The Circular Economy of AI Capex

On 10 August 2026, Nvidia put out a press release with one of those headlines that makes your eyes glaze over.

Strip away the jargon and it says something simple: Nvidia has convinced six of the world's biggest investment firms - Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR - to set up funds that will buy GPUs and rent them out. The target is over $500 billion.

Most of the coverage argued about whether this proves AI is a bubble. I think that's the wrong argument.

The real story is about who ends up holding the risk - and the answer might surprise you, because it's probably closer to your retirement account than you'd expect.

First, what actually happened

Think about how a taxi company grows.

Option one: it saves up profits and buys cars. Slow, but safe.

Option two: someone else buys the cars, and the taxi company rents them and pays out of the fares it earns. The taxi company grows much faster. The car owner earns steady rent.

That's the deal. Nvidia makes the chips that AI runs on. The companies that want those chips - AI labs, cloud providers, even governments - often can't afford to buy them outright. Now six enormous investment firms will buy them instead and rent the computing power out.

Nvidia sells more chips. The AI companies get capacity they couldn't afford. The investment firms collect rent.

One important detail: these are memorandums of understanding, not signed contracts. The press release says so in plain print. The $500 billion is a target, to be reached "over time," with no deadline. Treat it as an ambition, not a cheque.

Why Nvidia needed someone else to pay

It would be neat to say Nvidia fixed its supply problem and now has a demand problem. That isn't what happened, and the truth is more interesting.

Chips and capacity are still tight. Amazon's CEO told investors he expects to be capacity-constrained into 2027, and is already planning around demand in 2028. Nvidia does not have spare stock sitting in a warehouse.

What changed is how the bill gets paid - and there are at least four separate forces pushing in the same direction.

1. Spending has overtaken earning. The five biggest spenders are pouring money in faster than it comes through the door. Epoch AI's analysis of their SEC filings found capex growing about 70% a year while operating cash flow grows about 23% - lines that cross around the third quarter of 2026 (Epoch AI). Oracle crossed first. Alphabet posted its first negative free cash flow since going public in 2004, burning $5.9bn in one quarter. Microsoft is the only one of the big four still comfortably positive.

2. They can still borrow - it just costs them something. This is the bit I got wrong first time round. These companies absolutely can raise money, and have: borrowing rose from 9% of capex in 2024 to 32% by mid-2026, and Alphabet raised almost $85bn of fresh equity in June 2026 (FactSet). The cost isn't access, it's consequences - credit ratings, shareholder patience, and buybacks that collapsed to their lowest level since 2018.

3. Renting keeps debt off the books. If you borrow to buy computers, the debt shows up on your balance sheet and the rating agencies notice. If someone else buys them and you pay rent, largely it doesn't. That's a big part of the appeal, and it isn't new - Moody's has pointed out that lease commitments already sitting off hyperscaler balance sheets equal 113% of their reported debt.

4. The people who most need this aren't the giants at all. AI labs, smaller cloud providers and national governments want enormous amounts of computing power and have nothing like Microsoft's balance sheet. Nvidia had been helping fund them out of its own pocket - roughly $40bn of AI equity commitments in the first five months of 2026 alone. Handing that job to six professional investors lets Nvidia grow beyond a handful of mega-customers without carrying the risk itself.

So the money has to come from somewhere else. And there's a huge pool of it sitting around: pension funds, insurance companies, sovereign wealth funds. These groups hold trillions and desperately need somewhere to put it that pays a reliable income for years.

Nvidia's pitch is that GPUs are exactly that - a reliable income-producing asset, like a toll bridge or a wind farm.

That pitch is the part worth examining.

The catch: fast machines, slow money

Here's the thing about toll bridges. They last 50 years. You can borrow money for 30 years against one, because you're confident it'll still be standing and still collecting tolls.

AI chips are not like that.

Nvidia releases a significantly better chip roughly every year. Each new generation is dramatically faster and cheaper to run. So the chip you paid $40,000 for in 2026 is competing against something far better by 2028.

How long does an AI chip actually stay useful? Nobody agrees. And the disagreement is worth billions.

• Google, Microsoft and Oracle tell their accountants those chips last up to six years.
• Amazon looked at the same technology in 2025 and shortened its estimate, taking a $700 million hit to profits. Meta looked at it and went the other way.
• Investor Michael Burry - the man from The Big Short - argues the real answer is two to three years, and that the industry is overstating profits by around $176 billion between 2026 and 2028 by pretending otherwise.

Nobody actually knows, because the AI boom isn't old enough for anyone to have watched a chip reach the end of its life.

Now put those two facts together. You're lending money for 10, 15, maybe 20 years. The thing you're lending against might be worth very little in five. That gap is the entire trade.

Why this isn't simply "dot-com all over again"

Whenever a tech company helps its customers pay for its own products, people reach for the year 2000. It's a fair instinct, but the comparison needs one correction that most commentary misses.

Back then, telecom equipment makers like Lucent and Nortel lent money directly to their customers so those customers could buy their gear. When the customers went bust, the suppliers ate the losses. Lucent lost more than 90% of its value. Nortel effectively disappeared.

Nvidia has clearly studied that history, because it has done the opposite. It isn't lending anyone a cent. Independent funds put up the money. If borrowers fail, the losses land on those funds' investors, not on Nvidia.

That's a genuine improvement and the bears underrate it.

But notice what it does not fix. In 2000, the companies borrowing the money still went bust. Changing who writes the cheque doesn't change whether the underlying business works. It only changes who cries.

What happens after every building boom

We've done this before. Not with AI, but with railways, telephone lines, fibre-optic cable, shale oil. The pattern is remarkably consistent, and it isn't the one people expect.

Take fibre-optic cable around 2000. Companies borrowed roughly $90 billion to bury cable across America, convinced internet demand would soak it all up. Demand did come - just years later than the loan repayments were due.

The companies went bankrupt. Investors in Global Crossing and Level 3 were wiped out.

But here's the part people forget: the cable was still in the ground. Buyers picked it up for pennies and made fortunes over the following two decades. That fibre is carrying this article to you right now.

So the honest base case isn't "it all collapses" or "it all works." It's something in between and fairly predictable:

1. Money floods in while the story is exciting.
2. Returns get thinner as everyone crowds into the same trade.
3. Something triggers a reckoning - likely one chip generation that ages badly.
4. The people who lent at the peak take real losses.
5. The machines keep working, under new owners, and this quietly becomes a normal, boring asset class.

Every infrastructure boom punishes the first owner and rewards the second.

So who's actually holding the risk?

This is the question I'd want answered before anything else, and it's the one nobody seems to ask.

Work through it.

Nvidia? No. It sells the chips and walks away. It isn't lending.

The big tech companies? Less than before. Renting means the cost doesn't sit on their balance sheet - which is convenient timing, given that everyone is currently questioning how they account for chips they own.

The investment firms? Partly. But they mostly earn fees for managing other people's money. They do well whether the investments do or not.

So who? The investors in those funds. And a large share of that money comes from insurance companies and pension providers. Apollo owns the annuity business Athene. KKR owns Global Atlantic. These are firms that pay out retirement income, and they're hungry for investments that produce steady payments for decades.

Which means the risk that a 2026 AI chip is worth almost nothing in 2031 is drifting, quietly, toward ordinary people's retirement savings.

And here's my actual concern, stated plainly: this isn't a bubble in AI. It's a bubble in the word "infrastructure."

"Infrastructure" means bridges and power grids - things that last for generations and get a low interest rate because they're safe. GPUs are being sold using that word. But a bridge doesn't need to be demolished and rebuilt every five years to keep earning. These machines do. Calling them infrastructure doesn't make them behave like it.

What a normal investor should actually do

None of this means sell everything. AI demand is real and the chips are genuinely useful. It means being precise about what you own.

Sell the shovels, not the mines. In a gold rush, the person selling shovels does well whether or not anyone strikes gold. Here the shovel-sellers are Nvidia itself, the investment firms earning management fees, and the unglamorous power, cooling and electrical-equipment companies. They get paid regardless of how the AI bets turn out.

Be careful with the borrowers. The risky names are smaller AI cloud companies carrying heavy debt against short rental contracts. CoreWeave fell around 61% in six weeks in late 2025 when investors started asking exactly these questions. That's a preview, not a freak event.

Read the label on anything that says "infrastructure." If a fund or private-credit product offers you unusually high income from "digital infrastructure," find out what's inside. Renting out computers is a different risk from owning a power line, even when the brochure uses the same word. Higher yield is never free - it's payment for a risk someone has decided you should carry.

Keep some cash. If the reckoning comes, the machines don't stop working. Somebody buys them cheaply. Being able to be that buyer is worth more than being early.

What the used-chip market already tells us

I want to correct something I implied earlier. A market for second-hand AI chips does exist, and it's busier than most coverage suggests. Specialist resellers, IT-disposal firms and dedicated exchanges trade used data centre GPUs every day.

Here's roughly what a used H100 - Nvidia's 2023 flagship - holds onto, according to marketplace data compiled in mid-2026:

• 12 months · Share of original value retained: around 90%
• 24 months · Share of original value retained: around 80%
• 36 months · Share of original value retained: 50-65%

The realistic range at three years runs from about 30% if Blackwell chips ship quickly to around 70% if they don't (Mercatus). In cash terms, refurbished H100s were changing hands at roughly $18,000-$22,000 in 2026, and older A100s around $12,000-$18,000.

That's not nothing. It's a real asset with a real bid.

But look closer and two problems appear.

The prices are already volatile. One study of thousands of listings between June 2024 and December 2025 found used H100s trading as high as $50,000 in mid-2024 before falling sharply as supply arrived - while the official new price barely moved (Silicon Data). Most published figures are asking prices, not confirmed sales, which tends to flatter them.

And the market has only ever been tested in one direction. Every number above comes from a stretch where demand for computing power rose without a pause. Nobody has seen what a used AI chip fetches during a slump.

The one time we did see it

We have a precedent, and it's not a comfortable one. Bitcoin mining machines are the closest comparable asset: specialised computing hardware, often bought with borrowed money, whose entire value rests on one revenue stream.

That revenue stream turned in 2021-22. What happened to the machines:

• After China banned mining in mid-2021, second-hand rig prices fell around 40% in a single month (Hashrate Index).
• Through 2022 the decline kept going. The most efficient machines fell from a May 2021 peak of $119.25 per unit of computing power to $15.71 by December 2022 - down roughly 87%. Mid-tier machines fell 89%. The least efficient fell 91% (Hashrate Index data via FXStreet).
• Companies that had borrowed against those machines discovered their collateral had largely evaporated.

The machines still worked perfectly. They simply stopped being worth much, very fast, because the thing they earned money from stopped paying well.

Why aircraft are the wrong comfort blanket

When people argue that computers can be financed like infrastructure, they usually reach for aircraft leasing. It's a fair comparison in one way and misleading in another.

Aircraft leasing works because professional appraisers have more than forty years of sales data covering several booms and several busts. A lender can value a twelve-year-old jet within a few percentage points and be confident about it.

AI chips have about three years of pricing history, all of it gathered while demand only went up.

So the honest version of my earlier point is this: the market exists. The price history doesn't. And a residual value you've never watched fall is not the same as one you've measured.

Five things worth watching

You don't need a Bloomberg terminal. These will all show up in the financial press.

1. Do the deals actually get signed? Right now they're handshakes. Watch for real, sized agreements.
2. Does Nvidia end up guaranteeing anything? If it does, it has quietly become the lender after all - and that's the 2000 playbook.
3. Does another big tech company shorten its chip lifespan estimate? Amazon already did. A second one turns a debate into a trend.
4. Does a market appear for used AI chips? The market exists, as above - what it has never done is price these machines during a downturn. Watch for a large fleet being sold at a moment when demand is soft. That single event will tell you more than any forecast.
5. Do the returns on offer start shrinking? When a risky thing starts paying like a safe thing, the market has stopped paying attention.

Sources

• Nvidia press release, 10 August 2026 - the announcement itself, including the note that these are memorandums of understanding subject to final agreements.
• Epoch AI, Hyperscaler capex vs cash flow (16 June 2026) - capex growing ~70% a year against ~23% for operating cash flow, crossing around Q3 2026. Built from SEC filings.
• FactSet, Hyperscalers Tap External Financing (23 July 2026) - borrowing rising from 9% to 32% of capex, Alphabet's $84.75bn equity raise.
• TMT Finance, 2026 hyperscaler capex analysis (18 August 2026) - Alphabet's first negative free cash flow since 2004, Microsoft still positive, Amazon capacity-constrained into 2027.
• CNBC, How long before a GPU depreciates? - the six-year vs two-to-three-year argument, and Michael Burry's $176bn estimate.
• Let's Data Science summary of the Moody's and Nvidia figures (18 June 2026) - off-balance-sheet lease commitments at 113% of reported debt; Nvidia's ~$40bn of AI equity commitments in early 2026.
• Barclays research on data centre securitisation, via the Structured Finance Association (23 July 2026) - growth of the market that funds this kind of deal.
• Mercatus, H100 depreciation (May 2026) and H100 server price (July 2026) - retained-value curve at 12, 24 and 36 months, and refurbished price ranges.
• Silicon Data, H100 market value trends (January 2026) - analysis of thousands of listings showing secondary prices falling while list prices held flat.
• Servnet UK, GPU residual value study 2026 - cross-checked resale ranges, and the caution that most published figures are listings rather than confirmed sales.
• Hashrate Index on the 2021 mining-rig crash and ASIC prices at multi-year lows, December 2022 - the closest historical test of specialised computing hardware losing its revenue stream.

The fibre-optic and telecom comparisons are drawn from the widely documented history of the 2000-2003 period rather than a single source.

This is commentary, not financial advice. I don't know your situation, and you should talk to someone who does.