目录
I. Opening Remarks
II. Chief Executive Officer’s Remarks
Data Center Capacity
Manufacturing and Supply Chain
Support for GPT-5.6 Sol
Disaggregated Inference
Product Roadmap
Customer Progress
CEO Summary
III. Chief Financial Officer’s Remarks
Second-Quarter Revenue
Gross Margin
Operating Margin
Cash, Liquidity, and Capital Expenditures
Third-Quarter and Full-Year Guidance
CFO Summary
IV. Management Summary
V. Q&A
UBS: Customer Concentration and Manufacturing Capacity
TD Cowen: AWS Commercial Model and AMD Product Portfolio
Barclays: AMD Partnership Economics and Trainium Deployment
Needham: Throughput Gains and Heterogeneous Inference
Morgan Stanley: Commercialization Timeline for Disaggregated Inference
Mizuho: Data-Center Expansion and Neoclouds
VI. Closing Remarks
本内容基于公开资料和研报数据整理，不构成任何投资建议，不代表任何个人观点，仅供学习参考，请理性阅读
I. Opening Remarks
Operator: Good afternoon, and welcome to the Cerebras fiscal second-quarter 2026 earnings call. Following management’s prepared remarks, we will open the call for questions. This call is being recorded. I will now turn the call over to Sean Dorsey, Head of Investor Relations.
Sean Dorsey: Thank you. Good afternoon, and welcome to the Cerebras fiscal second-quarter 2026 earnings call.
Earlier today, we issued a press release and posted supplemental earnings presentation materials on our investor relations website. A replay of this webcast will also be available on the investor relations website after the call.
Joining me today are Andrew Feldman, our co-founder, CEO and President, and Bob Komin, our CFO.
Before we begin, I would like to remind everyone that today’s discussion will include forward-looking statements made under the safe-harbor provisions of the Private Securities Litigation Reform Act of 1995. These statements include, but are not limited to, our future financial performance, business strategy, market opportunities, customer demand, product roadmap, technology leadership, supply chain, operating model, and outlook for the third quarter and full fiscal year 2026.
Forward-looking statements are based on our current expectations and assumptions and are subject to various risks and uncertainties. Actual results may differ materially from those expressed or implied. These risks are described in our filings with the U.S. Securities and Exchange Commission, including the final prospectus relating to our initial public offering and our future periodic SEC filings. Except as required by law, we undertake no obligation to update these forward-looking statements.
During today’s call, we will also discuss certain non-GAAP financial measures. Reconciliations between GAAP and non-GAAP results are included in today’s press release and supplemental materials, which are available on our investor relations website.
With that, I will turn the call over to Andrew.
II. Chief Executive Officer’s Remarks
CEO: Thank you, Sean, and thank you all for joining us today.
The second quarter was strong. We completed our public offering without allowing it to distract from business execution. Core revenue reached a record high, while core revenue, core gross margin, and core operating margin all exceeded guidance.
Looking ahead, we see virtually unlimited demand for high-speed inference. The market is recognizing that speed is not merely a benchmark metric. It transforms user engagement, agent performance, and AI productivity. High-speed inference will unlock new applications and new markets.
As we have said previously, 2026 is a foundational year for Cerebras. In the 7 weeks since our last earnings call, we have made exceptional progress across several fronts, positioning us for tremendous growth in 2027, 2028, and 2029 as we progressively fulfill the $25 billion of remaining performance obligations currently on our books.
Based on this progress, we expect core revenue in 2027 to grow to more than 3 times the 2026 level, followed by several more years of multiple-fold growth. We measure our progress across 3 dimensions: capacity, capability, and customers.
We are expanding capacity to support extraordinary growth by signing new data center contracts worldwide, scaling manufacturing, and working with suppliers to secure supply.
We are expanding our capabilities by inventing new technologies that improve performance, throughput, and energy efficiency.
We are also broadening our customer base, both by increasing AI productivity in existing markets such as coding and agentic workflows and by opening new markets such as security, where speed creates entirely new opportunities.
Data Center Capacity
From a capacity perspective, data center space remains an industrywide bottleneck, and Cerebras is no exception. The faster we and our customers bring new data centers online, the faster we can grow.
Over the past 7 months, we have moved aggressively to secure and build data center capacity. We have 2 advantages.
First, because our products are used for inference, we do not need to find single sites with the gigawatt-scale footprints required for training clusters. This gives us greater flexibility to expand capacity across numerous locations worldwide.
Second, we have established a repeatable process for site selection, cluster deployment, and customer activation. This operational capability is essential for converting gigawatts of power capacity into production-grade tokens globally.
I am pleased to report that these efforts have been highly successful. We now have operational or contracted data centers in Alabama, Dallas, Denver, Minneapolis, Santa Clara, and Stockton in the United States, as well as France, Finland, Manitoba, Montreal, Norway, Saskatchewan, and Toronto outside the United States.
Over the past 7 months, we have secured more than 600 MW of data center capacity. Some of this capacity is already online, with the remainder scheduled for delivery by the end of 2027. While this remains far short of demand, our pipeline of potential expansion projects continues to grow and is now measured in gigawatts.
At scale, as we continue building our first-party cloud platform, it will become one of the world’s largest non-hyperscale AI clouds.
At the end of 2025, we were still on a steep learning curve. Today, I am pleased to say that we have become quite good at building data centers, with a clear path to becoming exceptional.
Manufacturing and Supply Chain
Manufacturing and the supply chain are the other critical components of capacity.
We have successfully expanded manufacturing and are building new facilities with Flex and Sanmina. We expect manufacturing capacity to increase to more than 10 times its previous level in 2026 and to expand further in 2027, positioning us for the rapid growth we anticipate over the coming years.
Our engagement with supply-chain partners has also become a significant advantage. TSMC has once again provided strong support, and we have secured the wafers required to support future growth.
Our wafer supply also benefits from the fact that Cerebras achieved industry-leading performance on TSMC’s 5-nanometer process, where wafer costs are lower and supply constraints are less severe.
Our decades-long relationships with supply partners further reinforce our confidence in delivering on future growth plans. Such relationships are particularly scarce and valuable during periods of supply shortages.
It is also worth emphasizing that most of the critical supply-chain constraints facing the industry do not apply to Cerebras. For example, we do not use HBM memory or CoWoS packaging, nor do we require 3-nanometer wafer-fabrication capacity.
Support for GPT-5.6 Sol
On product capabilities, during the second quarter we completed support for OpenAI’s GPT-5.6 Sol, the largest and most capable frontier model available today. Cerebras can serve GPT-5.6 Sol at 10 times the speed.
Support for GPT-5.6 Sol has eliminated any remaining doubt about Cerebras’ ability to host large frontier models. Serving as a delivery partner for GPT-5.6 Sol and offering the model through Cerebras Cloud demonstrate the maturity of our software stack.
Achieving the quality and reliability required at hyperscale takes millions of system-hours of production hardening. We are proud that Cerebras Inference Cloud can now meet the needs of the most demanding customers.
Our work serving frontier models also gives us an important strategic advantage previously available only to NVIDIA. Closed-source frontier models continuously introduce new insights and AI technologies. Serving these models gives us early visibility into where the industry is heading and enables us to prepare accordingly.
From hardware through the software stack, our product roadmap now reflects these frontier trends and will generate compounding advantages over the coming years.
Disaggregated Inference
Continuing with product capabilities, I will now discuss disaggregated inference.
We are currently working with 2 leading chip companies on disaggregated inference solutions: AMD with Helios and AWS with Trainium.
Disaggregated inference expands the market for both GPU vendors and Cerebras. It enables GPUs to participate in the high-speed inference market, where they have historically struggled, while allowing Cerebras to serve more price-sensitive customers and improve the economics of our data centers.
Here is how the solution works.
As in every computing market, opportunities for specialization emerge as the inference market evolves and matures. Disaggregation is one form of specialization, particularly well suited to workloads with predictable traffic patterns.
Under this approach, inference is divided into 2 stages—prefill and decode—with each stage running on a different processor.
Prefill processes the input submitted by a user or agent. Because this workload is parallelizable, it is well suited to GPUs and their HBM-based memory architecture.
Decode generates the output tokens. It is the more technically challenging stage and accounts for most of the computation in a disaggregated solution. Because decoding is sequential and requires high memory bandwidth, it is exceptionally well suited to the Cerebras Wafer-Scale Engine.
The prefill and decode processors must be interconnected to create an end-to-end solution. Cerebras’ standards-based I/O and open-partnership strategy make this integration straightforward.
Several weeks ago, we announced a partnership with AMD to build a disaggregated inference solution. The solution combines AMD Helios racks with Cerebras CS systems, increasing throughput to 5 times the previous level while preserving Cerebras speed.
Understanding the significance of this achievement requires distinguishing between speed and throughput.
Speed is measured at the individual-user level in tokens delivered per user per second. It determines how quickly a user’s question is answered or how long an agent takes to complete a task. On charts, speed typically appears on the horizontal axis.
Throughput is the total number of tokens a solution can generate per second, aggregated across all concurrent users. It is typically shown on the vertical axis.
Speed is critical to the user experience, while throughput determines the economics of inference.
GPU solutions can deliver high throughput, but only at low speeds. Once configured to support moderate speeds, GPU throughput falls sharply. This limitation applies not only to GPUs but also to ASICs and all HBM-based solutions.
HBM memory architectures force systems to trade off throughput against speed. Cerebras’ SRAM architecture has the opposite profile: it delivers extremely high token-generation speeds with comparatively moderate throughput.
GPU vendors want to improve speed without sacrificing throughput, while Cerebras wants to improve throughput without sacrificing speed. That is precisely the advantage of a disaggregated solution: it increases throughput 5-fold while preserving Cerebras speed.
Increasing throughput 5-fold while maintaining industry-leading speed has profound implications for token-generation economics. It means each Cerebras system can generate up to 5 times as many high-speed, high-value tokens.
