Nvidia Says Frontier Labs Are Only 25% to 30% of Its Compute as Content per Gigawatt Passes $40 Billion and Software Licensing and Revenue Share Emerge as New Growth Vectors
BG2 podcast with Jensen Huang, recorded days after the White House accord on superintelligence, October 15, 2026
The most useful disclosure in Jensen Huang's third BG2 conversation with Brad Gerstner and Clark Tang is a segmentation of Nvidia's demand, offered as a rebuttal to the concern that the AI buildout rests on two customers. Huang said Nvidia accounts for roughly 80% of the revenue share of AI infrastructure worldwide, and he divided the business into three roughly equal parts. About a third is classical computing: data processing, simulation, quantum chemistry, graphics, and traditional machine learning such as fraud detection at Visa and Mastercard. He said this segment would exist "in the absence of any AI labs" and is growing 30% to 40% a year. Another 25% to 30% of compute goes to frontier labs. Huang said he counts five or six, not two, naming Meta's Muse effort, xAI's Grok, Gemini, and Anthropic, which became a customer this year. The remaining third is thousands of startups renting Nvidia compute across physical, biological and materials AI, a business he likened to "AWS for the future."
The framing is deliberately aimed at the market's fixation on OpenAI and Anthropic revenue. Huang said the two labs "are going to be two of the most consequential technology companies ever built in history," but that "they don't have to be anywhere near a trillion dollars in revenues to be successful." He added that "we're putting way too much pressure on that." The caveat is that Huang did not directly answer Gerstner's central payback question. Gerstner estimated that frontier and inference lab revenue is about $200 billion this year and would need to double to $400 billion and then $800 billion to justify the capex. Huang instead redirected to a top-down argument, which investors should weigh accordingly.
Content per Gigawatt and Share Gains
Huang quantified how much of each data center dollar Nvidia now captures. He said the company's content per gigawatt rose from about $18 billion to the low $20 billions and is now "40 plus," and he expects it to keep rising as Nvidia adds networking, CPUs, storage and security software. Later, citing total cost, he put general-purpose computing at $5 billion per gigawatt and a Vera Rubin gigawatt at $50 billion. Huang described Nvidia as "the largest and fastest growing networking company" and said it will become "one of the largest CPU companies, surely the fastest growing one." He cited the 2 gigawatt Australia announcement as worth roughly $80 billion to Nvidia.
He also gave a rare view of share. Six years ago Nvidia's share of world capex was in single digits. It is now about 40% at the cloud service providers and about 80% outside them, in on-premises and neocloud deployments, with each half of the business roughly doubling annually. Huang's argument is that a power-constrained world pushes buyers toward the highest tokens per watt, which favors Nvidia. "Surprisingly to everybody," he said, "not only are we a large company and our market share is high, our market share is actually increasing." Nvidia's share of the capex dollar is also rising because it sells more content per gigawatt.
New Business Models: Software Licensing and Revenue Share
Huang laid out two growth vectors that are independent of hardware sales and tied to the installed base. The first is enterprise software licensing. Cloud providers can take Nvidia's software open source and run their own stacks, but most enterprises lack the IT depth and, as AI moves from training to deployment, want a vendor on the hook. "I'm on the SLA to support it and maintain it, keep it secure, keep it safe for as long as we shall live," he said. Using simple math of a couple thousand dollars per GPU per year, with Nvidia shipping around 20 million GPUs annually and half going to enterprises, he called it "very, very large."
The second is revenue share with clouds that rely on Nvidia for its stack, offtake and financing support. In those relationships, he said, Nvidia would "make an investment," capture the upfront sale, and also earn a share going forward, since it will provide offtake and software support for four to 10 years. Huang also signaled that the investment-heavy phase is ending. Investments will fall quickly as a share of cash generation, partly because the OpenAI and Anthropic stakes are an opportunity that "is not going to come around again," and partly because free cash flow is growing quickly. Nvidia returned more than 50% of free cash flow through buybacks, and Huang said the figure was 60-something percent last quarter. Gerstner estimated that Nvidia could retire up to 10% of the company over two years.
Demand Exceeds Supply, and the Constraint Is Land, Power and Shell
Asked about SemiAnalysis's estimate that 43 gigawatts of AI compute will come online next year and another 40 gigawatts the following year, Huang said "the demand is definitely there and more." He identified the obstacle as the physical world: land, permitting, power generation, skilled trades, and local moratoriums. His guess is that "land, power, shell" will be among the most important constraints on reaching that number. Gerstner noted that investors hear of roughly 50 gigawatts of announced chips against 35 to 40 gigawatts of powered capacity. Huang did not dispute the gap and said he does not "love" that the United States may not keep pace. He argued Nvidia is the most insulated vendor because it sells to every cloud, neocloud, OEM and region, so if one site stalls he will "adjust our supply chain accordingly."
On supply, Huang said orders have kept arriving "almost weekly and daily" since he gave his earlier forward view. He explained that he gave that view publicly because he was allocating far less than customers wanted and wanted to avoid disappointing them one by one. He named substrates as a binding upstream constraint, because larger packages and more chips per package increase material consumption, and organic materials are hard to make precisely given thermal coefficients and warpage. Silicon photonics is also supply constrained. His conclusion was blunt: "it's going to be hard to close that gap for a few years."
Gerstner presented a chart showing that consensus expectations for Nvidia have risen 100% to 200% in a year. A year ago consensus saw 2027 revenue flat to down, and it now sees growth of about 74%. Meanwhile the stock is up only about 20% and trades at 13 to 14 times earnings. Huang's explanation was the moat. He said it is increasing, and that customers will place $50 billion to $100 billion orders for Vera Rubin Ultra sight unseen, because systems of this complexity can no longer be benchmarked before purchase. "We're literally the only one," he said, comparing the trust to Nvidia's reliance on TSMC.
Capex Math: Fungibility, Moore's Law and a $3 Trillion to $4 Trillion Annual Market
Huang reiterated that annual AI infrastructure spending would reach $3 trillion to $4 trillion by the end of the decade, "closer to four," with about $1.5 trillion of that being compute. He noted compute was about $0.5 trillion before AI. His argument has three parts. First, Moore's law is over, so twice the capacity now costs twice as much. Second, a platform shift draws in new startups. Third, Nvidia's general-purpose architecture is fungible across workloads, models and geographies, which lowers risk. "Specialization is the enemy of fungibility," he said, "specialization is the enemy of durability." In his chain of logic, fungibility and durability produce rentability, financeability and the lowest total cost of ownership. This is also his case against custom silicon, though he did not name a competitor. He said total annual capacity building, including refreshes, could reach about $10 trillion without surprising him.
Open Models and the Vertical Superintelligence Thesis
Huang said about 80% of the startups on Nvidia compute use open models. He argued they are not competing with the frontier labs but customizing models, often Nvidia's Nemotron, with proprietary data to reach domain-specific superintelligence. His view is that closed labs will pursue general intelligence while vertical companies achieve "super intelligence for their domain." This is the commercial logic behind Nvidia's open-model advocacy: it broadens the base of compute renters beyond the labs.
Physical AI Is Already a Meaningful Business
Huang said he "understated" his earlier $100 billion estimate for physical AI. Nvidia generated $10 billion in the first six months of this year. It earned $10 billion in the last fiscal year, its first $10 billion year, after $5 billion the year before. The category spans medical imaging, sensing and radio networks, autonomous vehicles, robotics, inspection, and geospatial work. He said building a physical AI platform takes three computers: one to train the brain, one to run a digital twin in Omniverse where the system learns, and one embedded in the robot, car or surgical instrument. He argued the frontier labs will also need a digital twin of the internet so agents do not learn on the live web. He suggested Nvidia may eventually not be categorized as a chip company and described an automotive-style playbook applied across industries.
Data Centers in Space
Huang said Nvidia and SpaceX engineering teams are working "super closely" on a new server architecture designed for use in space and on Earth. His economic argument is that launch costs are a small share of the cost of a high-value system, so only the most productive compute makes sense in orbit. He said Vera Rubin would justify the launch cost and that fungibility matters because an idle satellite cannot be rented. He deferred to Elon Musk on timing when Gerstner asked about 500 megawatts by 2029. Huang also said Musk's operations are now reported under Nvidia's cloud service provider category instead of enterprise. Gerstner noted that Musk is raising roughly $40 billion, targeting 10 gigawatts by the end of 2027, and may buy more than a million Vera Rubin chips. Huang pointed to Musk's combination of training, deployment (Grok demand is "skyrocketing"), and renting compute to others as unusual, with Google the only comparable example.
The Cost of Intelligence and Where Compute Demand Comes From
Huang expects the cost of intelligence to fall quickly. Gerstner put the compute cost of the recent Navier-Stokes result, which involved 10,000 cooperating agents, at about $4 million, and Huang said a similar effort could be 5x to 10x cheaper within a few years. He noted that training the original GPT might have cost $10 million to $20 million, compared with perhaps $100 today, while model ambition has risen to 3 trillion, 10 trillion and 30 trillion parameters, and usage has grown perhaps a millionfold. He called demand for intelligence "fairly insatiable." Huang said OpenAI's result was built on years of prior mathematical work, and that many fields are on a similar "precipice" requiring exhaustive search. He credited memory, meaning working memory and knowledge graphs, and tool use as the key unlocks of the past year, and said agents are now moving beyond coding into design, creative and scientific work.
Safety Framework and the Policy Backdrop
Huang framed safety as inseparable from capability, comparing it to airbags and braking systems in a car, and said "there will be no AI technology industry if we don't build it safely and securely." His engineering prescription is zero-trust containment, minimal access rights expanded as confidence grows, and out-of-band monitoring that can escalate or shut a system down. Nvidia has launched OpenShell for this purpose, with more than 100 companies participating. Gerstner described a three-legged stool: federal frontier model review by the Commerce Department, the White House accord on superintelligence with audits and board oversight, and an industry coordination group. He noted that the government demonstrated it could pull a model, Mythos, from the market under export control authority earlier this year. Huang urged policymakers who want new laws to first "come and learn about" existing measures, and said he found no one at the White House table who wanted to be reckless. Huang also argued that "America winning" means every industry benefits, not only technology.
Optimism, With a Caveat on Narrative
Huang closed by saying he is "99.99%" confident that the next five years will go well, and that those predicting catastrophe "underestimate the human capacity." He also acknowledged a communications problem: AI is "soft matter," invisible to the public even though it already underpins search, retail and logistics. He disagreed with the idea that AI will give people more leisure, citing Keynes's 1930 forecast of a 15-hour work week. Huang said Keynes got the economy right but the human spirit wrong, and that he is working harder than ever. The conversation was notable for what Huang did not address, mainly a direct answer on whether lab revenue can double twice. His case for the durability of Nvidia demand rests on breadth of customers, supply scarcity and share gains, not on proof of end-market payback.