Cerebras Conference Call: Secures Major Deal with Microsoft, HBM Shortage Becomes an Opportunity, WSE-5 Chip to Enter Mass Production Next Summer

AI chip startup Cerebras reported its first net profit in the third quarter of 2026 and announced that Microsoft has become its third major cloud service client in six months. The company reiterated its strong guidance for a threefold revenue increase by 2027.

According to Wall Street Journal, on November 12 local time, Cerebras released its third-quarter financial report for 2026, with core revenue reaching $230.9 million, a 70% year-on-year increase, exceeding analysts' average expectation of $214.7 million.

For the first time as a public company, the company achieved a net profit of $30 million under Non-GAAP standards.

During the conference call, Cerebras co-founder and CEO Andrew Feldman stated that the company has raised its guidance for the fourth quarter and the full year, expecting core revenue to grow nearly 40% quarter-on-quarter. The company also maintains its target of "threefold growth" in revenue by 2027 and anticipates core gross margins to continue approaching the long-term target of 60%.

Andrew Feldman announced that the company has officially signed a cloud cooperation agreement with Microsoft, adding to the previously announced collaborations with AWS and Oracle, completing partnerships with three major cloud service providers in the past six months. He mentioned that more details about the Microsoft agreement will be disclosed in the first half of next year, and related revenue is expected to start being recognized during the same period.

Meanwhile, CFO Bob Komin revealed during the call that the company is experiencing a phased delay in the delivery of data center capacity to OpenAI, and has deducted the corresponding delayed compensation amount from the remaining performance obligations (RPO). He emphasized that the potential cancellation volume accounts for less than 5% of the total committed capacity under the OpenAI contract, and the company expects to return to normal delivery pace in the second half of 2027.

On the manufacturing side, Andrew Feldman stated that the company is collaborating with Flex and Sanmina to expand production capacity, expecting to increase manufacturing capacity by more than tenfold by the end of 2026, with further new factories added in 2027.

He also emphasized that since the company does not use high-bandwidth memory (HBM), it can effectively avoid the ongoing memory shortage and price pressure expected to last until 2028, creating a structural advantage in the supply chain.

Securing Major Deal with Microsoft, Completing Partnerships with Three Hyperscalers in Six Months

In addition to strong financial data, the market's most focused incremental information is Cerebras's significant breakthrough in customer expansion. Feldman officially announced at the meeting:

Today, we signed an important cloud agreement with Microsoft.

With this, along with the previously announced partnerships with AWS and Oracle, Cerebras has won the favor of three major cloud service providers in just six months. These agreements will greatly broaden its market channels.

Feldman revealed that the Oracle cloud inference service powered by Cerebras is expected to launch in the summer of 2027, while more details about the collaboration with Microsoft will be announced early next year.

CFO Bob Komin added that in the past few months, the company's sales channels have grown "far more than tenfold," with demand coming not only from hyperscale cloud service providers but also from sovereign clouds, enterprise clients, and AI-native companies.

Addressing OpenAI Contract Controversy: "Delays Due to Land and Power, Not Chip Shortage"

As Cerebras's largest client, the performance of the OpenAI contract is a focal point for investors. As of September 30, Cerebras's remaining performance obligations (RPO) reached $26.1 billion, with the vast majority coming from OpenAI.

In response to market concerns about delays in OpenAI order deliveries, management provided a transparent response. Due to the common bottlenecks in global data center construction, Cerebras has been unable to deliver infrastructure according to the originally aggressive timeline. CFO Bob Komin admitted:

If the delay reaches 90 days or more, both parties have the right to cancel that portion of capacity. However, we estimate that the capacity likely to be canceled this year will be less than 5% of the total committed megawatts.

Management has deducted potential late penalty fees from the RPO. Bob Komin emphasized that the impact of these penalties on revenue and gross margins is relatively limited, as the penalty amount accounts for a small proportion of the RPO and will be amortized over the next few years. The company expects to catch up on progress in the second half of 2027.

Feldman bluntly pointed out that the bottleneck is not due to silicon or system manufacturing:

We have capacity, we have all the materials. What we are waiting for is the completion of land, power, and shell.

He emphasized that not only Cerebras but also all industry giants, including OpenAI, are experiencing "comprehensive delays" in data center construction.

At the same time, Feldman firmly refuted rumors that Cerebras technology cannot support OpenAI's large models:

There are rumors that Cerebras cannot support Astra due to technical reasons. These are false. Our systems can fully support models of Astra scale or even larger.

Breaking the "Low Yield" Bias, Not Relying on HBM as a Strategic Advantage

In the face of Wall Street's skepticism about the yield of wafer-scale computing architecture, Feldman used a metaphor to counter:

Cerebras has invented a completely new computing architecture. Applying metrics and thinking models designed for other architectures to ours is like checking a Tesla and then asking: Where is the exhaust pipe? Electric vehicles require a different reference frame, and wafer-scale computing is no different.

He clearly stated that whether it is TSMC's foundry yield or parameter yield, "our yield is very high."

At the same time, Feldman pointed out the fatal weakness faced by competitors—memory shortages. Because Cerebras's architecture does not require high-bandwidth memory (HBM), it occupies a unique advantage in the global supply chain. Feldman added:

We believe that memory shortages and related price increases will continue until 2028, which will drive up competitors' costs and limit their supply. The memory challenges across the industry present an opportunity for Cerebras.

Andrew Feldman introduced during the call that the company has currently signed approximately 700 megawatts of total data center capacity, with plans to complete delivery by the end of 2027. Over 75% of the new capacity will be located outside the United States, with new facilities in Canada, France, and Norway set to come online soon, and additional contracts in Norway and Finland, as well as being in the final stages of signing the first data center in South America in São Paulo.

Bob Komin stated during the call that the larger-scale data center capacity increase will begin to be released in mid-2027, which is also the direct reason for the fourth-quarter cloud revenue remaining flat quarter-on-quarter, rather than pricing or demand issues. He added that the company is actively advancing negotiations for gigawatt-level capacity.

WSE-5 Chip Has Returned from Fab, Mass Production Expected Next Summer

In terms of technological iteration, Cerebras continues to maintain an aggressive pace.

The company announced that its next-generation processor, Wafer-Scale Engine 5 (WSE-5), has returned from the foundry ahead of schedule. Feldman proudly stated:

WSE-5 was taped out on June 15. The chip has now returned, has been powered on, and is running models, undergoing chassis testing. It is extraordinary execution to move from tape-out to system testing in less than five months.

The CS-5 system, based on TSMC's 5nm process, is expected to enter production in early summer 2027, with performance doubling compared to CS-4.

Additionally, the company has also launched a heterogeneous decoupled inference (DISAG) solution, using AMD's GPU for pre-fill and Cerebras's CS-3 system for decoding, significantly improving the power output and profit margin per megawatt.

In response to market discussions about low-batch GPUs potentially challenging Cerebras's inference speed, Andrew Feldman stated during the Q&A session that while GPUs can achieve some speed improvements by reducing batch sizes, the cost per unit token significantly increases.

He pointed out that at speeds above approximately 250 to 300 tokens per second, Cerebras's economics are clearly superior; at speeds of 800 to 2000 tokens per second, GPUs technically cannot achieve comparable performance.

Cerebras Q3 Earnings Call Transcript (AI-assisted Translation):

Operator: Hello everyone, welcome to the Cerebras Systems Q3 2026 earnings call. During the conference, all participants are in listen-only mode. After the presentation, there will be a question-and-answer session. Please note that this conference is being recorded.

Now, I will hand the call over to Sean Dorsey, Head of Investor Relations. Please go ahead, Sean.

Sean Dorsey (Head of Investor Relations): Thank you, Operator. Good afternoon everyone, and welcome to the Cerebras Systems Q3 2026 earnings call. Earlier today, we released a press release and will supplement the earnings report presentation on the investor relations section of the company website. A replay of this live webcast will also be available on the investor relations website after the call.

Joining me today are: Co-founder and CEO, President Andrew Feldman, and CFO Bob Komin.

Before we begin, I would like to remind everyone that today’s discussion will contain forward-looking statements under the safe harbor provisions of the Private Securities Litigation Reform Act of 1995. These statements involve risks and uncertainties that could cause actual results to differ materially from those expressed or implied. Relevant risks are outlined in our filings with the SEC, including the final prospectus related to the IPO and future periodic reports. Except as required by law, we do not undertake any obligation to update these forward-looking statements.

In today’s call, we will also discuss several non-GAAP financial metrics. The reconciliation of GAAP to non-GAAP performance has been included in today’s press release and in the supplemental materials on the investor relations website.

Now, I will hand the call over to Andrew.

Andrew Feldman (CEO and Co-founder): Thank you, Sean, and thank you all for joining us today.

Looking ahead over the next few months, we see two certain trends: rapid change and a shortage of computing power. In the past 90 days, the AI ecosystem has witnessed a number of leadership changes among cutting-edge models, the release of dozens of valuable open-source models, significant security incidents, the emergence of powerful new consumer agents, and many other developments.

No one can predict where AI capabilities will be in 90 days, let alone a year from now. However, amidst all these changes, intelligence and speed remain the two core elements of every application scenario—whether in programming, security, or personal agents.

In this rapidly changing world, Cerebras’ mission is to ensure that no matter how AI evolves, Cerebras continues to grow. As foundational models grow, we grow; as open-source models grow, we grow; as personal agents grow, we grow; whether enterprises or startups grow, we grow; whether customers need hardware or cloud computing resources, we can grow. Our goal is to ensure that no matter what drives industry development, Cerebras can seize the opportunity.

In the third quarter, we made significant progress towards this goal.

First, we exceeded all performance guidance targets. While we are pleased with these results, I think it is even more important to note that since our last earnings call, we have established significant partnerships with several leading companies in the AI industry, the latest being the announcement today of a major cloud services agreement with Microsoft. Along with previously announced collaborations with AWS and Oracle, three major cloud service providers have chosen to incorporate the Cerebras platform into their infrastructure strategies.

These new partnerships, combined with the continued growth of existing customers, allow us to raise our core performance guidance for Q4 and continue to expect a threefold increase in revenue by 2027, while gross margins continue to expand. Bob will provide a detailed overview of our Q3 performance and forward-looking guidance in his remarks.

About our business structure

As an AI accelerator company, our core innovations lie in chips, systems, and software that make AI computing faster. We are an AI hardware company with two market channels.

The first channel is system sales: installing Cerebras hardware racks in customers’ data centers. This is a high-quality business with quick revenue recognition and fast payments, but it also has some irregularity depending on whether the customer’s data center is ready to receive the hardware.

The second channel is our cloud services. The rapidly growing customer base of AI-native enterprises tends to consume computing power (tokens) rather than deploying hardware themselves. Given the extreme scarcity of data centers and computing resources, we have both systems and data centers, which constitutes a significant strategic advantage.

For companies that produce their own hardware, AI cloud business is particularly outstanding. Our cost per megawatt is far lower than that of emerging cloud service providers, allowing for higher capital efficiency. The business proportions of the two channels will vary by quarter, but we believe both will grow significantly over time.

Regarding market speculation about collaboration with OpenAI

OpenAI is a significant customer for us and a close partner, and I believe that our relationship will continue to be strong in the future. OpenAI experienced growth in Q3 compared to Q2, and we expect further growth in the future.

OpenAI’s UltraFast SOL5-6 preview, based on Cerebras, has achieved overwhelming success, strongly demonstrating the enormous market demand for fast tokens.

At the same time, Cerebras indeed does not have enough computing power to support the broader rollout of UltraFast, including Apsara. As the industry has noted, we believe that computing power is a key driver of business decisions in the industry. In fact, even if OpenAI has a massive supply of GPUs, it still needs to implement usage-based limitations.

After the agreement was signed, it took us longer than expected to find the contract and ramp up the substantial incremental data center capacity required to meet the agreement’s requirements. Therefore, we are currently lagging in execution at the initial stage.

Given the importance of this collaboration to our business, we understand that investors are highly concerned about it. Information regarding the details of the agreement or related news is unlikely to come from sources outside of Cerebras or OpenAI.

For example, there were previous rumors that Cerebras could not support Astra for technical reasons, which is incorrect. CS3, CS4, and CS5 are fully capable of supporting Astra-level models and even larger models.

At the same time, I expect and suggest that investors anticipate that OpenAI will build computing power using a variety of partnerships and diversified hardware solutions. OpenAI will phase out old models and deploy new models on these computing resources for both internal and external use. Their demand for computing power is enormous, and different solutions will have opportunities to play a role in their scaling process.

Regarding competitive position in inference speed

I would like to add one point: despite the existence of various experimental solutions in the market—low-batch GPUs, startups releasing, ultra-large computing cores, and custom chips from ultra-large cloud service providers—we remain the undisputed fastest inference solution. We believe our product roadmap will allow us to maintain this advantage over the next few years.

Business assessment framework: Customers, Computing Power, Capabilities

Our assessment framework for business progress remains unchanged, focusing on three dimensions: customers, computing power, and capabilities.

Customers

We are rapidly expanding our customer base to broaden our coverage and diversify our revenue sources. Q3 was a fruitful quarter. We won partnerships with two new ultra-large cloud service providers.

We signed an agreement with Oracle to bring Cerebras hardware into Oracle Cloud, with Cerebras being deployed in Oracle data centers to serve AI labs, enterprise customers, and developers building interactive intelligent applications. We expect the Cerebras-powered Oracle Cloud inference service to officially launch in the summer of 2027.

We also signed an exciting partnership agreement with Microsoft, with further details to be disclosed next year.

To date, we have secured partnerships with three ultra-large cloud service providers in the past six months, marking an important milestone achievement that lays a solid foundation for our growth in 2027 and 2028.

We have also expanded collaborations with several high-quality emerging cloud service providers, including Firework, further diversifying our market channels and providing customers with more access to Cerebras’ high-speed inference services.

Open-source weight models have also been a boost for us, continuously spawning new applications and expanding the boundaries of AI in daily life and various industries. For example, just this week, we announced that Cerebras’ fast inference service is supporting Built’s OS platform and personal assistant, bringing a rapid inference experience to millions of users.

Through consumer applications like Built, we are delivering Cerebras’ powerful computing power directly to users. This is precisely why fast tokens can command a premium—our cloud customers tell us that fast tokens yield excellent gross margins, and Cerebras’ fast tokens are their most profitable tokens to sell.

Computing Power

Computing power consists of two components: data center computing power and manufacturing capacity, both of which are crucial for business growth.

Data center availability is a persistent constraint faced by the entire industry, and we are no exception. Additionally, we believe that the evolution of the U.S. policy environment may pose greater challenges for data center construction within the United States.

By 2027, over 75% of our data center capacity will be online outside the United States. In fact, in the coming months, new facilities in Canada, France, and Norway will be put into operation. This quarter, we also signed contracts for additional capacity in Norway and Finland, further expanding our European business footprint. At the same time, we are in the final stages of signing a data center in São Paulo, which will be our first data center in South America.

To date, we have signed contracts for approximately 700 megawatts of data center capacity to be delivered by the end of 2027. Our own cloud has become one of the larger non-ultra-large cloud service providers, and we are currently advancing negotiations for gigawatt-level capacity.

The best way to enhance data center capacity is to enhance the token generation capability of data center hardware, that is, to generate more tokens per megawatt. To this end, disaggregated inference effectively expands our capacity by increasing the token delivery capability of data centers by up to five times. Generating more ultra-fast tokens per megawatt means that the revenue generated by each data center will significantly increase.

In terms of manufacturing and supply chain, we are working with Flex and Sanmina to scale up manufacturing, with manufacturing capacity expected to exceed ten times that of the beginning of the year by the end of 2026. In 2027, we plan to add factories and further increase capacity to support the anticipated significant growth. Our supply chain is also fully prepared to support this growth.

Most importantly, we have secured TSMC...

Sufficient wafer capacity commitment showcases TSMC's strength as a world-class partner.

It is also important to note what we do not need in our supply chain—this is equally crucial for us. Unlike other competitors, we do not use HBM memory. We believe that the memory shortage and the resulting price increases will persist until 2028, which will raise costs for competitors and limit their supply. The industry-wide memory challenges create opportunities for Cerebras.

Factories and supply are necessary conditions for expanding capacity, but they are not sufficient; we must also maintain high yields.

On Yield Issues

Cerebras has invented a completely new computing architecture. Measuring us with metrics and mental models designed for other architectures is like asking where the exhaust pipe is when inspecting a Tesla. Wafer-scale computing requires a different reference framework, just as electric vehicles require different evaluation dimensions.

We welcome questions about all aspects of the design and will strive to share architectural details more clearly.

However, there are some persistent misconceptions in the financial community: one is the claim that our systems cannot support large models—this is clearly false, as we are currently running large models; the other is the claim that our manufacturing yield is low. Both are untrue.

Regarding model scale, I have addressed this in previous remarks and discussions. Since manufacturing yield is difficult to observe externally, I will clarify it here.

Yields are divided into two categories: fab yield and parametric yield. Our fab yield from TSMC is very high, and so is our parametric yield. Our technical blog provides detailed information on this, and I will summarize it briefly below.

All wafers contain defects. For decades, these defects have been considered the core reason why wafer-scale processors were thought impossible. We have solved this problem through redundant design.

Our design includes about one million identical computing units (tiles), with a small portion serving as backups. When a defect causes a unit to fail, the hardware automatically bypasses it and activates the backup unit. This mechanism is effective during both the wafer production phase and the field usage phase, achieving high wafer yield and strong field reliability.

The second challenge is parametric variation. Even if silicon functions normally, its electrical characteristics can vary across the wafer—some areas may run faster or slower at the same voltage. This is a fundamental reality of chip design and one of the obstacles that must be overcome to achieve wafer-scale computing.

We address parametric variation in three ways:

First, removing the wafer's outer edge. The outer edge is the weakest part of the wafer and a major source of parametric variation. We use the largest square area suitable for 300mm wafers, eliminating a significant amount of edge silicon.

Second, designing and testing for a wide range of operating conditions. Our core can tolerate a broader range of variations. During the timing closure phase (the process of verifying that the circuit meets timing requirements), we validate the design under operating conditions that are much broader than industry norms, ensuring that expected performance can still be achieved despite variations across the wafer.

Third, independently adjusting voltage across different areas of the wafer. We have invented a technology that allows independent adjustment of core voltage across 84 partitions of the wafer—raising voltage in slower areas and lowering it in faster areas, allowing for precise performance control. To our knowledge, no other chip has this capability.

The end result is extremely high yields achieved at both the wafer and factory stages. We have solved problems that others believed made wafer-scale computing impossible through engineering means, achieving this accomplishment.

One final note: Some view our relatively low gross margins on hardware shipments as evidence of yield issues, but this merely reflects our ongoing improvement in manufacturing capabilities and the final batch shipment under special pricing agreements with early customers. We expect hardware gross margins to improve significantly in the coming quarters, ultimately surpassing our long-term target of 50%.

Capabilities

We are continuously enhancing our capabilities: new systems and processors, improved software, advanced manufacturing technologies, and partnerships that provide industry-leading solutions.

Despite the rapid pace of change in the AI industry, the importance of inference speed continues to rise, and Cerebras remains the undisputed leader. We believe this performance advantage will persist long-term, and our product roadmap continues to advance rapidly.

CS4 is about to enter full-scale production, with performance and throughput doubling.

Our next-generation processor—the Wafer Scale Engine 5 (WSE-5)—has been deployed in the fifth-generation CS5 system, returned from the fab on schedule, even ahead of plan. WSE-5 completed tape-out on June 15, the chip has returned to the factory, is running models, and has entered the system testing phase. The time from tape-out to system testing is less than five months, demonstrating excellent execution. The speed of CS5 will exceed that of CS4 by more than double, with mass production expected in early summer 2027. CS5 is based on TSMC's 5nm process node, which has relatively less competition for wafers, and wafer costs have significantly decreased.

Our innovations also extend to the complete solutions space. To our knowledge, Cerebras is the first company to achieve production-grade heterogeneous decomposed inference. Currently, our cloud service is using AMD GPUs for pre-fill and the Cerebras CS3 system for decode. To support production-grade deployment, we have built a decomposed inference software platform compatible with AMD and other GPU products, which will be deployed in the Cerebras inference cloud and among select data center customers.

Decomposed inference can provide more high-speed tokens at the same power consumption, meaning cloud service providers achieve higher profits per megawatt, and AI users benefit from lower costs per token. We believe decomposed inference will become one of the default configurations for mainstream cloud deployments.

Our software stack continues to mature, cloud services have achieved impressive high availability, system operating speeds are continually improving, and the startup time for new models has been reduced to just a few days.

Cerebras is the zero-day launch partner for Kimi K3.1—Kimi K3.1 is a 28 trillion parameter cutting-edge open-source weight model under the Moon's Dark Side, perhaps the most advanced open-source weight model on the market today. The Kimi K3.1 running on Cerebras is not only the fastest in the industry, but also runs at a speed of 2000 tokens per second.

After nearly 11 years, we continue to innovate at an astonishing pace. Our team has remained united through both good times and bad, and the energy and passion within the company have never been higher.

We take pride in solving the toughest problems in the industry.

Summary: We have exceeded all metrics, raised our guidance for the fourth quarter, and expect a threefold increase in revenue by 2027. We have made significant progress in all core areas, including adding two new hyperscale cloud service customers, and have won three hyperscale cloud service customers in the past six months. CS4 has begun shipping, the fifth-generation processor WSE-5 has returned to the factory, and CS5 system testing has been completed at a record pace. Production-grade decomposed inference with AMD has been launched, and zero-day support for another 28 trillion parameter model has been successfully achieved with industry-leading performance.

Now, I will pass the floor to Bob.

Bob Komin (Senior Vice President and Chief Financial Officer): Thank you, Andrew, and good afternoon, everyone. Before diving into the specifics of our third-quarter performance and fourth-quarter outlook, I would like to take a moment to review our journey towards scaling the business.

Business Strategy Background

We chose to launch cloud services in early 2025 because, in this rapidly evolving AI landscape, AI-native companies want to quickly scale compute power without building their own data centers. Since hyperscale cloud service providers often plan their data center compute allocations years in advance, they frequently lack immediate space in their own facilities for new hardware. To win large deployment contracts required by foundational model labs and hyperscale cloud service providers, Cerebras needed to demonstrate that our systems could deliver industry-leading inference performance in scaled scenarios. After fully evaluating our solutions in private clouds, these customers can choose to purchase hardware or procure our cloud services.

To quickly provide sufficient cloud compute power to meet these demands, we chose to rent data center capacity as soon as possible, pushing private cloud scale to the gigawatt level. This overall strategy has proven correct—data center capacity is currently extremely scarce, and we have already locked in significant capacity in contracts.

More importantly, having a private cloud allows us to demonstrate the value of rapid inference to the entire market and further stimulate market demand. Since new large data centers take 15 to 18 months from project initiation to operational use, to accelerate meeting our existing high demand, we chose to lease back some of our own equipment from one customer. As mentioned earlier, this has put downward pressure on gross margins in the short term, but it has also allowed us to serve customers promptly when they have needs, thereby expanding existing customers and winning important new ones.

Today, we have won cutting-edge labs like OpenAI and hyperscale cloud service providers like AWS, and we recently added two more hyperscale cloud service customers. Leading customers like Cognition, Figma, Block, and Built have also joined our cloud platform. We are customer demand-driven and adaptable—sometimes selling hardware, sometimes providing cloud services.

Previously, we also indicated that 2026 would be a year of building—expanding compute power, enhancing capabilities, and accumulating customers to lay the foundation for more than threefold revenue growth in 2027. With the strong progress we have made so far this year, we are confident in achieving this goal.

Third Quarter Performance

Now moving to the third-quarter performance, I am pleased to announce that the fourth quarter will be an important turning point for our growth and core financial trends.

We have exceeded core financial guidance for three consecutive quarters. We also achieved our first non-GAAP net profit earlier than expected.

A reminder that the definitions of core metrics and their reconciliation with GAAP are included in today's earnings release and can be found on our website.

Core revenue: Reached a record high of $231 million, a 10% increase quarter-over-quarter and a 70% increase year-over-year, exceeding our guidance midpoint of $215 million by $16 million, or 7%.

Core cloud services and other services revenue: Also reached a record high, up 272% year-over-year to $145 million. Third-quarter cloud services revenue grew 13% quarter-over-quarter. We currently expect fourth-quarter cloud services revenue to remain roughly flat. This trend is not driven by pricing or demand issues—the pricing environment remains healthy, and demand for our solutions continues to far exceed supply.

The bottleneck for cloud services revenue growth is entirely due to data center capacity. We will begin to increase data center capacity in December to build momentum for next year, but unfortunately, it is too late to have a substantial impact on the fourth quarter. We expect to see a more significant increase in data center capacity in 2027, with the majority of the incremental capacity expected to be released starting mid-next year, at which point we will deliver a large portion of the 700 megawatts of data center capacity already locked in contracts.

Core hardware revenue: This quarter was $86 million, a 5% increase quarter-over-quarter. We see customer demand for rapid inference services growing.

Demand for growth comes not only from our private cloud but also from hardware procurement contracts signed with several new customers. With these hardware shipments scheduled to begin in the fourth quarter, core total revenue is expected to see a significant leap.

The quarter-over-quarter growth in core revenue from Q3 to Q4 will be close to 40%, with a year-over-year increase of approximately 85%. These hardware orders, including those from Oracle, will continue to ship at least until the first half of 2027.

As Andrew mentioned earlier, we also announced today that Microsoft has become a new customer. Over the past few months, our sales pipeline has grown more than tenfold, driven by diverse opportunities, including demand from various new and existing customers such as hyperscale cloud providers, emerging cloud providers, sovereign clouds, enterprise clients, and AI-native companies.

Gross Margin

The core gross margin for Q3 was 41%, improving by 40 basis points quarter-over-quarter and exceeding our quarterly guidance by 200 basis points.

In addition to the impact of the leaseback system on gross margin, we faced some headwinds this quarter—specifically, the last batch of shipments from hardware procurement agreements signed years ago (at lower pricing) is being digested. We had previously guided that all these factors would lead to a temporary decline in gross margin, with the low point occurring in Q3. In fact, our performance exceeded expectations—the low point occurred in the previous quarter (Q2).

The breakdown of the core gross margin for Q3 is as follows:

Core hardware gross margin decreased quarter-over-quarter to 31%, primarily due to the last batch of shipments under customer agreements signed years ago. We are currently executing hardware orders at significantly increased current prices, which means that gross margins will improve significantly starting in Q4.

Core cloud services and other services gross margin increased significantly to 46.9%, up 510 basis points from Q2. The main driver was new business taken on at higher pricing and margins. This was partially offset by the higher costs of data center capacity and hardware leased back from customers. We expect this situation to persist at least until mid-2027, until we have sufficient owned data center capacity to meet the higher levels of cloud demand.

In summary, the core gross margin for Q4 is expected to see a stepwise improvement, increasing by 700 basis points to a midpoint value of 48%. We expect the core total gross margin to exceed 50% in the near future and continue progressing towards a long-term target of 60%.

Profitability Progress

In Q3 and year-to-date, we have made strong progress in profitability. The core operating margin improved by 170 basis points quarter-over-quarter, with the loss rate decreasing to 14.3%, exceeding the midpoint of our guidance by nearly 10 percentage points. Core operating loss remained flat compared to Q2, at approximately $33 million. Core net profit turned positive, reaching $30 million. Year-to-date adjusted core EBITDA also turned positive, at $2 million.

With significant growth in revenue and gross margin in Q4, we expect all core key profitability metrics (including operating margin) to turn positive in the next quarter.

Remaining Performance Obligations (RPO) and OpenAI Contract

As of September 30, remaining performance obligations increased to $26.1 billion. We update RPO quarterly, incorporating new signed business and adjusting the value of existing contracts based on changes in the expected amounts to be recognized during the remaining commitment period. The Q3 RPO does not include the impact of new contracts signed after September 30, including the large contract with Oracle.

There have been many investor inquiries regarding the status of the OpenAI contract, as it constitutes a significant portion of our reported RPO to date. Just a reminder, this contract was submitted with the S1 filing prior to the IPO and is a take-or-pay contract for 750 megawatts of computing power.

The contract includes numerous terms applicable to both parties, one of which states that if we experience delays in delivering computing power, OpenAI has the right to receive credits during that period, which can be used for future committed purchases. We have calculated the potential delay credits and have deducted them from the RPO as of September 30.

These credits have a minimal impact on revenue and gross margin for two reasons: first, they represent a relatively small proportion of the RPO; second, the vast majority of revenue will be recognized proportionally over the coming years, with the credit amounts being amortized accordingly. Future quarters will be updated based on the latest estimates and reflected in guidance as necessary.

If the delay in delivering computing power continues for 90 days or more, both OpenAI and Cerebras have the right to cancel that portion of the power commitment (limited to that portion). We estimate that the amount of computing power at risk of cancellation this year will be less than 5% of the total committed megawatts.

Given the strong global demand for data center capacity, we have spent additional time finding and bringing online initial data center capacity to meet delivery schedules, compared to a very tight timeline. Based on our team's frequent on-site inspections at various locations, we currently expect to catch up by the second half of 2027, at which point the risk of future delay credits should correspondingly decrease.

We have taken a conservative approach and have not assumed that any delays will be exempt due to external dependencies beyond our control. If exemptions occur, it will reduce the estimated amounts of delay credits we have projected.

Furthermore, even if OpenAI or Cerebras chooses to cancel a batch of delayed computing power, we believe that with sufficient market demand, we can easily find other customers to take over.

OpenAI also retains the option to purchase up to an additional 1.25 gigawatts of computing power beyond the initial 750 megawatts commitment, which can be brought online as early as 2029. Given that there are still three years until then and the rapid changes in the AI field, we have not yet begun discussions on this.

Balance Sheet and Capital Efficiency

We continue to maintain strong liquidity. At the end of Q3, cash, cash equivalents, restricted cash, and marketable securities totaled approximately $7.8 billion. We also have a revolving credit facility of up to $850 million, which has not yet been utilized.

The $1 billion prepayment working capital loan obtained in January this year, along with ongoing receipts from non-recurring pass-through costs used for the renovation of new data center capacity as part of the OpenAI contract, have funded most of our net working capital needs and capital expenditures to date this year.

We are not an emerging cloud service provider but rather a vertical integrator of the fastest inference solutions in the industry. Our business possesses the resilience afforded by technological differentiation, which we believe will drive higher profit potential and cash generation capabilities in the long term.

Our model also has lower capital investment in deployed computing equipment. For the large number of devices we manufacture in-house at data centers—similar to other companies that choose to develop and deploy AI chips—we do not have to pay third parties a 70% gross margin, nor do we need to use expensive HBM, only covering our bill of materials (BOM) costs. This reduces the relative capital expenditure per megawatt and allows us to produce the fastest and most valuable tokens.

This is reflected in our capital structure—despite the rapid expansion of our private cloud construction scale, the only debt on our balance sheet is the OpenAI working capital loan. We also continue to receive prepayments from customers, which is another lever to enhance capital efficiency.

Our current capital structure includes ample cash and investments, with minimal liabilities, providing us with ample flexibility to continue financing in various ways to support rapid growth—especially advantageous in the current capital market environment.

Outlook

Q4 2026 Outlook:

• Core Revenue: $310 million to $330 million

• Core Gross Margin: 47% to 49%

• Core Operating Margin: -1% to 1%

Full Year 2026 Outlook (Upward Revision):

• Core Revenue: $940 million to $960 million

• Core Gross Margin: 42% to 44%

• Core Operating Margin: -10% to -8%

Summary

Q3 was another record quarter for Cerebras, and Q4 is expected to be the same. Our flexible hardware and cloud services co-selling model is driving strong revenue growth, and the improvement in gross margin has come a full quarter ahead of expectations.

We have made significant progress in scaling manufacturing and data center capacity, which will enable us to achieve a threefold revenue increase by 2027 while continuing to push core gross margin towards the long-term target of 60%. We have demonstrated strong operational leverage and expect core operating profit to be close to breakeven in Q4, after which we anticipate a significant move towards the long-term operating margin target of 40% by 2027.

This combined advantage of scale, growth, and profitability makes us a uniquely positioned leading AI systems provider.

Lastly, just a reminder: we have been public for over 180 days, and the IPO lock-up period will fully expire on November 9.

Thank you for listening to our prepared remarks, and now I would like to open the line for questions.

Q&A Session

Operator: Thank you, everyone. We will now begin the Q&A session, please hold for a moment. The first question comes from Tom O'Malley at Park Place. Please go ahead.

Analyst (Tom O'Malley, Park Place):

Hi, Andrew, hi, Bob, thanks for taking my questions, and congratulations on the strong performance. My first question is about the new contracts—Oracle and Microsoft. Can you provide some insight into the relative scale of these two contracts? Also, Bob, you mentioned that the Oracle portion has been included in the RPO; can you share the specific amount that has been included?

Andrew Feldman:

Tom, great to talk to you. We will provide more details about the Microsoft collaboration and the agreements we have reached in the first half of next year, at which point we also expect to start generating revenue. Bob?

Bob Komin:

Tom, regarding the Oracle portion included in the RPO, just to clarify—we have included the estimated amounts for OpenAI, which is the total committed revenue under the take-or-pay contract, excluding any delay options or additional options, and has deducted the delay credits we believe may need to be paid. This also includes non-recurring fees paid by customers for facility construction for the data centers contracted for OpenAI, as well as recurring monthly fees during the contract term, all of which are recognized as pass-through revenue and pass-through expenses in the income statement.

Tom O'Malley:

Understood, Bob. I was referring to Oracle, not OpenAI, but we can discuss that later. Now, regarding the Oracle portion------

Bob Komin:

We have made significant efforts to win hyperscale customers. The situation with Oracle is more straightforward—currently, we have not included Oracle in the RPO because the agreement was signed after September 30.

Tom O'Malley:

Got it, thank you. The next question is: we see that cutting-edge AI labs can not only rapidly develop hardware (like the Jalapeno project) but also accelerate the deployment of models across different computing hardware, enhancing hardware versatility. How does your company benefit from this trend, and what are your expectations for the future development of heterogeneous infrastructure? What advantages does this bring to Cerebras?

Andrew Feldman:

That's a very good question. Profit companies mentioned in their S1 or in some tweet that they completed the migration to AMD GPUs in about three days.

Moving online.

This is a very meaningful and important development, especially for long-term observers. I believe the industry has not yet fully recognized what this means for our ecosystem.

Compilers and underlying software have historically been fundamental barriers faced by new architectures; they inherently favor incumbents and are a core source of what people often refer to as the "software moat." If Profit Corporation can truly achieve what they claim, it would mean that the value of underlying software will shrink and eventually approach zero over time. The software moat would also narrow and disappear.

If the value of underlying software trends towards zero, the only remaining source of value will be innovative hardware. This means that competitive differentiation will increasingly depend on the degree of hardware differentiation.

I believe this presents a huge advantage for innovative hardware companies like Cerebras—and other hardware pioneers adopting entirely new architectures—while simultaneously weakening the long-held advantages of incumbents.

This is a very good question and deserves close attention from the entire industry. Thank you.

Operator: Thank you for the question. The next question comes from Joe Moore at Morgan Stanley. Please go ahead.

Joe Moore (Morgan Stanley):

Thank you. Congratulations on the strong hardware performance. Could you provide any specific updates regarding the RPO issues related to the cloud business schedule? Additionally, how confident are you in resolving these issues before delays lead to additional penalties?

Andrew Feldman:

I believe the numbers we included in the RPO are extremely conservative estimates in the absolute worst-case scenario, so we are confident that the actual results will exceed this estimate. Such clauses are common in large data center contracts—we have similar agreements with data center builders, and delays will trigger corresponding clauses as well. Typically, both parties will decide whether to enforce these clauses based on commercial considerations at that time. I think this is the usual way these agreements operate and will continue to do so in the future.

Joe Moore:

Understood. Regarding the reasons for the delays, I think everyone can understand the challenges related to land, power, and construction. Are these similar reasons, or are there other factors?

Andrew Feldman:

Our current delays are entirely due to land, power, and construction issues.

Joe Moore:

Got it. We have the computing power and all the materials; we are just waiting for the completion of the land, power, and construction work.

Andrew Feldman:

Exactly. Thank you, Joe.

Operator: Thank you. The next question comes from Atif Malik at Citigroup. Please go ahead.

Atif Malik (Citigroup):

Thank you for taking my question. Andrew, I have a question about the product roadmap: When do you expect the crossover point between CS4 and CS3 to occur? Is it in 2027? Additionally, CS4 has shown significant improvements in throughput; if this path continues to CS5 and subsequent products, will your chips be able to provide sufficient speed and throughput on their own without needing to pair with GPUs?

Andrew Feldman:

Atif, hello, it's great to talk to you. First, just a few weeks ago, all our deployments were using CS3 exclusively; we have only recently begun deploying the decomposed inference solution, which has led to significant throughput improvements, which is very exciting. This is the first observation.

Clearly, CS3 performs very well on its own, but the decomposed inference (which we are currently deploying and has entered formal availability) can generate more high-speed tokens under high-speed token constraints.

Regarding the crossover point between CS3 and CS4, I expect it to occur in the first quarter of 2027. The product roadmap looks very good, and CS5 is ahead of schedule. Although I mentioned it in my remarks, it is indeed an extraordinary execution achievement that it only took five months from tape-out to complete system testing for the chip. CS5 is expected to begin mass production in early summer 2027, with performance doubling compared to CS4, while CS4's performance doubles compared to CS3.

Atif Malik:

Great. Bob, regarding the flat guidance for cloud revenue in the fourth quarter: Can you help us understand the business structure between cloud and hardware in the context of the goal of tripling revenue by 2027?

Bob Komin:

I think the key to understanding the hardware business structure is that the scale of these orders can be very large. We hope to provide some guidance on shipment status in the next couple of quarters. This largely depends on whether the customer data centers are ready. It can be said that in the fourth quarter and into the first quarter, the hardware proportion will be closer to 50% than it has been in the past few quarters.

As for the rest of 2027, we will have more information during the fourth quarter earnings call, at which point we will provide a more complete outlook for 2027 as a whole.

Atif Malik: Thank you.

Operator: Thank you. The next question comes from Quinn Bolton at Needham. Please go ahead.

Quinn Bolton (Needham):

Hello everyone, congratulations on the progress. Andrew, I want to confirm my understanding: In your RPO and assumptions, you have made provisions for potential delay offsets, but you have not yet received any cancellation notices regarding delayed computing power, is that correct?

Andrew Feldman: Correct.

Quinn Bolton:

Good. Second question: There is a concern in the market that low-batch GPUs can achieve fast inference. It sounds like a very costly token delivery method. Can you compare the cost per token of Cerebras with low-batch GPU configurations?

Andrew Feldman:

Certainly. GPUs can achieve higher speeds by reducing batch sizes, but at the cost of sacrificing throughput—throughput is very low at low batch sizes, and when you try to increase speed, GPUs enter the low-batch range, resulting in very small total throughput.

This is determined by architectural characteristics. It can be understood this way: GPUs can indeed achieve a certain level of performance improvement by reducing batch sizes, but doing so causes the cost per token to rise sharply—while the number of users served decreases significantly, and the total token production is extremely limited.

As speed requirements increase further, our advantages will become stronger. When speeds reach levels that GPUs cannot provide, Cerebras will be the only option—at a batch size of 1, GPUs still cannot compete with us.

Regarding the cost crossover point: around 250 to 300 tokens per second, we have a cost advantage; at 500, 800, and 1000 tokens per second, our advantage will far exceed that of competitors; in the range of 800 to 2000 tokens, GPUs cannot compete at all. This is the basic landscape of the current competitive situation. This question touches on a complex topic and is very well asked.

Quinn Bolton: Thank you, Andrew.

Operator: Thank you for the question. The next question comes from Joshua Bruckheimer at TD Cowen. Please go ahead.

Joshua Bruckheimer (TD Cowen):

Hello everyone, thank you for taking my question, and congratulations on the successful partnership with Microsoft. First, please discuss how you are allocating computing power in the context of land, power, and construction issues? It seems that chip and rack supply is not a problem, so how much room is there for incremental hardware sales? Additionally, given that cloud customers are facing challenges with land, power, and construction, have you discussed with OpenAI the possibility of OpenAI or its partners purchasing hardware and deploying it in their own data centers? Thank you.

Andrew Feldman: This question indeed has two layers. Josh, the first layer is: Are land, power, and construction and their construction truly limiting factors? The answer is yes. TSMC has made a commitment to us for ample wafer capacity in 2027, and we are very confident about that. Not using HBM is also a significant advantage, as HBM is currently in very tight supply.

Your second layer question is very good: Does this create opportunities for us to sell hardware during the data center construction period? The only thing to note is that, as far as we know, almost all data center constructions are delayed—including OpenAI and all the hyperscale cloud service providers we know. Data center delays are a global phenomenon and not unique to us; our customers are facing the same constraints. So, the answer is: Yes, we have product supply; yes, we are shipping to customers with data centers; but at the same time, the current widespread delays in data centers are indeed affecting actual progress. Is that clear?

Joshua Bruckheimer:

Understood, but the last question in the three-part question, which is the more difficult part of my question—Is there a possibility that the OpenAI contract could turn into hardware procurement?

Andrew Feldman:

Sorry, the contract does grant them that right, but as far as we understand, their own data center construction is also behind schedule.

Joshua Bruckheimer: Okay, understood. Thank you. So, a follow-up question—

Andrew Feldman: Josh, this is already the seventh question.

Joshua Bruckheimer: Indeed, my question, I admit. Can you talk about how Oracle and Microsoft are positioning Cerebras' products and services to their customers? Is this primarily reflected as hardware sales, and then driven by ultra-fast inference demand for customer usage? Thank you.

Andrew Feldman: Certainly. More details about the Oracle partnership have been released, and the specifics of the Microsoft partnership will be disclosed in 2027. Regarding Oracle, they are procuring hardware to be deployed in their data centers in the U.S., and we are collaborating; they are also building a commercial pipeline for this. Currently, there is a lot of interest, and we do not see this as a one-time collaboration; it marks the beginning of a large partnership, and our equipment will be deployed as part of OCI cloud in Oracle's data centers.

Quinn Bolton: Thank you, Andrew, I’ll take a question opportunity next quarter.

Andrew Feldman: A hamburger for a question opportunity next quarter?

Quinn Bolton: Deal.

Operator: Thank you very much. This concludes today's Q&A session, and now I will hand it over to Sean for closing remarks.

Sean Dorsey (Head of Investor Relations): Finally, I would like to remind everyone that we will be attending the TomorrowX Summit in Austin next week on November 17, and the Barclays Global Technology Conference in San Francisco on December 9.

Thank you all for participating today. Operator, please close the meeting.

Operator: Thank you everyone, this meeting is now concluded, please disconnect. The meeting has ended.