Cerebras (Trans): Q4 guide raised across the board; OpenAI deliveries remain behind schedule

Below is Dolphin Research's digest of $Cerebras(CBRS.US) FY26 Q3 earnings call

I. Key takeaways

1. Outlook (broadly raised)

a. Q4: core revenue of $310–330 mn, up nearly 40% QoQ and ~85% YoY. Core GPM of 47%–49%, core OPM of -1% to 1% (roughly breakeven), with all core profit metrics expected to turn positive.

b. FY26: core revenue of $940–960 mn, core GPM of 42%–44%, and core OPM of -10% to -8%. All three were raised.

c. 2027: revenue still targeted to more than triple, with GPM tracking toward the 60% LT goal. OPM is expected to make clear progress toward the 40% LT goal, and a full 2027 outlook will be provided on the Q4 call.

d. Revenue mix: Q4 cloud revenue is expected to be flat QoQ due solely to data center capacity constraints, not pricing or demand. New DCs will ramp from Dec, with the bulk coming after mid-2027; incremental growth will be driven by new hardware orders (incl. Oracle) beginning to ship, with shipments running at least through 1H27.

2. Margin inflection pulled forward by one quarter

a. Total GPM: Q3 core GPM was 41%, up 40 bps QoQ and 200 bps above guidance. The trough was expected in Q3, but actually occurred in Q2; Q4 GPM is guided to jump 700 bps to 48%, and to exceed 50% soon after.

b. Hardware GPM: fell to 31% in Q3 due to the final batch of legacy low-price hardware deliveries, which management emphasized was not a yield issue. From Q4 onward, shipments are at higher current pricing, and LT hardware GPM is expected to exceed the 60% target.

c. Cloud GPM: reached 46.9% in Q3, up 510 bps QoQ, mainly from higher-priced new workloads. Costs from leasing back owned equipment to customers will be a drag at least until mid-2027.

d. OP: Q3 core OPM was -14.3%, improving 170 bps QoQ and nearly 10 pts above guidance mid. Core net income turned positive for the first time at $30 mn, and YTD adj. core EBITDA was $2 mn.

3. RPO and OpenAI contract

a. RPO: $26.1 bn as of Sep 30, with OpenAI representing the majority. This excludes the Oracle contract signed after quarter-end, which is sizable.

b. Delay credits: The OpenAI agreement is a 750 MW take-or-pay; during delivery delays, OpenAI can credit future purchases. The company has conservatively estimated and deducted these credits from RPO, assuming no waivers for external causes; after pro-rata allocation, the impact on revenue and margins is minimal.

c. Termination clause: After a 90-day delay, either party can cancel the delayed capacity; any cancellations potentially occurring this year would be under 5% of the committed MW. Management is confident it can reallocate capacity to other customers, and expects the schedule to return to plan in 2H27.

d. Additional option: OpenAI holds an additional 1.25 GW option with the earliest online date in 2029. The parties have not started discussions.

4. Balance sheet and capital efficiency

a. Liquidity: Cash, cash equivalents, restricted cash, and marketable securities totaled $7.8 bn at Q3-end. The $850 mn revolver remains undrawn.

b. Funding sources: The $1 bn working capital loan from OpenAI in Jan, plus pass-through one-time DC build costs funded by OpenAI, cover most working capital and capex through 2026. This loan is the company's only debt.

c. Capex basis: For in-house systems, the company pays only BOM cost, not the 70% third-party gross margin, and it avoids high-priced HBM, resulting in lower capex per MW. The IPO lock-up fully expired on Nov 9.

II. Earnings call details

2.1 Management commentary

1. Biz. model: parallel channels of system sales and owned cloud, with mix fluctuating by quarter and both set to grow meaningfully. System sales are faster to recognize with faster cash conversion but are lumpy, depending on customer DC readiness; cloud serves AI-native token buyers, and in-house hardware drives far lower $/MW than neo-cloud.

b. From early 2025, the company will lease DCs to expand private cloud, allowing customers to validate on cloud before choosing hardware vs. cloud. It will also lease back a portion of previously sold systems to accelerate supply.

2. Customers: three hyperscalers signed within six months, and OpenAI revenue continued to grow QoQ in Q3. After AWS, Oracle was signed in Q3 (inference expected to go live in summer), and a Microsoft deal was announced today.

b. OpenAI sees very strong demand for Cerebras-based ultrafast model previews, but capacity is insufficient for broad rollout. DC procurement post-signing has been slower than expected, and initial batches are behind; management also expects OpenAI to use multiple hardware types.

c. Distribution is expanding via neo-cloud players such as Fireworks. Cloud customers report that Cerebras fast tokens are their most profitable tokens, and the sales pipeline has expanded more than 10x in recent months.

3. Capacity: DCs are an industry-wide bottleneck, and over 75% of new capacity in 2027 will be outside the U.S. About ~700 gross MW of DC capacity has been secured for delivery by end-2027; sites in Canada, France, and Norway will come online within months, Norway and Finland were signed in Q3, São Paulo is in the final signing stage, and GW-scale capacity is under discussion.

b. Manufacturing is ramping with Flex and Sanmina, taking 2026 year-end capacity to over 10x the level at the start of the year. Another plant will be added in 2027; avoiding HBM also sidesteps the tightest memory constraints, though management expects storage shortages and price inflation to persist into 2028.

4. Yield clarification: management stated both wafer and parametric yields are high, and CS3/CS4/CS5 all support large models. Among ~1 mn identical compute elements, redundancy is reserved and defective elements are bypassed; parameter variability is addressed by trimming wafer edges, wider-corner timing validation, and 84 partitions with independent voltage regulation.

5. Product roadmap: CS4 is entering full mass production with doubled performance and throughput, and WSE-5 returned from fab ahead of plan. WSE-5 was taped out on Jun 15 and reached full-system testing in under five months; CS5 delivers 2x+ CS4 performance, targets mass production in early summer 2027, uses TSMC 5 nm with limited wafer competition and lower cost.

b. Heterogeneous decoupled inference has launched: AMD GPUs handle prefill while Cerebras handles decode, lifting tokens produced per single DC by up to 5x. New model porting now takes days; Kimi K3.1 (2.8 tn parameters) was supported at first release, with peak throughput of up to 2,000 tokens/sec.

2.2 Q&A

Q: What is the size of the newly signed Oracle and Microsoft deals, and how much is in RPO?

A: Details of the Microsoft collaboration will be disclosed in 1H27, with revenue recognition beginning then. The company did not quantify deal sizes. The Oracle contract was signed after Sep 30 and is not included in Q3 RPO.

OpenAI's portion of RPO is based on the committed revenue under the take-or-pay terms, excluding extensions or the additional option and net of estimated delay credits. It also includes pass-through one-time DC build costs and monthly recurring costs for DCs signed for OpenAI, recognized as pass-through revenue and expense.

Q: As frontier labs rapidly deploy models across different hardware, making compute more substitutable, how does Cerebras benefit, and how will heterogeneous infra evolve?

A: The software moat at the low level will narrow, putting more weight on hardware innovation and favoring new architectures. Management cited a frontier lab that reportedly brought a model live on AMD GPUs in about three days.

Compilers and low-level software have historically been the main barrier for new architectures and the incumbent 'software moat'. If this capability proves durable, the value of low-level software will keep shrinking, potentially to zero, weakening incumbents' long-term advantage and warranting close industry attention.

Q: With delays in cloud delivery under the OpenAI contract, how confident are you in resolving issues before they trigger broader penalties, and what caused the delay?

A: RPO already reflects conservative worst-case deductions, and management expects actual outcomes to be better. Delay credits are common in large DC transactions, and the company has similar penalties with its DC builders; whether they are enforced will be a commercial decision at the time.

The delays are entirely about land, power, and DC shells. The company's own system production capacity and all materials are ready, and it is waiting for DC completion and power-up.

Q: When will CS4 overtake CS3 in shipments? Is that in 2027?

A: CS4 is expected to surpass CS3 in Q1 2027, with CS5 shipping in early summer next year. The roadmap is on track; CS5 is ahead of schedule and will begin shipping in early summer 2027, with WSE-5 moving from tape-out to system test in under five months and CS5 delivering 2x CS4 performance.

Q: Given CS4's large throughput uplift, can a Cerebras system run standalone without GPUs?

A: Systems already run standalone; until weeks ago, all deployments were CS3-only. The company only recently began deploying and formally launching the decoupled solution, which materially raises throughput. When the supply of high-speed tokens is constrained, decoupled designs yield more fast tokens, though management did not directly say GPUs would be unnecessary LT.

Q: With cloud revenue flat QoQ in Q4, how will cloud vs. hardware split in the plan to triple 2027 revenue?

A: Hardware will be close to 50% of revenue in Q4 and Q1 2027, above prior quarters. Single hardware deals can be very large, and shipments depend heavily on customer DC readiness, so management only provides directional color for the next one to two quarters; the 2027 full-year mix and outlook will be given on the Q4 call.

Q: Are delay credits just estimates for now, and have you received any cancellations due to delays?

A: Confirmed: there have been no cancellations due to delivery delays to date.

Q: The market worries low-batch GPUs can also do fast inference. How does Cerebras compare on per-token cost?

A: At 250–300 tokens/sec and above, Cerebras is more economical, with advantages growing at higher speeds. GPUs raise speed by lowering batch size, which slashes throughput and the number of concurrent customers served, pushing up cost per token; even batch size of 1 cannot match Cerebras speed.

Management said Cerebras has a large economic lead at 500–800 tokens/sec, and GPUs are largely non-competitive in the 800–2,000 tokens/sec range.

Q: If wafer and rack supply are fine, are land, power, and DC shells the bottleneck? How much room is left to sell hardware directly, and can the OpenAI contract pivot to hardware?

A: Land, power, and DC construction are indeed bottlenecks, and customers' DCs are also widely delayed. TSMC has committed sufficient wafer capacity for 2027, and avoiding HBM sidesteps the tightest memory bottleneck; the company is shipping to customers with ready DCs, but management understands OpenAI and hyperscalers are facing widespread DC delays globally.

The OpenAI agreement does allow conversion to hardware purchases, but its DCs are also behind schedule.

Q: How will Oracle and Microsoft bring Cerebras to their customers? Will it be mainly hardware sales scaling with demand for ultrafast inference?

A: Oracle is procuring hardware for its U.S. DCs as part of OCI cloud services. Details with Microsoft are not disclosed for now. The parties are jointly building the pipeline with strong customer interest, and management said this is not the last deal but the start of a large-scale partnership.

Risk disclosures and statements:Dolphin Research Disclaimer and General Disclosure