Nvidia’s quarterly report lists $108.5 billion of maximum exposure under guarantees that stand behind some of its customers, while seven large AI cloud and tech companies book their servers or GPUs to last up to six years, according to their own filings; five of the seven use five years or more. The report, for the period to 26 July 2026, lists $3.5 billion of guarantees on AI cloud partners’ data centre leases and guarantees capped at $105 billion for an OpenAI campus in Ohio. On the earnings call, finance chief Colette Kress described a model in which, she said, “we get paid twice—once on the hardware sale, and again through the share of rental revenue”, as Fortune reported.
A post on X from an account called Haircut argues that this makes the evidence that old chips keep earning harder to trust, because “some of it is the chip makers’ own money.” That is the poster’s view; the filings below are what can be checked.
A Business Pill Short that uses this story as its worked example. Facts: six-year lives from this piece; Microsoft and Oracle as reported by Bloomberg, in our Microsoft piece.
Business Pill · Financial Conditions
A short Business Engineer explainer on financial conditions: how the cost and availability of credit change. It teaches the general idea only and does not discuss Nvidia or any company in this story.
The key insight: Five of seven large AI cloud and tech companies we checked book servers for five years or more, and none for more than six, according to their filings, while Nvidia’s 10-Q lists $108.5 billion of maximum guarantee exposure and supply commitments up $160 billion in one quarter (our arithmetic). On $1 billion of GPUs, a six-year life shows about a third less annual depreciation than four (our arithmetic).
Guarantees. Nvidia’s 10-Q says it “entered into land, power, and shell guarantees for select AI cloud partners’ data center lease obligations in the event of their default”, with maximum gross exposure of $3.5 billion over terms of five to seven years, partly offset by $712 million the partners placed in escrow. Separately, guarantees capped at $105 billion support a buildout by affiliates of SB Energy on behalf of an OpenAI affiliate, for about 4.25 gigawatts in Pike County, Ohio.
Cloud agreements. Under its AI cloud agreements, which total $36 billion, the AI clouds “procure our data center infrastructure products and we commit to cloud service agreements”, and Nvidia says it will “participate in revenue share” from third-party customers if certain criteria are met.
Supply. The same filing says Nvidia increased “supply commitments from $119 billion last quarter to $279 billion”, commitments that are “primarily memory and manufacturing facilities.” That is $160 billion more, up about 134% in one quarter (our arithmetic).
The call. For neoclouds, Kress said Nvidia guarantees it will pay for a minimum portion of a data centre’s capacity and takes a cut of rental revenue above that level, according to Fortune. “Independent capital still underwrites every deal on its own merits. We’re not making loans,” she said.
A day after the call, the Wall Street Journal reported that Nvidia had paused some deals in that revenue-sharing program, and that some Nvidia employees had told customers it could draw antitrust scrutiny, according to Reuters. Nvidia said the model “is still in place and continues to evolve due to high demand.”
What Was Posted on X
$NVDA $AMD $AVGO
nvidia now lends to and guarantees the same buyers whose chips its next generation will make obsolete, and AMD and broadcom have joined in. that makes me trust the old-chips-keep-earning evidence a lot less, because some of it is the chip makers’ own money.
nvidia describes the tools itself:
— on the Q2 call colette kress said nvidia gives a take-or-pay commitment on part of a facility’s capacity, and that it gets paid twice, once on the hardware sale and again through a share of the rental revenue.
— at goldman, jensen’s test for circularity was the return on the stakes, put one in and get 100 back, which answers for nvidia’s equity and says nothing about whether the borrower’s rent covers the loan once its chip stops being the scarce one.
— in its september roadshow deck nvidia’s supply commitments went from $119bn to $279bn, mostly memory, so in an air pocket it’s exposed on memory it has promised to buy as well as on the chips it has guaranteed.
then there’s the software licence jensen floated in berlin, about $2,000 per GPU a year, as an example rather than a product.
over 8,760 hours that’s 23 cents an hour on top of the 43-cent floor an H100 already carries, a 53% rise in the running cost of an old chip if operators have to buy it, charged by the company arguing that old chips keep earning.
The account Haircut (@nodiligence) posted this at 21:31 UTC on 5 November. It argues that Nvidia now lends to and guarantees buyers whose chips its next generation will make obsolete, and that <a href=”https://fourweekmba.com/amd-business-model/”>AMD</a> and Broadcom have joined in, so some of the evidence that old chips keep earning is the chip makers’ own money. The post cites Colette Kress on the Q2 call, Jensen Huang at Goldman and Nvidia’s supply commitments of $119bn to $279bn. This piece checks those claims against Nvidia’s 10-Q and the buyers’ filings; the post’s own lender and software figures are its own.
financing new inference chips will be the hardest part of competing with nvidia
to get these chips to users, you have to get them into data centers. neoclouds borrow money to buy them but if a neocloud canโt repay these loans, the lender takes the chips and resells them
this works for nvidia because used nvidia gpus have a market. inference chips have a problem because they don’t have a resale market yet
I think lenders will hesitate to fund these chips, or maybe charge more to do it.
nvidia has a structural advantage even if other chips are actually faster
Jessie Dong, whose X bio reads “Compute & AI”, posted at 04:09 UTC on 6 November the view that financing new inference chips will be the hardest part of competing with Nvidia: neoclouds borrow to buy chips, a lender that is not repaid takes the chips and resells them, and used Nvidia GPUs have a resale market while inference chips do not yet, so Dong expects lenders to hesitate or charge more. The chart in the post, sourced on the image to the American Compute Rack Report (October 2026, listing-based estimates through August), puts used eight-GPU Nvidia servers at about $95K for A100, $195K for H100 and $349K for H200 (a limited sample), and shows no resale prices for inference chips from SambaNova, Cerebras, Etched, Positron or d-Matrix.
Depreciation life for servers or GPUs stated in each company’s latest annual or quarterly filing. Data: company filings. Microsoft’s bar shows the top of its two-to-six-year range.
CoreWeave, which says it deploys NVIDIA GB200 and GB300 NVL72 systems, depreciates technology equipment over six years; it moved computing equipment in its data centres from five to six years in January 2023.
Oracle lists servers and networking equipment at an estimated useful life of six years.
Alphabet depreciates servers and network equipment “generally over a period of six years.”
Microsoft gives servers and network equipment two to six years.
Meta raised the life of most servers and network assets to 5.5 years from January 2025.
Amazon cut a subset of servers and networking equipment from six years to five from January 2025, citing “the increased pace of technology development, particularly in the area of artificial intelligence and machine learning.”
Nebius, which says it relies on Nvidia for its GPU chips, says its GPUs “are currently depreciated over four years”; in January 2026 it concluded that servers and network equipment should move from four to five years.
The Arithmetic
On straight-line depreciation with no salvage value, $1 billion of GPUs costs $250 million a year over four years and about $167 million a year over six. The six-year book shows about $83 million, or a third, less depreciation a year on the same hardware (our arithmetic). Haircut’s posts say Nvidia puts GPU life at “eight years or more”, that “neoclouds book six and their banks underwrite three to four”; we could not check either figure. It also argues that the roughly $2,000 per GPU a year software figure Jensen Huang floated would add 23 cents an hour, a 53% rise on an H100’s running cost; those inputs are the poster’s.
As we read the filings, each life is the company’s own estimate, and the same class of hardware runs from four years for Nebius’s GPUs to six for CoreWeave’s technology equipment.
As we read Kress’s description, the same structure that earns Nvidia a share of rental revenue also puts it behind the borrower if that revenue falls short.
Colette Kress, Nvidia CFO, as quoted by Fortune
“In this model, we get paid twiceโonce on the hardware sale, and again through the share of rental revenue.”
Three Implications
FOR NEOCLOUD LENDERS Cloud filings show lives of up to six years for their servers; Nebius uses four for GPUs.
FOR NVIDIA SHAREHOLDERS The 10-Q lists $108.5 billion of maximum guarantee exposure and $279 billion of supply commitments.
FOR ACCOUNTING WATCHERSAmazon shortened part of its fleet to five years citing AI; Meta and CoreWeave lengthened theirs.
The framework says: “The $1T capex cycle is being underwritten by a $1T concentrated demand commitment from two counterparties whose combined free cash flow is deeply negative.”
As we read it, Nvidia’s guarantees add the chip supplier to the list of parties standing behind that demand. The framework does not discuss this research.
What Is Not Established
No filing we read links Nvidia’s guarantees to any buyer’s depreciation life, and the lives above cover servers broadly, not GPUs alone, except for Nebius. Kress’s words are as quoted by Fortune; we did not read a full transcript. The post’s claims that AMD and Broadcom have joined Nvidia in lending to buyers, and its figures for lenders and Nvidia, are its own and unchecked here.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Large AI clouds book their servers for up to six years, while Nvidia has guarantees and capacity commitments behind some of them and says it is paid twice where it shares rental revenue.
95,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
A note on sourcing. Nvidia’s commitments and guarantees are from its Form 10-Q for the quarter to 26 July 2026. Colette Kress’s remarks are as quoted by Fortune on 26 August 2026; the paused revenue-sharing deals were reported by the Wall Street Journal, as carried by Reuters. Useful lives are from each company’s latest annual or quarterly filing. The piece was prompted by a post on X from the account Haircut; its lender, Nvidia and software figures are the poster’s. Arithmetic is ours.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
