Nvidia and six of the world's largest financial firms announced plans on Monday to raise over $500 billion for AI "factories," the data centers that train and run AI models. The goal is to turn AI compute into a new investable asset class, but the entire venture hinges on contractual terms and on who absorbs the loss if hardware ages faster than the loans used to buy it.
The announcement, made on Aug 10, 2026, is not a done deal. The financing platforms are memorandums of understanding, or MOUs, not signed contracts. There is no timeline, no split of money among the six firms, and no first project named. The $500 billion is a target for capital to be raised over time, not Nvidia's revenue.
The six firms are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR. Each will judge every deal on its own merits. The key word is "independent." Nvidia supplies the computing platform, but investors decide what to fund.
A New Asset Class or a New Risk Class?
Nvidia CEO Jensen Huang argues that AI compute can be an investable asset class. "Those are the data centers that train and run AI models. CEO Jensen Huang argues that this compute can be an" investable asset class, he said. The company introduced its Vera Rubin platform at its GTC conference on March 16, 2026, along with the Rubin Ultra GPU architecture, signaling a rapid cadence of new hardware.
The likely borrowers are AI labs, big enterprises, and cloud companies. Investors will weigh customer demand, hardware usage, cash generation, and secondhand value. Residual value is what equipment is worth if a customer walks away. Huang said Nvidia may offer "residual-value support for up to 25% of an opportunity," case by case. That support is "limited" and adds to independent underwriting, not replaces it.
No details have been provided on how the support works or who is on the hook first. Calling AI compute an asset class does not make it one. Contracts do.
The Rental Market Tells a Complicated Story
Rental prices for Nvidia's H100 chips have moved in both directions. A one-year rental rose from about $1.70 per hour in October 2025 to $2.35 per hour in March 2026. That sounds like strength. But tracking firm Silicon Data reports the median H100 rental from big cloud providers was roughly $9.34 per hour in the second half of 2024, and it fell to about $6.26 per hour a year later.
Those numbers answer different questions. Rental income is not the same as resale value. A machine can earn steady rent and still lose most of its worth on the secondhand market when a newer chip arrives.
Amazon offers a warning. Effective Jan 1, 2025, the company shortened the useful life of some servers and networking gear from 6 to 5 years. Amazon cited "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." That change added roughly $1.4 billion to 2025 depreciation and cut net income by about $1 billion, mostly at AWS.
Amazon did not write down Nvidia chips specifically. But the move shows how quickly hardware is considered obsolete. Investor Michael Burry, in November 2025, estimated big cloud firms are understating AI depreciation by about $176 billion from 2026 through 2028. That is an estimate, not a reported loss, but it signals a real concern.
Who Bears the Loss When Hardware Ages?
Nvidia argues the opposite side of the ledger. The company says A100 chips released in 2020 still draw multi-year commitments, with useful life stretching toward a decade. Nvidia's CUDA software improves output on installed chips, which helps keep older machines productive.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Still, the gap between rental income and resale value is the crux. Airlines borrow against planes in a similar way, with contracted revenue paying down loans. But planes fly for decades. AI chips face a refresh cycle measured in years, sometimes months.
The article poses four questions that must be answered before signing off on GPU-backed financing. First, who has committed to use the capacity? Second, will the loan clear before the next hardware refresh? Third, who can redeploy the machines if the original customer fails? Fourth, who takes the first loss if resale value falls short?
These are not academic questions. Overbuilding worries have surfaced at Meta, and there is a fight over power and transformers to run all these data centers. Demand can be real and collateral can still disappoint.
Outside Money May Only Appear Because Nvidia Covers Part of the Downside
The structure of this deal suggests that outside money may only appear because Nvidia is willing to cover part of the downside. The residual-value support of up to 25% of an opportunity is a form of backstop. But it is not a guarantee. It is unclear who takes the first loss if a project fails.
The six firms are sophisticated. Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR have all sat through infrastructure financing reviews. They know the difference between a label and a contract. The word "independent" is meant to reassure investors that each deal will be judged on its own economics.
But the viability of the entire $500 billion target depends on those contractual terms. If a loan matures before the hardware is obsolete, the math works. If the hardware ages faster than the loan, someone eats the loss. Nvidia's support may soften that blow, but it does not eliminate it.
The Clock Is Ticking on Hardware Value
The timeline is unforgiving. Nvidia introduced Vera Rubin in March 2026. The H100 rental price rose from $1.70 to $2.35 per hour between October 2025 and March 2026. But the median big-cloud H100 rental fell from $9.34 to $6.26 per hour over a year. These are different measurements, but they all point to the same tension: hardware value is volatile.
Amazon's depreciation change, effective Jan 1, 2025, added $1.4 billion to its 2025 depreciation and cut net income by $1 billion. Burry's estimate of $176 billion in understated AI depreciation from 2026 to 2028 is a warning shot. Nvidia's claim that A100 chips from 2020 still draw multi-year commitments is a counterargument. Both cannot be fully right.
The financing platforms are MOUs, not contracts. There is no timeline, no split, no first project. The $500 billion is a target, not a commitment. Investors will need to answer the four questions before any money moves. Who is committed to use the capacity? Will the loan clear before the next refresh? Who can redeploy the machines? Who takes the first loss?
Those answers will determine whether AI factories become a new asset class or a new cautionary tale. Nvidia and its six partners are betting on the former. The fine print will decide.