Nvidia Bets Its Balance Sheet That $500 Billion in AI Chips Won't Age

Nvidia is backing up to 25% of the residual value of its AI chips in a $500 billion financing push with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, aiming to position GPUs as long-life infrastructure assets. However, critics like Michael Burry warn that depreciation assumptions are stretched, estimating a $176 billion gap between 2026 and 2028, and Amazon and Meta have moved in opposite directions on useful-life estimates for the same hardware.

Nvidia is trying to make AI chips look like long-life infrastructure, not fast-aging hardware. You can see why Wall Street wants the deal, and you can also see why investors are checking the depreciation math. Nvidia's new $500 billion financing push is not just a chip story. It is a balance-sheet story. Bloomberg, the Financial Times and Business Insider have all reported that Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are working with Nvidia on financing platforms meant to fund AI data centers with third-party capital. The money comes from Wall Street. The confidence comes from Nvidia. Here is the plain version. A data-center operator wants thousands of Nvidia GPUs but doesn't want to pay cash up front. A lender funds the purchase and uses the chips, the lease payments, or both as the asset behind the debt. That structure only works if the chips are still worth something when the loan has to be repaid. So Nvidia is offering residual-value support on some projects. Bloomberg and Barron's reported that Jensen Huang has put that support at up to 25% on a case-by-case basis. That is the guarantee. It is also the argument. Huang has been selling AI compute as an asset class. Business Insider reported that he has described Nvidia's AI infrastructure as something closer to a long-term, revenue-producing asset than a normal piece of IT equipment. You can understand the pitch. Railways, aircraft and data centers get financed because they throw off cash for years. Nvidia wants its GPUs, CUDA software and full AI systems to sit in that same bucket. The problem is simple. Chips are not railways. Nvidia's products improve fast, and its own roadmap is part of the risk lenders have to price. Hopper gave way to Blackwell, Rubin is already in the conversation, and every new generation makes the old one less attractive for the most demanding workloads. That doesn't mean older GPUs become worthless. It does mean you shouldn't pretend the resale market is settled science. The depreciation fight is now the real story Credit investors noticed. Axios, citing Bloomberg and ICE Data Services, reported in late July that the cost of protecting Nvidia debt through credit default swaps saw its biggest intraday increase since the contracts began actively trading in November, after reports that Nvidia could backstop as much as $250 billion tied to an OpenAI data-center project in Ohio. Bloomberg also reported that Nvidia was in talks to support OpenAI's lease of computing from a planned $500 billion, 10-gigawatt SoftBank-backed hub. That is why the phrase circular financing keeps coming back. Nvidia backs the customer. The customer buys Nvidia chips. The lender finances the deal because Nvidia is standing somewhere behind it. When demand is rising, the loop looks clever. When values wobble, the loop looks much less clever. Michael Burry has been blunt about this. Investing.com reported that Burry accused major AI spenders of understating depreciation by stretching the useful life of compute assets, and estimated the gap could reach $176 billion between 2026 and 2028. TipRanks reported the same criticism, including his view that companies buying Nvidia chips on a two-to-three-year product cycle shouldn't be extending useful lives to five or six years. That's not a rounding error. He has a point worth taking seriously, even if you don't buy the whole bear case. Amazon's 2025 filing shows it changed the useful life of a subset of servers and networking equipment from six years to five, effective January 1, 2025. Same hardware. Different math. Meta went the other way: its 2025 filing says it extended most servers and network assets to 5.5 years, also effective January 1, 2025. Two of the biggest AI spenders looked at the same hardware cycle and moved in opposite directions. That is not a footnote. It is the center of the financing model. Older GPUs still have a market The strongest argument for Nvidia is not theoretical. It is CoreWeave. Business Insider and Tom's Hardware reported this week that CoreWeave has signed contracts for Nvidia A100 GPUs running into 2029, nine years after the A100 architecture first arrived in 2020. CoreWeave said demand for the older chips remains strong because not every customer needs the newest rack-scale system, and not every data center can handle the power and cooling demands of the latest Nvidia gear. That helps Nvidia. A lot. If older GPUs can keep earning money for inference and less demanding workloads, residual-value support becomes less frightening. The used-chip market has something real under it. You can finance against cash flow, not only hype. But don't confuse that example with proof that every financed GPU will hold its value. A100s have survived partly because the infrastructure around them still fits. Future chips may age differently. A flood of used hardware from overbuilt AI projects would also test pricing in a way today's market has not. Nvidia's move is smart because it brings pension funds, insurers and private-credit firms into the buildout without forcing Nvidia to write every check itself. It is risky because the company is now asking lenders to believe that its chips are both the latest products and durable collateral. Those ideas can live together for a while. They won't always move together. If the resale value holds, Nvidia has turned its balance sheet into a weapon its rivals can't easily copy. If it doesn't, the company will have done more than sell the shovel in the AI gold rush. It will have helped finance the mine. Also read: Meta's Smart Glasses Trigger a Privacy Backlash It Can't Fix With Software https://startupfortune.com/metas-smart-glasses-trigger-a-privacy-backlash-it-cant-fix-with-software/ • IBM Partners With OpenAI To Put GPT-5.6 Inside Its Consulting Business https://startupfortune.com/ibm-partners-with-openai-to-put-gpt-56-inside-its-consulting-business/ • DeepSeek Is Raising API Prices by Up to 1,100% Ahead of a Possible IPO https://startupfortune.com/deepseek-is-raising-api-prices-by-up-to-1100-ahead-of-a-possible-ipo/