Cerebras Systems Inc. (CBRS) Earnings Call Transcript & Summary

Thank you, operator. Good afternoon, everyone, and welcome to the Cerebras Systems Q3 2026 Earnings Call. We released our press release earlier today and posted our supplemental earnings presentation to the Investor Relations section of our website. After the call, a replay of this webcast will also be available on our Investor Relations website. 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 want 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. The following contains forward-looking statements, which cover things like our expected financial results going forward, how we plan to run the business, the opportunities we see in the market, what customers will want, our product road map, our position as a technology leader, our supply chain, our operating model, and our expectations for Q4 and the full year 2026, though this list isn't exhaustive. Such statements rest on the expectations and assumptions we hold today, and they carry risks and uncertainties that might make actual outcomes differ materially from what those statements express or imply. Our SEC filings describe these risks, among them the final prospectus tied to our IPO and the periodic reports we will file with the SEC later on. Unless the law requires it, we assume no duty to revise these forward-looking statements. On today's call we will additionally go over certain financial measures that are non-GAAP. You can find the GAAP-to-non-GAAP reconciliations in today's press release and the supplemental materials posted on our Investor Relations site. Andrew, over to you.

Thank you, Sean, and I appreciate all of you joining us today. As we look ahead to the coming months, two things are guaranteed: rapid change and a shortage of compute. During the last 90 days, the leading frontier model has been overtaken multiple times throughout the AI landscape. It seems dozens of compelling open models were released, major security breaches occurred, powerful new consumer agents emerged, and numerous other events unfolded that none of us could even remember. At this point, no one is able to predict AI capabilities 90 days ahead, let alone a full year. Yet despite all this flux, intelligence and speed remain the 2 critical attributes for every use case, whether coding, security, or personal agents. In this rapidly changing environment, our mission at Cerebras is to make sure that Cerebras grows no matter how AI 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 prefer hardware or cloud capacity, we grow. We want to put ourselves in a position where Cerebras comes out ahead regardless of what drives the industry forward, and in Q3 we made outstanding progress toward that goal. First, I want to report that we beat every guided business metric. And although I'm happy with those results, I think it matters even more that since our last earnings call we closed several deals with key AI industry leaders. The most recent one is that we announced today we've signed a significant cloud agreement with Microsoft, and that, together with the AWS and Oracle announcements we made earlier, means there are now 3 hyperscalers who have opted to use the Cerebras platform within their inference strategy. Thanks to these new deals, plus ongoing growth and execution with customers we already have, we're able to lift Q4 core guidance across every metric and still expect revenue to triple in 2027 while gross margins expand. Bob will cover all the specifics on our Q3 results and the guidance looking ahead during his part. I'd like to spend a moment explaining how our business is shaped. We're an AI accelerator business, and what sets us apart is our groundbreaking work in chip systems and software that speed up AI computation. We're an AI hardware firm that reaches the market through 2 channels. The first is selling systems — we install racks and Cerebras hardware inside customer data centers. That's a strong business: revenue gets booked fast and cash comes in fast, though it's uneven and hinges on customers having data center capacity ready for the hardware. Our second channel is our cloud. A quickly expanding group of AI natives would rather consume tokens than deploy hardware themselves. Since compute and data centers are so scarce right now, owning both the systems and the facilities ourselves gives us a real strategic edge. When you build your own hardware, the AI cloud business becomes an especially attractive one. Our per-megawatt cost is much lower, which lets us achieve better capital efficiency than the neo-clouds. The blend between these two channels will vary from quarter to quarter, though we expect substantial growth in both over time. I'd also like to address the speculation about our OpenAI partnership. OpenAI is a very large customer and a close partner, and I believe that will remain the case going forward. OpenAI expanded from Q2 into Q3, and we expect that growth to continue. The response to OpenAI's ultrafast SOLV6 preview, which runs on Cerebras, was overwhelming and made it unmistakably clear that appetite for fast focus is huge. It's also the case that Cerebras didn't have sufficient capacity to back ultrafast across a wide rollout, Astra included. As others have noted, we see capacity as a central factor in business decisions within our industry. Indeed, despite the vast supply of GPUs OpenAI can access, restrictions tied to usage still had to be put in place. Once we had signed that agreement with OpenAI, the process of contracting for and activating the large added data center capacity required to satisfy the demanding terms of the deal took more time than we wanted. As a result, we are now behind where we planned to be on the first tranches. Because this partnership matters so much to our business, we recognize why investors are watching it so closely. Information about the agreement, or updates concerning it, will probably not emerge from anyone other than Cerebras or OpenAI. To give you an example, there were rumors going around that servers wouldn't be able to handle post-training for technical leads — those claims are wrong, since CSI, CS4 and CS5 are fully capable of handling post-scale models as well as larger ones. Also, I expect — and you should too — that OpenAI will build out capacity across a variety of hardware solutions. OpenAI is going to retire older models and roll out new ones on those compute options, both for internal and external use. Their compute requirements are enormous, and there's ample room for multiple solutions to contribute as they scale. Before I move on to our business framework, let me add one more general point. Even with all the experimentation going on—low batch GPU, the start-up announcement, Metacell, and hyperscalers building their own custom silicon—we're still the fastest inference solution partner out there. And we think our road map will keep that edge for years ahead. Our way of judging progress hasn't changed. We look at the business through three lenses: customers, capacity and capabilities. Starting with customers: we're growing our customer base quickly to extend our reach and make our revenue stream more diverse. Q3 was a good quarter on that front. We scored a significant win with two new hyperscalers. We've reached a deal to bring Cerebras hardware into Oracle Cloud, with Cerebras set to be deployed across Oracle data centers that serve AI labs, enterprises, and developers who are building interactive and intelligent applications. We expect this Cerebras-powered inference service on Oracle Cloud to launch during the summer. Beyond that, we've signed an exciting agreement with Microsoft, and we'll share more details on that next year. All told, that gives us 3 hyperscaler wins over the last 6 months. It's a significant accomplishment, and it puts us in a strong position for 2027 and 2028. We expanded our reach even more by working with a variety of neo-clouds and top-tier off-takers, one example being fireworks. That widens the channels through which we sell and gives customers more options for tapping into Cerebras high-speed inference. Open weight models have helped us too, because they continue to make new applications possible and push further what AI can do in daily life and across the industry. As an example, this very week we announced that Cerebras's Fast Inference is driving the built neighborhood OS platform and personal assistant, bringing extremely rapid inference to millions of users; through consumer apps such as built, we are putting Cerebras's capability right into people's pockets. That is the reason Fast Solos could charge a premium, and cloud clients report that fast tokens deliver outstanding gross margins and that the fast tokens from Cerebras are the most profitable ones they offer. Now let's discuss capacity, which breaks into two parts — data center capacity and manufacturing capacity — and each is essential for growth.

Data center availability is still a limiting factor throughout the industry, and we're no different. On top of that, we think the shifting policy environment in the U.S. may pose bigger obstacles for building data centers there. By 2027, over 75% of our data center capacity will come online outside the U.S. As a matter of fact, within the next several months, new facilities in Canada, France, and Norway will come online. And in the quarter, we signed contracts for extra capacity in Norway and Finland, which further expands our European footprint, and we are in the final stages of contracting for a data center in Sao Paulo. That's our first in South America; to date, we have secured roughly 700 gross megawatts of data center capacity for delivery by the end of 2027. Our first-party cloud already ranks among the larger non-hyperscaler AI clouds, and we are far along in negotiations for gigawatts of capacity. Last, the best method to raise data center capacity is to raise the token generating capability of the hardware in the data center, quite simply to generate more tokens per megawatt. Toward that end, this aggregated inference helps expand our capacity by boosting by up to 5x the number of tokens the data center can deliver. More super fast focus per megawatt means every data center generates noticeably more revenue. Moving now to the second element of capacity, manufacturing and supply; we are scaling manufacturing with Flex and Sanmina. We expect to exit 2026 with more than 10x the manufacturing capacity we had when the year began. In 2027, we intend to add factories and further raise capacity to support the substantial growth we anticipate. Our supply chain is well positioned to support that growth as well. Most importantly, we have safety, which keeps distinguishing itself as a world-class partner. What our supply chain does not need is also important; that matters to us too. Unlike other competitors, we do not use HBM memory; we believe memory shortages and the price increases tied to them will persist through 2028, raising competitors' costs and limiting supply. Memory challenges across the industry create an opportunity for Cerebras. Factories and supply are necessary but not sufficient to scale capacity. We must also have high deals. Cerebras has invented a fundamentally new computing architecture, and applying metrics and mental models designed for other architectures can be like examining a Tesla and asking where the exhaust pipe is — we're 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 we'll work to share architectural details more clearly, but there is also a set of persistent falls with prophetize the financial community; they believe that our systems can support large models, obviously, falls as they currently do and that our manufacturing yields are low. Both falls. I already addressed the model side earlier in my comments and in previous discussions. Because manufacturing yields are tougher to see, I'll spend a moment addressing that subject head-on. Yields come in 2 kinds: fab yield and parametric yield. The yields we get from TSMC — the fab yields — are high. Our parametric yields are high as well, and while our technical blog goes into this in detail, I'll give a quick summary. Every wafer has defects, and for decades those defects were the main reason the industry viewed wafer-scale processors as unworkable; we fixed that issue using redundancy. Roughly 1 million identical tiles sit in our design, with a small fraction kept as spares. If a defect takes out a tile, the hardware routes around it and switches in a spare. That mechanism functions at the fab and out in the field alike, giving us high wafer yield and solid in-field reliability. Parametric variation is the second challenge. Even when the silicon works, its electrical characteristics differ across the wafer. Certain regions run quicker or slower than others at a given voltage. That's a fundamental fact of chip design, and 1 of the hurdles we must clear to make wafer-scale computing feasible. We tackled parametric variation in 3 ways. First, we cut off the outer edges of the wafers. Those edges are the weakest portion of a wafer and account for a large share of the parametric variation. We use the biggest square that fits inside a 300-millimeter wafer, which retains much of the needed silicon. Second, we design and test for a wide range of conditions. Our cores put up with a broader spread of variation. During timing closure — the process of verifying that circuits meet timing requirements — we validate the design across operating conditions far broader than others in the industry use; that lets us hit expected performance despite variation across the wafer. Last, we tune voltage independently across the wafer — we invented technology that adjusts core voltage independently across 84 wafer sections; we can raise voltage in a slower section or lower it in a fast one, which gives us precise control over performance. We know of no other chip with this capability. The outcome is exceptionally high yield at both the fab and the factory. We got there by engineering solutions to the problems others say make wafer-scale computing impossible. As a final note, some have pointed to our relatively low gross margins on hardware shipments and suggested this proves a yield problem. In fact, it simply mirrors the ongoing ramp of our manufacturing capabilities and the final shipments of hardware covered under special pricing programs offered to 1 of our early customers. We expect our hardware margins to improve markedly over the next several quarters, eventually surpassing our long-term target of 60%. Turning now to capabilities: we are advancing them on every front — new systems and processors, better software, advanced manufacturing techniques, and partnerships that deliver best-in-industry solutions. Even with the rapid pace of change across the AI industry, inference speed keeps growing in importance. Fibers is still clearly number one, and we think that edge will last. With that in mind, our product roadmap keeps moving forward fast. Our CSI is about to reach full production, which doubles performance and throughput. Our next-generation processor, the wafer scale engine 5, set to be deployed in our fifth-generation CS5 system, has come back from the fab ahead of schedule. The WSC 5 taped out on June 15. The chip is back, has been brought up, and is already running models that is undergoing in chassis testing, and going from tape-out to system tests in under 5 months is exceptional execution. The CS5 will more than double the site in CS4, — it will enter production in early summer of 0.7 is built on TSMC's 5-nanometer node for competition for wafers is less intense and wafer costs are much lower. Our innovations also reach into complete solutions. As far as we know, Servers is the first to deliver production scale, heterogeneous disaggregated inference. Right now, our cloud serves customers using AMD GPUs for prefill and cerebral Cerebras systems for deCODE. To make the production launch possible, we built a disaggregated infant software platform. It is built to work with AMD as well as other GPUs.

That same solution is going to be rolled out across the Severence cloud as well as to certain data center customers. Delivering inference in this aggregated way produced considerably more high-speed tokens using the same amount of power. So cloud owners get more profit per megawatt, and AI users get a lower cost per token. We think DSG is going to become one of the default offerings in major cloud deployments. Our software stack keeps maturing. Our cloud services are delivering impressive uptime, our systems are continually getting faster, and the time it takes to bring up new models has dropped to only a few days. Cerebras served as a day 0 launch partner for KimiK3.1, the $2.8 trillion parameter frontier class open weight model from Moonshot AI, which may be the premier open weight model on the market; KM-1 on Cerebras isn't just the fastest in the industry, it also runs at up to 2,000 tokens per second. Nearly 11 years into our journey, we keep innovating at a tremendous pace. Our team has remained together through good times and tough times, and there's more energy and excitement in our building than ever; we take pride in tackling and solving some of the industry's hardest problems. To summarize, we beat on all metrics, and we're raising guidance for Q4 across the board and expect to triple revenue in 2027. We made important progress in each of our key domains, including at 2 hyperscale customers, and have now won 3 hyperscalers in the last 6 months. We began shipping our CSI, our fifth generation processor, and the WSC5 is back and in CS5 system test in record time. Production disag is now live with AMD at day 0 support for yet another $2.8 trillion parameter model, which successfully achieved an industry best performance; with that, let me turn things over to Bob.

Thanks, Andrew, and good afternoon to all of you. Before diving into the details of our Q3 results and Q4 outlook, I want to step back and talk about how we've grown the business. We made the decision to ramp a cloud offering in early 2025 because, in this fast-improving AI landscape, AI natives are eager to scale compute quickly and prefer not to own data centers. Since hyperscalers plan their data center compute allocation years ahead of time, they frequently lack near-term availability in their own facilities for new hardware. To land the large deployments that foundation labs and hyperscalers require, Cerebras needed to prove our systems could deliver our industry-leading inference performance at scale. After those customers have thoroughly evaluated what we offer in our private cloud, they’re then able to choose between purchasing our hardware and buying our cloud offering. In order to provide sufficient cloud capacity quickly and meet these demands, we made the decision to begin leasing data center capacity and to expand our private cloud toward gigawatt scale as rapidly as possible. This overall strategy appears to have been sound, because data center capacity is now very scarce, and we hold a great deal of it under contract. Even more importantly, having our private cloud in place allowed us to demonstrate the value of fast inference to the whole market and to generate more demand for it. Getting big new data center capacity up and running typically takes between 15 and 18 months. Instead of sitting around waiting for those facilities, and to accelerate how quickly we could address the overwhelming demand already in front of us, we leased back a portion of our own equipment from one of our customers. As we've noted previously, that step put short-term pressure on gross margin, but it also enabled us to expand business with existing customers and land major new ones, since we had capacity on hand when they were ready to purchase. We've now secured frontier labs like OpenAI and hyperscalers like AWS, and we recently brought 2 additional hyperscalers into our customer base. Additional top-tier clients—Cognition, Sigma, Block, and Belt, plus others—operate on our cloud platform. We engage these organizations on their own terms, sometimes providing hardware and other times offering our cloud-based services. Before long, newly signed neo-cloud partners will begin delivering rapid inference offerings powered by Cerebras equipment inside their own cloud environments. As we mentioned on earlier calls, 2026 is slated to be a period focused on expanding capacity and capabilities while bringing on more customers, which should enable us to grow revenue by more than threefold in 2027. Thanks to the substantial momentum we've achieved thus far this year, we are strongly positioned to reach that goal. Moving on to Q3 results, I want to talk about how we executed during the third quarter, and I'm glad to report that Q4 is set to mark a significant turning point for our growth and our core financial trends; for the third consecutive quarter, we came in above our core financial guidance on every measure. We also posted positive non-GAAP net income for the first time since we became a public company, and we reached that milestone earlier than planned. As a reminder, the definitions of our core metrics and the corresponding reconciliations to GAAP are included in today's earnings release and are available on our website. Core revenues hit a record $231 million — that's a 10% sequential increase and 70% year-over-year growth, and it came in $16 million, or 7%, above the $215 million midpoint of our guidance. Core cloud and other services revenue also reached a record, climbing 272% year-over-year to $145 million. In Q3, cloud revenue was up 13% sequentially; looking at Q4, our current expectation is that cloud revenue will be about flat sequentially. This trend isn't being driven by pricing or by demand. Pricing remains healthy, and demand for our solutions continues to outstrip supply by a wide margin. The only thing holding back our cloud revenue growth is data center capacity. We'll begin bringing that capacity online in December, building momentum as we head into next year. Sadly, that timing is too late to make any difference in Q4. We anticipate far bigger gains in data center capacity over the course of 2027. Most of that kicks off around the middle of next year, when we bring online larger shares of the 700 gross megawatts of data center capacity we currently have under contract. Core hardware revenue came in at $86 million for the quarter, a 5% rise sequentially. Demand from customers for rapid inference continues to climb, and we're meeting it both via our private cloud and via fresh hardware agreements signed with a number of new clients. Since those hardware deliveries are scheduled to start in Q4, overall core revenue is going to rise substantially. Between Q3 and Q4, core revenue growth on a sequential basis will come close to 40%, while the year-over-year figure will land around 85%. Shipments under these hardware orders — Oracle among them — will continue through at least the first half of 2027. We also disclosed today, as Andrew noted earlier, that Microsoft has become a new customer. During the last several months, the opportunities in our pipeline have expanded well beyond a tenfold increase. That growth came from a wide mix of chances involving new prospects as well as current customers, among them hyperscalers, Neo clouds, sovereigns, enterprises and AI natives. Now shifting to gross margin: core gross margin for Q3 came in at 41%. That marks a 40 basis point gain over Q2 and lands 200 basis points higher than the guidance we had set for the quarter. Beyond the effect on gross margin from running back our systems, we also encountered certain headwinds during the quarter as we worked through the final shipments under a hardware purchase agreement signed years earlier at lower pricing. We had told you that all of this would result in gross margin declines of a temporary nature, with the trough coming in Q3. Well, we outperformed that guidance, and the trough actually occurred in Q2, the quarter just ended. Now I'll separate Q3 core gross margin in total into its parts. Core hardware gross margin came down on a sequential basis to 31%, which was driven by the last shipments under the customer agreement that is a year old. At present we are fulfilling hardware orders at prices considerably higher than those before, and that likewise means significantly better margins starting in Q4. Gross margin from core cloud and other services increased substantially, hitting 46.9%, a figure that is 510 basis points above the Q2 level. The main factor behind that gain was new business that came in at better pricing and margin. That improvement was partly offset by our initiative of running back data center capacity as well as our hardware at elevated cost from a customer. We anticipate this will persist at least until the middle of 2027, when we will have sufficient data center capacity of our own to meet our strong cloud demand. Against that background, for Q4 we project total core gross margin to step up by 700 basis points to 48% at the midpoint; we expect total core margin to surpass 50% in the near term and keep climbing toward our long-term target of 60%. We also made good headway on profitability during Q3 and across the year so far. Core operating margin improved from the prior quarter by 170 basis points, landing at a negative 14.3% — a result that came in nearly 10 margin points better than the midpoint of our guidance.

Core operating loss came in flat for Q3 compared with Q2, at roughly $33 million. Core net income swung to positive, reaching $30 million. Adjusted core EBITDA for the year to date was positive as well, at $2 million. Given the strong revenue and gross increases we're seeing in Q4, we anticipate that every one of our core key profit measures, operating margin included, will be positive in the coming quarter. Moving now to RPO and the OpenAI contract. Remaining performance obligations rose to $26.1 billion as of September 30; we refresh RPO for new business signed during the quarter and revise existing contract values to reflect any changes in estimates of what remains to be recognized across the remaining committed term. Q3 RPO excluded the effect of new deals inked after September 30, Oracle among them, which is large. Many questions have come up regarding where our OpenAI contracts stand, since that relationship accounts for the bulk of reported RPO to date. As a reminder, the OpenAI contract, filed with our S-1 ahead of going public, is a take or pay arrangement covering 750 megawatts of capacity. Among the many contract requirements binding both parties, it contains a term under which, if we are late in delivering capacity, they are entitled during that period to credits usable against future committed purchases. We computed the figure for potential delay credits and subtracted it from RPO at September 30; the revenue and gross margin impact of these credits would be minimal, since relative to RPO they are estimated to be fairly small, and they would then be spread ratably as the bulk of revenue is recognized over the next few years. Any future revisions to these estimates will be made each quarter and reflected in both RPOs and guidance, if needed. Should a capacity delivery delay persist for 90 days or more, OpenAI and we both hold the option to cancel only that portion of capacity. The amount we estimate could be subject to cancellation this year would be under 5% of total committed megawatts. With demand for data center capacity relentless worldwide, finding and bringing up the initial data center capacity needed to hit the delivery schedule has taken us extra time against a very aggressive timeline. Based on our team carrying out frequent in-person inspections at each site, we now expect to be back on schedule by the second half of 2027, which should lower future delay credit risk. We've taken a conservative stance and have not assumed any of these delays might be excused because of dependencies beyond our control. If that occurs, it would shrink the delayed credits we have assumed; moreover, should OpenAI or Cerebras choose an option to cancel a delayed capacity tranche, we believe demand is sufficient and we would find other customers to absorb it. OpenAI still holds an option to buy another 1.25 gigawatts on top of the initial 750-megawatt commitment, and that extra capacity could begin coming online in 2029. Because that point is still 3 years off and AI is evolving at such a rapid pace, we have not even started discussing it yet. Moving now to the balance sheet and capital efficiency. Liquidity remains strong: we closed Q3 with $7.8 billion across cash, cash equivalents, restricted cash and marketable securities, and we also have a revolving credit facility of up to $850 million that has stayed unused so far. The $1 billion working capital loan received upfront in January, together with ongoing payments for nonrecurring pass-through expenses tied to fitting out new data center capacity to be delivered under the OpenAI contract, has covered most of our net working capital needs and capital expenses through 2026 to this point; we are not a neo-cloud. We are vertically integrated and the inventor of the industry's fastest inference solution. Our business carries a degree of durability and technology differentiation that we believe will support much greater long-term margin potential and cash generation over time. Our model also calls for less capital investment given the mix of compute equipment we deploy; for a substantial share of the equipment in our data centers that we manufacture ourselves, much like others who have chosen to build and deploy their own AI chips, we avoid paying a third party 70% gross margins. nor do we rely on high-priced HBM. We pay only our BOM cost, which lowers the relative capital expense per megawatt, while we also gain from producing the fastest and most valuable tokens. That is visible in our capital structure so far. The sole debt on our balance sheet, despite our fast private cloud build-out, is the open AI working capital loan. We have also received, and keep receiving, upfront payments from customers, which gives us another lever for capital efficiency. Our capital structure today features substantial cash and investments on hand with little leverage, giving us more flexibility and multiple paths to keep funding our rapid growth, which is an advantage, particularly in the present capital market environment; turning now to our outlook. For Q4 2026, we expect core revenue in the range of $310 million to $330 million, core gross margin in the range of 47% to 49% and core operating margin in the range of minus 1% to 1%. For the full year 2026, we are raising core revenue to the range of $940 million to $960 million; we are raising core gross margin to the range of 42% to 44% and we are raising core operating margin to the range of minus 10% to minus 8%. To close, Q3 was another record quarter for Cerebras and Q4 will be too. Selling both hardware and cloud through our flexible model is producing robust revenue growth along with gross margin gains that arrived one quarter sooner than we had anticipated. We have achieved considerable advancement in expanding our manufacturing capability and data center capacity, and that will let us triple revenue while core gross margin moves toward our 60% long-term target in 2027. Strong operating leverage has been demonstrated, and we project core operating profit at roughly breakeven in Q4. After that, during 2027, we expect substantial movement toward our long-term 40% operating margin target — with this combination of scale, growth and profitability, we stand out as a unique and leading provider of AI systems. As a final point, I want to remind everyone that we have now been a public company for over 180 days. Accordingly, our IPO stock lockup releases were entirely finished as of November 9. Thank you for listening to our prepared remarks. I would like to ask the operator to please open the line for questions.

Great job on those strong results. My first question concerns the new agreements with OpenAI and Microsoft. Could you share any metrics, maybe on how big those new deals are relative to each other? And then, Bob, you said part of OpenAI is already included in the RPO — could you tell us how much of it is already in the RPO?

Tom, Andrew here. Good to hear from you. We'll be sharing further details about the Microsoft opportunity in the deal we signed with them. We're clearly very excited about it; more detail will be announced during the first half of next year, and that's when we can expect revenue from the transaction to begin. Bob?

Tom, regarding the RPO, what we included for OpenAI is our estimate of the committed revenue amount under the take-or-pay contract, with no extensions and no additional options, adjusted by the delayed credits we think could be due. It also covers the data centers we've signed for them; it includes what they pay for nonrecurring setup costs as well as monthly recurring costs across the contract term. And those are the pass-through revenues and pass-through expenses we recognized in the — got you, Bob. I said opening, I meant Oracle, but we can come back to that. But I guess the other one is just ability from here — we fly really hard to win — scalers that the after forces even easier; so far, no Oracle in our view. We signed it up — all right.

Now for my follow-up: Frontier Airlines clearly can do more than just manufacture hardware quickly the way Calian does — they can also accelerate getting models deployed across different kinds of compute hardware, which increases fungibility. So how do you folks benefit from that deployment? And what do you expect heterogeneous infrastructure to look like down the road? In what ways does that play to your advantage?

Sure. Tom, that's a genuinely good question; and rate in there, I believe in their S-1 or on a — we talked about how standing up AMD GPUs took them about 3 days; and that's a very interesting and significant development, particularly for anyone with a long-term focus. What this implies for our ecosystem, I believe, hasn't truly sunk in for most people: compilers and low-level software have long stood as a core obstacle for architectures. They tilted the field toward the incumbent — they were that software bottleneck you kept hearing about. If the profit went toward doing what they claim to do, then the worth of this low-level software contracts and in time vanishes; and it also means that the most from that low-level software shrink as well, digitally speaking, and when low-level software holds only a little value, what remains to deliver value is innovative hardware. And it means that kind of differentiation will hinge more on hardware differentiation. In my view, that gives a big edge to hardware innovators such as Cerebras and similar firms creating novel or alternative architectures; and I believe it will take away a benefit the established players have held for a long time — that's an excellent question, and I think it's something the industry should watch closely. Thank you.

And well done on those solid hardware figures. Could you address the RPO matter regarding cloud scheduling — is there anything particular we should watch, and how would you describe your confidence in your capacity to sort that out before it becomes extra penalties?

What we included in the RPO, I believe, was a highly conservative figure representing the absolute worst case we could face. So we feel comfortable today that we will do better than that. Such provisions are common in large data center deals. We have comparable arrangements with our data center builders, their penalties when they are late — and I think it shifts over time whether you decide to apply them, and you make a business decision at that moment. I believe that is how these things operate, and that is how we anticipate them operating going forward.

All right. And what was behind the delay? I mean, I believe we're all aware of problems involving land power shell and similar things. Was that the kind of thing it was? Or did something different bring it about?

At the moment the delays come solely from land power shelf; we have capacity. Materials are all in place. What we're waiting for is the completion of and power shell.

Andrew, I've got a question about the earlier road map. At what point do you expect CSI and CSI to cross over? Is that going to happen in 2027? And then, just looking at your trial road map, in CSI you made a significant improvement to throughput. If you keep going down that path into CSI and past it, could your chip eventually provide so much speed and throughput that the system could be used on its own, with no GPUs required?

Good start — to answer the first part, up until a few weeks ago, every deployment we had was CS3 by itself. Only recently have we booked deployments of disaggregated solutions, and we observed a meaningful bump in throughput, which is really exciting. So that's my first observation. Obviously, they perform really well in aggregated form; but this aggregation, which we're now deploying and sort of GA with, delivers even more high-speed tokens when high-speed tokens are constrained. I'd say the crossover for CSI, where CSI sits in Q1. The road map looks really good at the moment; we're ahead of schedule on the CSV. And as I said in my prepared remarks, going from tape-out to chip in system in 5 months is extraordinary. And we expect shipping to begin in early summer, '27, and we are — that doubles performance over the floor double performance over.

Great. Then turning to Bob — you're steering toward flat cloud sales in Q4, so inside that tripling of sales in 2027, could you help us grasp how cloud versus hardware will be mixed?