NVIDIA reported fiscal 2027 second-quarter earnings for the period ending July 26, 2026, with quarterly revenue reaching $96.221 billion, up 18% sequentially and a massive 106% year-over-year, continuing to set new records. Data center revenue hit $89 billion, up 117% year-over-year. Following the earnings release, CEO Jensen Huang, Executive Vice President and CFO Colette Kress, and other executives participated in the earnings call to discuss key results and answer analyst questions. The following is the main content from the Q&A session.
Morgan Stanley analyst Joe Moore asked: I want to ask about the 70% guidance for fiscal 2028 revenue growth. What gives you confidence to provide such a full-year outlook, which you haven't done before? Also, between the 70% growth guidance and the 100% potential demand growth rate, what are the core constraints? Is there room for that gap to narrow in the future?
Jensen Huang responded: Thank you, Joe. As you've probably seen, AI has truly generated practical value, and AI agents are being deployed at scale across industries. Large language models have reached unprecedented scale, and model capabilities continue to strengthen. Agents perform reasoning, planning, and multi-round tool calls. Depending on the problem type, the compute required to run one agent ranges from 15 to 100 times that of a human directly using a model, creating an enormous overall compute demand that nearly everyone can see. However, what many people don't realize is that we are almost unique. Our delivery model is that globally, only we build and provide the complete full-stack AI factory platform. Customers are, of course, free to mix and match hardware from different vendors, but most enterprises lack both the technical capability and the desire to do so. Sovereign AI, regional clouds, new cloud service providers, AI startups, and enterprise customers form an entire market segment that now accounts for half of our business and is growing at 100% annually. Long-term, this market could even surpass traditional hyperscaler business. Additionally, building AI infrastructure today is no longer just about procuring hardware technology. You also need to secure land, power, and facility construction, which often requires two to three years of advance planning. The entire supply chain must coordinate construction, power supply, cooling, and a large workforce. AI infrastructure is creating massive numbers of jobs globally, but the upfront planning cycle is very long. That's why we've become deeply involved in upstream resource preparation. People asked me long ago why a chip company would work so closely with memory vendors; now everyone understands. We've gone further, collaborating deeply with power generation companies and firms responsible for land and facility construction worldwide, paving the way for our ecosystem customers' compute deployments. Whether upstream or downstream, our visibility is far greater than before. We've never provided full-year guidance a year ahead. Although actual demand far exceeds 70%, due to supply constraints, 70% is what we're confident we can achieve. We'll continue working with the supply chain to push supply higher, but we want to give customers, shareholders, and supply chain partners a unified, consistent expectation. Everyone is making massive investments and needs a common reference point. The coming year will be very exciting.
Cantor Fitzgerald analyst C.J. Muse asked: Thank you for taking my question. Investors are currently very focused on your share in the inference market. Can you discuss the evolving trends in general AI workloads and how you view future share changes? Considering the TAM expansion from each generation of full-stack platforms, the higher growth expectations for the ACIE segment, and the Grok-3 LPX product, I'd like to hear your thoughts.
Jensen Huang replied: Thank you. The full AI lifecycle has become far more complex, and this greatly leverages the advantages of NVIDIA's architecture. You can break the AI business into four phases: first, data preparation, including synthetic data, real data, and human-annotated data; second, model pre-training; third, post-training; and fourth, agent inference, which is extremely complex. Every stage has high barriers. When we introduced NV-Link 72, creating the world's first rack-scale system, the outside world was very surprised. Building the first-generation rack-scale system was incredibly difficult: we had to restructure the entire supply chain, redesign the hardware system, and rework the software — every aspect was challenging. But this system gives us a flexible, unified platform where data generation, data preprocessing, pre-training, post-training, and agent inference can all run. The value to customers is enormous. For example, the market opportunity per gigawatt of compute: in the Hopper era with InfiniBand, one gigawatt corresponded to about $18 billion; with Vera-Rubin, adding CPUs, multiple networking solutions to serve global data centers, plus Grok, the business opportunity per gigawatt now reaches $40 billion. Five years ago, capital investment for a one-gigawatt data center was around $30 billion; now it's $60 billion. Hardware performance has improved dramatically, but more critically: this massive investment can be reused across all stages of the AI lifecycle and runs nearly all model types — diffusion models, autoregressive models, state-space models, various hybrid architectures, diverse attention mechanisms, small and large models — all compatible. Customers' hardware investment lifecycle is longer, and utilization is higher. This is our massive competitive advantage in the new era and explains why our business, despite being large, continues to accelerate growth. Regarding Grok, I'm very optimistic about Grok-3 LPX. It achieves record token interaction rates and extremely low generation latency — the team has done outstanding work. Over the past few months, we've completed deep integration with the NV-Link architecture, which will be its core engine. For workloads demanding ultra-high interaction and rapid token generation, Grok accelerator cards can be attached. Such scenarios have relatively lower throughput and higher per-token cost, but they can support high-value services. However, most data centers globally remain primarily based on NV-Link 72 foundational compute hardware.
Bernstein Research analyst Stacy Rasgon asked: Thank you for the opportunity. Your 70% revenue growth for fiscal 2028 roughly corresponds to calendar 2027, implying incremental revenue of about $200 billion. Compared to prior guidance of $1 trillion over three years, this adds about $200 billion. Can you break down product contributions: Vera CPU, networking, and other segments? Also, you mentioned price increases take effect in Q1 — does this increment include pricing? And, is this 70% growth a supply-constrained number — what would potential growth be if supply were completely unconstrained?
Jensen Huang answered: If supply were completely unconstrained, potential growth would be much higher. We've already doubled year-over-year this fiscal year. Unconstrained demand is extremely strong. We're working hard to secure more capacity; our supply chain is massive, our ecosystem partners are strong, and we've locked in significant supply, but we need more. Breaking it down: most people only see hyperscalers, but that's only half the picture. The other half is what we call the ACIE segment: enterprise customers, new cloud providers, sovereign AI clients — a market many don't see. They don't purchase chips piecemeal; they need a complete AI factory platform, where we create enormous value. Of course, hyperscalers themselves are growing rapidly too. Their backlog has reached $2 trillion. Deploying NVIDIA compute boosts their revenue and profitability. Compute is now highly profitable, so everyone is racing to bring more NVIDIA compute online. But the real highlight is the massive opportunity beyond hyperscalers. Additionally, with each platform generation, the value per gigawatt continues to rise: Hopper era at $18 billion per gigawatt; Grace-Blackwell to $25 billion; Vera-Rubin to $40 billion — production efficiency doubles each generation. Customers want to transition to new platforms as quickly as possible. At the same time, NVIDIA compute generates tokens and rents GPU hours at very high profitability, with cloud providers showing strong margins. Multiple tailwinds are happening simultaneously. At the macro level, we're experiencing a computing platform transformation affecting every computer company and every industry globally. The computing paradigm is shifting from file retrieval to generative intelligence, all requiring compute, with tremendous returns. Every company worldwide wants to participate in this AI revolution.
Bank of America Merrill Lynch analyst Vivek Arya asked: Thank you for the high transparency and multi-year commitments. After reviewing the CFO's remarks, I calculate related investment of roughly $500 billion over the coming years. Is this number approximately the total ecosystem investment scale? Will there be additional equity investments? What's the cash level for fiscal 2028? Also, Jensen, many of these investments support frontier AI labs, especially OpenAI and Anthropic. But both are developing custom silicon — just days ago OpenAI released Jalapeño claiming performance surpassing Blackwell. How do you balance investing in ecosystem partners while some are developing competing chips?
Jensen Huang responded: Our products are positioned completely differently. Many custom XPU chips are tailored for one specific cloud or one inference service. NVIDIA is a complete AI factory platform covering the entire AI lifecycle, deployable on any cloud, already present across all global cloud environments. We can also help customers deploy anywhere. Those AI service providers, as their businesses go global, won't build all their data centers themselves. I fully believe their workloads will still largely run on NVIDIA platforms. From data preprocessing, training, post-training, to agent processing, I'm confident in our product economics — they'll remain our customers and partners long-term. Investing in these frontier labs is a once-in-a-lifetime opportunity. My only regret is not investing earlier and more. Several of these companies will go public soon, with more to follow — they'll become some of the most important technology companies in human history. I'm happy to be their friend and partner, pleased to see them build ecosystems on NVIDIA architecture. I'm confident that for a long time, they'll continue to use NVIDIA compute at scale.
Colette Kress added: Let me supplement regarding the various commitment amounts. A large portion of these are supply commitments, crucial for Rubin products and next year's supply. The bulk of commitments is concentrated in the first three years, which we use to plan production and secure capacity — that's the foundation of our confidence in growth and revenue.
UBS analyst Timothy Arcuri asked: Jensen, I want to discuss open-source models. There's much market discussion that open-source models will continue gaining share in U.S. workloads. We see you also have Nemotron open-source models. But a significant portion of end demand comes from leading frontier model companies. Many investors interpret the rise of open-source models as bearish for leading frontier labs. How do you view this? Are open-source models a positive or negative for NVIDIA?
Jensen Huang answered: The world needs both closed-source and open-source models. Usage of both is exploding. Nearly all open-source models run on NVIDIA. The reason is that NVIDIA's ecosystem has the broadest coverage and most universal architecture, spanning PCs, edge devices (DGX Spark performs very well), robotics, workstations, and hyperscale data centers. Open-source momentum is strong, and closed-source is equally thriving. Frontier labs are scaling rapidly with good profitability — token businesses can make money, and the only bottleneck is compute. The same applies to open-source. Our position in open source is very solid. The CUDA ecosystem is everywhere, and open-source models are the foundation for most AI startups and large enterprises. Companies can procure external AI capabilities, but nearly every large enterprise needs to build proprietary, domain-customized AI. Now that open-source models match frontier capabilities, this has become possible. Cybersecurity is a typical example: numerous cybersecurity companies leverage frontier-level open-source models to build large-scale, distributed, continuously running autonomous security systems — such startups are emerging constantly, which wouldn't be possible without open-source models. So open-source models have not only achieved great success and top-tier capability, but are also vital to the U.S. and global economy. Enterprise proprietary AI depends on open source. Closed and open source are both indispensable — both will flourish. Nearly all top models, closed or open, were initially developed on NVIDIA and run excellently on NVIDIA. Any model's success is good for us. Both closed and open source are driving our hardware sales.
Melius Research analyst Ben Reitzes asked: Jensen, let's discuss demand from another angle. You mentioned potential demand growth of 100% next year. I want to understand factors driving demand further, including recursive self-improvement. Anthropic and OpenAI are progressing rapidly on recursive self-improvement. OpenAI even claims it could achieve AGI by year-end. How might recursive self-improvement and AGI create inflection points in compute demand? What do they mean for NVIDIA? How do you view them as demand catalysts?
Jensen Huang replied: Compute demand will jump further upward. Currently, most AI is driven by human prompts. Just last month, AI usage shifted: the agent paradigm became mainstream. In the future, every enterprise will have numerous agents. We have about 40,000 employees today; in the future, the company might have 400,000 or 4 million agents running continuously in the background, collaborating to complete work for enterprises and people. Many people deploy personal AI agents on DGX Spark and DGX workstations — Dell sells these devices — and these agents execute tasks 24/7. We're already seeing coarse-grained recursive self-improvement: after completing tasks, agents review and optimize execution methods, update skill documentation, and perform better in the next run. In a sense, we already see capabilities approaching AGI. But I think fixating on these milestone labels isn't meaningful. The industry's three most important things: first, AI has genuinely produced valuable productivity; second, AI-generated tokens can be commercially profitable; third, if we have more compute, we can produce more profitable tokens, bringing higher margins to various services. We're at this stage, so the entire industry is going all-in.
Goldman Sachs analyst Jim Schneider asked: Jensen, given your comments about unconstrained demand near 100% versus supply-constrained 70%, can you rank the sharpest constraints: data center power and facility construction, DRAM memory, wafer foundry capacity, etc. — which are the biggest bottlenecks?
Jensen Huang answered: The entire supply chain is under pressure — every segment is running at full capacity. Capacity will continue to release, but not all at once — it's a gradual daily ramp with continuous yield improvements, and we're persistently advancing yield optimization. We're working tirelessly with every supplier. For now, our supply supports 70% growth; we actually have slightly more supply than 70%, but roughly at that level. Our real demand is far higher. We must push hard, otherwise we'll disappoint customers — and we don't want that; we want to give our all for customers. So I need the entire supply chain to help me. They all understand — what I just told you about next year's demand is exactly what I've communicated to them. Everyone is on the same page. We strive to be as transparent as possible, given we're discussing extremely large numbers.
Wells Fargo analyst Aaron Rakers asked: Returning to the $25-$40 billion per gigawatt value discussion. You mentioned at a recent conference that this number will continue rising. Looking beyond Vera-Rubin, say to Vera-Rubin Ultra, should we expect the market opportunity per gigawatt to grow to $60 billion or $80 billion? Deeper — is compute deployment scaling linear? Looking at fiscal 2027 and 2028, what variables could further unlock demand?
Jensen Huang replied: Good question. Very good. The logic is actually simple. Our goal is to pack as much compute as possible into the same land and power allocation. Obviously, the ideal ultimate answer is infinite compute per gigawatt. If we could truly achieve $1 trillion in compute scale within one gigawatt of power and the same land and facility conditions, that would be a perfect outcome. So the direction is toward that goal. In the Moore's Law era of general-purpose computing, pick any reference number — I'd say roughly $3-$5 billion per gigawatt. In the Hopper era, it reached $18 billion; Grace-Blackwell hit $25 billion; next, Vera-Rubin reaches $40 billion, and beyond that, the number will continue climbing. As long as compute output efficiency, durability, and versatility keep improving, this is a tremendous benefit for the industry and customers alike. Customers will be happy to invest in such assets — they generate revenue, generate profit, and pay back incredibly fast. I heard recently that capital payback periods are now under one year, even for data center projects at the $50 billion scale. That speaks to the output capability and leasable monetization potential of NVIDIA technology.
Comments