NVIDIA conference call: "The opportunity to bring Blackwell to the Chinese market" is real, and this year's gross margin is still expected to reach the mid-70% level

The latest financial report shows that NVIDIA's Q2 revenue maintained double-digit growth, with revenue from the new generation architecture Blackwell chips increasing by 17% quarter-on-quarter, seen by CEO Jensen Huang as a sign of "very strong demand." However, the company's core business, the data center, continued to underperform, partly due to a decrease in revenue from H20 chips, with no H20 sales in China during the quarter.

In the subsequent earnings call, CEO Jensen Huang emphasized, "The opportunities for the future are still enormous," and that in the coming years, certainly by the end of this decade, we will see very rapid growth and very significant growth opportunities.

Regarding sales in China, Executive Vice President and Chief Financial Officer Colette Kress stated, If geopolitical issues can be resolved, sales of its H20 chips to China in the third quarter are expected to reach $2-5 billion, and in the past few weeks, "some" Chinese customers have already obtained licenses.

China has a $50 billion business opportunity this year, with a year-on-year growth rate of up to 50%

Jensen Huang stated that China could bring a $50 billion business opportunity this year. If competitive products can meet this market, assuming this year is $50 billion, the market is expected to grow by 50% annually, just as the AI market is growing in other parts of the world.

He pointed out that China is a very large market, and NVIDIA is discussing with the U.S. government the importance of meeting the Chinese market, stating that "the opportunity to bring Blackwell to the Chinese market is real" and will continue to strive to sell Blackwell architecture GPUs to China.

Expectations of $3-4 trillion in AI capital expenditures are "quite reasonable"

Jensen Huang mentioned that by the end of this decade, AI infrastructure spending will reach $3 trillion to $4 trillion, and the scale and scope of these projects will bring significant long-term growth opportunities for NVIDIA.

He stated that the capital expenditures of the top four hyperscale cloud service providers have doubled in two years to $600 billion, and considering that cloud service providers around the world are building AI infrastructure, the expectation of "3 trillion to 4 trillion dollars" is quite reasonable in the next five years.

Kress pointed out that factors driving this sustained investment growth include: the need for significant increases in training and inference computing for inference agent AI, global construction of sovereign AI, enterprise AI adoption, and the arrival of physical AI and robotics technology.

Based on Blackwell, a 10x return on investment per token can be achieved

Kress stated that Blackwell has set a benchmark, establishing a new standard for AI inference performance. As inference and agent AI become more prevalent across industries, the AI inference market is rapidly expanding. Blackwell's rack-level NVLink and CUDA full-stack architecture address this issue by redefining the economics of inference Kress introduced that the new NVFP4 with 4-bit precision and NVLink72 on the GB300 platform has improved token efficiency by 50 times compared to Hopper, enabling the company to monetize its computing at an unprecedented scale. For example, a $3 million investment in GB200 infrastructure can generate $30 million in token revenue, a 10-fold return.

Gross margin still expected to reach mid-70% level

Kress stated that it continues to expect the non-GAAP gross margin to reach the mid-70% level by the end of the year. The company is accelerating investments in its business to address significant growth opportunities in the future.

Jensen Huang stated, "In many ways, the more you buy, the more you grow," our product cost-performance ratio is extremely outstanding, and customers can achieve excellent profit margins—therefore, the growth opportunities and gross profit opportunities using NVIDIA architecture are absolutely the best.

Expected to achieve $20 billion in sovereign AI revenue this year

Kress also mentioned at the meeting that its sovereign AI-related revenue this year could reach up to $20 billion, more than double last year.

Kress introduced that sovereign AI is on the rise, with countries leveraging local infrastructure, data, and talent to develop sovereign AI capabilities, presenting significant opportunities for NVIDIA. For example, the European Union plans to invest €20 billion to establish 20 AI factories in France, Germany, Italy, and Spain, including 5 super factories, increasing AI computing infrastructure tenfold.

Below is the transcript of Apple's earnings call (translated with AI tools):

Nvidia Q2 FY2026 Earnings Call

Event Date: August 27, 2025

Company Name: Nvidia

Conference Host:

Good afternoon. I’m Sarah, and I will be your conference host today. Now, I welcome everyone to the NVIDIA Q2 FY2026 financial performance conference call. All lines have been muted to prevent background noise. There will be a Q&A session after the speakers finish their remarks. (Host instructions). Thank you. Toshiya Hari, you may begin the meeting.

Investor Relations Toshiya Hari:

Thank you. Good afternoon, everyone, and welcome to the NVIDIA Q2 FY2026 conference call. Joining me today from NVIDIA management are: President and CEO Jensen Huang; Executive Vice President and CFO Colette Kress. I would like to remind everyone that our conference call is being webcast live on the NVIDIA Investor Relations website. The webcast will be available for replay until the meeting discussing our Q3 FY2026 financial performance takes place.

Executive Vice President and CFO Colette Kress:

Thank you, Toshiya. While continuing to navigate a dynamic external environment, we delivered another record quarter. Total revenue was $46.7 billion, exceeding our expectations, and we achieved sequential growth across all market platforms Data center revenue increased by 56% year-on-year. Despite a $4 billion decline in H20 revenue, data center revenue still achieved a quarter-on-quarter increase. NVIDIA's Blackwell platform reached record levels, growing 17% quarter-on-quarter. We began production and shipment of the GB300 in the second quarter. Our full-stack AI solutions provided to cloud service providers, emerging clouds, enterprises, and sovereign clients are all contributing to our growth. We are at the beginning of an industrial revolution that will change every industry. We expect AI infrastructure spending to reach $3 trillion to $4 trillion by the end of this decade. The scale and scope of these constructions provide NVIDIA with significant long-term growth opportunities. The GB200 NVL system is being widely adopted and deployed among cloud service providers and consumer internet companies. Lighthouse model builders, including OpenAI, Meta, and Mistral, are using GB200 NVL72 at data center scale for training next-generation models and providing inference model services in production. The new Blackwell Ultra platform is also performing strongly, generating billions in revenue. The transition to GB300 is seamless for major cloud service providers due to the shared architecture, software, and physical footprint with GB200, enabling them to easily build and deploy GB300 racks. The transition to the new GB300 rack architecture is seamless. Factory construction in late July and early August successfully transitioned to support GB300 capacity increases. Today, full production is underway.

Current production rates have returned to full capacity, producing approximately 1,000 racks per week. With additional capacity coming online in the third quarter, this output is expected to accelerate further. We anticipate widespread market availability in the second half of the year, with CoreWeave preparing to launch third-generation GB300 instances, which they have seen achieve 10 times the inference performance compared to H100 on inference models. Compared to the previous Hopper generation, the GB300 and VL72 AI factories promise a 10-fold improvement in energy efficiency per watt token, which will translate into revenue growth as data centers are constrained by power consumption.

Chips for the Rubin platform are in production at the wafer fab, including the Vera CPU, Rubin GPU, CX9 SuperNIC, NVLink 144 expansion switch, Spectrum-X horizontal scaling switch, and silicon photonic processors. Rubin is still on track to achieve mass production next year. Rubin will be our third-generation NVLink rack-scale AI supercomputer, with a mature and complete supply chain. This allows us to maintain an annual product cadence and continue innovating in computing, networking, systems, and software.

In late July, the U.S. government began reviewing licenses for the sale of H20 to Chinese customers. While some of our customers in China have obtained licenses in the past few weeks, we have not yet shipped any H20 products based on these licenses. U.S. government officials have stated that the government expects to receive a 15% share of the revenue from licensed H20 sales, but so far, the U.S. government has not issued regulations to codify this requirement Due to our continued handling of geopolitical issues, our third-quarter outlook does not include H20. If geopolitical issues are resolved, we should achieve H20 revenue of $2 billion to $5 billion in the third quarter, and if we have more orders, we can issue more invoices. We continue to advocate for the U.S. government to approve Blackwell's sales to China. Our products are designed and sold for beneficial commercial purposes, and every licensed sale we make will benefit the U.S. economy and its leadership. In a competitive market, we hope to win the support of every developer. If we compete globally, the U.S. AI technology stack can become the world standard.

Notably this quarter is the growth in shipments of Hopper 100 and H200. We also sold approximately $650 million worth of H20 products to non-restricted customers outside of China.

The quarter-over-quarter growth in Hopper demand indicates the breadth of data center workloads running on accelerated computing, as well as the powerful capabilities of the CUDA libraries and full-stack optimizations that continue to enhance the performance and economic value of our platform. As we continue to deliver Hopper and Blackwell GPUs, we focus on meeting the soaring global demand. This growth is driven by capital expenditures from cloud to enterprise, which are expected to invest $600 billion this year in data center infrastructure and computing, nearly doubling in two years. We expect annual AI infrastructure investment to continue to grow, driven by factors including: the need for significant increases in training and inference computing for inference agent AI, global construction of sovereign AI, enterprise AI adoption, and the arrival of physical AI and robotics.

Blackwell has set a benchmark as the new standard for AI inference performance. With the proliferation of inference and agent AI across industries, the AI inference market is rapidly expanding. Blackwell's rack-level NVLink and CUDA full-stack architecture address this by redefining the economics of inference. The new NVFP4 4-bit precision and NVLink72 on the GB300 platform improve token efficiency by 50 times compared to Hopper, enabling the company to monetize its computing at an unprecedented scale. For example, a $3 million investment in GB200 infrastructure can generate $30 million in token revenue, a 10x return.

NVIDIA's software innovations, combined with the strength of our developer ecosystem, have more than doubled Blackwell's performance since its launch. Advances in CUDA, TensorRT LLM, and Dynamo are unlocking maximum efficiency. Contributions from the open-source community to CUDA libraries, along with NVIDIA's open libraries and frameworks, are now integrated into millions of workflows. This powerful collaborative innovation flywheel between NVIDIA and the global community strengthens NVIDIA's performance leadership. NVIDIA is a major contributor to OpenAI models, data, and software Blackwell has introduced groundbreaking numerical methods for pre-training large language models. Using NVFP4, computations on GB300 can now achieve training speeds that are 7 times faster than using FP8 on H100. This innovation sets a new standard for AI factory efficiency and scalability, providing accuracy at 16-bit precision with the speed and efficiency of 4 bits. The AI industry is rapidly adopting this revolutionary technology, with major players such as AWS, Google Cloud, Microsoft Azure, and OpenAI, as well as companies like Cohere, Mistral, Kimi AI, Perplexity, Reflection, and Runway already embracing this technology.

NVIDIA's performance leadership has been further validated in the latest MLPerf training benchmarks, with GB200 achieving a comprehensive victory. Please look forward to the upcoming MLPerf inference results to be released in September, which will include benchmarks based on Blackwell Ultra.

NVIDIA RTX PRO servers have been fully produced for global system manufacturers. These are air-cooled, PCIe-based systems that can be seamlessly integrated into standard IT environments, running traditional enterprise IT applications as well as cutting-edge agent and physical AI applications. Nearly 90 companies, including many global leaders, have adopted RTX PRO servers. Hitachi uses them for real-time simulation and digital twins, Eli Lilly for drug discovery, Hyundai for factory design and autonomous driving validation, and Disney for immersive storytelling.

As enterprises modernize their data centers, RTX PRO servers are expected to become a multi-billion dollar product line. Sovereign AI is on the rise, with countries leveraging local infrastructure, data, and talent to develop autonomous AI capabilities, presenting significant opportunities for NVIDIA. NVIDIA is at the forefront of landmark initiatives in the UK and Europe. The European Union plans to invest €20 billion to establish 20 AI factories in France, Germany, Italy, and Spain, including 5 super factories, increasing AI computing infrastructure tenfold. In the UK, the NVIDIA-powered Isambard AI supercomputer debuts as the country's most powerful AI system, delivering 21 exaflops of AI performance, accelerating breakthroughs in areas such as drug discovery and climate modeling. We expect to achieve over $20 billion in sovereign AI revenue this year, more than doubling last year's figures.

The networking business achieved record revenues of $7.3 billion, with the growing demand for AI computing clusters requiring high efficiency and low-latency networks. This represents a 46% quarter-over-quarter growth and a 98% year-over-year growth, with strong demand for SpectrumX Ethernet, InfiniBand, and NVLink. Our SpectrumX enhanced Ethernet solutions provide the highest throughput and lowest latency networks for Ethernet AI workloads.

SpectrumX Ethernet has achieved double-digit growth both quarter-over-quarter and year-over-year, with annualized revenues exceeding $10 billion. At the Hot Chips conference, we launched Spectrum XGS Ethernet, a technology designed to unify distributed data centers into gigabit-level AI super factories CoreWeave is an early adopter of this solution, expected to double the GPU-to-GPU communication speed. InfiniBand revenue nearly doubled quarter-over-quarter, benefiting from the adoption of XDR technology, which offers double the bandwidth improvement compared to previous generations, particularly valuable for model builders. The world's fastest switch, NVLink, has 14 times the bandwidth of PCIe Gen 5, with strong growth as customers deploy Blackwell NVLink rack-scale systems. NVLink Fusion allows for semi-custom AI infrastructure, receiving widespread positive responses. Japan's upcoming FugakuNEXT will integrate Fujitsu's CPU with our architecture through NVLink Fusion. It will run a variety of workloads, including AI, supercomputing, and quantum computing.

FugakuNEXT joins the rapidly expanding list of leading quantum supercomputing and research centers running on the NVIDIA CUDEQ quantum platform, including ULIC, AIST, NNF, and NERSC, supported by over 300 ecosystem partners, including AWS, Google Quantum AI, Quantinium, Q-era, and SciQuantum. Our new robotic computing platform, Thor, is now available. Thor offers an order of magnitude higher AI performance and energy efficiency than NVIDIA AGX Orin. It runs the latest generative and inference AI models in real-time at the edge, achieving state-of-the-art robotics.

The adoption rate of the NVIDIA robotics full-stack platform is rapidly increasing. Over 2 million developers and more than 1,000 hardware, software applications, and sensor partners are bringing our platform to market. Leading companies across various industries have adopted Thor, including Agility Robotics, Amazon Robotics, Boston Dynamics, Caterpillar, Figure, Hexagon, Medtronic, and Meta. Robotic applications require exponentially more computing on devices and infrastructure, representing a significant long-term demand driver for our data center platform.

NVIDIA Omniverse and Cosmos are our data center physical AI digital twin platforms for robotics and robotic system development. This quarter, we announced a significant expansion of our partnership with Siemens to achieve AI automated factories. Leading European robotics companies, including Agile Robots, Neuro Robotics, and Universal Robots, are using the Omniverse platform to build their latest innovations.

Turning to a quick summary of revenue by geography, China saw a quarter-over-quarter decline to single-digit percentages of data center revenue. Note that our third-quarter outlook does not include H20 shipments to Chinese customers. Revenue from Singapore accounted for 22% of second-quarter billing revenue, as customers concentrated their invoicing in Singapore. Over 99% of the data center computing revenue billed to Singapore comes from U.S. customers

Our game revenue reached a record $4.3 billion, a quarter-over-quarter increase of 14% and a year-over-year jump of 49%. This is thanks to the growth of Blackwell GeForce GPUs, with strong sales continuing as we increase supply availability. In this quarter, we shipped the GeForce RTX 5060 desktop GPU. It delivers double the performance and advanced ray tracing, neural rendering, and AI-driven DLSS4 gaming experiences to millions of gamers worldwide. Blackwell will enter GeForce Now in September. This is the most significant upgrade to GeForce Now, offering RTX 5080-level performance, minimal latency, and 5K resolution at 120 frames per second.

We also doubled the GeForce NOW catalog to over 4,500 games, making it the largest game library among all cloud gaming services. For AI enthusiasts, on-device AI performs best on RTX GPUs. We partnered with OpenAI to optimize its open-source GPT model for millions of RTX-supported Windows devices, achieving high-quality, fast, and efficient inference. Through the RTX platform stack, Windows developers can create AI applications designed to run on the world's largest AI PC user base.

Professional visualization revenue reached $601 million, a year-over-year increase of 32%. The growth is driven by the adoption of high-end RTX workstation GPUs and AI-driven workloads such as design, simulation, and prototyping. Key customers are leveraging our solutions to accelerate their operations. Activision Blizzard uses RTX workstations to enhance creative workflows, while robotics innovator Figure AI powers its humanoid robots with RTX embedded GPUs.

Automotive revenue (including only in-vehicle computing revenue) was $586 million, a year-over-year increase of 69%, primarily driven by autonomous driving solutions. We have begun shipping the NVIDIA Thor SoC, the successor to Orin. The arrival of Thor coincides with the industry's accelerated shift towards vision, language, model architecture, generative AI, and higher levels of autonomy. Thor is the most successful robotics and autonomous driving computer we have created.

Our full-stack driving AV software platform is now in production, opening up billions of dollars in new revenue opportunities for NVIDIA while enhancing vehicle safety and autonomy. Now turning to the rest of our income statement, GAAP gross margin was 72.4%, and non-GAAP gross margin was 72.7%. These figures include a gain of $180 million or 40 basis points from the release of previously reserved H20 inventory.

Excluding this gain, the non-GAAP gross margin was 72.3%, still exceeding our outlook. GAAP operating expenses increased by 8% quarter-over-quarter, and on a non-GAAP basis, they grew by 6%. This increase was driven by higher computing and infrastructure costs, as well as increased compensation and benefits costs. To support the growth of Blackwell and Blackwell Ultra, inventory increased from $11 billion in the first quarter to $15 billion in the second quarter

Although we prioritized funding our growth and strategic initiatives in the second quarter, we returned $10 billion to shareholders through stock buybacks and cash dividends. Our board recently approved a $60 billion stock repurchase authorization to add to the remaining $14.7 billion authorization at the end of the second quarter.

Now let me turn to the outlook for the third quarter.

Total revenue for the third quarter is expected to be $54 billion, with a fluctuation of 2% up or down. This represents over $7 billion in sequential growth. Again, I want to emphasize that our outlook does not include any assumptions regarding shipments of H20 to Chinese customers. GAAP and non-GAAP gross margins are expected to be 73.3% and 73.5%, respectively, with a fluctuation of 50 basis points up or down.

We continue to expect non-GAAP gross margins to reach the mid-70% range by the end of the year. GAAP and non-GAAP operating expenses are expected to be approximately $5.9 billion and $4.2 billion, respectively. For the full year, we expect year-over-year growth in operating expenses to be in the high 30% range, higher than our previous expectation of the mid-30% range. We are accelerating investments in the business to address significant growth opportunities in the future.

GAAP and non-GAAP other income and expenses are expected to be approximately $500 million in income, excluding gains and losses from non-tradable securities and publicly held equity securities. GAAP and non-GAAP tax rates are expected to be 16.5%, with a fluctuation of 1% up or down, excluding any discrete items. More financial data is included in the CFO's comments and other information on our website.

Finally, let me highlight upcoming events in the financial community. We will participate in the Goldman Sachs Technology Conference in San Francisco on September 8. Our annual NDR will begin in early October. The GTC Data Center Conference starts on October 27, with Jensen's keynote scheduled for the 28th. We look forward to seeing you at these events.

Our earnings call discussing the third-quarter performance for fiscal year 2026 is scheduled for November 19. Now we will begin taking questions. Operator, please start the Q&A session.

Operator: Thank you. (Operational instructions) We will pause for a moment to organize the Q&A list. Just a reminder, please limit your questions to one. Thank you. Your first question comes from CJ Muse of Cantor Fitzgerald. Please go ahead.

CJ Muse:

Good afternoon. Thank you for taking my question. Considering the 12-month delivery cycle from wafer to rack, you confirmed in today’s call that Rubin is ramping up production as planned in the second half, and clearly many of these investments rely on multi-year projects involving power, cooling, and other conditions. I hope you can provide a high-level perspective on your vision for growth in 2026, and if you could comment on the interplay between networks and data centers, that would be very helpful. Thank you.

Jensen Huang, Founder, President, and CEO:

Thank you, CJ. At the highest level, the growth drivers will be the evolution and introduction of inference agent AI. Past chatbots were one-off; you give it a prompt, and it generates an answer Now AI can conduct research, think, and formulate plans; it may use tools. This is called long thinking, and the longer it thinks, the better the answers it usually produces. The computational load required for one-time generation versus reasoning agent AI models can differ by a factor of one hundred, one thousand, or even more, as it requires extensive research, reading, and understanding. Therefore, the computational demand for agent AI has increased significantly. Of course, effectiveness has also improved dramatically. Due to agent AI, the phenomenon of hallucination has significantly decreased. It can now use tools and perform tasks. This opens up new possibilities for businesses. With agent AI and visual language models, we are now seeing breakthroughs in physical AI, robotics, and autonomous systems. Over the past year, AI has made tremendous progress. Agent systems and reasoning systems are completely revolutionary. We have now built the Blackwell NVLink72 system, a rack-level computing system born for this moment. We have been developing it for several years. Last year, we transitioned from NVLink8 (node-level computing, where each node is a computer) to NVLink72, where each rack is a computer. Breaking down NVLink72 into a rack-level system is extremely challenging, but the results are outstanding. Thanks to NVLink72, we have seen an order-of-magnitude speed increase, leading to significant improvements in energy efficiency and cost-effectiveness of token generation. Therefore, in the coming years—you asked about the long-term situation. Over the next five years, we will expand into a $3 trillion to $4 trillion AI infrastructure opportunity through Blackwell, Rubin, and subsequent products. In the past few years, you have seen the capital expenditures of just the top four cloud service providers double, growing to about $600 billion. So we are at the beginning stage of this construction, and the advancements in AI technology have truly enabled AI to adopt and solve problems across many different industries.

Host: The next question comes from Vivek Arya of Bank of America Securities. Please go ahead.

Vivek Arya:

Thank you for taking my question. Colette, I want to clarify what needs to happen for the $2 billion to $5 billion business in China, and what the sustainable pace of the China business is as you enter the fourth quarter. Then Jensen, regarding the competitive landscape, several of your large customers already have or are planning many ASIC projects. I believe one of your ASIC competitors has signaled that their AI business may grow by 55% to 60% next year. Do you think there is a market shift more towards ASICs and away from NVIDIA GPUs? What are you hearing from customers? How are they managing the allocation between general-purpose chips and ASICs? Thank you.

Executive Vice President and Chief Financial Officer Colette Kress:

Thank you, Vivek. First, to answer your question about when the H20 chip will be able to ship. There is demand in the market for our H20 chip, and we have received the first batch of licenses, and we are also ready with supply. That is why we communicated that we might ship about $2 billion to $5 billion this quarter We are still waiting for some geopolitical issues being negotiated back and forth between the government and enterprises, as they are deciding on procurement plans and specific actions. The current situation remains unclear, and we cannot determine the exact shipment amount for this quarter. However, if more demand and licenses come in, we will also be able to continue producing more H20 chips and expand our shipment volume.

Founder, President, and CEO Jensen Huang:

The products built by NVIDIA are fundamentally different from Application-Specific Integrated Circuits (ASICs). Let's first talk about ASICs. Many projects have been launched, and many startups have been established, but very few products have actually gone into production. The reason is that it is indeed very difficult. Accelerated computing is different from general computing; you cannot just write software and then compile it to the processor.

Accelerated computing is a full-stack co-design problem. Over the past few years, AI factories have become more complex as the scale of the problems has significantly increased. This is indeed one of the most extreme computer science problems in the world. The entire technology stack is very complex.

Models are changing rapidly, from autoregressive generative models to diffusion-based generative models, to hybrid models, and then to multimodal models. Whether it is derivative or evolutionary versions of Transformers, the sheer number of emerging different models is astonishing. One of our advantages is that NVIDIA is available on every cloud platform, and every computer company offers our products.

From cloud to on-premises deployment, from edge computing to robotics, we use the same programming model. Therefore, it makes sense that every framework in the world supports NVIDIA. When you build new model architectures, it is the wisest choice to launch on NVIDIA. The diversity of our platform is reflected in its ability to evolve into any architecture; we are everywhere, and we accelerate the entire process, from data processing to pre-training, to post-training in reinforcement learning, all the way to inference.

Therefore, when you build a data center using the NVIDIA platform, its practicality is the best, and the utility over its lifecycle is much longer. In addition, this is now an extremely complex system problem. People are not just talking about the chips themselves; many are now talking about an ASIC, which is a GPU. But to build the Blackwell platform and the Rubin platform, we must build CPUs that connect to fast memory, low-power high-efficiency memory for the large caches required for intelligent AI, connected to GPUs, SuperNICs, and scalable switches (which we call NVLink, a completely revolutionary technology, and we are now in the fifth generation), extending to horizontal scalable switches, whether Quantum or Spectrum-X Ethernet, now extending across switches, so we can prepare for these AI super factories with multi-gigawatt computing power that are all interconnected. We call it Spectrum-XGS, and we just announced this technology at this week's Hot Chips conference. The complexity of everything we do is indeed very prominent

The current scale is indeed extremely large. Finally, I would like to add that there is a good reason for our presence on every cloud platform. We are not only the most energy-efficient, but our performance-to-power ratio is the best among all computing platforms. In a world where data centers are limited by power consumption, the performance-to-power ratio directly drives revenue growth.

As I mentioned before, in many ways, the more you buy, the more you grow. Due to our outstanding cost-performance ratio, you can also achieve excellent profit margins. Therefore, the growth opportunities and gross profit opportunities using NVIDIA architecture are absolutely the best.

This is why every cloud service provider, every startup, and every computer company chooses NVIDIA for many reasons. We truly are a comprehensive, holistic, full-stack solution for AI factories.

Host: The next question comes from Ben Reitzes of Melius. Please go ahead.

Ben Reitzes:

Thank you. Jensen, I want to ask you about the forecast for data center infrastructure spending reaching $3 trillion to $4 trillion by the end of this decade. Previously, you mentioned a scale of about $1 trillion, which I believe only referred to computing spending by 2028. If we refer to your past comments, $3 trillion to $4 trillion could imply that computing spending will exceed $2 trillion. I want to know if this understanding is correct and if this is what you see by the end of this decade. I would like to understand what you think NVIDIA's share of that will be. Currently, your share in overall infrastructure computing is very high. Also, are you concerned about any bottlenecks, such as power, that could affect achieving the $3 trillion to $4 trillion target? Thank you.

Jensen Huang, Founder, President, and CEO:

Thank you. As you know, the capital expenditures of the top four hyperscale cloud service providers have doubled in two years. With the full launch of the AI revolution, the AI race is now underway, and capital expenditures have doubled to $600 billion per year. There are five years left until the end of this decade, and $600 billion only represents the top four hyperscale cloud service providers. We also have other enterprise companies building local infrastructure. Cloud service providers around the world are building. The U.S. accounts for about 60% of global computing, and over time, you would think that AI should reflect the scale and growth of GDP, and of course, it will accelerate GDP growth. Our contribution to this is a significant part of AI infrastructure. In a gigawatt AI factory, costs may fluctuate within a range of about 10%, such as $50 billion to $60 billion, and we account for about 35% of that, which is about $35 billion out of the $50 billion in a gigawatt data center.

Of course, what you get is not just GPUs. I think people are familiar with our building and inventing GPUs, but as you know, over the past decade, we have truly transformed into an AI infrastructure company. Building a Rubin AI supercomputer alone requires six different types of chips, and to scale it to gigawatt levels, you need hundreds of thousands of GPU computing nodes and a large number of racks So we are truly an AI infrastructure company, and we hope to continue contributing to the growth of this industry, making AI more useful. It is very important to improve performance per watt because, as you mentioned, limiting factors may always exist globally, such as power limitations or AI infrastructure or AI building limitations. Therefore, we need to extract as much value as possible from that factory. The performance per unit of energy consumed by NVIDIA drives the revenue growth of that factory. This directly translates to, if you have a 100-megawatt factory, every 100 megawatts drives your revenue. This is the token count for a 100-megawatt factory. In our case, the performance per dollar spent is so high that your gross margin is also the best.

In any case, these are the limiting factors for the future, $3 trillion to $4 trillion in the next five years is quite reasonable.

Host: The next question comes from Joe Moore of Morgan Stanley. Please ask your question.

Joseph Moore:

Great. Thank you. Congratulations on reopening the opportunity in China. Can you talk about the long-term prospects there? You mentioned that about half of the AI software world is there; how much can NVIDIA grow in that business? How important is it for the Blackwell architecture to eventually get licensed there?

Jensen Huang, Founder, President, and CEO:

I estimate that the Chinese market presents about a $50 billion opportunity for us this year. If we can meet this market with competitive products, and if this year is $50 billion, you would expect it to grow by 50% annually, just like the AI market in other parts of the world is also growing.

This is the second-largest computing market in the world and a hub for AI researchers. About 50% of AI researchers globally are in China. The vast majority of leading open-source models are created in China. So I think it is quite important for American tech companies to meet this market.

As you know, open-source is created in one country but used worldwide. Open-source models from China are indeed excellent. DeepSeek has certainly gained global attention. Qwen is excellent. Kimi is excellent. And there is a large number of new models about to be released. They are multimodal, excellent language models that truly drive the adoption of enterprise AI globally, as enterprises want to build their own custom proprietary software stacks. So open-source models are really important for enterprises.

It is also very important for SaaS that also wants to build proprietary systems. This is truly incredible for global robotics. So open-source is really important, and it is important for American companies to meet this demand. This will be a very large market. We are discussing with the government the importance of American companies being able to meet the Chinese market. As you know, H20 has been approved for companies not on the entity list, and many licenses have been approved. So I believe the opportunity to bring Blackwell to the Chinese market is real. Therefore, we just need to continue advocating for the rationale and importance of American tech companies being able to lead and win the AI race.

Host: The next question comes from Aaron Rakers of Wells Fargo. Please go ahead.

Aaron Rakers:

Thank you, I want to return to the topic of the Spectrum XGS release this week, considering that Ethernet products now have an annualized revenue exceeding $10 billion. How do you view the opportunity space for Spectrum XGS? Should we consider it as part of the data center interconnect layer? What are your thoughts on the scale of this opportunity within the Ethernet product portfolio? Thank you.

Jensen Huang, Founder, President, and CEO:

We currently offer three types of networking technologies. One for vertical scaling, one for horizontal scaling, and one for cross-domain scaling.

Vertical scaling is aimed at building the largest possible virtual GPU, that is, virtual compute nodes. NVLink is revolutionary. NVLink 72 achieves such an outstanding generational leap for Blackwell compared to Hopper's NVLink 8. Now we have long thinking, thinking models, agent AI, and reasoning systems; NVLink essentially amplifies memory bandwidth, which is critical for reasoning systems.

NVLink 72 is exceptional. Then we horizontally scale through networking, and we have two options. We have InfiniBand, which undoubtedly has the lowest latency, lowest jitter, and best horizontal scaling network. It does require more expertise to manage these networks.

For supercomputing, for leading model manufacturers, InfiniBand and Quantum InfiniBand are the clear choices. If you want to benchmark an AI factory, the performance of a factory using InfiniBand is the best. For those who wish to use Ethernet because their entire data center is built on Ethernet, we have a new type of Ethernet called Spectrum Ethernet. Spectrum Ethernet is not off-the-shelf.

It has a complete set of new technologies designed for low latency, low jitter, and congestion control. It has the capability to be much closer to InfiniBand than any product on the market. This is what we call Spectrum X Ethernet. Finally, we have Spectrum XGS, gigabit-scale, for connecting multiple data centers and multiple AI factories into a super factory, a massive system.

Networking is obviously very important in AI factories. In fact, choosing the right network can improve performance and throughput from 65% to 85% or 90%; this improvement brought about by network capability essentially makes the cost of the network negligible. By choosing the right network, you are basically paying, but you will receive incredible returns because, as I mentioned earlier, a gigawatt AI factory could be worth $50 billion. Therefore, improving the efficiency of that factory by several percentage points can yield effective returns of $10 billion to $20 billion.

The network is a very important part of this. This is why NVIDIA has invested so much in networking. This is also the reason we acquired Mellanox five and a half years ago. As we mentioned earlier, Spectrum-X has now become a fairly substantial business.

It has only been around for about a year and a half. So Spectrum-X is a home run. All three will be outstanding. NVLink is used for vertical scaling, Spectrum-X and InfiniBand for horizontal scaling, and then Spectrum-XGS for cross-domain scaling.

Host: The next question comes from Stacy Rasgon of Bernstein Research. Please go ahead.

Stacy Rasgon:

Hello, thank you for taking my question. I would like to ask Colette a more specific question. Regarding the over $7 billion growth, most of which will come from data centers. How should I allocate this $7 billion between Blackwell, Hopper, and networking? It looks like Blackwell might be $27 billion this quarter, up from $23 billion last quarter. Hopper still has $6 billion or $7 billion after H20.

Will Hopper's strong performance continue? How should I understand the allocation of this $7 billion among these three different components?

Colette Kress, Executive Vice President and Chief Financial Officer:

Thank you for the question, Stacy. First, looking at our growth from Q2 to Q3, Blackwell will still be a major part of our data center business. But keep in mind that this helps both our computing and networking aspects, as we are selling those important systems that integrate NVLink, which Jensen just mentioned. We are still selling Hopper.

We are still selling H100 and H200. But similarly, they are HDX systems, and I still believe Blackwell will be the main part of our business. We will continue. We don’t have more specific details on how to complete this quarter, but you should expect Blackwell to be a growth driver again.

Host: The next question comes from Jim Schneider of Goldman Sachs. Please go ahead.

Jim Schneider:

Good afternoon. Thank you for taking my question. You have clearly articulated the reasoning model opportunities you see and have relatively clearly outlined the technical specifications of Rubin. But perhaps you could provide some background on how you view the future transformation of Rubin products. What incremental capabilities does this provide to customers? From a performance perspective, do you see Rubin as a larger, smaller, or similar improvement compared to what we see with Blackwell? Thank you

Jensen Huang, Founder, President, and CEO:

Okay, thank you. Rubin, we operate on an annual cycle. The reason we adopt an annual cycle is that it accelerates cost reductions and maximizes revenue generation for our customers. When we improve the performance per watt, that is, the token generation per unit of energy used, we are actually driving revenue growth for our customers. For inference systems, Blackwell's performance per watt will be an order of magnitude higher than Hopper.

Therefore, for the same amount of energy consumption, every data center is inherently energy-constrained, and for any data center using Blackwell, compared to anything we've done in the past and anything available in the world today, you will be able to maximize your revenue. With such outstanding performance, the cost-effectiveness of invested capital will also enable you to improve your gross margins. As long as we have good ideas for each generation of products, we can enhance customer revenue generation, AI capabilities, and profit margins by releasing new architectures. Therefore, we advise our partners and customers to adjust their pace and build these data centers according to an annual rhythm.

Rubin will have a range of new ideas. I pause because I have enough time from now until a year later to tell you all the breakthroughs that Rubin will bring. However, Rubin has many good ideas. I would love to tell you, but I can't do that right now.

I will share more relevant information at GTC. But in any case, over the next year, we are making significant progress with Grace Blackwell, GB200, and then Blackwell Ultra, GB300. We are making significant strides in data center construction. This year is clearly a record year.

I expect next year to also be a record year. As we continue to improve the performance of AI capabilities on one hand, moving towards artificial superintelligence, and on the other hand, continue to enhance the revenue generation capabilities of hyperscale cloud service providers.

Host: The last question comes from Timothy Arcuri at UBS. Please go ahead.

Tim Arcuri:

Thank you very much. Jensen, I want to ask you, you mentioned a number when answering a question earlier, saying that the AI market has a 50% compound annual growth rate. I would like to know how much visibility you have for next year. Is this a reasonable target for how much your data center revenue should grow next year? I think you will at least grow in line with that. Thank you.

Jensen Huang, Founder, President, and CEO:

I think the best view is that we have reasonable forecasts for large customers next year. Very, very important forecasts. We still have many businesses competing for, and many startups being created.

Don't forget, we are still in the early stages of growth. Don't forget that last year, AI-native startups raised $100 billion. This year, the year isn't over yet, and they have already raised $180 billion. If you look at AI-native, the top AI-native startups generating revenue, last year it was $2 billion This year is 20 billion USD.

It is not unimaginable that next year could be 10 times higher than this year. Open-source models are now enabling large enterprises, SaaS companies, industrial companies, and robotics companies to join the AI revolution, which is another source of growth. Whether it is AI-native companies, enterprise SaaS, industrial AI, or startups, we are seeing tremendous interest and demand for AI. The current hot topic is, as I believe you all know, that everything is sold out.

H100 is sold out, H200 is sold out. Large cloud service providers are renting capacity from other cloud service providers. AI-native startups are truly scrambling for capacity so they can train inference models. So the demand is really very high.

But from the long-term outlook we have today, capital expenditure has doubled in two years, now at about 600 billion USD per year, just among large hyperscale cloud service providers. Growing to this 600 billion USD annual scale represents a significant portion that is not unreasonable. Therefore, I believe that in the coming years, certainly in this decade, we will see very rapid growth and very important growth opportunities. Let me conclude with this: Blackwell is the next-generation AI platform the world has been waiting for. It offers an extraordinary generational leap.

NVIDIA's NVLink72 rack-level computing is revolutionary. As the demand for training and inference performance driven by inference AI models grows exponentially, the arrival of this technology is timely. Blackwell Ultra is advancing at full speed, and demand is extremely strong. Our next platform, Rubin, has already gone into production at the wafer fab.

We have six new chips representing the Rubin platform, all of which have completed tape-out at TSMC. Rubin will become our third-generation NVLink rack-level AI supercomputer. Therefore, we expect to have a more mature and fully scalable supply chain. The Blackwell and Rubin AI factory platforms will expand into a global AI factory construction worth 30 trillion to 40 trillion USD by the end of this decade.

Customers are building increasingly large AI factories. From thousands of Hopper GPUs in data centers of tens of megawatts to now hundreds of thousands of Blackwell in 100-megawatt facilities, we will soon build millions of Rubin GPU platforms to power multi-gigawatt, multi-site AI super factories. The demand for each generation of products is growing. One-time chatbots have evolved into reasoning intelligent agents capable of researching, planning, and using tools.

This has driven a leap in the demand for training and inference computing. Intelligent agents are maturing, opening up opportunities for building domain and company-specific AI agents for enterprise workflows, products, and services. The era of physical AI has arrived, opening up a new industry in robotics and industrial automation. Every industrial company needs to build two factories: one for manufacturing machines and another for building their robotic AI

This quarter, NVIDIA achieved record revenue, marking an extraordinary milestone in our development journey. The future opportunities are immense.

A new industrial revolution has begun. The AI race is underway. Thank you all for participating today, and I look forward to communicating with everyone in the next earnings call. Thank you.

Host: Today's conference call has concluded. You may now disconnect. The meeting is over