NVIDIA announced after the U.S. market closed on Wednesday, Eastern Time, that its revenue for the first quarter ended April 26, 2026, reached a new record of $81.6 billion. This represents a 20% increase from the previous quarter and an 85% increase compared to the same period last year.
Following the earnings release, NVIDIA's President and CEO Jensen Huang and Executive Vice President and CFO Colette Kress, along with other executives, participated in the subsequent earnings conference call to discuss key financial points and answer analyst questions.
The following are the main points from the analyst Q&A session of the call:
**Morgan Stanley analyst Joe Moore:** I'd like to ask about the rationale behind the segment reorganization. What is the philosophy behind providing data in this way? Also, could you discuss the competitive differences between the two segments, and regarding the impressive CPU data you mentioned, how do you view its performance across these segments? Thank you.
**Jensen Huang:** Thank you, Joe. First, a quick correction: Colette misspoke earlier. We increased the quarterly dividend from 1 cent to 25 cents. I believe that extra 5 cents is quite significant for major shareholders. Now, back to your question, Joe. Regarding segment reporting and business description, we want you to better understand our business.
AI and computing are highly diverse, and this diversity manifests in several ways. First, AI spans multiple domains. Depending on the industry, it could involve 3D graphics, manufacturing, industrial robotics, protein research in life sciences, small molecule research in life or materials science, physical sciences including energy, research labs, higher education, and so on. The application areas for AI are extremely broad.
Second, the deployment environments are diverse: it could be in enterprises, energy, manufacturing, and other sectors. The operating environments are also diverse, ranging from hyperscale clouds, AI-native companies—a large number of which are emerging globally—to on-premises enterprise deployments, industrial factory deployments, supercomputing centers, and the edge. This, of course, includes well-known areas like autonomous vehicles, robotics, and the growing computer networks within manufacturing plants, such as chip factories, packaging plants, and computer factories. In the future, every base station and every radio network will become AI-driven radio networks.
Finally, the management models are diverse. Operations might be run by public clouds, but some cannot operate on regulated clouds due to industry regulations, confidential computing, national security reasons, etc., requiring different construction models for different data centers.
NVIDIA's uniqueness lies in being the only company that builds all technological components. We create products with extreme co-design, complete end-to-end, full-stack approaches and adapt them to various environments. Some environments, like those for enterprise customers, require a company with a full suite of technologies so they can purchase and use the system without building it themselves. Therefore, the data center market has multiple segments, and NVIDIA's fully integrated, full-stack solutions—while remaining open—are crucial for this product delivery model.
If we look at our segment breakdown, the simplest way to categorize everything I just described is into three major segments: Hyperscale Cloud, AI-Native / Enterprise / Industrial On-Premises, and Robotics & Edge. Hyperscale Cloud is the first major segment. We collaborate with them in three modes: first, accelerating their data processing and machine learning workloads to support their internal AI processing; second, bringing significant business and NVIDIA ecosystem business to their public clouds.
The second major segment is AI-Native companies, enterprise on-premises, and industrial on-premises. This segment is growing extremely rapidly because every industry, every country, and every company needs AI, each with different construction needs. We provide a complete solution, enabling them to build AI systems easily.
The third major segment is Robotics & Edge. Past computing was primarily personal computing; the future will be personal AI. Autonomous vehicles are a classic example of personal AI—it's a robotic system, essentially your personal AI. In the future, various robotic systems will emerge, and as I mentioned, every base station radio network will move in this direction.
This is the simplest way to understand our business. Each segment has different characteristics in many aspects, including operating systems, business models, and market expansion methods. Among them, market expansion for hyperscale cloud providers is the simplest, as there are only about five or six globally. Other industries encompass 250,000 companies worldwide, making market expansion complex and diverse.
It's essential to have a highly diverse understanding of AI. As is well known, NVIDIA possesses the world's largest suite of acceleration libraries, spanning computational photography, fluid dynamics, particle physics, molecular dynamics, and many other fields. All these libraries are key to our service in vertical industries, which form the core of the second and third segments. In summary, our business has grown to such a large scale that segment reporting helps you better understand our business model.
**Melius Research analyst Ben Reitzes:** I'd like to ask Mr. Huang about your growth philosophy. Excluding China, the Data Center business grew approximately 120% year-over-year this quarter. Many, including myself, predict hyperscaler capital expenditures will grow 90% to 100% this year. You mentioned earlier that the Data Center business is still expected to reach a $3-4 trillion scale by the end of the decade. I'm wondering if you still believe the company's growth rate can outpace hyperscaler capital expenditure? Do you think hyperscaler capital expenditure will continue to grow at a high rate after this year? Thank you very much.
**Jensen Huang:** Thank you, Ben. First, our growth rate should outpace hyperscaler capital expenditure, and the reason is illustrated by the segment breakdown I just described. Our Data Center business is divided into two major parts. The reality is more complex, but for simplicity, we can think of it as two main components.
The first part is the hyperscalers, which is the hyperscale capital expenditure you mentioned, expected to reach $1 trillion this year. I fully expect this number to continue growing for good reason: this is the direction of future computing. Without computing power, they have no revenue. It's clear that computing power is revenue; computing power is profit.
The world is changing. Software and SaaS (Software-as-a-Service) previously didn't require massive computing power, but AI demands enormous computing power while delivering far greater value than before. This is why frontier AI companies are developing. AI companies like Anthropic and OpenAI are growing at an astonishing rate—growth in one month equivalent to a decade of growth for some SaaS companies. That speaks volumes.
The first segment is hyperscalers, with trillion-dollar capital expenditure moving towards a $3-4 trillion scale. The second segment includes all AI-native cloud providers, which are spread across global regions and supported by numerous startups; also, 250,000 enterprises worldwide, many of which must or want to build their own AI factories; numerous industrial enterprises that must deploy computing power on-site and cannot rely on the cloud, requiring reliable, fast, and stable response—scenarios like chip factories connecting to cloud services are completely unrealistic. Additionally, there are sovereign AI clouds. This entire data center segment is not suitable for semi-custom chips because these data centers need to purchase and run systems directly, not design and build them themselves.
The second segment is extremely diverse. The first segment's revenue comes from about five or six companies, while the second segment involves thousands, and in the future, hundreds of thousands of companies, with relatively smaller individual deployment scales. This segment will continue to grow at an astonishing pace.
What I refer to as Physical AI, and the hundred-trillion-dollar industries not digitized over the past 30 years, which are about to be reshaped by AI, pertains to this second segment. This segment is growing extremely rapidly, and our market share in it is very large. Our uniqueness lies in our ability to serve this industry. Our platform is designed with vertical integration, all components work together, while also supporting flexible configurations—customers can purchase and assemble as needed. The second segment is not yet fully understood because it contains a large number of small and medium-sized businesses, with individual deployment scales being smaller compared to hyperscalers.
Judging from segment size and breakdown, our market share in the hyperscaler market is increasing, as we've added an important partner, Anthropic, and will strongly support its computing power expansion in the coming years. As for the second segment, due to our platform solution, few other companies can enter it.
**Cantor Fitzgerald analyst C.J. Muse:** Good afternoon, thank you for taking my question. With the upcoming AI superchip platform Vera Rubin, you clearly have deep insights into frontier model updates and new technologies optimizing diverse AI workloads. Investors are very focused on your inference market share. How will Vera Rubin and "extreme co-design" impact your inference market share in late 2026 through 2027?
**Jensen Huang:** Our market share in inference is growing, and it's growing very fast. The reason is that the number of frontier model companies has increased this year. For example, Cursor, Perplexity, and some new model companies like TML, Reflection, etc. The number of frontier model companies keeps growing. This year, we established a partnership with Anthropic. Their expansion is very rapid. We are working with them to secure computing capacity on platforms like Azure, AWS, CoreWeave, etc. I forgot which other companies we announced partnerships with, but there is a long list of others, and we are going live for them.
Therefore, the capacity we are launching for Anthropic this year and next will be quite substantial, very substantial. Our business is growing, and until recently, our coverage of Anthropic was essentially zero. So, our inference market share is growing very fast. In that regard, Vera Rubin will be even more successful than Grace Blackwell. I cannot think of any frontier model company that would not choose Vera Rubin from the start, which was not the case with Blackwell before. So, Vera Rubin is off to a good start; it will definitely be more successful than Grace Blackwell.
Returning to Ben's earlier question. Everything I explained in the inference question is actually concentrated in the hyperscaler segment. Don't forget, we have an entire second segment of AI data centers, almost exclusively served by us. This segment is highly fragmented, requiring highly integrated platform solutions and extensive market channels. The majority of inference business in this segment is located in Asia. Additionally, there is Physical AI, where NVIDIA is currently almost the only vendor serving it. We have been deeply involved in this field for a long time, and the business continues to grow. Therefore, our inference market share is rapidly increasing.
**UBS analyst Timothy Arcuri:** I'd like to ask about your progress in custom commercial products, such as CPX and LPX. You previously mentioned such products might capture 20% of the market. I assume you've made good progress with LPX. Could you discuss the situation and how this fits into your broader platform strategy?
**Jensen Huang:** LPX is designed for low latency and high token rate, but its throughput is lower, model capacity is limited, and context processing capability is weaker. For example, it underperforms in scenarios requiring handling large contexts, like software coding, agent workloads, etc. So its challenges are clear.
As I explained before, LPX's application scenarios are not broad. It mainly targets customers with a large portfolio of different token services, focusing on high-end services with high token rates. The number of such customers is not large, but their token rates are extremely high. This is entirely consistent with my previous statements, and I maintain this view. I expect LPX and other accelerators based on static random-access memory architecture, focusing on high token rate generation, will remain niche products for some time.
As you know, Grace Blackwell and Vera Rubin support the entire AI lifecycle, from data processing, training preparation, pre-training, fine-tuning, reinforcement learning, all the way to inference. Grace Blackwell is the best platform globally for completing all these tasks. In some cases, if a service provider already has high token rate services, we can pair it with LPX to make their service performance even better.
That's my view of the market. Whether the market share is 20% or 10% depends on the stage of AI development. Currently, it's far below 20%. In the future, high-end token scenarios might reach 20%. We are ready to work with service providers to achieve this capability, and I'm excited about it.
**Bank of America Merrill Lynch analyst Vivek Arya:** Thank you for taking my question. Regarding CPUs for agent applications, there's a lot of exciting news, and also discussion about CPUs potentially outnumbering GPUs. I'd like to hear your perspective: First, is this incremental workload, or will it cannibalize workloads that GPUs should handle? Second, the $20 billion figure you gave—is that for standalone Vera CPUs, or does it include the Vera portion within Vera Rubin?
**Jensen Huang:** The $20 billion is for standalone CPUs. Remember, our Vera has three usage modes. Actually, as a standalone CPU, it has four usage modes. Let me start with the first one you already know. The first is Vera Rubin. We will sell millions of Rubin, and every two of them connect to a Vera. Of course, we will price them, and reasonably. That's the first use case. The second use case is Vera standalone CPU. The third is Vera paired with CX9, for the software stack for storage.
The fourth is Vera paired with CX9, for the software stack for security and compute isolation, confidential computing. Each application scenario is based on Vera, and I expect Vera Rubin to face supply constraints throughout its entire lifecycle. In summary, the answer to your question is: the $20 billion is standalone CPU revenue.
Regarding CPUs, agents are essentially what we call scheduling frameworks. The scheduling framework handles input/output, orchestration, memory management, tool calling, like browsers, C compilers, Python compilers, etc. The scheduling framework runs on the CPU, and tool calling also runs on the CPU. For example, when an AI performs a search or uses a browser, these operations run on the CPU.
There are 1 billion human users globally. I believe there will be tens of billions of agents in the future—not now, but gradually developing. These tens of billions of agents will all use tools; their tools will be like the personal computers we humans use now. In the future, agents will work like humans using computers. Each agent will spawn sub-agents. Every spawn requires inference; the thinking process is entirely on the GPU, and all orchestration work basically runs on the CPU. When sub-agents launch, and when agents use simulators, these tasks can run on the CPU or GPU. That's why we closely collaborate with companies like Cadence and Synopsys to accelerate the operation of all tools globally.
We are accelerating all tools globally, data processing engines, and database engines because agents use these tools, and they are less patient than humans, demanding fast responses. Therefore, we are driving all tools to run on GPUs based on CUDA for faster speed. We will need a large number of CPUs in the future, and Vera is the CPU designed for agents.
Past CPU designs had many cores, convenient for leasing; the traditional economic model of cloud computing charged per core. The future economic model for AI is charging per token—how many tokens can be generated per dollar. What we need to do in the future is generate and process tokens as fast as possible, and Vera excels in this regard. Therefore, we anticipate NVLink 72 will be very successful. It requires extremely strong security and confidential computing capabilities, which is why Vera Rubin is the world's first end-to-end confidential computing platform. And, as you know, it also needs a powerful CPU. We are ready; all aspects are covered.
**Bernstein Research analyst Stacy Rasgon:** Hello everyone, thank you for taking my question. I'd like to return to the segment question. First, I'm wondering which segment emerging cloud providers belong to—Hyperscale business or AI Cloud business? I lean towards the latter, but I'm not entirely sure. Also, their sizes are currently comparable. You seemed to imply earlier that the AI Cloud business might grow faster than the Hyperscale business in the future. Is that your view? Or do you think both will maintain similar growth rates?
**Jensen Huang:** First, you are correct. AI-native cloud providers do not develop or design their own chips, nor can they assemble disparate components into an AI factory. They have extremely high demands for time-to-first-token, require high architectural compatibility, need to run all models, and serve global customers. Therefore, NVIDIA's architecture is perfect for them.
We provide all components. Even if we don't, our ecosystem partners do. All components are fully integrated and work together. Almost all AI developers globally, AI-native startups, SaaS companies, enterprises, and industrial companies can lease computing power from AI-native cloud providers. Our architecture is the world's easiest-to-lease, highest-performance, easiest-to-deploy, lowest total cost of ownership, and easiest-to-finance computing platform. These characteristics perfectly match the needs of AI-native companies. They are similar to large enterprises, so they fall into the second segment.
This segment started growing after the AI ecosystem and hyperscale business developed, for many reasons: they have excellent computer science talent, strong data center capabilities, and primarily focus on consumer applications. Even if minor issues occur, the consequences are not severe; they just need to improve service quality. In contrast, other application scenarios like industrial and enterprise require AI to have strong capabilities, create real value, and generate benefits and revenue safely and reliably before being truly adopted. Therefore, the second segment developed slower than the hyperscale business, and data reflects this.
But in the long term, the industrial and enterprise sectors are clearly the core of the future economy because they correspond to a $50-80 trillion global economy, and AI will further expand this scale. Therefore, I expect the second segment to be larger in the long term. In the short term over the next few years, both segments will achieve astonishing growth. I expect the second segment to grow faster, but both will grow at high rates. I hope that within the next five years, the Physical AI and Robotics segment will also experience explosive growth.
**Goldman Sachs analyst Jim Schneider:** Good afternoon, thank you for taking my question. I recall at the GPU Technology Conference (GTC), you mentioned that Rubin and Blackwell platform revenue is expected to reach $1 trillion. But I believe this expectation does not include LIX, Rubin, CPX, and Vera CPU Rack. Could you tell us if Vera CPU will be the largest growth source beyond this $1 trillion? Are you considering other product portfolios, including CPUs, to further expand overall market share? Thank you.
**Jensen Huang:** Regarding incremental sources beyond the $1 trillion expectation, I think the first is the continued increase in market share for frontier large models. I expect the share to keep growing. Second, we did not include any standalone CPU revenue in that number, so that will be the second largest increment. The market size for agent systems is extremely large. All customers are excited about Vera; we will sell a large number of Vera CPUs. Third is LPX. As I explained earlier, LPX is designed based on static random-access memory architecture. Its advantages are low latency, high bandwidth, and high interactivity, but throughput and context processing capabilities are limited—an inherent characteristic of static random-access memory architecture. The combination of Vera, Rubin, and LPX will cover full-scenario AI needs from pre-training, fine-tuning, inference, to agent systems.
**TD Cowen analyst Joshua Buchalter:** Hello everyone, thank you for taking my question, and congratulations on the excellent performance. Colette, you mentioned in your remarks that GB300 is the fastest ramp in the company's history. How should we gauge Vera Rubin's ramp pace? It's not a completely new architecture at the chip level, but the rack design is similar. Does this mean Vera Rubin's ramp pace is similar to GB300, or is it more gradual due to the new chip?
**Colette Kress:** We have stated multiple times before that Vera Rubin will launch in the second half of this year, with initial deliveries starting in Q3, and ramp pace continuing to accelerate in Q4. It's difficult to judge which product ramps faster at this point, but we already have clear demand plans and orders; all major customers are ready. These are extremely complex systems requiring integration and debugging, so I believe the key is the product's time-to-market. It's too early to tell now, but it's certain that deliveries start in Q3, ramp continues in Q4, and large-scale deliveries will also occur in Q1 next year.
**Jensen Huang's closing remarks for the conference call:**
**Jensen Huang:** This was an extraordinary quarter. Demand grew parabolically for a simple reason: Agent AI has arrived. AI can now perform efficient and valuable work; tokens are starting to generate profit, so model developers are racing to expand capacity. In the AI era, computing power is revenue and profit, and NVIDIA is the core platform of this era.
I'll summarize five key points:
First, NVIDIA is the only platform running all frontier AI models. With Anthropic joining our existing partners, including OpenAI, xAI, Meta, Gemini, and many others, our share in frontier AI continues to grow.
Second, we cover all hyperscale cloud providers, supporting their core data processing, machine learning workloads, internal AI services, and demand for NVIDIA GPUs in their public cloud services.
Third, our full-stack AI factory solutions and responsive global ecosystem allow us to exclusively serve the new AI data center segment, new AI-native clouds, sovereign AI clouds, and enterprise and industrial on-premises infrastructure. This is the second segment I mentioned earlier.
Fourth, the NVIDIA CUDA ecosystem extends to the edge, covering robotics, autonomous vehicles, embedded medical devices, and AI wireless communication base stations. The next wave is Physical AI, where billions of autonomous robotic systems will operate in the physical world. This is the third segment I mentioned.
Fifth, we have spent three decades building the NVIDIA computing platform, with a unified architecture, vast ecosystem, and extreme chip-system-network-software co-design. We prepared in advance precisely for the arrival of the Agent AI era, and that moment has come. I look forward to speaking with you all next time.
Comments