NVIDIA (Minutes): Target gross margin for next year is 75%, collaboration with OpenAI is not blind
The following are the minutes of NVIDIA's Q3 FY2026 earnings call organized by Dolphin Research. For an interpretation of the earnings report, please refer to "More Important than Non-Farm Payrolls! Can NVIDIA Save the U.S. Stock Market Again?"
More important than the earnings report itself is the information conveyed in NVIDIA's earnings call:
① Data Center Revenue: GB300 sales surpassed GB200, contributing about two-thirds of Blackwell's total revenue. Hopper revenue was approximately $2 billion, with H20 around $50 million;
Regarding guidance: The company has not included revenue from AI GPUs in the Chinese market in its next quarter revenue guidance. Driven by new demand from KSA, Anthropic, etc., the company may ultimately exceed its previous guidance (Blackwell+Rubin's $500 billion revenue by the end of 2026)
② Value per GW for NVIDIA: Hopper is approximately $20-25 billion/GW, Blackwell generation (especially Grace Blackwell) is around $30 billion/GW, Rubin will be higher;
③ Target Gross Margin: Despite rising costs, the target gross margin for FY2027 (i.e., 2026) remains around 75%;
④ Cooperation with OpenAI: The core is to help build and deploy AI data centers of at least 10 gigawatts, with an opportunity to invest in the company. The company is currently working towards a final agreement.
The company provides services through OpenAI's cloud service partners—Microsoft Azure, Oracle Cloud Infrastructure (OCI), and CoreWeave. The company will continue this cooperation model for the foreseeable future.
⑤ Cooperation with Anthropic: For the first time, NVIDIA technology will be used to optimize Anthropic's models for CUDA technology, achieving optimal performance, efficiency, and total cost of ownership (TCO). The initial computing power commitment includes 1GW of capacity, based on the Grace Blackwell and Vera Rubin systems.
⑥ Response to AI Bubble Theory: NVIDIA is different from any other accelerator manufacturer, excelling at every stage of AI, whether pre-training, post-training, or inference. Moreover, with 20 years of investment in the CUDA X acceleration library, the company also has outstanding capabilities in scientific and engineering simulation, computer graphics, structured data processing, and traditional machine learning
⑦ Response to ASIC Competition: It's not just about randomly placing an ASIC in a data center, but about where demand comes from, where diversity comes from, and where resilience comes from. The multifunctionality of the company's architecture, the diversity of capabilities, and the vast ecosystem bring extremely good demand. These five aspects—every stage of accelerated transformation, every stage of AI, every model, coverage from cloud to local deployment—all lead to real demand.
In summary, NVIDIA has responded to market concerns:
a) Gross Margin Target: Next year, the company aims for a gross margin of around 75% for FY2027, stabilizing market confidence (especially concerning the risk of a gross margin decline in the second half of next year);
b) ASIC Market Competition: The company is confident in its product strength and coverage, offering more value and capability than ASICs;
c) Cooperation with OpenAI: The company's cooperation with OpenAI is not blind, but through partnerships with Microsoft, Oracle, CoreWeave, etc., under certain conditions and constraints;
d) Driven by new demand from KSA, Anthropic, etc., the company believes AI revenue may actually exceed previous guidance (Blackwell+Rubin's $500 billion revenue by the end of 2026).
Overall, after delivering an earnings report that exceeded expectations, NVIDIA's management also addressed market concerns. The clear guidance on gross margin and the cooperation plan with OpenAI injected some confidence into the market. After the management's communication, the company's stock price continued to rise to around 5%.
As for ASIC competition, the company has not provided a very clear view, which remains a factor suppressing the company's stock price increase. If the company can respond more clearly to competition issues, NVIDIA's valuation is expected to rise further.
I. $NVIDIA(NVDA.US) Core Financial Data Review
Quarterly Revenue: $57 billion, up 62% year-over-year, up $10 billion (22%) quarter-over-quarter, a record high.
Data Center Revenue: $51 billion, up 66% year-over-year, with computing business up 56% year-over-year (mainly benefiting from GB300 mass production), and network business more than doubled (due to NVLink scale deployment).
Revenue by Business:
Gaming Business: $4.3 billion, up 30% year-over-year.
Professional Visualization Business: $760 million, up 56% year-over-year.
Automotive Business: $592 million, up 32% year-over-year.
Gross Margin: GAAP gross margin 73.4%, non-GAAP gross margin 73.6%.
Operating Expenses: GAAP operating expenses up 8% quarter-over-quarter, non-GAAP operating expenses up 11%.
Inventory and Supply Chain: Inventory up 32% quarter-over-quarter, supply commitments up 63% quarter-over-quarter.
Tax Rate: Non-GAAP effective tax rate 17%, slightly above the expected 16.5%.
- Detailed Information from NVIDIA's Earnings Call
2.1 Key Information from Senior Management
Strategy and Outlook:
Customers are driving the transformation of three major platforms (accelerated computing, AI models, and intelligent agents), which are in the early stages and will impact all industries.
From early 2024 to the end of 2026, the Blackwell and Ruben architectures are expected to generate $500 billion in revenue; by the end of 2030, the AI infrastructure market is expected to reach $3 trillion to $4 trillion, with NVIDIA as the preferred solution.
The three expansion laws (pre-training, post-training, and inference) create a virtuous cycle: access to computing resources drives intelligence enhancement, which in turn promotes application proliferation and profit growth.
The world is experiencing a triple platform transformation: from CPU to GPU accelerated computing (accelerated computing), generative AI transformation (generative AI is gradually replacing traditional machine learning), and physical AI transformation (intelligent agents, etc.). NVIDIA's architecture can support all three transformations simultaneously.
Business Progress and Outlook:
Data Center: Demand exceeded expectations, AI GPUs from cloud service providers sold out, and all generations of GPU product lines (including Blackwell, Hopper, and Ampere) are running at full capacity;
Network Business: Revenue of $8.2 billion, up 162% year-over-year; the launch of NVLink scale deployment, Spectrum X Ethernet, and Quantum X InfiniBand products all achieved strong double-digit growth.
As for the expectations for leading cloud service providers (CSPs) and hyperscale cloud providers in 2026, their total capital expenditure (CapEx) continues to grow, currently reaching about $600 billion, an increase of more than $200 billion since the beginning of the year. The transition to accelerated computing in the generative AI field accounts for about half of the company's long-term opportunities in the current workload of hyperscale cloud providers.
Product Progress:
Blackwell Products: Strong momentum in the third quarter, GB300 shipments surpassed GB200, contributing about two-thirds of Blackwell's total revenue; the transition to GB300 is smooth, with major cloud service providers receiving mass production shipments.
Blackwell Ultra's training speed is 5 times faster than Hopper (platform); Blackwell achieves the highest performance under all models and use cases, with the lowest total cost of ownership (TCO).
Hopper Products: Quarterly revenue reached approximately $2 billion this quarter. H20 product sales this quarter were about $50 million; geopolitical factors and increasing competition in the Chinese market led to several large procurement orders not being fulfilled. Despite the current situation preventing us from delivering more competitive data center computing products to the Chinese market, the company remains committed to ongoing communication and collaboration with the U.S. and Chinese governments.
Rubin Products: Mass production will start in the second half of 2026. The Vera Rubin platform is driven by seven chips, achieving another breakthrough in performance compared to the Blackwell (platform). The company has already retrieved wafers from supply chain partners and is smoothly advancing debugging work.
The Rubin platform is the company's third-generation RAC-scale system, enhancing performance while maintaining compatibility with the Grace Blackwell platform. The company is ready for a rapid production ramp-up of the Rubin platform.
NVLink Fusion: Strategic cooperation with Fujitsu to integrate Fujitsu's central processing unit (CPU) and NVIDIA's graphics processing unit (GPU) through NVLink Fusion technology, connecting their vast ecosystems.
The company also announced cooperation with Intel to jointly develop multi-generation customized data center and personal computer (PC) products, connecting NVIDIA and Intel ecosystems through NVLink technology.
ARM also announced the integration of NVLink technology licensing (IP) for customers to build CPU system-on-chips (SoCs) that can connect to NVIDIA GPUs.
Most accelerators that lack CUDA technology and NVIDIA's time-tested general architecture will be eliminated within a few years as model technology evolves. Thanks to CUDA technology, the A100 GPU shipped six years ago is still running at full capacity.
AI Ecosystem and Cooperation:
Cooperation with OpenAI: The core is to help build and deploy AI data centers of at least 10 gigawatts, with an opportunity to invest in the company. The company is currently working towards a final agreement.
The company provides services through OpenAI's cloud service partners—Microsoft Azure, Oracle Cloud Infrastructure (OCI), and CoreWeave. The company will continue this cooperation model for the foreseeable future.
Cooperation with Anthropic: For the first time, NVIDIA technology will be used to optimize Anthropic's models for CUDA technology, achieving optimal performance, efficiency, and total cost of ownership (TCO). The initial computing power commitment includes 1GW of capacity, based on the Grace Blackwell and Vera Rubin systems.
New Customers and Partners: This quarter, the AI factory and infrastructure projects announced by the company involve a total of 5 million GPUs. This demand covers various market entities, including sovereign state-related agencies, new computing power builders, enterprises, and supercomputing centers, including several milestone construction projects—such as XAI's Colossus 2 and Lilly's AI factory for drug development.
Amazon Web Services (AWS) and Humane expanded their cooperation, including deploying up to 150,000 AI accelerators (including our GB300 chips); in addition, X AI and Humane also announced a partnership to jointly build a world-class GPU data center network, centered around a 500-megawatt flagship data center;
Physical AI: Has become a multi-billion-dollar business for NVIDIA, bringing trillions of dollars in opportunities for its next phase of growth. Collaborating with PTC, Siemens, Bentley, Caterpillar, and other companies to build Omniverse digital twin factories; Agility Robotics, Amazon Robotics, and others are developing intelligent robots based on the NVIDIA platform.
Supply Chain and Manufacturing: Collaborating with TSMC to produce Blackwell wafers domestically in the U.S.; expanding U.S. operations with Foxconn, Wistron, and others over the next four years.
Fourth Quarter Outlook:
Total revenue is expected to be $65 billion (±2%); GAAP and non-GAAP gross margins are 74.8% and 75% (±50 basis points), respectively. Consistent with the previous quarter, the company has not included any data center computing business revenue from the Chinese market in this quarter's expectations. GAAP and non-GAAP operating expenses are approximately $6.7 billion and $5 billion, respectively. GAAP and non-GAAP other income (excluding gains and losses on non-circulating and publicly held securities) are expected to generate approximately $500 million in income. The effective tax rate is expected to be 17% (±1%). Excluding any special items.
FY2027 Outlook: Despite rising costs, the company strives to maintain a gross margin of around 75%;
Other Important Discussions:
Response to AI Bubble: NVIDIA is different from any other accelerator manufacturer, excelling at every stage of AI, whether pre-training, post-training, or inference. Moreover, with 20 years of investment in the CUDA X acceleration library, the company also has outstanding capabilities in scientific and engineering simulation, computer graphics, structured data processing, and traditional machine learning.
2.2 Q&A
Q: Previously, you mentioned that the Blackwell and Ruben architectures will generate $500 billion in revenue in 2025 and 2026, with $150 billion already shipped. After this quarter, do you still expect to achieve approximately $350 billion in revenue over the next 14 months?
Is there any upside potential for these numbers?
A: Yes, we are on track to achieve the $500 billion forecast and have completed some quarterly targets. This quarter, we shipped $50 billion, and we will continue to advance over the next few quarters until the end of 2026. This number will grow, and the company is confident that it will meet more deliverable computing demand in FY2026. For example, the company's agreement with KSA added 400,000 to 600,000 GPUs over three years, and Anthropic is also a new demand. Therefore, we clearly have the opportunity to generate more revenue beyond the announced $500 billion.
Q: Given concerns about the scale of AI infrastructure construction, financing capacity, and return on investment, while GPUs are in short supply, and the huge benefits of new generation technologies (such as B300, Rubin) have not yet fully manifested, can supply catch up with demand in the next 12-18 months? Or will this situation last longer?
A: We have done an excellent job in supply chain planning, working closely with global technology companies such as TSMC, memory suppliers, and system ODMs, and have planned for major years. We are witnessing three major transformations—shifting from general computing to accelerated computing, generative AI replacing traditional machine learning, and the new intelligent agent category—advancing simultaneously. All these applications run on and are accelerated by NVIDIA GPUs. Meanwhile, as the quality of AI models rapidly improves, their adoption in various use cases (such as code assistance, healthcare, video editing, etc.) is growing rapidly. These exponential growth factors are all at play, and it is crucial to examine the progress of each dynamic from first principles.
Q: In the $500 billion revenue forecast, what is the assumed value of NVIDIA equipment per GW installed?
In the long term, you mentioned that the data center scale will reach $3-4 trillion by 2030. How much of this requires supplier financing?
How much can be supported by the cash flow of large customers, governments, or enterprises?
A: In each generation of products, our value in data centers is increasing. The Hopper generation is approximately $20-25 billion/GW, the Blackwell generation (especially Grace Blackwell) is around $30 billion/GW, and the Rubin generation will be higher. Each generation's performance improvement is several times, and the total cost of ownership for customers also improves accordingly. The most critical factor is that data center power is limited (e.g., 1 gigawatt), so performance per watt (energy efficiency) is crucial, directly determining revenue. We optimize energy efficiency through full-stack collaborative design.
Regarding financing, customers decide their financing methods. Currently, most investments are concentrated in hyperscale enterprises, primarily supported by their cash flow. Investing in NVIDIA GPUs not only reduces costs by replacing general computing with accelerated computing (as Moore's Law has slowed) but also enhances existing business revenue through generative AI upgrades to recommendation systems. On top of that, intelligent agent AI is a new application and consumption, and these are historically the fastest-growing applications. Additionally, each country will fund its infrastructure, and many industries worldwide (such as autonomous driving, digital twin factories, biotechnology) are just beginning to apply intelligent AI, and they will finance themselves. Do not view future construction solely from the perspective of hyperscale enterprises.
Q: Regarding the approximately $0.5 trillion in free cash flow that may be generated in the coming years, how do you plan to use it? How much will be used for stock buybacks, and how much for ecosystem investment? How do you invest in the ecosystem?
There is confusion about how these transactions (such as investments in Anthropic, OpenAI) are conducted and the standards.
A: First, we need a strong balance sheet to support growth and ensure a resilient supply chain, which is the foundation our customers rely on us for. Second, we will continue to conduct stock buybacks.
Regarding investment, all our investments are related to expanding the CUDA ecosystem. For example, our investment in OpenAI began in 2016, aiming to deepen technical cooperation to support its rapid growth and jointly develop the ecosystem. As a return on investment, we obtained shares in the company, which is a groundbreaking intelligent agent company, and we expect this to bring extraordinary returns.
Another example is Anthropic, which is being introduced to the NVIDIA architecture for the first time. Through this deep cooperation, the company will bring Anthropic's Cloud platform into the NVIDIA ecosystem.
Our investment logic is: the NVIDIA platform is the only platform in the world that can run all AI models. Through ecosystem investment, we engage in deep technical cooperation with the world's top companies, which not only expands the coverage of our ecosystem but also allows us to invest in these often groundbreaking, highly successful future companies and hold their shares. This is our investment strategy.
Q: In the past, you mentioned that about 40% of shipments are related to AI inference. Looking ahead to next year, how do you expect this proportion to change?
Additionally, can you discuss the Rubin CPX product expected to be launched next year, its potential total addressable market (TAM), and the target customer applications for this product?
A: Regarding inference, the three expansion laws (pre-training, post-training, inference) are all working simultaneously. Inference, due to chain of thought and inference capabilities, has seen its computing demand grow exponentially. It is difficult to predict the exact percentage at any given time, but we hope inference will become a significant part of the market, as this means people are using it more frequently and in more applications. Our Grace Blackwell platform leads the world by an order of magnitude in this regard, with GB200 and GB300 achieving 10 to 15 times performance improvement through NVLink 72, and our leadership in this field is undoubtedly for years to come.
Regarding Rubin CPX, it is designed for long-context workloads. These workloads require absorbing a large amount of context information first, which could be a batch of PDFs, watching a series of videos, studying 3D images, etc. CPX is designed for such workloads, with excellent performance per dollar.
Q: Many customers are seeking self-supplied power, but what is your biggest concern, the biggest bottleneck that may limit growth? Is it power, financing, or other issues such as memory or wafer fabs?
A: These are all issues and constraints. At our growth rate and scale, nothing is easy. What NVIDIA is doing is unprecedented; we have created a whole new industry. On one hand, we are shifting computing from general computing and traditional computing to accelerated computing and AI; on the other hand, we have created a new industry called "AI factory".
This entire transformation requires enormous scale, involving all aspects from supply chain to land, power, construction, and even financing. These things are not easy, but they are all attractive and solvable. We are very confident in the supply chain part because we are very good at managing the supply chain and have excellent partners we have worked with for 33 years.
At the same time, we have established many partnerships in land, power, and financing, with multiple market channels. Most importantly, we must ensure that our architecture provides the best value for customers. Currently, I am very confident that NVIDIA's architecture can provide the best performance, and for any given energy input, our architecture will drive the highest revenue. Our success rate is increasing, and the number of customers and platforms choosing us is increasing rather than decreasing.
Q: Regarding the goal of maintaining a gross margin of around 75% next year, where is the biggest cost growth coming from? Is it memory or other factors?
How will you achieve this goal through cost optimization, pre-purchasing, or pricing?
Additionally, considering that revenue may grow significantly, how should we expect operating expense (OpEx) growth next year?
A: We have achieved the goal of reaching a gross margin of around 75% by the end of this fiscal year through cost improvements and product mix optimization, and we will continue to execute in the fourth quarter. Looking ahead to next year, the rising input costs in the industry are a known challenge, and our systems contain many complex components. By continuing to focus on cost improvements, cycle time reduction, and product mix optimization, we plan to maintain a gross margin of around 75%. Regarding operating expenses, our goal is to ensure that engineering teams and all business teams can continue to innovate and develop more systems for the market. We are busy with the launch of new architectures, so we will continue to invest in innovation to fully leverage this competitive environment. In terms of the supply chain, we have been forecasting, planning, and negotiating with supply chain partners early on, and due to long-term cooperation and clear demand, we have secured a large supply for ourselves.
Q: With the announcement of the Anthropic deal and the broad customer base, has your view on the role of AI-specific integrated circuits (ASICs) or specialized XPUs in architecture construction changed? In the past, you thought these projects were difficult to truly deploy; is the situation now more inclined towards GPU architecture?
A: First, as a company, you are competing with teams, and there are not many teams in the world that are good at building such complex systems. Today, building an AI system is far more than a single chip; it requires an entire rack, three different types of switches (scale-up, scale-out, cross-platform), and AI now requires huge memory and context capacity, with model diversity increasing dramatically.
The complexity of the problem is higher, and the diversity of AI models is enormous. Our uniqueness is reflected in five aspects:
First, we accelerate every stage of transformation. CUDA allows us to transition from general computing to accelerated computing, and we excel in generative AI and intelligent agent AI. You can invest in one architecture that applies to all scenarios without worrying about workload changes at different stages.
Second, we excel at every stage of AI. It is well known that we are very good at pre-training, post-training is also excellent, and it turns out we are very strong in inference because inference is actually very difficult.
Third, we are the only architecture in the world that can run all AI models, including every cutting-edge AI model, open-source model, scientific model, biological model, robotic model, etc. No matter the model type, we can run it.
Fourth, we cover every cloud platform. Developers love us because we are everywhere, from the cloud to local deployment, to robotic systems, edge devices, personal computers, etc. One architecture, applicable everywhere, is amazing.
Fifth, and most importantly, for cloud service providers or new companies, choosing the NVIDIA platform is because our demand sources are extremely diverse. We can help solve demand problems. It's not just about randomly placing an ASIC in a data center, but about where demand comes from, where diversity comes from, and where resilience comes from. The multifunctionality of our architecture, the diversity of capabilities, and our vast ecosystem bring extremely good demand. These five aspects—every stage of accelerated transformation, every stage of AI, every model, coverage from cloud to local deployment—all lead to real demand.
The risk disclosure and statement of this article: Dolphin Research Disclaimer and General Disclosure
