Cerebras (CBRS.US) Q3 Earnings Call: CEO Acknowledges OpenAI Delivery Delays, Raises Guidance and Expects Revenue to Triple by 2027

Cerebras Systems (CBRS.US), an AI chip and cloud service provider, reported fiscal 2026 third-quarter results after the U.S. market close on November 12. Core revenue for the quarter rose 70% year over year to $230.9 million, above the $214.7 million analyst average estimate compiled by LSEG; adjusted net income was $29.82 million, compared with a loss of $10.86 million a year earlier, while the market had originally expected a loss of $38.4 million. Fourth-quarter core revenue guidance is $310 million to $330 million, with the midpoint about 16% above the $274.9 million market estimate; full-year core revenue guidance was raised from $880 million to $890 million to $940 million to $960 million. The real focus of the call landed on the OpenAI contract underpinning $26.1 billion in remaining performance obligations (RPO): due to data center delivery delays, the company has deducted potential delay compensation from RPO. Co-founder and Chief Executive Officer Andrew Feldman acknowledged that the company is behind schedule on the first batches of capacity deliveries, attributing the cause to land, power and building construction rather than chips or materials. He reiterated that revenue will grow to about triple this year's level in 2027. Focus and response: OpenAI contract delays, RPO has accrued compensation The most noteworthy item in the meeting was not the already better-than-expected third-quarter report, but the OpenAI contract that accounts for the bulk of RPO. Chief Financial Officer Bob Komin disclosed that the contract is a 750-megawatt take-or-pay agreement, under which OpenAI has the right to receive credits usable for future committed purchases during any period of delayed delivery; the company has estimated the potential delay compensation amount and deducted it from RPO as of September 30. If delivery delays persist for 90 days or more, either party may choose to cancel the corresponding portion of capacity, and the company estimates that the scale that could be canceled this year is less than 5% of total committed megawatts. Komin said the company accrued on a conservative "absolute worst case" basis and expects to return to the delivery schedule in the second half of 2027, thereby reducing future delay compensation risk. Feldman added that the current delays are "entirely from land, power and buildings," and that "capacity and materials are all in place," with the company waiting for power and building completion. He also said that because data center delays are widespread globally, the company has sufficient demand to take on capacity even if cancellations occur. Faced with Morgan Stanley's Joseph Moore pressing on the causes of the delay, Feldman gave the above answer; Needham's Quinn Bolton confirmed that the company has so far not received any cancellation notice for delayed capacity. Q4 guidance well above market expectations The company set fourth-quarter core revenue guidance at $310 million to $330 million, up about 85% year over year and about 40% quarter over quarter at the midpoint; core gross margin guidance is 47% to 49%, and core operating margin guidance is negative 1% to positive 1%, implying the core business is expected to approach break-even in the fourth quarter. Komin said the fourth quarter will be an important inflection point for the company's growth and core financial trends, with the revenue jump mainly coming from hardware orders beginning delivery in the fourth quarter. Cloud revenue flat quarter over quarter in Q4, capacity the only constraint Third-quarter core cloud and other services revenue was $144.9 million, up 272% year over year and 13% quarter over quarter, accounting for about 63% of core revenue. However, the company expects fourth-quarter cloud revenue to be roughly flat quarter over quarter. Komin stressed that this is not a pricing or demand issue"the pricing environment is healthy, and demand still far exceeds supply"and that the cloud revenue ramp is "entirely constrained by data center capacity," with new capacity only beginning to come online in December, too late to be reflected in the fourth quarter. Hardware revenue declines, gross margin trough has passed The hardware business posted third-quarter core revenue of $85.96 million, down about 11% year over year; on a GAAP basis it fell to $68.88 million from $96.79 million a year earlier. Komin explained that the lower hardware gross margin stems from the last batch of low-priced orders under procurement agreements signed years ago, as well as manufacturing capacity still ramping; as new orders signed at current prices begin delivery from the fourth quarter, hardware gross margin will improve significantly. He noted that the core gross margin trough actually occurred in the second quarter, not the third quarter as previously guided. Data center capacity: 700 megawatts, more than 70% outside the U.S. The company has contracted about 700 gross megawatts of data center capacity, planned for delivery by the end of 2027. Feldman said more than 75% of 2027 capacity will come online outside the U.S., with new facilities in Canada, France and Norway coming into operation in the coming months; in the quarter the company signed additional capacity contracts in Norway and Finland and has entered the final signing stage for a data center in So Paulo, Brazilwhich will be its first data center in South America. Three hyperscale customers won in six months In the quarter the company added two hyperscale cloud service providers: Oracle Corporation will deploy Cerebras hardware in its OCI cloud to serve AI labs, enterprises and developers, expected to go live in summer 2027; the other is Microsoft Corporation, with details to be disclosed next year. Together with previously announced AWS, the company has won three hyperscale customers in the past six months. Feldman said this lays the foundation for 2027 and 2028. In addition, Bilt's Neighborhood OS platform and personal assistant are already powered by Cerebras for inference. Product roadmap: CS-4 ramping, WSE-5 back from fab ahead of schedule CS-4 is about to enter full mass production, with both performance and throughput set to double. The fifth-generation wafer-scale engine WSE-5 completed tape-out on June 15 and has now returned from the fab, been powered on and begun running models, and is undergoing full-system testing, taking less than five months from tape-out to system testing. The CS-5 system featuring WSE-5 will be more than twice the scale of CS-4, uses Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR's 5-nanometer process, and is planned for production in early summer 2027. First production-grade disaggregated heterogeneous inference The company says it is the first to deliver production-grade disaggregated heterogeneous inference: currently in its own cloud, AMD GPUs handle prefill and Cerebras systems handle decode. Feldman said the approach can increase high-speed token output per unit of power by up to 5x, making it a key lever for expanding the company's data centers, and he believes disaggregated inference will become one of the default options in mainstream cloud deployments. Responding to yield concerns: defects bypassed through redundancy In response to external questions about wafer-scale chip yields, Feldman said the claims are "all wrong." He explained that the company bypasses defects through a redundant design of about 1 million identical units (tiles), with bad spots switched by hardware to spare units, a mechanism that works both at the fab and in the field; parametric yield is addressed through three measures: removing the wafer edge, expanding design validation conditions, and independently regulating voltage across 84 wafer partitions. He stressed that the lower hardware gross margin reflects capacity ramp and special pricing for early customers, not yield issues, and he expects significant improvement over the next several quarters and ultimately exceeding the long-term target of 60%. $7.8 billion in cash, only debt is the OpenAI revolving loan As of the end of September, the company had $7.8 billion in cash, cash equivalents, restricted cash and short-term investments, plus an unused $850 million revolving credit facility. Komin stressed that the company is "not a neocloud," but a vertically integrated maker of proprietary hardware that does not need to pay a 70% gross margin to third parties and does not use expensive HBM, "only paying material costs"; despite rapid expansion of private cloud buildout, the only debt on the balance sheet is OpenAI's $1 billion revolving loan. Full meeting transcript I. Executive remarks Andrew Feldman (Co-founder, Chief Executive Officer and President): Thank you, Sean, and thank you all for joining today. Looking ahead to the coming months, we see two certainties: rapid change and a shortage of compute. Over the past 90 days, the title of leading frontier model in the AI ecosystem has changed hands multiple times, dozens of interesting open models have been released, major security incidents have occurred, and powerful new consumer agents have emerged. No one can predict AI capabilities 90 days from now, let alone a year from now. But amid all this change, intelligence and speed remain the two core elements of every use casefrom coding to security to personal agents. In such a fast-changing world, Cerebras's mission is: however AI grows, Cerebras grows. Foundation models grow, we grow; open models grow, we grow; personal agents grow, we grow; whether enterprises or startups grow, we grow; whether customers want hardware or cloud capacity, we grow. Our goal is to position ourselves so that whatever drives the industry, Cerebras wins. In the third quarter, we made significant progress toward that goal. First, I want to report that we exceeded all guidance metrics across the board. While I am pleased with these results, I think more important is that since the last earnings call, we have struck a series of deals with key AI industry leaders. The latest is the important cloud partnership with Microsoft Corporation announced today. Together with previously announced AWS and Oracle Corporation, we now have three hyperscale cloud service providers choosing to incorporate the Cerebras platform into their inference strategies. These new deals, together with continued growth and execution with existing customers, allow us to raise fourth-quarter core guidance across the board and continue to expect revenue to triple in 2027 with gross margin expansion. Bob will share all the details. Let me take a moment to explain the shape of our business. As an AI accelerator company, our pioneering innovation is the chips, systems and software that make AI computing fast. We are an AI hardware company with two go-to-market paths. First, we sell systems, installing racks and Cerebras hardware in customers' data centers. This is a good business: fast revenue recognition, fast cash collection; but it is also volatile, depending on whether customers already have data center capacity ready to receive hardware. The second path is through our cloud. There is a fast-growing class of AI-native customers who want to consume tokens rather than deploy hardware. Given the extreme shortage of data centers and compute, owning both systems and data centers is itself a strategic advantage. The AI cloud business is especially good for a company that makes its own hardwareour cost per megawatt is much lower, so we can deliver higher capital efficiency than neoclouds. The revenue mix between the two paths will shift quarter to quarter, but we believe both will grow significantly over the long term. I also want to address speculation about our partnership with OpenAI. OpenAI is a very large customer and a close partner, and I believe it will remain so. OpenAI's usage in the third quarter grew versus the second quarter, and we expect further growth. OpenAI's GPT-5.6 Sol Ultrafast preview based on Cerebras has been an overwhelming success, clearly and unmistakably proving that market demand for fast tokens is enormous. It is equally true that Cerebras does not have enough capacity to support broad rollout of Ultrafast. As others have said, we believe capacity is a key driver of commercial decisions in the industry. In fact, even with its enormous GPU supply, OpenAI still had to impose usage limits. After signing the agreement with OpenAI, it took us longer than we would have liked to sign and bring online the substantially increased data center capacity needed to meet the aggressive requirements in the agreement. As a result, we are behind schedule on the first batches of deliveries. Given the importance of this partnership to our business, we understand investors are watching closely. Details or news about the agreement are unlikely to come from any source other than Cerebras or OpenAI. For example, there have been rumors that Cerebras cannot support certain modelsthese are wrong. CS-3, CS-4 and CS-5 are all fully capable of supporting frontier-scale and even larger models. At the same time, I expect, and you should expect, that OpenAI will build capacity using multiple hardware options. OpenAI will retire old models and deploy new models across these different compute options for internal and external use. Their compute needs are enormous, and there is ample opportunity for different options to participate as they expand. Before getting into the business framework, one more point. Despite announcements from low-batch GPU startups, Meta and hyperscalers' in-house chips, we remain the fastest inference partner. We believe our roadmap will allow us to maintain this advantage for years to come. Our framework for evaluating progress is unchanged. We evaluate the business across three dimensions: customers, capacity and capability. Customers. We are rapidly expanding our customer base to broaden reach and diversify revenue sources. The third quarter was a successful quarter on this front. We scored an important winadding two hyperscale cloud service providers. We signed an agreement to bring Cerebras hardware into Oracle Cloud, with Cerebras to be deployed in Oracle Corporation's data centers to serve AI labs, enterprises and developers building interactive intelligent applications. We expect this Cerebras-powered Oracle Corporation cloud inference service to go live in summer. We also signed an exciting agreement with Microsoft Corporation, with more information to be shared next year. This brings us to three hyperscale customers won in the past six months. This is a significant achievement that lays a good foundation for 2027 and 2028. We are also expanding reach through multiple neoclouds, including high-caliber customers such as Fireworks, thereby diversifying market paths and giving customers more ways to access Cerebras high-speed inference. We also benefit from open-weight models. For example, just this week we announced that Cerebras fast inference is powering the Bilt platform and personal assistant, bringing ultrafast inference to millions of people. With consumer apps like Bilt, we put Cerebras capability directly into people's pockets. That is why fast tokens command a premiumour cloud customers tell us fast tokens deliver excellent gross margins, and Cerebras fast tokens are the most profitable tokens they sell. Capacity. Capacity has two parts: data center capacity and manufacturing capacity, both critical to growth. Data center availability remains a constraint across the industry, and we are no exception. In addition, we believe changes in the U.S. policy environment may pose greater challenges to U.S. data center development. In 2027, more than 75% of our data center capacity will come online outside the U.S. In the coming months, new facilities in Canada, France and Norway will come into operation. In the quarter, we signed additional capacity contracts in Norway and Finland and have entered the final signing stage for a data center in So Paulo, Brazil, our first data center in South America. To date, we have locked in about 700 gross megawatts of data center capacity, to be delivered by the end of 2027. Our first-party cloud is already one of the larger AI clouds among non-hyperscalers, and we are well along in negotiations for multi-gigawatt-scale capacity. The best way to increase data center capacity is to increase the token generation capability of the hardwaremore tokens per megawatt. Disaggregated inference plays a role here, increasing the number of tokens a data center can deliver by up to 5x. On manufacturing and supply, we are working with Flex and Sanmina to expand capacity, and we expect manufacturing capacity by the end of 2026 to be more than 10x the start of the year. In 2027 we plan to add plants and further increase capacity. Our supply chain is well positioned. Most importantly, we have Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR, which continues to demonstrate world-class partnership. We do not use HBM memory. We believe the memory shortage and the resulting price increases will persist through 2028, raising competitors' costs and limiting their supply. The industry-wide memory challenge is an opportunity for Cerebras. Plants and supply are necessary but not sufficient to support capacity expansion. We must also have high yields. Cerebras invented an entirely new computing architecture, and applying metrics and mental models designed for other architectures is like looking at a Tesla, Inc. and asking where its exhaust pipe is. Electric vehicles require a different frame of reference, and so does wafer-scale computing. We welcome questions about all aspects of the design, but there is also a set of incorrect claims circulating repeatedly in financial circles. They believe our systems cannot support large modelsclearly wrong, they are doing so right now; and that our manufacturing yields are low. Both are wrong. Yields fall into two categories: fab yield and parametric yield. Our yields from Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADRthat is, fab yieldare high, and parametric yield is also high. All wafers have defects, and for decades these defects were the core reason the industry considered wafer-scale processors impossible. We solved this through redundancy. Our design contains about 1 million identical tiles, a small fraction of which serve as spares. When a defect disables a tile, the hardware bypasses it and activates a spare unit. This mechanism works both at the fab and in the field, enabling high wafer yield and good field reliability. The second challenge is parametric variation. Even when the silicon functions, its electrical characteristics vary across the wafer, with some regions running faster or slower at the same voltage. We address parametric variation in three ways: first, removing the wafer edgethe edge is the most fragile part of the wafer and the main source of parametric variation; we use the largest square that fits in a 300mm wafer, preserving most of the usable silicon. Second, designing and testing for a wide range of conditionsour cores tolerate a wider range of variation, and during timing closure we validate designs across a much wider range of conditions than others. Third, independently regulating voltage across wafer partitionswe invented technology to independently adjust core voltage across 84 wafer partitions, raising voltage in slower regions and lowering it in faster regions to precisely control performance. To our knowledge, no other chip has this capability. The result is very high yields at both the fab and the factory. One final point: some have pointed to the relatively low gross margin on our hardware shipments and concluded this proves a yield problem. In fact, it merely reflects the continued ramp of our manufacturing capacity and delivery of the last batch of hardware under a special pricing program for an early customer. We expect hardware gross margin to improve significantly over the next several quarters, ultimately exceeding our long-term target of 60%. Capability. We are advancing capability on all frontsnew systems and new processors, improved software, advanced manufacturing technologies and partnerships. CS-4 is about to enter full mass production, with both performance and throughput set to double. Our next-generation processor WSE-5, featured in the fifth-generation system CS-5, has returned from the fab ahead of schedule. WSE-5 taped out on June 15, and the chip has returned, been powered on and is already running models, and is now undergoing full-system testing. Less than five months from tape-out to system testing is extraordinary execution. CS-5 will be more than twice the scale of CS-4 and will go into production in early summer 2027. It is based on Taiwan Semiconductor Manufacturing Co., Ltd. Sponsored ADR's 5-nanometer process, where wafer competition is less intense and wafer costs are much lower. On disaggregated inference: to our knowledge, Cerebras is the first company to deliver production-grade disaggregated heterogeneous inference. Today, our cloud serves customers using AMD GPUs for prefill and Cerebras systems for decode. We have built a disaggregated inference software platform that can adapt to AMD and other GPUs. The same approach will be deployed in the Cerebras cloud and at some data center customers. This delivers significantly more high-speed tokens per unit of powercloud owners get higher profit per megawatt, and AI users get lower cost per token. We believe disaggregated inference will become one of the default options in major cloud deployments. On software, our software stack continues to mature. Cloud services have achieved impressive uptime; systems keep getting faster; onboarding time for new models has shrunk to just days. Cerebras is a Day 0 launch partner for Kimi K3.1, Moonshot AI's 2.8-trillion-parameter frontier-class open-weight model. Kimi K3.1 on Cerebras is not only the fastest in the industry, running at up to 2,000 tokens per second. To sum up. Nearly 11 years since we started, we continue to innovate at an extremely rapid pace. In summary: we exceeded expectations on all metrics and raised fourth-quarter guidance across the board, while expecting revenue to triple in 2027; we made important progress in all key areas, including adding two hyperscale customers for a total of three in the past six months; we have begun delivering CS-4; the fifth-generation processor WSE-5 returned from the fab at record speed and entered CS-5 system testing; production-grade disaggregated inference is already running on AMD and supports another 2.8-trillion-parameter model on Day 0 with industry-best performance. Bob Komin (Chief Financial Officer): Thank you, Andrew. Before getting into the specifics of third-quarter results and fourth-quarter outlook, I want to review how we have expanded the business. We chose to do cloud first in early 2025 because, in this rapidly evolving AI landscape, AI-native customers want to scale compute quickly rather than build their own data centers. Because hyperscalers plan their data center compute allocation years in advance, they often do not have near-term available capacity in their own facilities to accommodate new hardware. To win the large-scale deployments needed by foundation model labs and hyperscalers, Cerebras needed to prove our systems can deliver industry-leading inference performance at scale. After fully evaluating our solution in our private cloud, customers can choose to buy our hardware or purchase our cloud services. To quickly provide enough cloud capacity, we chose to begin leasing data center capacity and expand our private cloud to gigawatt scale as quickly as possible. This overall strategy appears correct. Because data center capacity is now very scarce, and we have a large number of contracts in hand. More importantly, once we have a private cloud, we can demonstrate the value of fast inference to the entire market and drive more demand. Bringing large new data center capacity online can take 15 to 18 months. To avoid waiting and to accelerate meeting already enormous demand, we leased some of our own equipment back to a customer. As we have said before, this pressures gross margin in the short term, but it also gives us some capacity when customers are ready to buy, enabling growth with existing customers and winning important new customers. We have now won frontier labs like OpenAI and hyperscalers like AWS, and most recently added two more hyperscale customers. We have other leading customers on the cloud, including Cognition, Sigma, Block and Bilt. We serve customers where they areselling hardware in some cases and delivering cloud services in others. Soon, our signed neocloud customers will begin distributing fast inference running on Cerebras hardware in their clouds. We also mentioned on previous calls that 2026 will be a year of building capacity and capability and adding customers, laying the foundation for revenue to more than triple in 2027. With strong progress year to date, we are well positioned to achieve that goal. Now to third-quarter results. The fourth quarter will be an important inflection point for our growth and core financial trends. We exceeded core financial guidance across the board for the third consecutive quarter and achieved our first positive non-GAAP net income since listing earlier than planned. Core revenue was $231 million, up 10% quarter over quarter and 70% year over year, $16 million, or 7%, above the midpoint of guidance ($215 million). Core cloud and other services revenue was $145 million, up 272% year over year and 13% quarter over quarter. We expect fourth-quarter cloud revenue to be roughly flat quarter over quarter. This trend is not due to pricing or demand issues. The pricing environment is healthy, and demand still far exceeds supply. Our cloud revenue ramp is entirely constrained by data center capacity, with capacity only beginning to come online in December, too late for the fourth quarter. We expect data center capacity to increase substantially throughout 2027, with most of it starting mid-year next year as we deliver a larger portion of the 700 gross megawatts we have contracted. Core hardware revenue was $86 million, up 5% quarter over quarter. We see rising customer demand for fast inference, not only through our private cloud but also through new hardware deals. As these hardware shipments begin in the fourth quarter, we will see a significant jump in total core revenue. From the third quarter to the fourth quarter, core revenue grows nearly 40% quarter over quarter and about 85% year over year. These hardware orders (including Oracle Corporation) will continue delivering at least through the first half of 2027. Today we also announced Microsoft Corporation as a new customer. Over the past few months, our pipeline has grown more than 10x, DRIVE by diversified opportunities from new and existing customers, including hyperscalers, neoclouds, sovereign customers, enterprises and AI-native customers. Third-quarter core gross margin was 41%, up 40 basis points quarter over quarter and 200 basis points above guidance. In addition to the gross margin impact from leased-back systems, we also faced pressure from clearing the last shipments under hardware procurement agreements signed at lower prices years ago. We had guided that these factors would cause a temporary decline in gross margin, with the trough in the third quarter. In fact we did better, and the trough actually occurred in the previous quarter, the second quarter. Core hardware gross margin declined quarter over quarter due to the last shipments under that customer agreement signed a year ago. We are now delivering hardware orders at significantly higher current prices, which also means gross margin will improve significantly starting in the fourth quarter. Core cloud and other services gross margin rose to 46.9%, up 510 basis points quarter over quarter. The main driver was new business delivered at higher pricing and gross margin, partly offset by leasing data center capacity and hardware back to customers at higher cost. We expect this situation to persist at least through mid-2027, until we have enough of our own data center capacity. Fourth-quarter guidance: total core gross margin up about 700 basis points to 48% (midpoint). We expect total core gross margin to exceed 50% in the near term and continue improving toward our long-term target of 60%. Core operating margin improved 170 basis points quarter over quarter to negative 14.3%, nearly 10 full percentage points above the midpoint of our guidance. Core operating loss was roughly flat quarter over quarter at about $33 million. Core net income turned positive at $30 million. Year to date, adjusted core EBITDA was also positive at $2 million. With strong fourth-quarter revenue and gross profit growth, we expect all core key profit metrics, including operating margin, to turn positive next quarter. As of September 30, RPO increased to $26.1 billion. We update RPO for newly signed business quarterly and adjust existing contract value based on estimated changes in amounts to be recognized over the remaining commitment period. Third-quarter RPO did not include new deals signed after September 30, including the very large Oracle Corporation deal. Since the OpenAI contract became the main component of the RPO we have reported to date, there have been many questions about its status. As a reminder, the OpenAI contract filed with the S-1 before listing is a 750-megawatt take-or-pay contract. Among the many contractual requirements applicable to both parties is a clause: if we delay capacity delivery, OpenAI has the right to receive credits during that period usable for future committed purchases. We have calculated the potential delay compensation amount and deducted it from RPO as of September 30. The impact of these compensations on revenue and gross margin is small, because their proportion relative to RPO is estimated to be small and will be amortized as most revenue is recognized proportionally over the coming years. Future changes to these estimates will be updated quarterly and reflected in RPO and guidance as needed. If capacity delivery delays persist for 90 days or more, both OpenAI and we may choose to cancel only that portion of capacity. We estimate that the scale that could be canceled this year is less than 5% of total committed megawatts. Given strong global demand for data center capacity, it took us a little longer to find and bring online the initial data center capacity needed to meet the delivery schedule under a very aggressive timeline. Based on the team's frequent site inspections at each location, we now expect to return to the planned schedule in the second half of 2027, which should reduce future delay compensation risk. We took a conservative approach and did not assume any delay would be excused due to factors beyond our control. In addition, if OpenAI or Cerebras chooses to cancel a delayed capacity batch, we believe there is enough demand to find other customers to take it on. OpenAI also retains an option to purchase an additional 1.25 gigawatts of capacity beyond its initial 750-megawatt commitment, which could come online starting in 2029. Since that is three years away, we have not even begun discussions. Our liquidity remains strong, with $7.8 billion in cash, cash equivalents, restricted cash and marketable securities at the end of the third quarter. We also have a revolving credit facility of up to $850 million, unused to date. The $1 billion prepaid revolving loan we obtained in January, along with ongoing non-recurring pass-through fees charged for fitting out new data center capacity, has funded most of our net working capital needs and capital expenditures year to date in 2026. We are not a neocloud. We are vertically integrated and the inventor of the industry's fastest inference solution. Our business has a degree of durability and technological differentiation that we believe will drive much higher long-term margin potential and cash flow. Our model also requires lower capital investment for the compute equipment portfolio we deploy. A substantial portion of the equipment in our data centers is self-produced, similar to other companies that develop and deploy their own AI chips; we do not need to pay a 70% gross margin to third parties and do not use expensive HBM, only paying our own bill of materials costs, which lowers relative capital expenditure per megawatt while we also benefit from producing the fastest, most valuable tokens. Despite rapid expansion of private cloud buildout, the only debt on our balance sheet is the OpenAI revolving loan. We have also received and continue to receive customer prepayments. Our capital structure today includes substantial cash and investments on hand and very low leverage. Fourth-quarter 2026 outlook: core revenue $310 million to $330 million; core gross margin 47% to 49%; core operating margin negative 1% to positive 1%. Full-year 2026: core revenue raised to $940 million to $960 million; core gross margin raised to 42% to 44%; core operating margin raised to negative 10% to negative 8%. In summary, the third quarter was another record quarter for Cerebras, and so will be the fourth. Our flexible model of selling both hardware and cloud is driving strong revenue growth and gross margin improvement, with improvement arriving a quarter earlier than expected, enabling us to triple revenue in 2027 while driving core gross margin toward our long-term target of 60%. We have demonstrated strong operating leverage and expect fourth-quarter core operating profit to approach break-even. Subsequently, we expect 2027 to make significant progress toward our long-term operating margin target of 40%. One final point: we have been public for more than 180 days. As a result, our IPO share lock-up fully expired on November 9. Thank you for listening. I will ask the host to open the line for questions. II. Q&A session Thomas O'Malley (Barclays Research): Congratulations on such strong results. My first question is about the new deals, OpenAI and Microsoft Corporation. Can you give some metrics on the relative scale of these new deals? Also Bob, you mentioned part of OpenAI is already in RPO; can you say how much has been included? Andrew Feldman: Tom, this is Andrew, good to hear your voice. Regarding the Microsoft Corporation opportunity and the deal we struck with them, we will disclose more details. We are clearly very excited about it and will share more details in the first half of next year, at which point you can expect the deal to begin generating revenue. Bob? Bob Komin: Tom, in RPO, what we included for OpenAI is our estimate of the committed revenue amount under this take-or-pay contract, excluding extensions and excluding additional options, adjusted for delay compensation we believe likely to occur; it also includes the data centers we contracted for them, the non-recurring fees they paid for buildout, and the monthly recurring fees over the contract term. These are the pass-through revenue and pass-through costs we recognize. Thomas O'Malley (Barclays Research): My follow-up is that we see frontier labs not only able to build hardware quickly but also to accelerate deployment of models across different compute hardware, with increasing substitutability. So how do you benefit from this? What do you expect future heterogeneous infrastructure to look like? What are the pros and cons for you? Andrew Feldman: That's a great question, Tom. OpenAI has said in filings or on calls that it took them about 3 days to bring AMD GPUs online. This is a very interesting and important development, especially for long-term thinkers. I think people have not really understood what this means for our ecosystem. Compilers and low-level software have historically been a fundamental barrier to architectures, favoring incumbentsthat is the software moat you hear so much about. If frontier labs have truly done what they say, it means the value of this kind of low-level software will shrink and disappear over time. And if low-level software has little value, what remains to provide value is innovative hardware. That means differentiation will depend more on hardware differentiation. I think this is a huge advantage for Cerebras and other companies building pioneering hardware, new architectures or different architectures. I think it will erode the advantages incumbents hold. It's a great question, and I think the industry needs to keep watching it. Thank you. Joseph Moore (Morgan Stanley Research): Congratulations on the strong hardware numbers. Can you talk about how the RPO issue relates to the cloud schedule, is there anything specific we can watch? And your confidence in resolving this before it turns into additional penalties? Andrew Feldman: I think what we included in RPO is a very conservative number, the absolute worst case we could encounter. So we are comfortable today that results will be better than that number. These are common in large data center deals. We have similar agreements with data center builders, and their penalties related to their own delaysI think over time whether to exercise those clauses will depend on commercial decisions at the time. I think that is how these clauses work, and how we expect them to work in the future. Joseph Moore (Morgan Stanley Research): What is the cause of the delay? I think everyone understands land, power, buildings and so on. Is it these kinds of reasons, or something else? Andrew Feldman: The current delays are entirely from land, power and buildings. We have capacity, all materials are available. We are waiting for power and building completion. Atif Malik (Citi Research): Andrew, I want to ask about the product roadmap. When does the CS-4 crossover (replacement) versus CS-3 occur? Is it in 2027? Also in your roadmap, CS-3 significantly improved throughput. If you continue down this path into CS-4 and beyond, will there come a day when your chips provide enough speed and throughput that systems can be used alone without GPUs? Andrew Feldman: Good question. First, until a few weeks ago, all our deployments were pure CS-3. Only recently did we begin deploying the disaggregated approach, and we are seeing significant throughput gains, which is very exciting. So that is the first observation. Clearly they run very well in aggregated form, but disaggregationwhich we are deploying and is now largely generally availableproduces more high-speed tokens when high-speed tokens are constrained. I think the CS-4 crossover (replacement) is in the first quarter. The roadmap currently looks very good. We are ahead of schedule on CS-5. As I mentioned in my prepared remarks, it took only 5 months from tape-out to the chip entering the system, which is extraordinary. We expect it to begin shipping in early summer 2027 with double the performance. Atif Malik (Citi Research): Then a question for Bob: you guided fourth-quarter cloud sales flat, so within the framework of sales tripling in 2027, can you help us understand the mix of cloud and hardware? Bob Komin: I think the best way to understand the hardware mix isthese deals can be very, very large. We hope to be able to give some indication on shipments over the next one or two quarters. It depends heavily on whether customers' data centers are ready. So what I can say is that we think the hardware proportion in the fourth quarter and into the first quarter will be closer to 50% than in recent quarters. Looking at the rest of 2027, we will provide more information on the fourth-quarter call, and I think at that point we will have a more complete view of the overall 2027 outlook. Quinn Bolton (Needham & Company): Andrew, I just want to confirm I heard correctly. In RPO and your assumptions, you assumed potential delay compensation, but to date you have not received any cancellation for delayed capacity. Is that correct? Andrew Feldman: Correct. Quinn Bolton (Needham & Company): The market is worried about fast inference offered by low-batch GPUs, which sounds like an expensive way to serve tokens. So I wonder if you can talk about Cerebras's cost per token relative to low-batch GPU configurations? Andrew Feldman: Of course. GPUs allow you to trade throughput for extra speed by lowering batch size. So you have high throughput at low speed. When you try to get higher speed, throughput drops off a lot, until you enter the low-batch range. That is the characteristic of the architecture. They can get some limited performance by going to low batch, but in doing so, cost per token jumps. They also serve fewer customers at once, so the total number of tokens generated is very small. As you keep pursuing faster, we get stronger and stronger. At some point, no matter what you do to the GPU, they still cannot do it at batch size 1. The question is where the crossover is. I think at around 250 to 300 tokens per second