Global tech stocks suffered a massive sell-off this week, with the world's top twenty technology companies collectively losing over $1 trillion in market capitalization, as doubts about the return on investment in AI infrastructure intensified sharply.
At this critical juncture, NVIDIA CEO Jensen Huang made consecutive moves—first publicly endorsing CoreWeave in a Bloomberg interview in Stockholm, then releasing a nearly 140-minute in-depth interview on the BG2 podcast, addressing the market's most core anxiety: who exactly is going to pay for the money being spent on AI compute buildout?
It is worth noting that the BG2 interview with Jensen Huang, a well-known Silicon Valley investment podcast, was released before the U.S. market opened on October 15 Eastern Time, around 7 AM, just as Wall Street traders, analysts, and quantitative funds were preparing with heightened tension for the U.S. market open.
In the Bloomberg interview, Huang explicitly stated that CoreWeave "does not need NVIDIA to provide backstop support," and noted that its "cost of capital is actually declining." These remarks directly drove the new cloud computing stock to rebound intraday, with CoreWeave closing up 5.94% against the broader market trend that day. The Nasdaq index subsequently recovered all its losses, and Barron's noted in a retrospective report that the market "only needed a few reassuring words from the 'Godfather of AI' Jensen Huang."
In the BG2 interview, Huang went further, directly identifying the market's core misjudgment: betting trillions of dollars in compute buildout on the revenue of just two AI labs is "asking the wrong question." He broke down NVIDIA's compute demand into three roughly equal segments, with frontier labs accounting for only about one-quarter to one-third, and involving five or six institutions rather than the two the market typically assumes. This assessment has direct market reference value for understanding the true logic of AI infrastructure investment.
The CoreWeave Turmoil
CoreWeave was one of the most closely watched pressure points in this round of tech stock sell-offs. According to ICE Data Services, the five-year credit default swap (CDS) spread insuring against CoreWeave's debt default risk surged by as much as 74.5 basis points intraday, hitting an intraday historical high of 1,194.5 basis points, implying a five-year default probability exceeding 60%.
This signal put the market on extremely high alert. CoreWeave has borrowed heavily to expand compute capacity, and NVIDIA is both its equipment supplier and holds approximately a 10% stake, making it the second-largest shareholder. This "supplier financing the buyer, buyer then purchasing the supplier's products" circular financing structure has triggered widespread doubts about the financial sustainability of AI infrastructure expansion.
Huang's remarks in Stockholm directly offset the pricing in the credit markets. "CoreWeave is getting stronger every day," he told Bloomberg. "I'm very optimistic. Their cost of capital is actually declining." He also explicitly stated that NVIDIA does not need to provide any form of backstop support for CoreWeave.
The weight of these words lies in Huang's dual identity—he is both CoreWeave's most important equipment supplier and its second-largest shareholder. The market's reaction was immediate: CoreWeave's stock closed up 5.94% that day, halting a two-day cumulative plunge of over 18%.
Breaking Down the Misjudgment: Compute Demand Goes Far Beyond Two Labs
One of the deeper logics behind this week's market sell-off was systematic doubt about the return on AI infrastructure investment—if OpenAI and Anthropic's revenue growth slows, who will pay for trillions of dollars in compute buildout?
In the BG2 interview, Huang characterized this logic as "asking the wrong question." He broke down NVIDIA's compute demand into three roughly equal segments:
The first segment, about one-third, is foundational computing completely unrelated to AI labs—molecular dynamics, physics simulation, quantum chemistry, classical machine learning (such as Visa and Mastercard's fraud detection), recommendation systems, and more. "Even with no AI at all, NVIDIA is not zero," he said. "This business is growing 30% to 40% a year."
The second segment, about one-quarter to one-third, is compute demand from frontier labs. But Huang emphasized that the market typically only sees two, while he sees five or six—Meta, Grok, Gemini, Anthropic, and other emerging institutions. "If you tell me this represents less than a third of our business, about a quarter, that's fine."
The third segment, also about one-third, is thousands of startups renting compute—which he compared to "the next-generation AWS." About 80% of these companies use open-source models, building specialized intelligence in vertical domains such as physical AI, biological AI, and materials science. "Think of this large cloud computing block as the next-generation AWS," he said. "This is today's AWS, just oriented toward the future."
"Betting trillions of dollars in buildout on the revenue of two labs is a big mistake," Huang said. "I'm happy to work with them and support them. They will become one of the most important tech companies ever and achieve brilliant success. But their revenue doesn't need to approach a trillion dollars for this to work."
The Underlying Logic of Accelerating Demand
In the interview, Huang systematically explained the internal mechanisms of why compute demand will not decelerate linearly but will continue to accelerate.
The core lies in the multiplication of two compounding factors. First, inference (thinking) is orders of magnitude more expensive than recitation-style generation—AI needs to explore multiple options, plan, reflect, prune, and retain "strategic options," which itself consumes substantial compute. Second, the more capable the model, the more scenarios it gets "hired" for. "It uses more compute to do the work, but it does it well, so we hire it to do more work," he said. "These two things compound, and demand accelerates rather than decelerates."
The key unlock last year, which Huang called "harness"—working memory management, knowledge graph construction (similar to offline compression during sleep), and tool use—these three things transformed large language models from "chatbots" into intelligent agents capable of doing work in enterprises. "Coding was the first landing scenario," he said. "Design and creativity are the next wave."
In his view, OpenAI and Anthropic currently each have only about two to three gigawatts of compute, "which is very small relative to the scale that is coming." And it is at such a small compute scale that AI has already solved the Navier-Stokes equations, a Millennium Prize mathematical problem. "This is the worst intelligence we will ever have from now on," he said. "From here, it only gets better."
Power Scarcity Reverses into NVIDIA's Share, 2027 Bottleneck in the Physical World
In the interview, Huang gave a clear assessment of the compute buildout bottleneck for 2027: demand will be "no less, only more," but land, permits, power, and construction labor will be the main obstacles.
According to BG2 host Brad Gerstner, citing SemiAnalysis data, 43 gigawatts of AI compute will be added next year, exceeding the total global installed base today, with another 40 gigawatts added the following year. Huang agreed with the demand-side assessment of these figures but pointed out that physical world constraints will limit the actual pace of deployment. "Land, power, shells," he said. "This will be one of the most critical obstacles."
It is worth noting that in Huang's logical framework, power scarcity actually constitutes a driving force for further increases in NVIDIA's market share. His argument is that NVIDIA's architecture delivers the highest tokens per watt globally, and the tighter the power supply, the more customers tend to choose the platform with the highest output per unit of energy. "When power is reduced, the quality of compute you put in must be the highest," he said. "We are the most rentable, the most interchangeable, the most durable."
He further disclosed that NVIDIA's market value per gigawatt of compute has risen from about $18 billion to over $20 billion, now exceeding $40 billion, and is still rising. In global cloud service providers' (CSP) capital expenditure, NVIDIA accounts for about 40%; on the non-CSP side (including on-premises deployments and new cloud computing), it accounts for about 80%. "Six years ago, our share of global capital expenditure was still in the single digits," he said.
By the End of This Decade: $3-4 Trillion Annual Buildout Scale, "Not Surprised at All by $10 Trillion"
In the interview, Huang provided his framework for judging the long-term scale of AI infrastructure, which is also the core basis for his view that the market is underestimating the persistence of AI compute demand.
His calculation path is as follows: before AI emerged, global computing infrastructure annual buildout was approximately $500 billion and on a natural growth trajectory. Combined with the impact of the end of Moore's Law—"if you want twice the transistors, no matter how long you wait, you'll pay twice the price"—the semiconductor industry was already on a path to becoming the world's largest industry. Adding AI-driven new demand and a new wave of entrepreneurship, "by the end of this decade, annual buildout will be between $3 trillion and $4 trillion, closer to $4 trillion, with pure computing at about $1.5 trillion," he said.
For a longer-term outlook, he stated: "If someone tells me annual infrastructure construction and refresh reaches $10 trillion, I wouldn't be surprised at all."
Regarding NVIDIA's own financial strategy, Huang revealed that the company currently allocates over 50% of free cash flow to stock buybacks, with that ratio reaching about 60% last quarter. He also mentioned two new growth vectors: enterprise software licensing (charging annual fees based on GPU count) and revenue-sharing (Revshare) arrangements with some cloud service providers—both paths are relatively independent of hardware sales and rely on NVIDIA's massive installed base and ecosystem.
Comments