트럼프 행정부 고위 고문이 Nvidia의 $500 billion AI 금융 로드맵이 심각한 GPU 공급 과잉 위기를 야기할 수 있다고 경고했다.

David Sacks, a member of President Trump’s Council of Advisors on Science and Technology, warned on the “All In Podcast” that the greatest threat to Nvidia’s $500 billion AI financing plan is not insufficient demand but rather an oversupply of compute capacity. Developed in partnership with investment institutions to transform GPUs into financeable, yield-generating assets, the initiative faces the risk of recreating the “dark fiber” crisis that followed the dot-com bubble—this time manifesting as “dark GPUs.” Sacks cautioned that such a scenario would deliver a systemic shock to the entire AI infrastructure investment chain. His remarks represent the most direct risk warning to date on this AI financing boom from within Trump administration circles.

Drawing a parallel to the early 2000s telecom overbuild, Sacks explained that when fiber optic cables were laid at massive scale, demand fell far short of expectations, leaving vast amounts idle and triggering a price collapse. He stated directly on the show: “To me, the biggest risk is not on the demand side. The biggest risk is that compute becomes oversupplied, that there's overbuilding. Just like after the dot-com bust there was dark fiber, if there are dark GPUs, that would be a disaster for everybody—especially for people who are building out compute infrastructure expecting spot prices of $30 to $50 per watt.” That valuation range references Elon Musk’s public estimate of AI compute worth, which he cited during an internal SpaceX conference call. According to reports, Musk projected that one gigawatt of compute capacity could generate a $300 billion to $500 billion annualized revenue run rate by the end of 2027, entering 2028—not actual revenue booked within the calendar year. However, cloud infrastructure provider Nebius subsequently disclosed that its multi-year agreements carry annual contract values of approximately $20 million to $25 million per megawatt. Further details from Nebius’s Q2 2026 earnings call revealed a two-tier pricing structure: one-to-three-year mid-term contracts at $20 million to $25 million per megawatt annually, and short-term contracts under six months reaching $40 million to $50 million per megawatt. This indicates that Musk’s $30–$50 per watt estimate aligns with the premium tier of short-term agreements rather than mainstream industry pricing, lending concrete weight to Sacks’s argument that a “premium bubble” exists between short-term spot rates and long-term contractual norms.

Beyond issuing warnings, Sacks offered a counterintuitive assessment: heavy political resistance to data center construction may paradoxically shield the market from oversupply. “If suddenly too many people are racing to supply compute and there's a glut and the market crashes, that's the real risk,” he said. “But weirdly, all of this political resistance actually serves as insurance against that outcome. Because building data centers is just so hard—for all the reasons we've discussed, and now there's a moral panic, hysteria, even a circus playing out. It's this political resistance that I think almost guarantees that supply won't outstrip exponentially growing demand. So in a strange way, you're actually protected.” Higher construction barriers, he reasoned, slow supply expansion and mitigate oversupply risks. Sacks also outlined the fundamental driver behind Nvidia’s financing strategy: the capital expenditure gap among downstream buyers has grown so large that traditional equity and debt financing can no longer cover it. Using Musk’s expansion plans as an example, Sacks noted that adding roughly six to eight gigawatts of compute next year requires $300 billion to $400 billion in capital expenditures. While SpaceX recently raised $100 billion through equity and debt, additional financing mechanisms remain necessary. Sacks identified vendor financing as the direct solution, explaining that Jensen Huang is leveraging major banks and private equity firms to establish credit lines available to downstream buyers, thereby alleviating the current infrastructure funding bottleneck.

At the core of Nvidia’s plan is the transformation of GPUs into an asset class resembling financial securities, with Nvidia providing residual value support to endow them with financeable, yield-generating properties. Recent developments regarding these guarantees have drawn significant market scrutiny. According to market sources, Nvidia has committed to covering residual value shortfalls for certain projects up to 25% of the project’s scale if customer-held compute assets depreciate below expectations at exit. The specific calculation methodology for this guarantee ratio has not yet been clearly defined. Given that GPUs are extremely expensive, rapidly obsolete, and increasingly scrutinized by cautious investors and banks, analysts warn that Nvidia’s pivot from a “chip vendor” to a “compute insurer” could expose it to a dual threat of declining revenues and substantial guarantee payouts. Additionally, Nvidia has scaled back its financial guarantee to OpenAI from $250 billion to no more than $120 billion. OpenAI had previously announced a $500 billion investment plan for Ohio to build the world’s largest compute center and agreed to purchase all its GPUs from Nvidia, prompting the initial guarantee commitment. However, with OpenAI still lacking stable profitability and facing payment capacity questions, alongside rising competitive momentum from rivals like Anthropic, Nvidia faced pressure from overcommitment. According to The Wall Street Journal, Nvidia’s stock price fell nearly 5% at one point after the initial $250 billion guarantee proposal was exposed, as investors expressed concern over excessive balance sheet risk concentration. On the demand front, however, indicators remain robust: Anthropic’s second-quarter revenue surged 13 times year-over-year to $11.5 billion, and Tencent prepaid over 51 billion yuan (approximately $7.6 billion) in Q2 for compute procurement deposits, signaling continued acceleration in enterprise and AI-driven demand.