Good day, and thank you for standing by. Welcome to the Cerebras Systems third quarter 2026 earnings conference call. At this time, all participants are in listen-only mode. After the speaker's presentation, there'll be a question and answer session. To ask a question during the session, you'll need to press star one one on your telephone. You will then hear an automated message advising your hand is raised. To withdraw your question, please press star one one again. Please be advised, Today's conference is being recorded. I would now like to turn the conference over to Sean Dorsey, head of investor relations. Please go ahead, Sean.

Thank you, operator.

Good afternoon, everyone, and welcome to Cerebras Systems Q3 2026 earnings call. Earlier today, we issued our press release and posted our supplemental earnings presentation to the investor relations section of our website. A replay of this webcast will also be available on our Investor Relations website following the call. Joining me today are Andrew Feldman, our co-founder, chief executive officer, and president, and Bob Komin, our chief financial officer. Before we begin, I would like to remind everyone that today's discussion will include forward-looking statements under the safe harbor of the Private Securities Litigation Reform Act of 1995. These statements include, but are not limited to, statements regarding our future financial performance, business strategy, market opportunity, customer demand, product roadmap, technology leadership, supply chain, operating model, and outlook for Q4 and full year 2026. Forward-looking statements are based on our current expectations and assumptions and are subject to risks and uncertainties that could cause actual results to differ materially from those expressed or implied. These risks are described in our SEC filings, including our final prospectus related to our IPO and our future periodic filings with the SEC. We undertake no obligation to update these forward-looking statements except as required by law. During today's call, we will also discuss certain non-GAAP financial measures. Reconciliations between GAAP and non-GAAP results are included in today's press release and supplemental materials available on our investor relations website. With that, I'll turn the call over to Andrew.

Thank you, Sean, and thank you all today for joining us. As we look to the months ahead, we see two certainties, rapid change and compute shortages. In the last 90 days, the AI ecosystem has seen the title of leading frontier model change hands multiple times. It's seen dozens of interesting open models released, major security incidents occurred, and powerful new consumer agents emerged, and many other developments that none of us can even remember. It's impossible for anyone to predict AI capabilities 90 days from now, let alone a year from now. Yet through all that change, intelligence and speed remain the two essential features for every use case, from coding to security to personal agents. At Cerebras, our job in this rapidly changing world to make sure that no matter how AI grows, Cerebras grows. As foundation models grow, we grow. As open models grow, we grow. As personal agents grow, we grow. Whether enterprises or startups grow, we grow. And whether customers want hardware or cloud capacity, we grow. Our goal is to position ourselves so that whatever drives the industry, Cerebras wins. And in Q3, we made remarkable progress towards that goal. First, I would like to report that we exceeded all guided business metrics. And while I'm pleased with these results, I believe even greater importance is the fact that we closed a number of deals with key AI industry leaders since our last earnings call. With the latest being our announcement that today we signed an important cloud deal with Microsoft. which along with the AWS and Oracle announcements previously made bring us to three hyperscalers who have chosen to leverage the Cerebras platform as part of their imprint strategy. These new deals alongside continued growth and execution with our existing customers enable us to raise Q4 core guidance across the board and continue to see revenue tripling in 2027 with expanding gross margins. Bob will share all of the details regarding our Q3 performance and portal looking guidance in his section. Let me take a minute to clarify the shape of our business. As an AI accelerator company, our pioneering innovations are in the chips, systems, and software that make AI compute fast. We are an AI hardware company with two channels to market. First, we sell systems, installing racks of Cerebras hardware in our customer's data center. This is a great business. Revenue is recognized quickly and cash is received quickly. It is also lumpy and depends on customers having data center capacity ready to receive the hardware. Our second route to market is through our cloud. There's a very rapidly growing category of AI natives who want to consume tokens rather than deploy hardware. And given the fact that data centers and compute are in very short supply, the fact that we have both the system and the data center is a strategic advantage. The AI cloud business is a particularly good business if you make your own hardware. Our cost per megawatt is much lower, and as such, we can bring more capital efficiency than NeoClouds. The mix of these two channels will vary from quarter to quarter, but we believe both will grow significantly over time. Let me also address speculation about our OpenAI partnership. OpenAI is a very large customer and a close partner, and I believe they will continue to be so going forward. OpenAI grew in Q3 as compared to Q2, and we expect more growth in the future. The success of OpenAI's ultra-fast SALT56 preview based on Cerebris was overwhelming, demonstrating unequivocally that the demand for fast tokens is enormous. It is also true that Cerebras didn't have sufficient capacity to support UltraFast in a broad rollout, including for Astra. As others have stated, we believe that capacity is a key driver for business decisions in our industry. In fact, even with the enormous supply of GPUs available to OpenAI, usage-based restrictions still needed to be implemented. After signing, The agreement with OpenAI took us longer than we would have liked to find contracts for and bring online the substantial increase in data center capacity required to meet the aggressive requirements in the agreement. And as such, we find ourselves running behind schedule on the initial tranches. Given the importance of this partnership to our business, we understand the close scrutiny it gets from investors. The details of the agreement or news about it are unlikely to come from any source other than Cerberus or OpenAI. For example, there were rumors that Cerebras couldn't support Astra for technical reasons. These are false. RCS3, CS4, and CS5 are more than capable of supporting Astra scale models and larger models as well. At the same time, I expect, and you should expect, OpenAI to build capacity using a range of hardware solutions. OpenAI will retire older models, deploy new ones across those compute options for internal and external use. Their compute needs are vast. Plenty of opportunities for different solutions to play a role as they scale. Before turning to our business framework, let me make one more general point. Despite the experimentation, low-batch GPU, startup announcement, megakernels, and hyperscaler of internal custom silicon, we continue to be the fastest inference solution bar none. We believe our roadmap will enable us to maintain that advantage for years to come. Our framework for evaluating progress remains unchanged. We assess our business across three dimensions, customers, capacity, and capabilities. Let's begin with customers. We are rapidly expanding our customer base to broaden our reach and diversify our revenue stream. Q3 was a successful quarter to that end. We had an important win with two new hyperscalers. We signed an agreement to bring Cerebras hardware into Oracle Cloud, where Cerebras will be deployed in Oracle data centers serving AI labs, enterprises, and developers building interactive and intelligent applications. We expect this Cerebras-powered Oracle Cloud inference service to launch in the summer of 2027. We also signed an exciting deal with Microsoft, about which we will share more next year. That brings us to three hyperscaler wins in the last six months. This is an important achievement, and it positions us well for 2027 and 2028. We also expanded our reach through a variety of neoclouds with off-takers of high caliber like Firework, diversifying our routes to market and giving customers more ways to access three versus high-speed inference. We also benefited from open weight models, which continue to enable new applications and expand what AI can do in everyday life and across industries. For example, just this week, we announced that Cerebras Fast Inference is powering the built neighborhood OS platform and personal assistant, bringing blisteringly fast inference to millions of people. With consumer apps like built, we're putting the power of Cerebras directly into people's pockets. That is why Fast Tokens command a premium. And our cloud customers tell us that Fast Tokens generate exceptional growth margins and that Cerebras Fast Tokens are the most profitable tokens they sell. Let's now turn to capacity. Capacity has two components, data center capacity and manufacturing capacity. Both are critical to growth. Data center availability remains a constant We are no exception. Moreover, we believe the evolving U.S. policy landscape could create greater challenges for data center development in the U.S. In 2027, more than 75% of our data center capacity comes online outside of the U.S. In fact, in the next several months, new facilities in Canada, France, and Norway come online. and during the quarter we signed contracts for additional capacity in Norway and Finland, further expanding our European footprint and we're in the final stages of contracting for a data center in Sao Paulo, our first in South America. So far we've secured approximately 700 gross megawatts of data center capacity for delivery by the end of 2027. Our first party cloud is already among the larger non-hyperscalar AI clouds and we're far down the path to negotiations for gigawatts of capacity. Finally, the best way to increase data center capacity is by increasing the token generating capability of the hardware in the data center, but simply to generate more tokens per megawatt. To this end, this aggregated inference plays a role in expanding our capacity by increasing by up to 5x the number of tokens that a data center can deliver. More super fast tokens per megawatt We're scaling manufacturing with Flex and Sandmina. We expect to exit 2026 with more than 10 times the manufacturing capacity we had at the start of the year. In 2027, we plan to add factories and further increase capacity to support the substantial growth we anticipate. Our supply chain is well positioned to support this growth as well. Most importantly, we have sufficient wafer commitments from TSMC, which continues to distinguish itself as a world-class partner. It is also important what our supply chain does not need. That is also key to us. Unlike other competitors, we do not use HPM memory. We believe memory shortages and the price increases associated with them will continue through 2028 raising competitors' costs and limiting supply. Industry-wide memory challenges create an opportunity for Cerebras. Factories and supply are necessary but not sufficient to scale capacity. We must also have high yields. Cerebras has invented a fundamentally new computing architecture. Applying metrics and mental models designed for other architectures can be like examining a Tesla and asking, where's the exhaust pipe? are trying to measure its tailpipe emissions. Electric cars require a different frame of reference, and so does wafer-scale computing. We welcome questions about every aspect of our design and will work to share architectural details more clearly. But there are also a set of persistent falsehoods proselytized to the financial community, namely that our systems can't support large models, obviously false, as they currently do, and that our manufacturing yields are low. Both are false. I already addressed model size earlier in my comments and in previous discussions. Since manufacturing yields are harder to observe, I'll take some time here to address the topic directly. There are two types of yields, fab yield and parametric yield. Our yields from TSMC, which are the fab yields, are high. So are our parametric yields. While our technical blog covers this subject in detail, I'll provide a brief overview. All wafers contain defects. For decades, those defects were a central reason that industry considered wafer-scale processors impossible. We solved that problem through redundancy. Our design contains roughly a million identical tiles with a small portion held in reserve When a defect disables a tile, the hardware routes around it and activates a spare. This mechanism works both at the fab and in the field, delivering high wafer yields and strong infield reliability. The second challenge is parametric variation. Even when silicon functions, its electrical characteristics vary across the wafer. Some regions run faster. or slower than other regions at the same voltage. This is a basic fact of chip design and one of the obstacles we have to solve to make wafer-scale computing possible. We address parametric variation in three ways. First, we remove the wafer's outer edges. The edges are the weakest part of a wafer and account for much of the parametric variation. We use the largest square that fits into a 300 millimeter wafer, excluding and much of the lead to silicon. Second, we design and test for a broad range of conditions. Our cores tolerate a wider range of variation. During timing closure, the process of verifying that circuits meet timing requirements, we validate the design across operating conditions much broader than those used by others in the industry. This enables expected performance despite variations across the wafer. Finally, we tuned voltage independently across the wafer. We invented technology to adjust core voltage independently across 84 wafer sections. We can raise voltage in slower sections or lower it in faster sections, giving us precise control over performance. We know of no other chip with this capability. The result is exceptionally high in yield at both the fab and the factory. We achieved this by engineering solutions to the problems others said made wafer-scale computing impossible. As a final note, there are those who have pointed to our relatively low gross margins on hardware shipments and suggested that this is evidence of a yield problem, when in fact, it simply reflects the ongoing ramp of our manufacturing capabilities and the final shipments of hardware covered under special pricing programs offered to one of our early customers. We expect our hardware margins to improve markedly over the next several quarters, eventually exceeding our long-term target of 60%. Turning now to capabilities. We are advancing our capabilities on every front. New systems and processors, improved software, advanced manufacturing techniques, and partnerships that deliver best-in-industry solutions. Despite the rapid pace of change across the AI industry, the importance We believe this performance advantage will endure over time, and to that end, our product roadmap continues to advance rapidly. Our CS4 will soon be in full production, doubling both performance and throughput. Our next generation processor, the Wafer Scale Engine 5, which we've deployed in our fifth generation CS5 system is back from the fab ahead of schedule. The WSC5 taped out on June 15. The chip is back, has been brought up, and is already running models and is undergoing in-chassis testing. Moving from tape out to system test in less than five months is exceptional execution. The CS5 will more than double the speed of the CS4 We will enter production in early summer of 27. It is built on TSMC's 5 nanometer node, but competition for wafers is less intense and wafer costs are much lower. Our innovations also extend to complete solutions. To our knowledge, Cerebras is the first to deliver production-scale, heterogeneous, disaggregated Today, our cloud serves customers using AMD GPUs for pre-fill and Cerebras CS3 systems for decode. To enable the production launch, we built a disaggregated inference software platform. It is built to work with AMD as well as other GPUs. The same solution will be deployed across the Cerebras Inference Cloud and to select data center customers. This aggregated inference delivered substantially more high-speed tokens for the same power. That means more profit per megawatt for cloud owners and lower cost per token for AI users. We believe DISAG will become one of the default offerings in major cloud deployments. Our software stack continues to mature. Our cloud services deliver impressive uptime. Our systems keep getting faster. and the time needed to bring up new models has fallen to just a few days. Cerebras was a day zero launch partner for Kimi K3.1, Moonshot AI's 2.8 trillion parameter frontier class open weight model. Perhaps the premier open weight model in the market, Kimi 3.1 on Cerebras is not only the fastest in the industry, but it runs at up to 2,000 tokens per second. Nearly 11 years into our journey, we continue to innovate at a tremendous pace. Our team has stayed together through good times and tough times, and there is more energy and excitement in our building than ever. We take pride in tackling and solving some of the industry's hardest problems. By way of summary, we beat on all metrics, and we're raising guidance for Q4 across the board and expect triple revenue in 2027. We made important progress in each of our key domains, including adding two hyperscale customers and have now won three hyperscalers in the last six months. We began shipping our CS4. Our fifth generation processor, the WSE5, is back and in CS5 system test in record time. Production DISAG is now live with AMD at day zero support for yet another 2.8 trillion parameter model were successfully achieved at industry best performance. With that, let me turn things over to Bob.

Thank you, Andrew, and good afternoon, everyone. Before I get into the specifics on our Q3 results and Q4 outlook, I'd like to reflect on how we've scaled the business. We chose to ramp a cloud offering in early 2025 because in this rapidly advancing AI landscape, AI natives want to scale compute quickly and don't want to own data centers. And because hyperscalers plan their data center compute allocation years in advance, they often don't have near-term availability in their own facilities for new hardware. In order to win the large deployments that Foundation Labs and hyperscalers require, Terebras needed to demonstrate that our systems could deliver our industry-leading inference performance at scale. After having the opportunity to thoroughly evaluate our solution in our private cloud, these customers can then choose to buy our hardware or purchase our cloud offering. To be able to provide enough cloud capacity quickly to meet these demands, we chose to begin to lease data center capacity and grow our private cloud as quickly as possible towards gigawatt scale. This general strategy appears to have been sound. As data center capacity is now very scarce and we have a lot of it under contract. Even more importantly, once we had our private cloud, we were able to demonstrate the value of fast inference to the entire market and drive more demand for it. A large new data center capacity can take 15 to 18 months to bring online. Instead of waiting for these data centers and to accelerate our ability to meet the overwhelming demand we already had, We rented some of our own equipment back from one of our customers. As we've previously described, this applied downward pressure to gross margin in the short term. But it also allowed us to grow existing customers and win significant new ones because we had some capacity when they were ready to buy. We have now won frontier labs like OpenAI and hyperscalers like AWS, and recently added two new hyperscalers to our customer base. We have other leading customers like Cognition, Figma, Block and Built among others on our cloud. We meet these customers where they are, in some cases selling them hardware and in others delivering them our cloud services. Soon, new NeoCloud customers we have signed will begin distributing fast inference services running on Cerebras hardware in their clouds. In our prior calls, we also shared that 2026 is going to be a year of building out capacity, capabilities, and adding customers to set us up to more than triple our revenue in 2027. With our strong progress so far this year, we are well positioned to achieve that target. Now turning to Q3 results. I'd like to talk about our execution in the third quarter, and I'm pleased to say that Q4 will be an important inflection point for our growth and core financial trends. We've exceeded our core financial guidance across the board for the third quarter in a row. We also delivered positive non-GAAP net income for the first time as a public company and ahead of schedule. As a reminder, the definition of our core metrics and the associated reconciliations to GAAP are included in today's earnings release and on our website. We delivered record core revenues of $231 million, up 10% sequentially and 70% year-over-year, exceeding the $215 million midpoint of our guidance by $16 million or 7%. Revenue from our core cloud and other services was also a record, increasing by 272% year-over-year to $145 million. Sequential cloud revenue growth in Q3 was 13%. We currently expect flattish sequential cloud revenue in Q4. This trend is not driven by a pricing or demand issue. The pricing environment is healthy and demand for our solutions continues to outpace supply by a wide margin. Our ramp in cloud revenue is limited entirely by our data center capacity, which we will begin to ramp in December. as we build momentum going into next year, but unfortunately that's too late to affect Q4. We expect much more substantial increases in data center capacity across 2027. The bulk of this starts in the middle of next year as we deliver larger portions of the 700 gross megawatts of data center capacity currently under contract. Core hardware revenue was $86 million in the quarter, increasing We are seeing increasing customer demand for fast inference delivered not just through our private cloud, but also through new hardware deals that we've closed with several new customers. With those hardware shipments scheduled to begin in Q4, we will see a significant step up in total core revenue. Sequential core revenue growth from Q3 to Q4 will approach 40%, and year-over-year growth will be about 85%. These hardware orders, including Oracle, will continue to ship through at least the first half of 2027. As Andrew mentioned earlier, today we also announced Microsoft as a new customer. Over the last few months, our pipeline has grown by much more than 10x, driven by a diverse set of opportunities from both new prospects and existing customers including hyperscalers, neoclouds, sovereigns, enterprises, and AI natives. Turning now to gross margin. Q3 core gross margin was 41%, a sequential improvement of 40 basis points as compared to Q2 and 200 basis points above our guidance for the quarter. In addition to the gross margin impact from running back our systems, We also faced some headwinds in the quarter as we were burning off the last shipments from a hardware purchase agreement signed years ago at lower pricing. We guided that all of this would create temporary gross margin declines and the low point would be reached in Q3. Well, we did better than that and the low point was actually last quarter in Q2. Now I'll break total Q3 core gross margin into its components. The core hardware gross margin declined sequentially to 31%, driven by the final shipments under the year-old customer agreement. We are now shipping hardware orders at significantly higher current prices, which also means significantly better margins starting in Q4. Gross margin from core cloud and other services improved significantly to 46.9%, 510 basis points higher than in Q2. The primary driver of the increase was new business delivered at higher pricing and margin. This was partially offset by our program of running back data center capacity and our hardware at higher costs from a customer. We expect this to continue through at least the middle of 2027 until we have enough of our own data center capacity to fill our high levels of cloud demand. With all of that as a backdrop, in Q4, we expect total core gross margin to improve by a step function of 700 basis points to 48% at the midpoint. We expect total core margin to exceed 50% in the near term and to continue to improve towards our 60% long-term target. We also made strong progress towards profitability in Q3 and year-to-date. Core operating margin improved sequentially by 170 basis points to a loss of 14.3%, which exceeded our guidance midpoint by nearly 10 full margin points. Core operating loss was flat in Q3 versus Q2 at approximately 33 million. Core net income turned positive and was 30 million. Adjusted core EBITDA year-to-date was also positive at $2 million. With strong increases in revenue and gross margin forecast for Q4, we expect that all of our core key profit measures, including operating margin, will be positive next quarter. Now turning to RPO and the OpenAI contract. Remaining performance obligations increased We update RPO for new business that was signed by quarter and adjust existing contract values for any changes to estimates of what remains to be recognized over the remaining committed term. Q3 RPO did not include the impact of new deals signed after September 30th, including Oracle, which is large. There have been a lot of questions about the status of our OpenAI contract since it represents the majority of our reported RPO to date. As a reminder, OpenAI contract, which was filed with our S-1 before we went public, is a take or pay contract for 750 megawatts of capacity. Among many contract requirements that apply to both parties, it includes a term that if we experience delays in delivering capacity They are entitled to credits during that time to use against future committed purchases. We have calculated the amount for potential delay credits and deducted them from RPO at September 30th. These credits would have minimal impact on revenue and gross margin because as a percentage of RPO, they are estimated to be relatively small, and then they would get spread ratably as the majority of the revenue gets recognized over the next few years. Future changes to these estimates will be updated each quarter and reflected in both RPO and guidance if needed. If a delay in delivering capacity continues for 90 days or more, both OpenAI and us have the option to cancel that portion of capacity only. The amount we estimate could be subject to cancellation this year would be less than 5% of the total committed megawatts. Given the relentless demand for data center capacity around the world, it has taken us some additional time against a very aggressive schedule to find and bring up the initial data center capacity required to meet the delivery schedule. Based on our team executing frequent in-person inspections of each site, We now project we will be back on schedule by the second half of 2027 and this should reduce future delay credit risk. We've taken a conservative approach and have not assumed that any of these delays may get excused due to dependencies that are outside our control. If that happens, it would reduce the amount of the delay credits we have assumed. If OpenAI or Cerebras were to elect an option to cancel a delayed capacity tranche, we believe there is sufficient demand and we would find other customers to absorb it. OpenAI also retains an option to purchase an additional 1.25 gigawatts beyond their initial 750 megawatt commitment that could come online starting in 2029. Since that's still three years away and the world is changing as quickly as it is in AI, we've not even begun to discuss this yet. Now, turning to our balance sheet and capital efficiency. We continue to have strong liquidity, ending Q3 with $7.8 billion in cash, cash equivalents, restricted cash, and marketable securities. We also have a revolving credit facility of up to $850 million that remains unused to date. The upfront $1 billion working capital loan we received in January and continuing payments we received for non-recurring pass-through expenses to fit out new data center capacity to be delivered as part of the OpenAI contract have funded the majority of our net working capital requirements and capital expenses. So far in 2026. We are not a neocloud. We are vertically integrated inventor of the fastest inference solution in the industry. Our business comes with a level of durability in technology differentiation that we believe will drive much higher long-term margin potential and cash generation over the long run. Our model also requires lower capital investment for the mix of compute equipment we deploy. For a significant portion of the equipment used in our data centers that we manufacture, similar to others who've chosen to make and deploy their own AI chips, we do not pay a third party 70% gross margins, nor do we use high priced HBM. We only pay our BOM cost, which reduces the relative capital expense per megawatt, while we also benefit from producing the fastest and most valuable tokens. This shows up in our capital structure to date where the only debt on our balance sheet, despite our rapid growth and private cloud build out, is the OpenAI working capital loan. We've also received and continue to get upfront payments from customers, which is another lever for capital efficiency. Our capital structure today includes significant cash and investments on hand and little leverage which provides us with more flexibility to have multiple ways to continue to fund our rapid growth, which is an advantage, especially in the current capital market environment. Now turning to our outlook. For Q4 2026, we expect to raise core revenue to the range of 310 to 330 million, core gross margin to the range of 47 to 49%, and core operating margin to the range of minus 1% to 1%. For the full year 2026, we're raising core revenue to the range of 940 million to $960 million. We're raising core gross margin to the range of 42 to 44%. And we're raising core operating margin to the range of minus 10% to minus 8%. In closing, Q3 was another record quarter for Cerebras, and Q4 will be too. Our flexible model of selling both hardware and cloud is driving strong growth in revenue and improvements in gross margin that began a quarter earlier than expected. We've made significant progress scaling our manufacturing and data center capacity, which will enable us to triple revenue with core gross margin progressing towards our 60% long-term target in 2027. We've demonstrated strong operating leverage and are projecting core operating profit to be approximately break-even in Q4. Then we expect to make significant progress towards our long-term 40% operating margin target during 2027. This mix of scale Gross, and profitability positions us as a unique and leading AI systems provider. Also, as a final note, I'd like to remind everyone that we've now been a public company for more than 180 days, and as such, our IPO stock lockup releases were fully complete as of November 9th. Thank you for listening to our prepared remarks. I'd like to ask the operator to please open the line for questions.

Thank you. As a reminder, if you would like to ask a question, please press star 11 on your telephone. You will then hear that automated message advising your hand is raised. If you would like to remove yourself, please press star 11 again. We also ask that you limit yourself to one question and one follow-up. As well, wait for your name and company to be announced before proceeding with your question. One moment while we compile the Q&A roster. First question will be coming from Tom O'Malley of Barclays. Please go ahead.

Hey, Andrew. Hey, Bob. Thanks for taking the questions, and congrats on the really good results. First question is in regards to the new deals, OpenAI and Microsoft. If you could just give us any metrics perhaps on the relative sizes of those new deals, and then, Bob, you mentioned that some of OpenAI is already in the RPO, perhaps giving us how much is already in the RPO.

Hey, Tom. This is Andrew. Good to hear from you. We'll be announcing more details about the Microsoft opportunity and the deal we struck with them. We are really obviously very excited about that. We'll be announcing that in more detail in the first half of next year, and that's when you can expect revenue to start from that transaction. Bob?

Hey, Tom. In the RPO, what we've included for OpenAI is our estimate of the committed amount of revenue under the contract, under the take or pay contract, no extensions, no additional options, adjusted by the delay credits that we think would be due, could be due. And it also includes, for data centers that we've signed for them, it includes the amounts that they pay for non-recurring costs to set them up and also for monthly recurring costs over the term of the contract. And those are the pass-through revenues and pass-through expenses that we recognize in the P&L.

Gotcha, Bob. I said OpenAI. I meant Oracle, but we can circle to that. But I guess the other one is just we've seen the ability from frontier.

We fight really hard to win. We're hyperscalers. We can't show them that.

The answer for Oracle is even easier. So far, no Oracle in our field. We signed it up. Gotcha. Thank you.

All right, and then the follow-up is, so we've seen the ability from Frontier Airlines to not only to build the hardware quickly, like Jalapeno, but to also accelerate model deployment across different compute hardware and increase fungibility. So how do you guys benefit from this deployment, and what do you expect in terms of the future of heterogeneous infrastructure to look like? How does that play into your hands?

Sure, that's a really good question.

Anthropic in there.

I think in their S-1 or on a tweet talked about how it took them about three days to bring up AMD GPUs. And this is a very interesting and important development. Particularly those who have a long-term focus, I don't think people have really grokked what this What this means to our ecosystem. Compilers and low-level software have historically been a fundamental impediment to new architectures. They have favored the incumbent. They were the software moat that you heard so much about. If the in-profit can in fact do what they say they do, it means that The value of this low-level software shrinks and over time disappears. And it means that the boats from that low-level software also shrink and disappear. And if the low-level software has little value, the thing that is left to provide value is innovative hardware. And it means that differentiation will be more dependent on hardware differentiation. I think this is a tremendous advantage to innovative hardware companies like Cerebras and others who build pioneering hardware and use new and different architectures. And I think this will remove an advantage that the incumbents have long had. But that's a really good question and I think something that the industry needs to keep an eye on. Thank you.

Thank you for your question. One moment for the next question. That question will be coming from the line of Joe Moore of Morgan Stanley. Please go ahead.

Great. Thank you and congratulations on the strong Harvard numbers. Can you talk about the RPO issue with the scheduling on the cloud? You know, is there anything specific that we can focus on and your sort of ability, your confidence in your ability to resolve that before it turns into additional penalties? I think what we put in the RPO was a very conservative number of what is the absolute worst case we could be hit with. And so I think we are comfortable today that we would be better than that. These are commonplace in big data center deals. We have similar agreements with our with our data center builders. There are penalties when they're late. And I think it varies over time whether you choose to implement them and you make a business decision at that time. I think that's the way these work and that's the way we expect these to work going forward.

Okay, and was there a cause of the delay?

I think we all recognize issues with land power shell and things like that. Was it something like that or was it something else that sort of caused it?

Our delays right now are exclusively Land Power Shell. We have capacity. We have all materials. We are waiting on the completion of Land Power Shell.

Great. Thank you.

Thanks, Joe.

Thank you. The next question will be coming from the line of... Malik of Citi. Please go ahead.

Hi. Thank you for taking my questions, Andrew. I have a question on the product roadmap. When are you expecting crossover of CS4 versus CS3? And then on just your trial roadmap, in CS4, you improved throughput significantly. And if you stay on that path into CS5 and beyond, can at some point your chip deliver enough fast speed and throughput that your system cannot be used, its system can be used alone and no GPUs are needed?

Well, so, hey, Adil, how are you? Good to talk to you. First, Until a few weeks ago, all our deployments were CS3 alone. It is only recently that we've begun deploying disaggregated solutions and we saw a meaningful bump in throughput, which is really exciting. So that's observation number one. So obviously they work really well in an aggregated form. But this aggregation, which we are now deploying at a GA width, produces even more high-speed tokens when high-speed tokens are constrained. I would say the crossover for CS3 versus CS4 is in T127. The roadmap is Looking really good right now. We are ahead of schedule on the CS5. And while I mentioned in my prepared remarks, getting from tape out to chip in system in five months is extraordinary. And we expect that to begin shipping in early summer, 27. That doubles performance over the four, and the four doubles performance over the three.

Great. And then for Bob, you're guiding two flat cloud sales for Q4. So within that tripling of sales in 2027, can you just help us understand the mix of cloud versus hardware?

I think the best way to think about the mix of hardware is the size of those deals can be very, very large. What we would like to do is provide some color on the shipments over the next quarter or two. It's very reliant on our customers having their data centers ready for us. So what I can say is we think that the hardware mix in Q4 and into Q1 will be closer to 50% than where it's been over the last few quarters. As we look forward into the rest of 2027, we'll have more information for the Q4 call, and I think we'll give a more complete view of our outlook on sort of everything for 2027 then. Thank you.

Thank you. One moment for the next question. Our next question is coming from the line of Quinn Bolton of Needham and Company. Please go ahead.

Hey, guys. Let me add my congratulations on the progress. Andrew, just wanted to make sure I heard you right. In the RPO and your assumptions, you're making assumptions on potential delay credits, but you have not received any cancellations for delayed capacity at this point. Is that correct?

That is correct.

Excellent. Okay, good. The second question is, you know, there's, you know, fears in the market about, you know, fast inference from low-batch GPUs. That sounds like an expensive way to serve tokens. So I'm wondering if you can address sort of the cost per token for Cerberus versus deployment on the low-batch GPU configuration.

Sure. Yeah. The GPU allows you to reduce throughput by trading off for additional speed. And you do that by bringing down batch. So you have a lot of throughput at low speed. And as you try and get higher speed, your throughput emerges to the point where you're in the low batch domain and have very low speed. Now, that's the characteristic of the architecture. The way to think about that is they can get to some modest performance by going to low batch. But when they do that, the cost per token, the cost to generate a token jumps. They have fewer simultaneous customers. So the total number of tokens generated is very small. As you continue to go faster, we become stronger and stronger and stronger. And we get to a point where no matter what you do with the GPU, they can't deliver the performance we can. At low batch, at batch size one, they still can't do it. And so the question is where do you cross over? I think around 250, 300 tokens per second, we are more economical. By the time you get to 500, 800, 1,000 tokens, we are vastly more economical. Somewhere between 800 and 2,000 tokens, the GPU can't play. And that's the way it is right now. That's sort of the lay of the land. That's a good question on a complicated topic.

Thank you, Andrew.

Thank you for your question. One moment for the next question. Our next question is coming from the line of Joshua Beckhammer of TD Cal. Please go ahead.

Hey, guys. Thank you for taking my question, and congrats on the Microsoft deal. Maybe to start, can you speak to us about how you're allocating capacity in light of the land power shell issues? There's a couple layers to that. It doesn't sound like silicon or rack availability is an issue. So I guess how much outlet do you see is there for incremental hardware sales to customers now? And then I guess given the challenges with getting land power and shell for cloud customers, Has there been any discussions with OpenAI about them potentially taking hardware sales and putting them in their own or their partners' data centers? Thank you.

Sure. The question really has two parts, right, Josh? The first is, is it really the case that land power, shell, and the construction thereof is the limiting factor? The answer is absolutely. TSMC has committed... The fact that we don't use HBM, obviously hugely beneficial as that's in very, very tight supply. You are correct in the second part of your question, which is, doesn't that create opportunity for you to sell hardware while your data centers are being built. I think the only caveat to that is that to the best of our knowledge, everybody's data centers are late. That includes open AIs, that includes all the data centers we know being built for hyperscalers. Data centers are late across the board and across the world. So the constraint of data centers isn't unique to us, It is also manifest in our customers. And so, yes, we have supplies for them. Yes, we are shipping to our customers with data centers. Yes, it's the case that they, too, are very late. Make sense?

Yeah, no, totally. But the last part of the annoying three-part question was, you know, is there any Potential for the OpenAI contract to switch to hardware sales in light of the land power show.

Sorry. It is in the contract. They have that right. But as I'm saying, our understanding from them is their data centers are also behind schedule.

Okay. Got it. Thank you. And then for my follow-up, can you help us out?

Seven parts, Josh?

It wasn't my best, I'm sorry. I acknowledge that. Can you maybe talk about how Oracle and Microsoft are positioning your offering with their customers? Is this a case of it'll be hardware sales and then it'll just basically track based on how the demand for ultra-fast inference is for their customers? Thank you.

We've announced more information about the Oracle engagement and partnership that we will set aside for 2027 for details on Microsoft. In Oracle, they are purchasing hardware that is being deployed in their data centers. Those data centers are in the U.S. We are collaborating and they are building pipelines. For that now, there's a ton of interest. and so we do not think this is by any stretch of the imagination the last deal we do together. It is the start of a large partnership in which our equipment will be deployed in Oracle Data Centers and part of the OCI cloud.

Thank you, Andrew, and I'll give up a question next quarter.

You're trading me a hamburger today for a question next quarter?

Fair deal. Thank you so much. That does conclude today's Q&A session. I would like to turn it back to Sean for closing remarks. Go ahead, Sean.

In closing, please note we are attending the Tomorrow X Summit in Austin next week on November 17th, as well as the Barclays Global Technology Conference in San Francisco on December 9th. Thank you for joining us today. Operator, please go ahead and close the call.

Thank you. This concludes today's programming. You may all disconnect.