A nineteenth-century pattern, rebuilt for silicon
In July 2026 Nvidia announced a new line of business. It carried no fancy technical name, but its nature was clear: exchange compute for profit. AI startups no longer have to pay cash up front for GPUs. Nvidia will accept a share of their future revenue or profit in return for the scarce chips they need right now.
Two partners are already in place. Australia’s Sharon AI will deploy up to 40,000 GPUs. Singapore’s Firmus Technologies is building a 360-megawatt data centre in Batam, Indonesia, expected to hold 170,000 GPUs. Those 210,000 GPUs are the first collateral behind Nvidia’s compute loan. For the first time in semiconductor-industry history, a GPU has become a hard currency that can be mortgaged, amortised and repaid with future earnings.
The customer cannot pay, and that is the point
In the past a chip was a commodity, cash on delivery, the deal closed, both sides square. Nvidia discovered its customers could no longer afford what it builds. AI startups queue to train models but cannot buy GPUs, spot prices are bid to the sky, and venture money thins through the rate-hike cycle. The shop is full of customers who cannot open their wallets.
The traditional fixes are two: cut the price, or wait for the customer’s payday. Nvidia chose a third road and turned the chip into a financial product. It now accepts a customer’s future wages to settle today’s bill.
This resembles the subprime form but differs sharply in risk. Subprime lent money to people who could not repay a mortgage, and the bank lost when prices fell. The compute loan’s collateral is an AI company’s future profit, earned from model training, inference and enterprise subscriptions. Nvidia is betting some of these companies become the next OpenAI or Anthropic, and with every GPU it lends it pre-emptively locks in a cut of those future winners.
Why be so generous with credit
The first reason is that the customers are genuinely out of money. For two years global AI-infrastructure investment has swollen exponentially while application-layer revenue lags far behind. OpenAI and Anthropic remain deeply loss-making, compute lessors lose money, and cloud vendors only reach break-even by amortising huge capital spending across five or six years. The industry’s cash is drained upstream, leaving a dry riverbed downstream.
Nvidia itself announced in June a plan to issue at least USD 20 billion of debt. It wants to sell chips to the downstream, but the downstream’s cash has already been absorbed by the supplier. Like a winery that buys up every grape for miles, bottles it at a premium, then finds the villagers cannot afford the wine and offers an instalment plan secured against next year’s crop.
The second reason is sharpening competition. AMD is closing in, Intel is pivoting, Google’s TPU eats into inference, and Amazon’s Trainium reduces GPU dependence. The moment an alternative exists, even a slightly weaker one, price decides. Nvidia must carpet every data centre on earth with its GPUs before substitutes mature. The compute loan is the tool to buy time and share.
The hidden motive sits in the revenue structure
Nvidia chose profit-sharing over simple instalments because it sees through the hardware ceiling. A chip is a one-time sale. But turn that chip into a share of a mint and the mint pays tribute on every note. This is a financial-engineering move from hardware company to infrastructure platform, disguised as a customer-care programme.
Nvidia is no longer just the pick-seller. It is taking an equity slice of the gold mine, sharing in every ounce extracted. The business model stops being hardware and becomes finance. It no longer merely sells chips. It exchanges them for the customer’s future cash flow, turning its balance sheet from inventory into equity. And the elegant part: when those AI startups finally turn profitable, the first party they owe is not their angel investor but their hardware vendor.
The deeper shift in production relations
The most far-reaching effect is that it quietly changes the production relations of the AI industry. Compute used to be a means of production a firm owned or leased. The compute loan makes it a resource exchanged for profit share, a kind of equity-like input. Future AI startups no longer need to raise, then buy GPUs, then train. They can pay today’s compute cost directly with tomorrow’s profit.
It sounds like a financing democracy, but the ending may read more like a dystopia. Nvidia is becoming a super-hub that holds both the allocation of compute and the profit claim on future AI winners. The firms that accept the loan have already handed the first line of their profit statement to Nvidia before their model works. Nvidia need not guess who wins. It only needs to ensure that whoever wins, the winner owes it.
Every gold rush leaves behind not the first to find gold but those who traded denim for gold dust, shovels for shares, and boat tickets for deeds. They never panned. They only shared in the fate of the panners. When the last flake is gone and the panners leave, they fold the canvas, put away the tools, and open a bank on the empty claim. Nvidia is becoming the fate of every panner.
Editor’s note: This is an adapted translation of the original OFweek Robotics report. It has been trimmed and restructured for readability for an international business audience.