very thin filing record.
What it does
Cerebras builds wafer-scale AI computers. Instead of dicing a silicon wafer into hundreds of chips, it uses the entire wafer as one processor — the Wafer Scale Engine — paired with on-chip SRAM rather than external HBM. The pitch is to eliminate the interconnect and memory-bandwidth bottlenecks that force GPU clusters to talk to each other over networks, producing dramatically lower inference latency. Fabless via TSMC, system integration via Flex. 784 employees, Sunnyvale.
This is not a GPU knock-off. It is a real engineering bet, and inference is NVIDIA's weakest flank — training is where the CUDA moat is deepest.
One number governs how to read everything else. RPO is $24.6B, but management guides only ~15% of it to be recognised across 2026 and 2027 combined, roughly $1.85B annualised. The headline and the near-term P&L are very different things, and management named datacentre capacity as the constraint on conversion.
A frontier lab choosing you for production inference at 750MW is genuine validation. Two customers being 86% of revenue is genuine risk. Both are true and neither cancels the other.
Bull case
- $24.6B remaining performance obligation — contracted, not aspirational. Roughly 48x 2025 revenue
- The OpenAI deal is transformational: >$20B multi-year, with OpenAI planning to deploy 750MW of Cerebras inference compute. About 80% of the RPO
- Q1 2026 revenue $193.4M GAAP, core revenue $191.3M, +92% YoY — beating consensus by 7%. EPS −$0.04 against −$0.14 expected
- Core gross margin 46.5% (from 42.1%); operating loss narrowed 82% to −$3.5M, close to breakeven
- Hardware $111.6M (+60%); cloud and services $79.8M (+167%) — the faster-growing and more strategic line
- Distribution is arriving: AWS pairing Trainium3 for prefill with CS-3 for decode, revenue from 2027; a CrowdStrike partnership; AMD pairing Helios with the Wafer-Scale Engine for low-latency inference via Cerebras Cloud in H2 2026; Meta referenced as a partner
- $6.4B of IPO cash — fully funded for the European buildout (200MW by end-2027) plus US capacity
Bear case
- 86% of revenue comes from two entities — OpenAI and G42. The backlog is presumably concentrated the same way. This is the whole risk in one number
- The G42 history has not gone away. The 2024 IPO attempt was pulled after a CFIUS review of the G42 relationship (>80% of H1'24 revenue, UAE, historical Huawei ties). CFIUS cleared the 2026 deal, but concentration was replaced, not reduced
- Gross margin guided sharply down — core GM from 46.5% to 36–38% in Q2, narrowing further through 2026, with operating margin guided to −30% to −32%. The stock fell ~20%. The explanation involves renting equipment back from a large customer, which is not a reassuring shape
- The CEO went on CNBC the next day to say the guide had been "misunderstood." Re-explaining guidance after a drawdown is a communication failure at minimum
- ~57x FY26 revenue guidance for a company with a shrinking gross margin and a −31% operating margin guide
- The lockup is a wall: ~171M shares — about 5x the IPO float — become tradeable by 9 November 2026, with over 60M unlocking around the Q2 print. Pre-IPO holders bought between ~$0.85 and $107
- Sole-source TSMC with no long-term allocation agreement; 25kW and ~$3M per node limits deployable footprint
- A securities-fraud investigation was announced in July — routine after a 20% drop, no wrongdoing established
Major customers
- OpenAI — >$20B multi-year, 750MW planned deployment, ~80% of the $24.6B RPO
- G42 (UAE) — the other half of the 86% concentration, and the CFIUS history
- AWS — partnership pairing Trainium3 with CS-3; revenue from 2027
- AMD — Helios paired with the Wafer-Scale Engine for ultra-low-latency inference, H2 2026
- CrowdStrike — enterprise partnership
- Meta — referenced as a partner
Read-through. Cerebras is a customer of TSM and a competitor to NVDA and AMD in accelerators and to AVGO and MRVL in custom silicon — while also being an AMD partner through the Helios tie-up, which is worth holding both ways. Its datacentre buildout drives demand for VRT and the power names. ⚠️ Note it is a bottleneck buyer, not a bottleneck owner — downstream of TSMC wafer allocation and datacentre power, both of which management named as its own constraints. The most useful cross-read is that Cerebras and AMD are both attacking inference specifically, which says something about where NVIDIA's position is contestable and where it is not.
What would change the view
- Core gross margin trajectory — 46.5% guided to 36–38%. Whether it stabilises above 40% is the thesis
- The 9 November 2026 lockup expiry — ~171M shares, roughly 5x the IPO float
- Any named third anchor customer outside OpenAI and G42 — the single thing that would most change the risk profile
- AWS volume disclosure, with revenue from 2027
- European datacentre capacity on schedule for late 2026
- Any modification to the OpenAI commitment
- RPO conversion running ahead of or behind the ~15%-over-two-years guide
Update log
19 Sep 2026 — Converted to the standard template, and a Stack Layer assigned: Semi Components. The page previously had none, and was described only as "Bellwethers-adjacent — no clean category fit." Semi Components is the closest existing home, since Cerebras designs silicon rather than operating infrastructure — but it is a genuinely imperfect fit, and if an AI Systems category is ever created for SMCI and CLS, this belongs there instead. Also added the AMD Helios partnership, disclosed on AMD's own Q2 call and missing from this page.
Q1 2026 — First public quarter. Revenue $193.4M GAAP, core revenue $191.3M (+92%), beating consensus by 7%; EPS −$0.04 against −$0.14. Core gross margin 46.5%; operating loss −$3.5M, narrowed 82%. Hardware $111.6M (+60%), cloud and services $79.8M (+167%). RPO $24.6B. Q2 core gross margin guided to 36–38% and operating margin to −30% to −32%; the stock fell ~20%.
14 May 2026 — IPO at $185.
Research and education only — not investment advice.
Made with Bullet