How far will AI infrastructure spending keep climbing? Many readers likely feel uneasy about a figure that keeps expanding with no ceiling in sight. NVIDIA's announcement on August 10, 2026 of a lending facility exceeding $500 billion looks like one answer to that unease. But what tends to get overlooked in this plan isn't the sheer size of the number—it's the fact that NVIDIA itself is taking on residual value guarantees of up to 25% on a portion of these deals. A figure deeply involved in the administration's AI policy named, in a podcast, an oversupply risk that runs counter to intuition—more so than any shortage of demand. It echoes the memory of "dark fiber," when nearly 90% of the fiber-optic network built two decades ago sat unused.
What NVIDIA's $500 Billion Plan Actually Is
On August 10, NVIDIA announced it would team up with six firms—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—to launch an independent lending platform to finance the construction of AI compute infrastructure. The goal is to mobilize more than $500 billion in third-party capital. All six participating firms are asset managers or private equity firms with large assets under management, and this move is positioned as part of a broader trend of incorporating AI compute as a new investment asset class. What's important to note here is that the $500 billion figure is the sum of the mobilization targets each of the six firms aims for under its own independently operated platform, and the whole framework is based on memoranda of understanding (MOUs). As the word "mobilize" suggests, this is a framework representing something close to an upper bound each firm is aiming for—it does not represent the total value of finalized lending contracts.
On top of this, NVIDIA is reported to be offering residual value guarantees of up to 25% on some of these deals. If you mechanically apply this publicly disclosed 25% ratio to the entire $500 billion mobilization target, you arrive at a theoretical ceiling of roughly $125 billion. However, the guarantee reportedly applies only to "a portion of the deals," and the specific transaction amounts and number of deals covered remain undisclosed. The $125 billion figure is merely an upper bound derived from applying the disclosed ratio to the entire mobilization target—it is not the actual guarantee balance NVIDIA is currently and definitively on the hook for. A residual value guarantee is a mechanism whereby, if the future resale or repurposing value of an invested asset (in this case, GPUs) falls below a predetermined level, NVIDIA compensates for the shortfall.
From the perspective of the lending financial institutions, this lets them offload onto NVIDIA the risk of committing long-term capital to a fast-depreciating asset like GPUs. That's precisely why financial institutions can extend capital with greater confidence, and why the $500 billion mobilization target starts to feel achievable. Flip that around, though, and this guarantee is also evidence that NVIDIA itself has designed the deal while anticipating the very scenario in which "the future value of GPUs might decline." Built into the very foundation of this plan is the paradox that the party providing the guarantee is the one that most precisely understands where the risk actually lies.
By placing construction financing in vehicles outside NVIDIA's own balance sheet—on the financial institutions' side—the uncertain variable of data center utilization rates can be separated from NVIDIA's earnings. This is the reason each firm chose the design of operating independently. GPU sales are booked as revenue on NVIDIA's own books, while much of the risk over whether the constructed data centers actually get utilized sits with the lending platforms instead. The residual value guarantee is also, in a sense, an admission that this separation cannot be made complete.
What Sacks's Warning About "Dark GPUs" Really Means
The sharpest skepticism toward this plan has come from David Sacks, co-chair of the President's Council of Advisors on Science and Technology (PCAST). Sacks served as the White House's AI and crypto "czar" overseeing AI and cryptocurrency policy from January 2025, before stepping down from that role on March 26, 2026 to become PCAST co-chair. He is also a co-founder of Craft Ventures and someone positioned close to AI and data center investing.
Sacks is reported to have said on the All-In Podcast that the biggest risk to NVIDIA's $500 billion plan isn't insufficient demand but oversupply. The gist of his remark was that if a huge volume of "dark GPUs"—GPUs with no assigned workload—were to emerge, it would be a disaster. If it were merely a shortfall in demand, prices would simply stay elevated and the only consequence would be a delay in recouping the investment.
But oversupply is different. GPUs that sit idle just keep depreciating, and if you try to resell them, buyers' bids come in below the guaranteed floor. Covering that gap is exactly what NVIDIA's up-to-25% residual value guarantee is for. What Sacks's warning zeroes in on is where the losses land the moment demand stalls.
In the same podcast, Sacks reportedly referenced the high electricity cost estimates put forward by Elon Musk, while pointing out that actual contracted prices—such as those from Nebius—are significantly lower than that assumption. If supply plans are being stacked on top of market realities that diverge sharply from these high-cost assumptions, it suggests the cost structure itself may rest on overly optimistic premises. If the input estimate for electricity costs is too lenient, then the assumptions underlying the entire edifice of construction plans and guarantee structures built on top of it become shaky as well.
What the Dark Fiber of the 2000s Teaches Us About the Cost of Oversupply
What Sacks pointed to as a precedent was the "dark fiber" phenomenon that emerged from the telecom bubble collapse of the early 2000s. In the late 1990s, telecom carriers, anticipating explosive growth in internet traffic, raced to lay continent-spanning fiber-optic networks. But demand growth didn't come as fast as predicted, and after a wave of telecom carrier bankruptcies led by WorldCom, the vast majority of the fiber that had been laid was left "dark"—unused, with no light even passed through it. According to estimates from the tech media outlet technostatecraft, the unused rate peaked at around 97.5% in 2002, and even by 2005 roughly 85% remained unused.
Flip that figure around, and it means that as of 2002, only about 2.5% of laid lines were actually in operation. This oversupply also triggered price collapse simultaneously, with bandwidth costs falling sharply and making it even harder for telecom carriers to recoup their investments. The miscalculation investors made at the time concerned the timeline over which demand would catch up.
This lesson applies to GPU investment because both share the same structure: making advance investments premised on long lead times, betting on future exponential growth. Building data centers, procuring semiconductors, and securing power infrastructure all take multiple years. A GPU cluster ordered today won't come online for one to two years, and there's no guarantee that AI demand at that future point will have grown in line with what was assumed at the time of ordering. GPU investment carries the same kind of structural time lag that fiber optics did—the gap between "laying it down" and "actually putting it to use."
The Full Picture of Circular Financing and Burry's Warning
Alongside Sacks's warning, prominent investor Michael Burry is also reported to have criticized the AI industry's circular capital structure on X on August 13. The circular pattern goes like this: hyperscalers invest in AI development companies, those AI companies use that capital to purchase compute from the hyperscalers, and that purchase then gets booked as revenue on the hyperscaler side. Burry reportedly pointed out that off-balance-sheet debt held by the five major hyperscalers totals approximately $1.65 trillion combined.
This concern itself isn't new. Bloomberg has analyzed a circular structure among ten major companies involving roughly $46 billion in direct cross-investments and $879 billion in interlocking purchase agreements, and NVIDIA's $500 billion plan was announced right in the middle of these spreading concerns.
It's worth distinguishing that this circular structure and NVIDIA's $500 billion lending platform are separate deals, with different capital providers and different mechanisms. NVIDIA itself has announced more than $540 billion in similar deals cumulatively for fiscal year 2026, including acquiring $30 billion worth of equity in February 2026 as part of OpenAI's restructuring, and investing up to $10 billion in Anthropic in November 2025. These are transactions in which NVIDIA is directly creating the very demand of the companies it invests in—something quite different from the $500 billion lending platform's mechanism of mobilizing capital from financial institutions to fund the construction of compute infrastructure. The November 2025 investment in Anthropic was up to $10 billion from NVIDIA alone, but reportedly totals around $15 billion when including Microsoft.
That said, the two are not unrelated. Whether the data centers built through the lending platform ultimately get utilized depends on whether AI development companies like OpenAI and Anthropic can keep buying compute—and if that latter circular structure weakens, it will spill over into the former's utilization rates. NVIDIA's guarantee covers up to 25% of the appraised value of the covered deals, and any decline beyond that is borne directly by the six firms and the investors receiving the financing—a structure whose asymmetry shows up in the accounting as well. The guarantee NVIDIA provides is a contingent liability tied to the future value of GPUs, while the profit from GPU sales has already been booked as revenue on NVIDIA's own books—meaning the guarantee's potential losses won't surface until they're actually triggered. Revenue is locked in today; losses only ever materialize in the future—and that very mismatch in timing is the essence of this structure.
This scale of numbers isn't a distant matter for Japanese readers either. Converted at the assumed exchange rate as of August 2026 (¥150 to the dollar), $500 billion comes to roughly ¥75 trillion, and the same kind of residual value guarantee and off-balance-sheet debt structures could well be brought into play in the same form whenever domestic companies participate financially in the global AI supply chain.
How much the six firms—Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR—have actually allocated, under what conditions the residual value guarantee would be triggered, and how the rates are set: none of this has been disclosed. Utilization rate, arguably the single most crucial metric, likewise remains unpublished. The very state of affairs—where only the scale of the dollar figures gets reported while capital keeps being mobilized without visibility into the conditions that would trigger the risk—is what shapes the precariousness of this plan.
If Demand Is Strong, Why Talk of Oversupply?
At first glance, this warning appears to contradict the actual numbers. NVIDIA's data center revenue for fiscal Q3 2026 (ending October 26, 2025) was $51.2 billion, up 66% from roughly $30.9 billion in the same period a year earlier—a quarterly record. There's no sign anywhere of demand cooling off.
Comparing the estimated ceiling of $125 billion on the residual value guarantees NVIDIA has taken on against that record quarterly revenue of $51.2 billion, the guarantee amount works out to roughly 2.4 quarters' worth of revenue—a bit more than half a year. In other words, this guarantee is sized for a scenario in which "demand collapses to near-zero for more than half a year." While the $125 billion figure itself is merely an estimated ceiling, its scale reflects just how large an assumption is baked into the guarantee's design.
There's a substantial gap between the current 66% growth actuals and the scale of loss the guarantee anticipates. That gap is precisely what the oversupply risk is made of. GPUs ordered today won't come online until the future, and demand at that future point can't be assumed to reflect today's 66% growth.
There's also a decisive difference from dark fiber. The unused rate of telecom carriers' fiber could be gauged by simply counting the physical lines that had been laid, but GPU utilization rates are not regularly disclosed by either the hyperscalers or NVIDIA. While dark fiber was "visible inventory," dark GPUs are, for now, "invisible inventory"—one that will only become visible once demand growth stalls. What Sacks's warning is pointing to is precisely this invisibility itself.
Sacks's Paradox, and What to Watch for Next
Sacks is known as a proponent of deregulation, and even during his time as AI/crypto czar he leaned toward light-touch oversight that wouldn't stifle innovation. That the very same person is now singling out an oversupply risk that market discipline alone cannot prevent is, one could say, an ironic position to be in. As long as capital keeps being mobilized without disclosure of information like utilization rates or the conditions that would trigger the guarantees, market participants will have no way of knowing that demand has stalled until after the fact. The remarks of someone who has himself championed deregulation implicitly underscore the need for mechanisms—like disclosure—that supplement market discipline.
The odds that these numbers will be voluntarily disclosed seem slim, but there are indicators that could serve as clues. One is the contingent liabilities footnote in NVIDIA's quarterly earnings, where the actual usage of the guarantees may seep through as numbers. The other is any downward revision in hyperscalers' capital expenditure guidance—signs of demand slowdown tend to show up here before they show up in the utilization rate itself. When these two indicators start to move, that's when the true picture of dark GPUs will finally become visible from the outside.
It took nearly a decade for dark fiber to eventually be repurposed as the backbone of broadband and cloud infrastructure. Even if a temporary oversupply does occur, one difference from the dark fiber era is that GPUs have a broader range of outlets available for repurposing—beyond training, there's inference workloads and reallocation to smaller AI companies. How quickly that repurposing proceeds, accompanied by price adjustments, versus whether it instead goes through a decade-long stagnation, is what will determine the real success or failure of this plan.