Stay updated with the latest stock market news and insights
Weekly insights + SMS alerts
Nvidia Corp. (NVDA) released its third-quarter earnings after Wednesday's close, and the subsequent earnings call provided a window into how CEO Jensen Huang is thinking about AI infrastructure demand amid mounting questions about return on investment and sustainability.
The conversation started where many Wall Street debates currently live: Is this all a bubble?
Huang opened his remarks by addressing bubble concerns head-on. From Nvidia's vantage point, he said, something entirely different is happening. The world is experiencing three massive platform shifts simultaneously—the first time this has occurred since the dawn of Moore's Law.
"The first transition is from CPU general purpose computing to GPU accelerated computing," Huang explained. As Moore's Law slows, the world has hundreds of billions of dollars invested annually in non-AI software, from data processing to engineering simulations. These applications, which once ran exclusively on CPUs, are now rapidly shifting to CUDA and GPU acceleration. "GPUs accelerated computing has reached a tipping point," he said.
The second shift involves generative AI transforming existing applications while enabling entirely new ones. Huang pointed to Meta Platforms, where the company's Gem foundation model for ad recommendations—trained on large-scale GPU clusters—drove over 5% higher ad conversions on Instagram and 3% gains on Facebook feed in Q2. "Transitioning to generative AI represents substantial revenue gains for hyperscalers," he noted.
Then there's the third wave: agentic AI systems capable of reasoning, planning, and using tools. Huang rattled off examples spanning coding assistants like Cursor and Quadcode, radiology tools like iDoc, legal assistants like Harvey, and autonomous driving systems from Tesla and Waymo. "The fastest growing companies in the world today—OpenAI, Anthropic, XAI, Google, Cursor, Lovable, Replit, Cognition AI, Tesla—are pioneering agentic AI," he said.
His framework positions these transitions as layered, not sequential. The move to accelerated computing is foundational and necessary in a post-Moore's Law era. Generative AI is transformational and necessary, supercharging existing business models. And agentic AI will be revolutionary, giving rise to entirely new applications, companies, products, and services.
When Morgan Stanley's Joseph Moore asked about Nvidia's previously disclosed $500 billion revenue forecast for Blackwell and Rubin through calendar year 2026, Huang confirmed the company remains on track. Nvidia shipped $50 billion in the most recent quarter, and with several quarters ahead, additional compute needs will likely be filled by fiscal year 2026.
"We'll probably be taking more orders," Huang said. "Just even today our announcements with KSA and that agreement in itself is 400 to 600,000 more GPUs over three years. Anthropic is also net new. So there's definitely an opportunity for us to have more on top of the 500 billion that we announced."
The Saudi Arabia deal he referenced represents a significant expansion of Nvidia's customer base, while the Anthropic partnership marks the first time Claude will run on Nvidia's architecture at scale.
Cantor Fitzgerald's CJ Muse raised what might be the central tension in AI infrastructure right now: massive consternation about funding and ROI coexisting with Nvidia being completely sold out. "Every stood up GPU is taken," Muse noted, asking whether supply could realistically catch up with demand over the next 12 to 18 months.
Huang's answer meandered through the different layers of demand, which seems to be the point. "It's really important to recognize that AI is not just agentic AI," he said. Generative AI is transforming how hyperscalers handle search, recommender systems, ad targeting, and content moderation—work that used to run on CPUs. "Whether you installed Nvidia GPUs for data processing or you did it for generative AI for your recommender system, or you're building it for agentic chatbots, all of those applications are accelerated by Nvidia."
He pointed to the explosive growth of coding assistants—Cursor, Claude Code, OpenAI's Codex, GitHub Copilot—as the fastest-growing applications in history. And they're not just for software engineers anymore. "Because of vibe coding, it's used by engineers and marketeers all over companies, supply chain planners all over companies," Huang said.
The underlying theme: demand isn't a single monolithic thing. It's coming from multiple directions simultaneously, each with its own growth trajectory.
Bank of America's Vivek Arya asked about Nvidia's content assumptions per gigawatt of power, noting industry estimates ranging from $25 billion to $40 billion per gigawatt.
Huang walked through the generations. Hopper was probably $20 billion to $25 billion per gigawatt. Blackwell, particularly Grace Blackwell, is around $30 billion, maybe a bit more. Rubin will be higher still. "And in each one of these generations the speed up is X factors and therefore their TCO, the customer TCO improves by X factors," he said.
But here's the critical constraint: "In the end you still only have 1 gigawatt of power." Performance per watt translates directly to revenue for customers. "That 1 gigawatt translates directly, absolutely directly to your revenues. Which is the reason why choosing the right architecture matters so much."
On the financing question, Huang argued that much of the investment is already cash-flow funded. Hundreds of billions in capex will go toward cost reduction as Moore's Law scaling slows, plus revenue-boosting improvements to recommender systems that now use generative AI. "The first two things that I just said, hundreds of billions of dollars of capex is going to have to be invested is fully cash flow funded," he said. What sits above that is agentic AI—net new consumption and net new applications.
He also pushed back on the narrow focus on American cloud service providers. "Each country will fund their own infrastructure. You have multiple countries, you have multiple industries. Most of the world's industries haven't really engaged agentic AI yet and they're about to," Huang said, citing autonomous vehicles, digital twins for factories, and digital biology for drug discovery.
Melius analyst Ben Reitzes asked about capital allocation, noting Nvidia may generate roughly half a trillion dollars in free cash flow over the next couple of years.
Huang's first answer: funding growth. "No company has grown at the scale that we're talking about and have the connection and the depth and the breadth of supply chain that Nvidia has," he said. The company's balance sheet gives its supply chain partners confidence to invest alongside Nvidia's forecasts. "When we make purchases, our suppliers can take it to the bank."
Stock buybacks will continue, but Huang spent more time on ecosystem investments. The OpenAI relationship dates back to 2016, when Huang delivered the first AI supercomputer to the company. "Everything that OpenAI does runs on Nvidia today," he said. The investment allows deeper technical partnership and co-development while giving Nvidia a stake in "one of the most consequential once in a generation companies."
The Anthropic deal has different strategic value. "This is the first time that Anthropic will be on Nvidia's architecture," Huang said. Claude is the second most successful AI globally by user count and performs exceptionally well in enterprise settings. "What do we have now? Nvidia's architecture, Nvidia's platform is the singular platform in the world that runs every AI model. We run OpenAI, we run Anthropic, we run XAI, we run Gemini, we run thinking machines... We run them all."
The investment thesis boils down to: partner deeply with the best companies on a technical basis, expand the reach of CUDA's ecosystem, and get equity stakes in what will often be generational companies.
Goldman Sachs analyst Jim Schneider asked about inference workloads, which Nvidia has previously said represent roughly 40% of shipments, and about the new Rubin CPX product.
CPX is designed for long-context workloads, Huang explained—situations where an AI needs to absorb substantial information before generating answers. "It could be a bunch of PDFs, it could be watching a bunch of videos, studying 3D images," he said. The product offers excellent performance per dollar for this specific use case.
On inference more broadly, Huang described three scaling laws operating simultaneously. Pre-training continues to be effective. Post-training has found algorithms for improving AI's ability to break problems down step by step, and this capability is scaling exponentially. And inference itself is scaling because of chain-of-thought reasoning—AIs are essentially thinking before they answer.
"The amount of computation necessary as a result of those three things has gone completely exponential," Huang said. He expressed hope that inference becomes a very large part of the market, because that would indicate people are using AI in more applications and more frequently.
This is where Grace Blackwell's advantages become clearest. In the largest single inference benchmark ever conducted by Semianalysis, GB200 with NVLink 72 delivered 10 to 15 times higher performance than H200. "That's a big step up. It's going to take a long time before somebody is able to take that on. And our leadership there is surely multi year," Huang said.
UBS analyst Timothy Arcuri asked about the single biggest bottleneck that could constrain Nvidia's growth—power, financing, memory, or foundry capacity.
Huang's answer was refreshingly direct: "These are all issues and they're all constraints." When growing at Nvidia's rate and scale, how could anything be easy? The company is simultaneously transitioning computing from general purpose to accelerated architectures while creating an entirely new industry—AI factories that generate tokens on demand rather than retrieving pre-created information.
"This whole transition requires extraordinary scale," Huang acknowledged. But the supply chain is where Nvidia has the most visibility and control, having worked with partners for 33 years. The company has established partnerships across land, power, and shell capacity, plus financing structures. "None of these things are easy but they're all tractable and they're all solvable things."
What matters most is planning—up and down the supply chain, with multiple routes to market, and with an architecture that delivers the best value to customers. "At this point, I'm very confident that Nvidia's architecture is the best performance per TCO. It is the best performance per watt," he said. For any amount of delivered energy, Nvidia's architecture will drive the most revenue.
And then this: "The number of customers coming to us and the number of platforms coming to us after they've explored others is increasing, not decreasing."
Bernstein's Stacy Rasgon asked CFO Colette Kress about gross margins, given her guidance that Nvidia is working to hold them in the mid-70% range next year.
Kress noted that earlier in fiscal 2026, the company indicated it would exit the year with gross margins in the mid-70s through cost improvements and mix, which it has achieved. Looking to next year, "there are input prices that are well known in industries that we need to work through," she said. Nvidia's systems involve a tremendous number of components across many different parts.
"We do believe if we look at working again on cost improvement, cycle time and mix, that we will work to try and hold at our gross margins in the mid seven days," Kress said.
On operating expenses, the focus is innovation. "Right now we have a new architecture coming out, and that means they are quite busy in order to meet that goal," Kress explained. Investment will continue in engineering teams and business operations to create more systems for the market.
Huang added that Nvidia plans, forecasts, and negotiates with its supply chain well in advance. Suppliers have known requirements and demand for a long time. "In many cases we've secured a lot of supply for ourselves because, obviously, they're working with the largest company in the world in doing so."
Wells Fargo's Aaron Rakers asked whether Huang's views on AI ASICs or dedicated XPUs have changed, noting the CEO has been fairly adamant that many of these programs never see deployment.
Huang's answer focused on complexity and team capability. "You're not competing against teams, excuse me, against a company. You're competing against teams. And there just aren't that many teams in the world who are built, who are extraordinary at building these incredibly complicated things."
In the Hopper and Ampere days, Nvidia built one GPU—that was the definition of an AI system. Today the company builds entire racks with three different types of switches: scale-up, scale-out, and scale-across. AI now requires memory, where it didn't before. The context it needs to maintain is gigantic. The diversity of models has exploded—mixture of experts, dense models, diffusion models, autoregressive models, biological models that obey laws of physics.
"The challenge is the complexity of the problem is much higher. The diversity of AI models is incredibly, incredibly large," Huang said.
He outlined five things that make Nvidia special. First, the company accelerates every phase of the transition from general purpose computing to accelerated computing to generative AI to agentic AI. "You can invest in one architecture, use it across the board, you can use one architecture and not worry about the changes in the workload across those three phases."
Second, Nvidia excels at every phase of AI—pre-training, post-training, and inference. Inference is particularly hard, Huang noted. "People think that inference is one shot and therefore it's easy. Anybody could approach the market that way. But it turns out to be the hardest of all, because thinking, as it turns out, is quite hard."
Third, Nvidia is now the only architecture running every AI model. "We run open source AI models incredibly well. We run science models, biology models, robotics models, we run every single model." It doesn't matter if the model is autoregressive or diffusion-based.
Fourth, Nvidia is in every cloud. "We're literally everywhere. We're in every cloud, we could even make you a little tiny cloud called DGX Spark." The architecture spans from cloud to on-premises to robotic systems, edge devices, and PCs. "One architecture, things just work."
Fifth—and Huang called this probably most important—Nvidia's diverse offtake helps cloud service providers and new companies. "If you're a new company like Humane, if you're a new company like CoreWeave, N Scale, NVS or Oracle Cloud Infrastructure for that matter, the reason why Nvidia is the best platform for you is because our offtake is so diverse. We can help you with offtake."
It's not about putting a random ASIC into a data center, he argued. Where does the offtake come from? Where's the diversity, resilience, versatility? "Nvidia has such incredibly good offtake because our ecosystem is so large."
Nvidia Vice President of Investor Relations Toshiya Hari closed the call by noting the company will appear at the UBS Global Technology and AI Conference on December 2nd, with its fiscal Q4 2026 earnings call scheduled for February 25th.
The broader narrative Huang constructed during the call positioned current AI infrastructure spending not as speculative bubble activity but as the convergence of three distinct, mutually reinforcing transitions. Whether that framework holds up depends largely on whether the applications he described—from coding assistants to autonomous systems to enterprise AI agents—actually deliver the productivity gains and new revenue streams required to justify the investment.
What's clear is that Nvidia sees demand coming from far more sources than just a handful of hyperscalers building out frontier model training capacity. The company's bets on Anthropic and continued partnership with OpenAI suggest it's positioning to be the infrastructure layer regardless of which specific AI companies or applications win. And the technical claims about inference performance and energy efficiency per watt reflect an awareness that as these systems move from research to production, economics and practical constraints will matter as much as raw capability.
For now, at least, Nvidia remains sold out. Every GPU that gets stood up is taken, as Cantor Fitzgerald's analyst put it. That could change if one of Huang's three transitions stalls, or if the complexity of building competitive AI accelerators turns out to be less daunting than he suggests. But based on this call, Nvidia's working assumption is that all three waves continue building simultaneously, that inference becomes an increasingly large workload as AI reasoning capabilities improve, and that the company's architectural choices and ecosystem depth create sufficient switching costs to maintain multi-year leadership.
The market will get more data points on February 25th.
Real-time alerts on price moves, news, and trading opportunities.
Join 20,000+ investors. No spam, ever.
