Is AI Compute Becoming a New Asset Class? Testing Jensen Huang's Claim

The short answer

Jensen Huang argues AI compute is becoming a new, financeable asset class because data center systems earn steady, predictable revenue and, in his words, are "productive," "long-lived," "fungible," and "flexible" instead of quickly obsolete. The evidence supports part of that: revenue is predictable when it is under long-term contract with strong customers, and lenders are already financing it. But hardware depreciation is unproven, and what lenders are really financing is the contract, not the chip.

Accounting lives for servers run about five to six years, banks reportedly underwrite three to four, and CoreWeave's contract-backed loans cost SOFR plus 4% or more while its unsecured notes recently yielded over 15%. Calling it a new asset class is premature, but the financing structure is real.

Hey everyone,

Nvidia CEO Jensen Huang has been making an unusual argument to Wall Street: that AI computing capacity is becoming a brand-new asset class, something lenders can finance the way they finance aircraft or railroads, because data center systems produce steady, predictable revenue and don't lose value like ordinary tech hardware. He said it again this week on CNBC at GTC Berlin. It's a big claim, and the timing is interesting, because this same week the companies that rent out Nvidia's chips saw their stocks and bonds fall on debt worries.

This post tests the claim piece by piece. I'm leaning on primary documents wherever I can: Nvidia's own SEC filing and press release, annual reports from CoreWeave and four of the biggest tech companies, and reporting from Reuters, TechCrunch, and others. My short version: Huang is partly right, and the part he's right about is not quite the part he emphasizes.

What exactly is Jensen Huang claiming?

On August 10, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build financing platforms that aim to mobilize more than $500 billion of third-party capital for AI infrastructure. Nvidia's release says the partnerships "remain subject to execution of the final agreements," so it's a framework, not a funded pool. In a CNBC interview the same day, Huang said:

"This is really the first time that technology chips have become an investable asset class."
Jensen Huang, CNBC, August 10, 2026, as reported by Yahoo Finance

He described the hardware as "productive," "long-lived," "fungible," and "flexible." Nvidia's release puts the revenue point in four words, "In AI, compute is revenue," and Apollo's Jim Zelter said "modern compute has emerged as a scarce, mission-critical asset class." On October 22, Huang told CNBC that Nvidia is trying to make its compute a "financeable asset," that demand at GPU-rental companies called neoclouds is "off the charts," and that "what is missing in the middle is the working capital to fuel that growth."

TechCrunch adds that Huang argues AI servers are closer to railroads or airlines, which last a long time, than to quickly obsolete PCs, and that a data center's compute can be reused by "another customer, another cloud or another operator." Strip it down and there are three testable claims:

The revenue is steady and predictable.
The hardware doesn't lose value quickly.
Therefore it can be financed, at good rates, like infrastructure.

What makes something an "asset class" in the first place?

An asset class isn't just something you can borrow against. Think of aircraft, railcars, or shipping containers. They generate contracted cash flows, they're standardized enough to be priced and swapped, there's a deep secondary market to sell them into, residual values are measurable from past data, and there's a track record through at least one downturn. That last part matters most. Lenders and rating agencies are comfortable with aircraft leasing because there are decades of data on what planes sell for.

AI compute has some of these traits and lacks others. That's the heart of the debate.

Claim 1: Is the revenue steady and predictable?

For capacity under long-term contract, yes, and the evidence is strong. CoreWeave, the largest independent GPU cloud, reported in its 2025 annual report that it had $60.7 billion of remaining performance obligations, up from $15.1 billion a year earlier, with committed contracts averaging about five years. It borrows against that through delayed-draw term loans "collateralized with the assets underlying the contributed contracts and the pledged contractual cash flows, generally from investment grade counterparties," which amortize as contracted cash flows arrive "in a regular and predictable manner." Reuters reports that Moody's rated one $8.5 billion CoreWeave facility A3, largely because of Meta's contractual payments.

You can see the other side of those contracts in our breakdown of Anthropic's S-1: Anthropic committed to spend at least $110 billion with Amazon and about $114.5 billion with Google regardless of actual usage, on non-cancelable terms. That's the kind of take-or-pay arrangement that makes revenue predictable.

CoreWeave's growth, 2024 to 2025 (multiple of the prior year)

Source: CoreWeave 2025 annual report (Form 10-K, filed March 2, 2026). Multiples are our arithmetic from the reported figures. Revenue for 2024 is the rounded $1.9 billion the company reports.

The catch: the predictability belongs to the contract and its counterparty, not to the chip. CoreWeave's net interest expense grew faster than its revenue (3.4 times versus 2.7 times), and its net loss in the 2025 report was $1.2 billion. Capacity sold on demand, or to customers with weaker credit, is less predictable. My verdict on claim 1: valid for contracted capacity with strong customers.

Claim 2: Does AI hardware really hold its value?

This is the weakest link, and it's where the real disagreement is. Companies decide how many years to spread a server's cost over, called its useful life, and the filings show a range:

Server useful lives assumed in company filings and by lenders (years)

Sources: SEC annual reports of Alphabet (FY2025), Microsoft (FY2026), Meta (FY2025), Amazon (FY2025), and CoreWeave (FY2025). Bank underwriting life per Impax's Tony Trzcinka in Reuters. The 9-to-10-year figure is an estimate by valuation firm Barkr for Nvidia's GB300 NVL72 systems, cited by Nvidia in Reuters reporting. Accounting lives are management estimates and are not the same as resale value.

Several things stand out.

The accounting is a judgment call, and it moves earnings. Meta lengthened most server lives to 5.5 years effective January 2025, which, per its filing, cut 2025 depreciation by $2.92 billion and added $2.59 billion, or $1.00 a share, to net income. Amazon went the other way: it shortened a subset of servers from six years to five, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning."
Critics say the real lives are shorter. Investor Michael Burry argued in November 2025 that hyperscalers understate depreciation, estimating about $176 billion across 2026 to 2028, because chips bought on a two-to-three-year product cycle are being depreciated over five to six years. Nvidia disputed this in a memo to analysts, saying customers depreciate GPUs over four to six years and that older chips like the A100 still run at high utilization.
Lenders are more cautious than the accounting. Reuters reports that banks typically underwrite GPUs over a three- to four-year schedule, and that S&P Global Ratings says GPUs have proven useful beyond five years but that it takes a conservative view of chip value. Sources told Reuters there isn't enough history to confidently underwrite long-term residual value.
Rental prices for older chips have held up so far. Silicon Data's H100 rental index rose about 10% in four weeks to $2.20 an hour in January 2026 even though the H100 is a 2022-era design. That's one encouraging data point, from a market where supply has been tight.
Demand for a specific chip can vanish. Even Nvidia took a $0.4 billion charge in the first half of its fiscal 2027 for excess H200 inventory and purchase obligations as demand for those products diminished, according to its latest quarterly report.

The key distinction, as SiliconANGLE put it: "Functional life is not the same as economic residual value." A GPU can keep working for ten years while being worth very little. My verdict on claim 2: unproven. The experience so far with older chips is encouraging, but it comes from a supply-constrained boom, not a downturn, and nobody has yet watched a large pool of AI chips try to sell into a weak market.

Claim 3: Can it be financed like infrastructure, and at what price?

It is being financed, so in that sense the answer is already yes. CoreWeave had $9.7 billion outstanding under its delayed-draw loans and $21.6 billion of total debt at the end of 2025. But the price tells you how lenders see the risk:

Contract-backed loans (2025)

SOFR + 4.0-4.25%

Unsecured notes (2025)

9.0-9.25%

Yield on June 2026 notes, Oct.

15%+

5-yr credit default swap

1,115 bp

The first two figures come from CoreWeave's annual report. The last two come from Yahoo Finance's reporting this week: CoreWeave's unsecured notes issued in June fell below 80 cents on the dollar, pushing the yield above 15%, its five-year credit default swaps reached 1,115 basis points, and its stock closed on October 23 at $54.24, down about 38% from September 30. Musk added to the pressure on October 25 by writing that the 200-plus neoclouds will see "guppy-level death and consolidation." Yahoo also reports that Gartner said on October 14 that more than 80% of neocloud providers may stop operating independently by 2030. CoreWeave's CEO argues that its roughly $104 billion contracted backlog more than covers its debt.

Two lessons here. First, financeable does not mean cheap. These are priced like high-yield credit, with the best terms reserved for the strongest contracts. Second, what lenders are really financing is the contract. Reuters reports that Moody's A3 rating on CoreWeave's $8.5 billion facility rested largely on Meta's payments, that Broadcom backstopped more than 80% of a $35 billion financing structure tied to Anthropic, and that three banking sources said Nvidia may need to guarantee all of its deals or back them with revenue from investment-grade customers.

What is Nvidia's own role, and is it circular?

Here's the awkward part for the "independent asset class" story. Nvidia sells the chips, finances the buyers, and is itself a customer of those buyers. Its latest quarterly report (the quarter ended July 26, 2026) shows revenue of $96.2 billion, up 106%, and discloses that Nvidia:

commits to buy cloud services from its AI-cloud customers, typically for six years, totaling about $36 billion in the filing's discussion (the commitments table shows $29 billion), while noting those clouds "can unilaterally stop providing" the service to Nvidia and sell it to others at better rates;
has guaranteed some AI clouds' data center lease obligations, with $3.5 billion of notional exposure and $712 million placed in escrow by the partners;
has $25 billion of data center leases not yet commenced and $25 billion of equity investment commitments; and
holds warrants in public companies with a fair value of $824 million.

On the new $500 billion plan, reporting says Nvidia may cover up to 25% of the shortfall if collateral chips fall below expected value and a lender has to sell them. SiliconANGLE puts that at a potential $125 billion. Huang says the support is limited and designed to complement independent underwriting, and Nvidia's release does not itself describe a 25% backstop, so treat that figure as reported, not confirmed. TechCrunch flags a "wrong-way risk": Nvidia's guarantee gets more costly exactly when demand weakens and its sales are under pressure. It also notes the Lucent comparison, a company that lent customers money to buy its equipment before the dot-com crash, is "not entirely unfair." Huang has said the plan brings in independent institutional capital and isn't circular financing. As SiliconANGLE puts it, "independent capital does not create independent demand."

So how valid is the claim? A scorecard

Huang's three claims against the evidence
Claim	Evidence for	Evidence against	My verdict
Revenue is steady and predictable	Multi-year take-or-pay contracts; $60.7B backlog at CoreWeave; A3-rated, contract-backed loan	Depends on a few big customers; on-demand and weaker-credit demand is less stable	Valid for contracted capacity
Hardware doesn't depreciate quickly	5 to 6 year accounting lives; H100 rents held up in 2026; A100s still used	Lenders use 3 to 4 years; annual chip cadence; no downturn test; some demand vanished (H200)	Unproven
So it can be financed like infrastructure	Billions already lent; big asset managers signed on	Priced like high-yield credit; this week's bond and stock sell-off; guarantees needed	Financeable, but not cheap or proven

My overall read: what's being financed is the contract, not the chip. Calling it a new asset class is premature, because asset classes are defined by track records and deep secondary markets that don't exist yet. But "a financeable stream of contracted cash flows with a depreciating asset as secondary collateral" is a real and legitimate structure. That's closer to how equipment leasing works than to a revolution.

What could it mean for the future of finance?

If Huang is right, or partly right, here's what could change:

New credit products. Compute-backed loans, leases, and eventually securitizations, which package many loans into bonds, could become a large corner of private credit and infrastructure investing, with pension funds and insurers as buyers.
A pricing layer for compute. Indexes like Silicon Data's already track GPU rental prices. If they become trusted benchmarks, they could support hedging and residual-value markets, the way indexes support other commodities.
Lower cost of capital, faster buildout. Huang's point is that capital, not demand, is the bottleneck. Cheaper financing would speed the data center buildout we described in our post on data centers and the economy.
The vendor as banker. Chip makers and cloud companies are becoming financiers of their own customers, which blurs the line between suppliers, lenders, and buyers.

And if he's wrong, history gives a template. In the late 1990s, a San Francisco Fed study found that telecom service providers' capital spending rose from about $47 billion in 1995 to a peak of about $121 billion in 2000, driven partly by expectations that proved too high. Capital spending by the new competitive carriers then fell by nearly 80% by 2002. The IMF's October Global Financial Stability Report flagged AI-linked leverage and circular financing among the risks that could amplify a correction, as we noted in that post. The difference this time is that much of the demand comes from large, well-capitalized companies, but the weaker links, the smaller neoclouds, are exactly where this week's stress showed up.

The most likely outcome, in my view, is somewhere in between: financing follows contract quality. Strong contracts with strong customers get cheap, long-term money, perhaps with short amortization so lenders don't depend on resale value. Weaker ones pay high-yield prices or can't borrow. Over time, if enough data builds up on what used chips actually sell for, a true secondary market could emerge. Until then, who bears obsolescence risk, the lender, the lessor, Nvidia, or the customer, is the central question in every deal.

What does this mean for investors?

Read the contract, not the pitch. The quality of the customer and the length and terms of the contract matter far more than the chip model.
Watch the cost of capital. Bond yields, credit default swaps, and the spreads on new financings are real-time evidence of how lenders view this asset class.
Be wary of circularity. When a supplier funds its customers, strong demand can be partly manufactured by the financing. Check who's guaranteeing what.
Depreciation assumptions affect earnings. A longer assumed life boosts reported profit, as Meta's change showed. Compare companies' assumptions.
Diversify. A concentrated bet on one theme or one customer is the main risk if the thesis fails.

For the broader framework, see our P/E post on why standard metrics break down for capital-hungry companies.

Key takeaways

Huang's claim has three parts: steady revenue, slow depreciation, and therefore financeability.
Revenue is predictable when it's contracted, and the contract is what lenders really lean on.
Depreciation is the unproven part. Lenders assume three to four years, the companies' accounting assumes five to six, and Nvidia cites nine to ten.
Financeable doesn't mean cheap. This week's bond and stock sell-off shows how quickly the price of that financing can change.
"New asset class" is premature, but the financing structure is real and could reshape private credit if it holds up through a downturn.

More soon,
Learn to Love Money

Frequently asked questions

What is Jensen Huang's claim about AI compute as an asset class?

Huang argues that AI computing capacity is becoming a new, financeable infrastructure asset class because data center systems generate steady, predictable revenue and are productive, long-lived, fungible, and flexible, unlike rapidly depreciating tech components. He said on August 10, 2026 that it is the first time technology chips have become an investable asset class, and Nvidia announced memorandums of understanding with six financial firms to mobilize more than $500 billion of third-party capital.

Is AI compute revenue really steady and predictable?

It is when the capacity is under long-term contract with strong customers. CoreWeave reported $60.7 billion of remaining performance obligations at the end of 2025 with contracts averaging about five years, and Reuters reports Moody's rated one $8.5 billion CoreWeave facility A3 largely because of Meta's contractual payments. Capacity sold on demand or to weaker-credit customers is less predictable.

How fast do GPUs lose value, and how long do they last?

It is unsettled. Company filings use useful lives of about five to six years for servers, Reuters reports banks typically underwrite GPUs over three to four years, and Nvidia cites an estimate of nine to ten years for its newest systems. Functional life is not the same as economic resale value, and there is no full-cycle track record of what large pools of used AI chips sell for.

What is Nvidia's $500 billion financing plan?

On August 10, 2026, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to build financing platforms that aim to mobilize more than $500 billion of third-party capital for AI infrastructure. The partnerships remain subject to final agreements, and reporting says Nvidia may cover up to 25% of a shortfall if collateral chips are worth less than expected.

How expensive is it to finance AI compute?

It depends on contract quality. In 2025 CoreWeave priced contract-backed delayed-draw loans at SOFR plus 4.00% to 4.25% and unsecured notes at 9.00% to 9.25%. In October 2026, Yahoo Finance reported its unsecured notes issued in June traded below 80 cents on the dollar, with yields above 15%.

Is Nvidia's financing of AI companies circular?

Critics say it can be. Nvidia's latest quarterly report shows it commits to buy cloud services from AI-cloud customers, about $36 billion in the filing's discussion, has guaranteed some of their data center leases with $3.5 billion of notional exposure, and has other lease and equity commitments. Huang says the plan brings in independent institutional capital and is not circular financing, while critics note independent lenders do not create independent demand.

What could AI compute as an asset class mean for finance?

If it works, it could create new compute-backed loans, leases, and securitizations, price benchmarks for GPU rentals, and cheaper capital for the AI buildout, with chip makers acting as financiers. If it fails, the late-1990s telecom boom is the cautionary template, when capital spending peaked around 2000 and then fell sharply. The most likely outcome is that financing follows contract quality.

NOT FINANCIAL ADVICE. This post is an opinion and analysis piece for education only and is not a recommendation to buy or sell any security, including Nvidia, CoreWeave, or any debt. Market data is as of late October 2026 and changes quickly. Executive statements are reported as published by the cited outlets, and the Nvidia financing plan is a framework subject to final agreements. Do your own research or talk to a licensed advisor before investing.