Glonce

Nvidia bankrolls its own customers, The Economist reports

businessThe Glonce editorsHacker News🔥82Glonce rating82Impact75Novelty65Relevance100Surprise70Players85Credibility80
Nvidia bankrolls its own customers, The Economist reports

The Economist reports that Nvidia's growth now rests on more than chip sales: CEO Jensen Huang has turned the company into a financier for much of the AI industry. Nvidia's valuation took thirty years to reach $1trn, nine more months to reach $2trn, and under two years after that to pass $5trn; it now stands at about $5.4trn, making Nvidia the world's most valuable company. Some analysts expect it to take in $1trn in annual revenue by 2029. To keep that growth going, Nvidia has moved into financial engineering. In mid-August it agreed to provide a backstop worth up to $105bn for a data centre SB Energy, a SoftBank unit, is building in Ohio for tenant OpenAI; the project will use 1.5m Nvidia processors. A week earlier Nvidia unveiled a plan to mobilise more than $500bn of AI infrastructure investment with six big Wall Street firms, guaranteeing the value of the hardware sold into such projects and potentially underwriting up to a quarter of some investments' cost. In July it began offering some neocloud customers income guarantees: over roughly six-year deals, Nvidia promises to pay a set price for compute and takes a share if the neocloud can sell capacity for more, a safety net meant to make the neocloud's revenue predictable enough to borrow against. Over the past three years Nvidia has pledged more than $70bn to startups and offered $300bn in financial support to customers; in theory its backstops could expose it to costs of nearly $300bn. Jay Goldberg of Seaport Research Partners says Nvidia is walking a fine line between "enabling demand" and "creating it", and though he does not think it has crossed that line yet, it is "getting pretty close". Michael Burry, known for betting against mortgage-backed securities before the 2007-09 financial crisis, is among the more sceptical, questioning what happens if AI-chip demand grows more slowly than expected while supply and prices move against Nvidia. Critics draw a parallel to the dotcom era, when Cisco and Lucent lent billions to telecom firms buying their equipment, then took big losses when demand fell short. The financial push is partly a response to hyperscalers, Nvidia's biggest customers, becoming rivals. Amazon, Google, Meta and Microsoft together account for roughly half of Nvidia's revenue and are projected to invest around $800bn this year, mostly on AI infrastructure, but most now design their own chips, which cost between a fifth and a third as much as Nvidia's and can be tailored to their own software. Google has already sold specialised processors to Anthropic, and Amazon expects its custom-chip business to become a significant revenue source. Bloomberg Intelligence forecasts custom chips rising from about 40% of the AI-processor market this year to roughly 50% by the end of the decade. Hyperscalers borrow cheaply on investment-grade credit; Alphabet sold $2.75bn of 50-year bonds in November at 5.7% annual interest. Upstart neoclouds pay far more: CoreWeave, the biggest of them, borrowed $2.6bn in July at almost double that rate. Nvidia's financing is aimed at narrowing that gap, partly by taking equity stakes, roughly 90 last year, nearly double the count two years earlier, and 60-odd more already agreed this year, in firms that will become customers or otherwise drive demand for its chips. Some of that money backs open-weight AI, seen as a way to counterbalance the hyperscalers: Nvidia agreed in August to pay Poolside $6bn to license its coding-model software plus $1bn for a stake, and separately agreed to buy Hugging Face, the open-weight model platform, for $12.9bn. The arrangements often interlock. Sharon AI, an Australian neocloud, expects to use a roughly $4.9bn backstop to deploy about 40,000 Nvidia chips; Firmus plans to buy as many as 170,000. Nvidia has invested more than $2bn in CoreWeave, owns about 11% of it, and has agreed to buy up to $6.3bn of CoreWeave's unused data-centre capacity through 2032. Even Anthropic, which buys chips from Amazon and Google, is entangled: it has reportedly signed a $35bn deal to rent cloud capacity from Lambda, a neocloud Nvidia holds a stake in, and Lambda in turn uses capacity from Hut 8, another neocloud whose entire capacity Nvidia had leased and may have largely sublet. In the $500bn Wall Street plan, Nvidia will put up no cash and take on no debt itself, but will offer up to a quarter of a deal's value in "residual-value support" against hardware depreciation; Huang describes the goal as making compute "an investable asset class". All of it rests on two assumptions: that Nvidia's chips hold their value and that compute demand keeps growing fast. Huang calls the chips "fungible" and durable enough to serve as loan collateral; Burry has argued cloud providers inflate profits by depreciating chips over five or six years instead of two or three. So far demand for older chips has held up: CoreWeave signed a contract in August covering Nvidia's 2020-vintage A100 chips through 2029, and SemiAnalysis estimates a one-year H100 rental costs about $2.80 an hour, only about a tenth below its early-2023 launch price. Analysts caution that both durability and value could simply reflect scarcity, since older chips are increasingly used for the less demanding task of inference rather than training. Meanwhile Nvidia keeps releasing new chips meant to supersede the old ones, and as hyperscalers and labs roll out their own inference-focused chips, competition could compress Nvidia's roughly 75% gross margin, well above AMD's roughly 55%. The article concludes that the biggest risk to the whole structure is a shortfall in AI demand itself, which would strain both Nvidia's sales and the guarantees it has extended across the industry.

Key facts

  • Nvidia backed a $105bn Ohio data-centre project (SB Energy/OpenAI) and a $500bn AI-infrastructure plan with six Wall Street firms, potentially underwriting up to a quarter of some deals' cost.
  • Over the past three years Nvidia has pledged over $70bn to startups and offered $300bn in financial support to customers, exposing it to theoretical costs of nearly $300bn.
  • Hyperscalers, roughly half of Nvidia's revenue, are building custom chips costing a fifth to a third as much as Nvidia's, projected to reach about 50% of the AI-processor market by decade's end, up from 40% now.
  • Nvidia took equity stakes in about 90 startups last year and 60-odd more this year, including a $6bn Poolside licence plus $1bn stake and a $12.9bn Hugging Face acquisition.
  • Critics including Michael Burry compare the arrangements to the 1990s dotcom lending boom that later hurt Cisco and Lucent when telecom demand fell short.

Why it matters

Nvidia's chip sales alone no longer explain its growth or the pace of AI infrastructure spending. By backstopping loans, guaranteeing hardware values and taking equity in its own customers, Nvidia has become a financier for large parts of the industry it supplies, which means the health of the AI buildout is now partly a reflection of Nvidia's own balance sheet and risk appetite rather than pure end-customer demand.

Who it affects

Hyperscalers (Amazon, Google, Meta, Microsoft) that are Nvidia's biggest customers but also its emerging chip rivals; neoclouds such as CoreWeave, Sharon AI, Firmus, Lambda and Hut 8 that depend on Nvidia backstops to finance data centres; AI labs including OpenAI and Anthropic that rely on this financed capacity; startups like Poolside and Hugging Face that Nvidia has bought stakes in or acquired outright; and investors evaluating whether Nvidia's reported margins and revenue reflect organic demand or vendor financing.

How to use it

There is no product or price here; the piece is a financial explainer. Investors and operators can use it to check exposure: how much of a company's AI-related revenue or capacity commitment traces back to Nvidia financing, equity, or guarantees, since that changes how independent the demand signal actually is.

How solid is it

The reporting draws on named, on-record figures (Nvidia's own disclosed deals, Bloomberg Intelligence forecasts, SemiAnalysis pricing estimates, Alphabet's public bond terms) and named analysts including Jay Goldberg, Michael Burry, Dan Nishball and Tim Davis. Some elements are explicitly reported rather than confirmed, notably Anthropic's $35bn Lambda deal ("reportedly signed"), and the article does not state how much of Nvidia's $300bn in offered support has actually been drawn.

Risks and caveats

The arrangements assume Nvidia chips keep their value and that AI compute demand keeps growing quickly; neither is guaranteed. Michael Burry argues cloud providers depreciate chips too slowly, overstating profits, and both the current resale value and long service life of older Nvidia chips may simply reflect a supply shortage rather than durable demand. If AI-chip demand grows more slowly than expected while supply and price competition (including from hyperscalers' own custom chips) increase, Nvidia's guarantees and equity stakes could turn from a growth engine into a source of real losses, a dynamic the piece compares to Cisco and Lucent's vendor-financing losses in the dotcom bust.

“Nvidia is walking a fine line between "enabling demand" and "creating it"”

— Jay Goldberg, Seaport Research Partners