Cerebras (CBRS.US) Q3 conference call: CEO admits OpenAI delivery lags behind, upward guidelines expect 2027 revenue to triple
The Zhitong Finance App learned that AI chip and cloud service provider Cerebras Systems (CBRS.US) announced the results for the third quarter of the 2026 fiscal year after the market on November 12 EST. Core revenue for the quarter increased 70% year over year to US$230.9 million, higher than LSEG's average analysts' estimate of US$214.7 million; adjusted net profit of US$29.82 million, a loss of US$10.86 million for the same period last year, and the market originally expected a loss of US$38.4 million. The core revenue guide for the fourth quarter was US$310 million to US$330 million, about 16% higher than the market forecast of US$274.9 million; the full year core revenue guidance was raised from US$880 million to US$890 million to US$960 million.
The real focus of this call was on the OpenAI contract that supports the $26.1 billion remaining performance obligation (RPO): due to delays in data center delivery, the company has deducted potential delay compensation from the RPO. Co-founder and CEO Andrew Feldman acknowledged that the company was behind schedule in the first few capacity deliveries and attributed the reason to land, electricity, and plant construction rather than chips or materials. He reiterated that revenue in 2027 will approximately triple this year.
Focus and response: The OpenAI contract was delayed, and the RPO has already calculated compensation
What is most noteworthy about the conference is not the three-quarter report, which has already exceeded expectations, but the OpenAI contract, which accounts for the majority of RPO. Chief Financial Officer Bob Komin revealed that the contract is a 750 megawatt take-or-pay (take-or-pay) agreement, which stipulates that if the company's delivery is delayed, OpenAI is entitled to a deduction amount that can be used for future promised purchases during the delay period; the company has estimated the potential delay compensation amount and deducted it from the September 30 RPO. If delivery delays continue for 90 days or more, both parties can choose to cancel the corresponding portion of production capacity. The company estimates that the scale of cancellation this year may be less than 5% of the promised total megawatts.
According to Komin, the company uses the conservative “absolute worst case scenario” approach and expects to return to delivery plans by the second half of 2027, thereby reducing the risk of future delayed compensation. Feldman added that the current delay “comes entirely from land, electricity and plant”, “production capacity and materials are ready”, and the electricity and plant are awaiting completion. At the same time, he said that due to the general extension of data centers around the world, even if cancelled, the company has enough demand to handle it.
Faced with Morgan Stanley Joseph Moore's question about the cause of the delay, Feldman gave the above answer; Needham's Quinn Bolton confirmed that the company has not received any cancellation notices for delayed production capacity so far.
Q4 guidance significantly exceeded market expectations
The company set the core revenue guide for the fourth quarter at 310 million to 330 million US dollars, an increase of about 85% year-on-year and 40% month-on-month according to the median value; the core gross margin guide is 47% to 49%, and the core operating margin guide is negative 1% to positive 1%, which means that the core business is expected to approach break-even in the fourth quarter. Komin said that the fourth quarter will be an important inflection point for the company's growth and core financial trends. The jump in revenue is mainly due to hardware orders that began to be delivered in the fourth quarter.
Cloud revenue leveled off month-on-month in Q4, and production capacity was the only constraint
Core cloud and other services revenue for the third quarter was US$144.9 million, up 272% year over year and 13% month on month, accounting for about 63% of core revenue. However, the company expects cloud revenue to remain roughly flat month-on-month in the fourth quarter. Komin stressed that this is not a pricing or demand issue — “the pricing environment is healthy, and demand still far exceeds supply”; the cloud revenue climb is “completely limited by data center capacity”. Additional production capacity will not be released until December, too late to be reflected in the fourth quarter.
Hardware revenue has declined, and the low level of gross margin has passed
The hardware business's core revenue for the third quarter was US$85.96 million, down about 11% year on year; under GAAP, it fell from US$96.79 million in the same period last year to US$68.88 million. Komin explained that the low gross margin of hardware stems from the last batch of low price orders under the purchase agreement signed many years ago, and manufacturing capacity is still climbing; as new orders signed at current prices are delivered from the fourth quarter, hardware gross margin will improve significantly. He pointed out that the low core gross margin actually occurred in the second quarter, not the third quarter as previously indicated.
Data center capacity: 700 megawatts, over 70% outside the US
The company has contracted about 700 gross megawatts (gross megawatt) of data center capacity and plans to deliver it by the end of 2027. Feldman said that in 2027, more than 75% of production capacity will be launched outside the US, and new facilities in Canada, France, and Norway will be put into operation in the next few months; the company also signed an expansion contract between Norway and Finland this quarter, and has entered the final signing stage for the Sao Paulo data center in Brazil — this will be its first data center in South America.
Acquired three hyperscale customers in six months
The company added two new hyperscale cloud service providers this quarter: Oracle will deploy Cerebras hardware in its OCI cloud to provide services for AI laboratories, enterprises, and developers, which is expected to be launched in the summer of 2027; the other is Microsoft, with details to be disclosed next year. In addition to the previously announced AWS, the company has won a total of three hyperscale customers in the past six months. Feldman said this laid the foundation for 2027 and 2028. Additionally, Bilt's Neighborhood OS platform and personal assistant are already inference supported by Cerebras.
Product route: CS-4 release, WSE-5 early return film
The CS-4 is about to enter full mass production, and both performance and throughput will double. The fifth-generation wafer-level engine WSE-5 completed streaming on June 15. The model has now been screened, completed, and the model has started operation. The whole machine is being tested. It takes less than five months from the flow film to entering the system. The CS-5 system equipped with WSE-5 will be more than double the size of CS-4. Using TSMC's 5-nm process, it is planned to be put into operation in early summer 2027.
Implementing production-level heterogeneous decoupling inference for the first time
The company called it the first company to deliver production-level heterogeneous decoupling inference (disaggregated inference): it currently uses AMD GPUs for prefilling (prefilling) and Cerebras systems for decoding (decoding) in its own cloud. Feldman said that the solution can increase high-speed token output under the same power by up to 5 times, which is a key lever for the company's data center expansion, and believes that decoupled reasoning will become one of the default options for mainstream cloud deployments.
Responding to yield questions: Defects are circumvented by redundancy
In response to questions from the outside world about the yield of wafer-level chips, Feldman said the relevant claims were “all wrong.” He explained that the company circumvented the defects through a redundant design of about 1 million identical units (tiles) and switched the bad points from hardware to backup units. The mechanism was effective in both the fab factory and the field; the parameter yield was solved by three methods: removing the outer edge of the wafer, expanding the design verification conditions, and independently regulating the 84 wafer partitions. He stressed that the low gross margin of hardware reflects rising production capacity and special pricing for early customers rather than yield issues. It is expected to improve significantly over the next few quarters and eventually exceed the long-term target of 60%.
$7.8 billion in cash, the only debt is the OpenAI revolving loan
As of the end of September, the company's cash, cash equivalents, restricted cash and short-term investments totaled US$7.8 billion, while an additional US$850 million in revolving credit had not been used. Komin stressed that the company “is not a new cloud service provider (neocloud),” but rather a vertically integrated self-developed hardware vendor. It does not need to pay 70% gross profit margin to a third party, does not use expensive HBM, and “only pays material costs”; despite the rapid expansion of private cloud construction, the only debt on its balance sheet is OpenAI's $1 billion revolving loan.
Full minutes of the meeting
I. Executive speeches
Andrew Feldman (Co-founder, CEO and President): Thank you Sean and thank you all for attending today. Looking ahead to the next few months, we see two certainties: rapid change and a shortage of computing power. Over the past 90 days, the title of the leading cutting-edge model in the AI ecosystem has changed hands several times, dozens of interesting open models have been released, major security incidents have occurred, and powerful new consumer intelligence has emerged. No one can predict AI capabilities after 90 days, let alone a year from now. But in the midst of all this change, intelligence and speed remain two core elements of every use case—from programming to security to personal intelligence.
In such a rapidly changing world, Cerebras's mission is: no matter how AI grows, Cerebras can grow. The basic model grows, we grow; the open model grows; the individual intelligence grows, we grow; whether the enterprise or startup grows, we grow; whether the customer wants hardware or cloud capacity, we grow. Our goal is to position ourselves in such a position: no matter what the industry is driven by, Cerebras can win.
In the third quarter, we made significant progress towards this goal. First, I would like to report that we have completely surpassed all the guiding indicators. While I'm happy with these results, I think more importantly, since our last performance meeting, we've reached a series of deals with a number of key AI industry leaders.
The latest is an important cloud cooperation agreement with Microsoft announced today. In addition to the previously announced AWS and Oracle, we have already chosen three hyperscale cloud service providers to include the Cerebras platform in their reasoning strategies. These new deals, along with continued growth and execution with existing customers, enabled us to comprehensively raise our core guidance for the fourth quarter and continue to expect revenue to triple and gross margin to expand in 2027. Bob shares all the details. Let me take a moment to explain the shape of our business. As an AI accelerator company, our pioneering innovations are chips, systems, and software that make AI computation faster. We are an AI hardware company with two market channels.
First, we sell systems to install racks and Cerebras hardware to customer data centers. It's a good business: revenue recognition is fast and cash recovery is fast; but it's also volatile, depending on whether the customer already has the data center capacity to receive hardware.
The second pathway is through our clouds. There is a rapidly growing group of AI native customers who want to consume tokens rather than deploy hardware. Considering the extreme shortage of data centers and computing power, having a system and data center at the same time is a strategic advantage. The AI cloud business is a particularly good business for companies that manufacture their own hardware — our cost per megawatt is much lower, so it can bring higher capital efficiency than new cloud service providers (neocloud). The revenue structure of both channels will change from quarter to quarter, but we believe both will grow significantly over the long term.
I'd also like to respond to speculations about our collaboration with OpenAI. OpenAI is a very big customer and a close partner, and I believe this will continue to be the case in the future. OpenAI usage increased in the third quarter compared to the second quarter, and we expect more growth in the future. OpenAI's Cerebras-based GPT-5.6 Sol Ultrafast (Extreme Speed) preview was an overwhelming success, unmistakably proving that the market demand for fast tokens is huge.
It is also true that Cerebras does not have enough capacity to support the widespread rollout of Ultrafast. As others have said, we believe capacity is a key driver of the industry's business decisions. In fact, even though OpenAI has a huge GPU supply, it still has to implement usage limits. After signing the agreement with OpenAI, it took us longer than we hoped to sign and launch the significantly increased data center capacity required to meet the aggressive requirements of the agreement. As a result, we fell behind schedule on the first few deliveries.
Given the importance of this partnership to our business, we understand investors are paying close attention to it. 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 models — these are all wrong. CS-3, CS-4, and CS-5 are all fully capable of supporting leading-edge and even larger models.
At the same time, I expect, and you should expect, that OpenAI will use a variety of hardware solutions to build capacity. OpenAI will decommission the old model and deploy the new model to these different computing power options for internal and external scenarios. Their computing power requirements are huge, and in the process of their expansion, different solutions have plenty of opportunities to participate.
Before I get into the business framework, I would like to say one more thing. Despite announcements from low-volume GPU startups, Meta, and self-developed chips from hyperscale manufacturers, we are still the fastest inference solution partner. We are confident that our roadmap will enable us to maintain this advantage for years to come.
Our framework for evaluating progress has not changed. We evaluate the business from three dimensions: customer, production capacity, and capacity.
customer. We're rapidly growing our customer base to broaden our reach and diversify our revenue streams. The third quarter was a successful one in this regard. We have achieved an important victory — adding two new hyperscale cloud service providers. We signed an agreement to bring Cerebras hardware to Oracle Cloud. Cerebras will be deployed in Oracle's data centers to serve AI laboratories, enterprises, and developers building interactive intelligent applications. We expect this Cerebras-powered Oracle Cloud Inference service to go live in the summer. We've also signed an exciting agreement with Microsoft, and more information will be shared next year.
This has enabled us to win a total of three hyperscale customers over the past six months. This is an important achievement that lays a good foundation for 2027 and 2028. We are also expanding our coverage through a number of new cloud service providers. Our undertakers include high-level customers such as Fireworks, thereby diversifying market channels and giving customers more access to Cerebras high-speed inference methods.
We also benefit from an open weighting model. Just this week, for example, we announced that Cerebras Quick Inference is supporting the Bilt platform and personal assistants, bringing extreme inference to millions of people. With consumer-grade apps like Bilt, we put the power of Cerebras directly into people's pockets. That's why fast tokens get a premium — our cloud customers tell us that fast tokens provide excellent gross profit margins, and Cerebras' fast tokens are the most profitable tokens they sell.
Production capacity. Production capacity consists of two parts: data center capacity and manufacturing capacity, both of which are critical to growth.
Data center availability continues to be a constraint across the industry, and we are no exception. Furthermore, we believe that changes in the US policy environment may bring greater challenges to US data center development. In 2027, more than 75% of our data center capacity will be launched outside of the US. New facilities in Canada, France, and Norway will be put into operation in the coming months. This quarter, we signed contracts for additional capacity in Norway and Finland, and have reached the final stage of signing the data center in Sao Paulo, Brazil, which is our first data center in South America.
To date, we have locked in approximately 700 megawatts of data center capacity, which will be delivered by the end of 2027. Our first-party cloud is already one of the larger AI clouds among non-hyperscale vendors, and we are already quite advanced in multi-gigawatt capacity negotiations.
The best way to increase data center capacity is to increase the hardware's ability to generate tokens — use each megawatt to produce more tokens. Decoupled reasoning plays a role in this, increasing the number of tokens a data center can deliver by up to 5 times.
In terms of manufacturing and supply, we are cooperating with Flex and Sanmina to expand production capacity, and we expect manufacturing capacity to exceed 10 times that of the beginning of the year by the end of 2026. In 2027, we plan to add new plants and further increase production capacity. Our supply chain is well-positioned. Most importantly, we have TSMC, which continues to demonstrate the standards of a world-class partner. We don't use HBM memory. We believe the memory shortage and resulting price increases will continue until 2028, which will drive up competitors' costs and limit their supply. The industry-wide memory challenge is an opportunity for Cerebras.
Factories and supplies are necessary, but they are insufficient to support the expansion of production capacity. We must also have high yield rates. Cerebras invented a new computing architecture, applying metrics and thinking models designed for other architectures, like examining a Tesla and asking where its exhaust pipes are. Electric vehicles require different reference systems, and the same goes for wafer-level calculations. We welcome questions about all aspects of design, but there are also a number of repeated misconceptions circulating in the financial world. They think our system can't support large models — this is clearly wrong; they currently support it; and they think our manufacturing yield is low. Both are wrong.
Yield is divided into two categories: fab yield and parameter yield. Our yield from TSMC — that is, the yield of the wafer factory — is high, and so is the parameter yield. All wafers are flawed, and for decades these defects were at the core of why the industry thought wafer-level processors were impossible. We solved this problem through redundancy. Our design contains approximately 1 million identical tiles, a small portion of which is used as a backup. When a defect causes a tile to fail, the hardware will bypass it and enable a backup unit. The mechanism is effective both in the fab and in the field to achieve high wafer yield and good field reliability.
The second challenge is parameter fluctuations. Even when silicon wafers function properly, their electrical properties vary across the entire wafer, and some regions run faster or slower at the same voltage. We deal with parameter fluctuations from three aspects: First, we remove the outer edge of the wafer — the outer edge is the weakest part of the wafer and the main source of parameter fluctuations; we use the largest square that can fit into a 300 mm wafer, retaining most of the usable silicon wafers. Second, design and test for a wide range of conditions — our core can withstand a wider range of fluctuations, and during the timing convergence phase, we verify the design within a much wider range of operating conditions than others. Third, independent voltage regulation of wafers - We have invented a technology that can independently adjust the core voltage in 84 wafer partitions, which can increase the voltage in slower regions and lower the voltage in faster regions, thereby accurately controlling performance. As far as we know, no other chip has this capability.
As a result, both the fab and plant sides achieved extremely high yield rates. One last point: some people point out that the gross margin of our hardware shipments is relatively low, and based on this, they think this proves that there is a yield problem. In fact, this is simply a reflection of the continued decline in our manufacturing capacity and the delivery of the last batch of hardware under a special pricing plan offered to an early customer. We expect hardware gross margins to improve significantly over the next few quarters, eventually exceeding our long-term target of 60%.
Competence. We are advancing capacity building on all fronts—new systems and processors, improved software, advanced manufacturing technology, and partnerships.
The CS-4 is about to enter full mass production, and both performance and throughput will double. Our next generation processor, WSE-5, is equipped with a fifth-generation system CS-5 and has been returned from the fab ahead of schedule. WSE-5 was released on June 15. The chip has been released, the lighting is complete, and the model is already running. The whole machine is currently being tested. It was an extraordinary execution in less than five months from streaming to entering system testing. CS-5 will be more than double the size of CS-4 and will be put into operation in early summer 2027. It is based on TSMC's 5 nm process. At this node, wafer competition is less intense, and wafer costs are much lower.
About decoupled reasoning: As far as we know, Cerebras was the first company to deliver production-grade heterogeneous decoupled inference. Today, our cloud serves customers using AMD GPUs for pre-filling and Cerebras systems for decoding. We build an analytical inference software platform that is compatible with AMD and other GPUs. The same solution will be deployed to the Cerebras cloud and select data center customers. This results in significantly more high-speed tokens with the same amount of power — cloud owners get higher profits per megawatt, and AI users get lower cost per token. We believe decoupled inference will be one of the default options in major cloud deployments.
On the software side, our software stack continues to mature. The cloud service has achieved impressive uptime; systems are getting faster; and the time to go live for new models has been reduced to just a few days. Cerebras is Kimi K3.1's Day 0 launch partner, a cutting-edge open weighting model with 2.8 trillion parameters for Moonshot AI (Moonshot AI). Kimi K3.1 is not only the fastest in the industry on Cerebras, it can run up to 2,000 tokens per second.
Let's sum it up. It's been almost 11 years since we started, and we continue to innovate at an extremely fast pace. In summary: We have surpassed expectations on all metrics and raised our fourth-quarter guidance across the board, and expect revenue to triple in 2027; we have made significant progress in all key areas, including adding two new hyperscale customers and winning three hyperscale customers in the past six months; we have begun delivering CS-4; the fifth-generation processor WSE-5 has returned at record speed and entered CS-5 system testing; production-level decoupling reasoning is already running on AMD, and using Day 0 to support another 2.8 trillion parameter model to achieve the best performance in the industry.
Bob Komin (CFO): Thanks Andrew. Before getting into the details of our third quarter results and fourth quarter outlook, I'd like to review how we expanded our business. We chose to start a cloud business in early 2025 because in this rapidly evolving AI landscape, AI-native customers want to rapidly expand their computing power rather than build their own data centers. Since hyperscale vendors plan their data center computing power allocations several years in advance, they often don't have the recent available capacity to accommodate new hardware in their own facilities. To win the large-scale deployment required by basic model labs and hyperscale vendors, Cerebras needed to prove that our system could deliver industry-leading inference performance at scale. After fully evaluating our solutions in our private cloud, customers can choose to buy our hardware or purchase our cloud services. In order to quickly provide sufficient cloud capacity, we chose to start leasing data center capacity and expand our private cloud to gigawatts as soon as possible. This overall strategy appears to be correct. Because data center capacity is scarce today, we have a large number of contracts.
More importantly, once we have a private cloud, we can demonstrate the value of rapid reasoning to the entire market and drive more demand.
Large new data center capacity may take 15 to 18 months to bring online. In order not to wait and meet the huge demand that was already there, we leased back some of our own equipment from a customer. As we've mentioned before, this puts downward pressure on gross margins in the short term, but it also allows us to have some capacity when customers are ready to buy, thereby growing existing customers and winning over important new customers.
We've now won cutting-edge labs like OpenAI, and hyperscale vendors like AWS, and recently added two more hyperscale customers. We have other leading customers in the cloud, including Cognition, Sigma, Block, and Bilt. We serve our customers based on where they are—in some cases selling hardware and in others delivering cloud services. Soon, our contracted emerging cloud service provider customers will begin distributing fast inference services running on Cerebras hardware in their clouds.
In our previous conference call, we also mentioned that 2026 will be a year to build production capacity, capacity, and increase customers, laying the foundation for more than tripling revenue in 2027. With strong progress so far this year, we are well positioned to achieve that goal.
Now let's talk about third-quarter results. The fourth quarter will be an important inflection point for our growth and core financial trends. We fully surpassed core financial guidance for the third consecutive quarter and achieved our first non-GAAP net profit correction earlier than planned since listing.
Core revenue of US$231 million, up 10% month-on-month and 70% year-on-year, was US$16 million, or 7% higher than the median guideline value (US$215 million). Core cloud and other services revenue was US$145 million, up 272% year over year and 13% month on month. We expect cloud revenue to remain essentially flat in the fourth quarter compared to the previous 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 climb was entirely limited by data center production capacity. Production capacity will not be released until December, which is too late for the fourth quarter. We expect a significant increase in data center capacity throughout 2027, most of which will begin mid-next year, as we deliver a larger portion of the 700 megawatts we have contracted.
Core hardware revenue was US$86 million, an increase of 5% over the previous month. We're seeing a rise in customer demand for fast reasoning, not only through our private cloud, but also through new hardware deals. As these hardware shipments begin in the fourth quarter, we'll see a significant jump in core gross revenue. From the third quarter to the fourth quarter, core revenue increased nearly 40% month-on-month and about 85% year-on-year. These hardware orders (including Oracle) will continue to be delivered until at least the first half of 2027. Today we're also announcing Microsoft as a new customer.
Over the past few months, our order pipeline has grown more than tenfold, driven by diversified opportunities for new and existing customers, including hyperscale vendors, new cloud service providers, sovereign customers, enterprises, and AI-native customers.
The core gross margin for the third quarter was 41%, up 40 basis points from month to month, and 200 basis points higher than the guideline. In addition to the gross margin impact of the leaseback system, we are also under pressure to clear the last batch of shipments of hardware purchase agreements signed many years ago at lower prices. We have indicated that these factors will cause a temporary decline in gross margin, with a low in the third quarter. We actually did better, and the low actually occurred in the previous quarter, the second quarter.
The gross margin of core hardware declined month-on-month due to the last batch of shipments under the customer agreement that had been signed for a year. We are now delivering hardware orders at significantly higher current prices, which also means that gross margins will improve significantly starting in the fourth quarter.
The gross margin of core cloud and other services increased to 46.9%, an increase of 510 basis points over the previous month. The main driver was new business delivered at higher pricing and gross margins, partially offset by leasing back data center capacity and hardware to customers at a higher cost. We expect this to continue until at least mid-2027 until we have sufficient in-house data center capacity.
Fourth quarter guidance: Total core gross margin increased by approximately 700 basis points to 48% (median). We expect total core gross margin to exceed 50% in the short term and continue to improve towards the long-term target of 60%.
The core operating margin improved 170 basis points month-on-month to minus 14.3%, nearly 10 full percentage points higher than our guideline median. Core operating losses remained essentially flat month-on-month, at around US$33 million. Core net profit changed to $30 million. The year-to-date adjusted core EBITDA is also positive, at US$2 million. With strong fourth-quarter revenue and gross profit growth, we expect all core key profit indicators, including operating margins, to be corrected in the next quarter.
As of September 30, the RPO increased to US$26.1 billion. We update the RPO for newly signed businesses on a quarterly basis and adjust the value of existing contracts based on estimated changes in the amounts to be confirmed during the remaining commitment period. The third quarter RPO did not include new deals signed after September 30, including large-scale Oracle.
Since the OpenAI contract became a major part of our reported RPO to date, there have been many questions surrounding its status. It is important to note that this OpenAI contract, which was submitted with S-1 before listing, is a 750 megawatt capacity non-negotiable contract. Among the many contractual requirements applicable to both parties, there is a clause: if there is a delay in our capacity delivery, OpenAI is entitled to a credit during that period that can be used for future promised purchases. We have calculated the amount of potential delay compensation and deducted it from the September 30 RPO. These compensations have little impact on revenue and gross margin because they are estimated to be small compared to RPO and will be amortized as most revenue is recognized in proportion over the next few years. Future changes to these estimates will be updated quarterly and reflected in RPO and guidance as required.
If capacity delivery is delayed for 90 days or more, both OpenAI and we can choose to cancel only that portion of capacity. We estimate that less than 5% of the total promised megawatts may be cancelled this year. Given the high demand for data center capacity around the world, it took us a little longer to find and bring online the initial data center capacity needed to meet the delivery plan at a very aggressive pace. Based on the team's frequent on-site inspections at each site, we now expect to return to schedule in the second half of 2027, which should reduce the risk of compensation for future delays. We have taken a conservative approach and have not assumed that any delays will be exempted due to factors beyond our control. Furthermore, if OpenAI or Cerebras choose to cancel a delayed batch of capacity, we believe there is enough demand to find other customers to accept it.
OpenAI also reserves the option to purchase an additional 1.25 gigawatts of capacity in addition to its initial 750 megawatt commitment, which could be launched starting in 2029. Since that is still three years away, we haven't even started discussing it.
Our liquidity remains strong, with cash, cash equivalents, restricted cash and marketable securities totaling $7.8 billion at the end of the third quarter. We also have a revolving credit line of up to $850 million, which has not been used to date. The $1 billion upfront working capital loan we received in January and the ongoing collection of non-recurring handling fees for new data center capacity improvements have funded most of our net working capital requirements and capital expenses so far in 2026.
We're not a new cloud service provider. We are vertically integrated and are the inventors of the fastest reasoning solutions in the industry. Our business has a certain degree of durability and technological differentiation, which we believe will bring much higher long-term profit margin potential and cash flow. Our model also requires a lower capital investment for the computing power equipment portfolio we have deployed. We produce quite a bit of our own equipment in our data center. Similar to other companies that choose to develop and deploy their own AI chips, we don't need to pay 70% gross profit margin to a third party, we don't use expensive HBM, and only pay our own bill of materials costs. This reduces our relative capital expenses per megawatt, and we also benefit from producing the fastest and most valuable tokens. Despite the rapid expansion of private cloud construction, the only debt on our balance sheet is OpenAI's revolving loan. We have also received and continue to receive advance payments from our customers. Our capital structure today includes large amounts of cash and investments, and very low leverage.
Fourth quarter 2026 outlook: core revenue of US$310 million to US$330 million; core gross profit margin of 47% to 49%; core operating margin of negative 1% to positive 1%. Full year 2026: Core revenue increased to US$940 million to US$960 million; core gross margin increased to 42% to 44%; core operating margin increased to minus 10% to negative 8%.
In summary, the third quarter was another record quarter for Cerebras, and so will the fourth quarter. Our flexible model of selling hardware and cloud simultaneously is driving strong revenue growth and improved gross margin, and the improvements arrived a quarter earlier than expected, enabling us to triple our revenue in 2027 while driving our long-term target of 60% in core gross margin. We have shown strong operating leverage, and our core operating profit for the fourth quarter is expected to be close to breakeven. Subsequently, we expect significant progress towards the 40% long-term operating margin target in 2027.
And last but not least, we've been on the market for over 180 days. As a result, our IPO stock lockdown period completely ended on November 9th. Thank you all for listening. Please keep the host open to questions.
II. Q&A session
Thomas O'Malley (Barclays Research): Congratulations on such great results. The first question is about the new deal, OpenAI, and Microsoft. Can you give some indicators of the relative size of these new deals? Also, Bob, you mentioned that part of OpenAI is already in the RPO, can you explain how much has been included?
Andrew Feldman: Tom, I'm Andrew and I'm happy to hear from you. We'll be announcing more details about Microsoft's opportunity and the deal we've made with it. We're clearly excited about this and will be announcing more details in the first half of next year, when we can expect the deal to start generating revenue. Bob?
Bob Komin: Tom, in the RPO, we included our estimate for OpenAI of the amount of revenue promised under this non-negotiable contract, without extensions, without additional options, and adjusted for delays we think might occur; it also included the data centers we signed up for, the extraordinary expenses they paid to prepare for construction, and recurring monthly expenses during the contract period. These are our confirmed transaction income and handling fees.
Thomas O'Malley (Barclays Research): My question is that we have seen that cutting-edge laboratories can not only quickly build hardware, but also accelerate the deployment of models on hardware with different computing power, and interchangeability is improving. So how do you guys benefit from it? What do you expect the future of heterogeneous infrastructure to look like? What are the pros and cons to you guys?
Andrew Feldman: That's a great question, Tom. OpenAI mentioned in a file or conference call that it took about 3 days to complete the launch of the AMD GPU. This is a very interesting and important development, especially for long-term activists. I don't think people really understand what this means to our ecology. Compilers and underlying software have always been fundamental barriers to architecture. They benefit those in place; that is, the software moat you hear a lot about. If Frontier Labs actually does what they say, that means the value of this underlying software will shrink and disappear over time. And if the underlying software has little value, what's left to provide value is innovative hardware. This means that differentiation will rely more on hardware differentiation. I think this is a huge advantage for Cerebras and other companies that build pioneering hardware, new architectures, or different architectures. I think this will weaken the advantage that those in power have. This is a great question, and I think the industry needs continued attention. Thank you. Joseph Moore (Morgan Stanley Research): Congratulations on the strong hardware data. Can you talk about the relationship between RPO issues and cloud scheduling, is there anything specific we can pay attention to? And how confident are you to resolve this issue before it evolves into additional penalties?
Andrew Feldman: I think what we included in RPO was a very conservative number, and it was the absolute worst situation that could happen. So today we are relieved that the results will be better than that number. These are common in big data center deals. We have similar agreements with data center builders, and they have penalties associated with their own delays — I think over time, choosing to implement these terms depends on the commercial decisions at the time. I think this is how these kinds of terms work, and how we expect them to work in the future.
Joseph Moore (Morgan Stanley Research): What's the reason for the delay? I think everyone is aware of issues such as land, electricity, factory buildings, etc. Is that the reason or something else?
Andrew Feldman: The current delays are entirely due to land, electricity, plants. We have production capacity, and we have all the materials. We are waiting for the electricity and plant to be completed.
Atif Malik (Citi Research): Andrew, I'd like to ask about the product roadmap. When will the intersection (replacement) of CS-4 and CS-3 occur? Is it 2027? Additionally, in your roadmap, CS-3 has significantly increased throughput. If you follow this path to CS-4 and beyond, will your chips one day provide enough speed and throughput so that the system can be used alone without the need for a GPU?
Andrew Feldman: Great question. First, until a few weeks ago, all of our deployments were pure CS-3. We've only recently started deploying decoupling solutions, and we've seen a significant increase in throughput, which is very exciting. So that's the first observation. Obviously they work well in an aggregated form, but decoupling — we're deploying and almost fully usable — can produce more high-speed tokens when high-speed tokens are limited. I think the CS-4 crossover (replacement) was in the first quarter. The roadmap looks really good right now. We are ahead of the plan at CS-5. As I mentioned in my preparation statement, it only took 5 months from the streaming film to the chip entering the system, which is extraordinary. We expect it to start shipping in early summer 2027, doubling its performance.
Atif Malik (Citi Research): Then ask Bob, if you guide cloud sales to remain flat in the fourth quarter, can you help us understand the structure of cloud and hardware within the framework of tripling sales in 2027?
Bob Komin: I think the best way to understand the hardware structure is — these deals are probably very, very large. We hope to be able to give some instructions on shipments in the next quarter or two. This largely depends on whether the customer's data center is ready. So what I can say is that we think the share of hardware in the fourth quarter and into the first quarter will be closer to 50%, rather than the level of previous quarters. Looking ahead to the rest of 2027, we'll provide more information during the fourth quarter call, and I think we'll give a more complete view of the overall outlook for 2027 at that time.
Quinn Bolton (Needham & Company): Andrew, I just wanted to make sure I was listening right. In the RPO and your assumptions, you made assumptions about potential delay compensation, but so far you haven't received any cancellations for latency capacity. Is this statement correct?
Andrew Feldman: Right.
Quinn Bolton (Needham & Company): The market is concerned about the fast reasoning provided by low-volume GPUs; this sounds like an expensive token service. So I was wondering, can you guys talk about the cost per token for Cerebras relatively low batch GPU configuration deployments?
Andrew Feldman: Of course. GPUs allow you to trade throughput for extra speed by reducing batch size. So you have very high throughput at low speeds. As you try to get higher speeds, you'll lose a lot of throughput until you reach the lower batch range. This is a characteristic of this architecture. They can get some limited performance by moving towards lower batches, but when they do, the cost per token jumps up. They are serving fewer customers at the same time, so the total amount of tokens generated is very small. As you keep on trying to be faster, we get stronger. At some point, no matter what you do with the GPUs, they still can't do it when the batch is 1.
The question is where the intersection is. I think at around 250 to 300 tokens per second, we're more economical. When you reach 500,000 to 800,000 tokens per second, our economy is far superior. At around 800 to 2,000 tokens per second, GPUs can also compete. This is the current situation. This is the current landscape. This is a great question about a complicated topic.
Joshua Buchalter (TD Cowen Research): Congratulations on winning the Microsoft deal. First, considering land, electricity, and plant issues, can you talk about how you allocate production capacity? It sounds like silicon or rack supply isn't an issue. So how much room do you see for incremental hardware sales to customers? Also, considering the challenges of obtaining land, electricity, and plant for cloud customers, have they discussed with OpenAI that they might instead purchase hardware and put it in their own or partner data centers?
Andrew Feldman: Of course. Josh, there are actually two parts to this question. First, are land, electricity, plants, and their construction really the limiting factors? The answer is absolutely. TSMC has promised us sufficient production capacity and wafers for 2027, and we are very relieved about this. The fact that we don't use HBM is clearly beneficial because the supply of HBM is extremely tight.
Second, you're right — did this create an opportunity for you to sell hardware during data center construction? I think the only reservation is that, as far as we know, everyone's data centers are being extended, including OpenAI's. As we know, data centers built for hyperscale vendors are generally being postponed around the world. Therefore, data center constraints are not unique to us; they are also reflected in our customers. So — yes, we have supplies to them; yes, we ship to customers who have data centers; yes, the truth is that they are also very lagging behind. Does that make sense?
Joshua Buchalter (TD Cowen Research): So, considering land, electricity, and plant issues, how likely is the OpenAI contract to switch to hardware sales?
Andrew Feldman: It's in the contract. They have that right. But as I said, we've learned from them that their data center is also behind schedule.
Joshua Buchalter (TD Cowen Research): Can you talk about how Oracle and Microsoft are positioning your solutions to their customers? Did they first sell hardware and then basically follow their customers' needs for quick reasoning?
Andrew Feldman: We've already announced more information about the Oracle partnership, but we'll keep the details of Microsoft for now. At Oracle, they are procuring hardware to deploy at its data center in the US. We're collaborating, and they're currently building a pipeline for this. There is strong market interest. So we don't think this is the last deal we've made together — it's the beginning of a major partnership where our devices will be deployed in Oracle's data centers as part of the OCI cloud.