Article — "NVIDIA Turns AI Compute Into an Asset Class, Mobilizing $500B With Wall Street Giants": Article — "NVIDIA Turns AI Compute Into an Asset Class, Mobilizing $500B With Wall Street Giants": "One-year H100 rental pricing rose from about $1.70 per GPU-hour in October 2025 to about $2.35 per GPU-hour in March 2026."
That single line, buried in NVIDIA's own blog post, is the load-bearing beam of a $500 billion structure. Rising rents mean the collateral appreciates. Appreciating collateral means you can lend against it. Everything else is scaffolding around that number.
On August 10, per CNBC, executives from the seven companies sat for what CNBC called a "rare, live joint interview" to bless the idea. Jensen Huang put it plainly: "This is really the first time that technology chips have become an investable asset class. These are revenue-generating assets now. They're productive, they're long-lived, they're fungible, they're flexible."
The mechanism: memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize over $500 billion in third-party capital, so NVIDIA's customers can buy NVIDIA hardware without loading it onto their own balance sheets.
Then Larry Fink reached for his analogy. He reportedly called this the start of the "next future for financial engineering," akin to the creation of mortgage-backed securities in the 1970s.
He meant it as praise. Read it as a tell.
Mortgage-backed securities were a genuine financial innovation that worked well for three decades and then, in 2008, discovered that the underlying asset—home values—did not in fact only go up. The entire edifice had been underwritten on a residual-value assumption that turned out to be a correlation, not a law.
Fink, of all people, knows this. BlackRock was born from the wreckage. That he reached for MBS as the flattering comparison tells you exactly which assumption this whole structure rests on: that the collateral holds its value.
So interrogate the collateral. NVIDIA's pitch is that a GPU is not a depreciating chip but a "fungible, redeployable" asset, kept alive by CUDA. Every software generation, the argument goes, makes already-installed hardware produce "more intelligence at lower cost throughout its life."
The proof point is the A100: introduced in 2020, still in commercial use six years later, extending A100's economic life "toward a decade." Depreciate a chip over ten years instead of three and the math on any financing deal transforms.
That is the claim underwriting decade-long depreciation schedules. And it may well be true—for now, while demand outstrips supply and rents climb.
But watch what NVIDIA does, not what it says. Buried in the same announcement: "In some cases, NVIDIA may provide a residual-value support mechanism for up to 25% of an opportunity, assessed carefully on a project-by-project basis."
Sit with that. NVIDIA is telling institutional investors that this compute is fungible, universally adopted, and holds its value across generations—and then quietly offering to backstop up to a quarter of the residual value itself.
If the market for used GPUs were as deep and liquid as claimed, why would the manufacturer need to guarantee the floor? You don't backstop the residual value of an asset you genuinely believe will only appreciate.
If the market for used GPUs were as deep and liquid as claimed, why would the manufacturer need to guarantee the floor?
NVIDIA frames the backstop as generosity—"substantially lower than other compute-financing arrangements," a favor it can afford because its compute is so uniquely valuable. But the more honest reading is that even NVIDIA doesn't fully trust the market it's inventing. The 25% is not confidence. It's a hedge against its own thesis.
This is the circularity the market already flagged. A supplier financing its customers to buy the supplier's product, then insuring the resale value of that product—if any one large customer wobbles, the ripple runs through the whole structure. Fink wants to raise the money "as fast as possible." Gray reportedly says AI use at Blackstone portfolio companies surged sevenfold this year. Nobody in that live interview was underwriting the downside.
But suppose the residual-value bet is sound. Suppose CUDA really does keep a Blackwell rack productive for a decade. There is still one thing NVIDIA cannot securitize, cannot upgrade, and cannot backstop.
The grid.
An AI factory is only "productive infrastructure" if it can plug in. Compute is revenue only when the megawatts arrive. And the megawatts are not arriving on the schedule the capital is.
On July 14, New York became the first U.S. state to ban data centers, reportedly imposing a moratorium of up to a year on any facility drawing 50 megawatts or more. Texas followed days later, ordering its grid operator to pause new connections pending an audit.
Here is the number that should unsettle every underwriter in that CNBC studio. About 14 GW of large load sits in NYISO's interconnection queue—12 GW from data centers and crypto across 40 projects, all but two of them above the 50 MW threshold now banned.
NYISO plans for roughly 3 GW of the total queue to actually come online within the next decade. The rest withdraws, delays, or dies in study.
About 14 GW in the queue; NYISO expects only ~3 GW to come online within a decade. That is not a financing gap you can close with $500 billion of institutional capital. It is a physical gap, and physics does not respond to attractive rates.
New York's own grid explains why. Since 2019, per NYISO, the state added 2.9 GW of generation and lost 4.4 GW. Its summer reserve margin this year was 417 MW—the thinnest since 2017. NYISO's Kevin Lanahan describes routinely relying on "old steam, peaking units that were built sometimes 70 years ago."
His diagnosis of what's needed: "new dispatchable generation. Hard stop." That is a decade-long, permitting-bound, capital-intensive problem that no securitization structure accelerates.
Notice where the capital actually flows in response. Firebird reportedly built the CIS region's largest AI factory in Armenia in "just over six months," and plans roughly 2 gigawatts across Armenia, Kazakhstan and other "frontier markets." Capital routes to where the power is, not where the demand is.
And notice what NVIDIA's engineering answer addresses. The 800 VDC architecture—the MGX power rack, the row power center, the DC power block—is entirely about getting power more efficiently from the grid to the GPU, so existing sites aren't "stranded."
It is a sophisticated solution to the last hundred feet. It does nothing about the interconnection queue, the retired baseload, or the state that just imposed a moratorium on your 200 MW campus.
This is where the asset class meets its collateral problem. You can financialize a GPU's residual value; there is a global market of offtakers and a manufacturer willing to backstop the floor.
You cannot financialize a grid interconnection that does not exist. There is no secondary market for a megawatt New York refuses to connect, no CUDA upgrade that conjures dispatchable generation, no tranche structure that overrides a governor's executive order.
The $500 billion is real, and it will move fast—Fink will see to that. The megawatts will move at the speed of transformers, transmission lines, and permitting fights that are only beginning.
That divergence is the whole story. Capital is being mobilized faster than the power it depends on, underwritten on a residual-value assumption its own author is quietly hedging, blessed with an analogy to the last asset class that discovered its collateral could fall.
The moratorium is not a footnote to this deal. It is the place where "compute is revenue" runs into a fact you cannot securitize: an AI factory that can't plug in produces no intelligence, no tokens, and no cash flow—just a depreciating asset waiting for a grid that isn't coming.