Debt has a due date; demand doesn't

On 10 August 2026 Nvidia signed memoranda with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to raise what its press release calls “independent compute financing platforms”, meant to mobilise more than five hundred billion dollars of third-party capital ^[1]. The declared purpose fits in a line: to create dedicated pools of capital “at significant scale at attractive rates” for Nvidia's customers ^[1]. Jensen Huang summed it up in a sentence worth reading twice. “We began by building chips; today, we are helping create a new class of productive, investable infrastructure: AI factories” ^[1].

The larger phenomenon of which this agreement is the best available example is not the size of the figure. What decides where intelligence gets built stopped being what a country can offer and became what a creditor can repossess. Electricity, land and tax exemptions are things a state grants. The resale value of a hundred thousand graphics cards is granted by nobody. Latin America has spent two years competing with the wrong instrument.

The guarantee is not in the press release

The 10 August text never once uses the words guarantee, backstop or residual value ^[1]. The missing piece was published the next day, signed by Huang himself: “In some cases, NVIDIA may provide a residual-value support mechanism for up to 25 % of an opportunity, assessed carefully on a project-by-project basis” ^[2]. The rest of that text builds the commercial argument. An AI factory deserves to be financed as infrastructure because it “produces revenue, serves a broad market, improves in performance over time and can be redeployed” ^[2]. The support, the same text clarifies, complements rather than replaces each lender's own analysis, which weighs the customer, demand, utilisation, cash flow and the residual value of the collateral for itself ^[2].

The centrepiece of the reasoning is fungibility. “When needs change, the factory can be used by another customer, another cloud or another operator”, Huang writes, and from there deduces “a deep market of potential users and offtakers” that protects residual value ^[2]. The example he picks is the A100, still in commercial use six years later and with an economic life that he argues stretches “toward a decade” ^[2]. Not everyone buys the argument whole. David Sacks, of the White House science and technology advisory council, described the plan's biggest risk with a phrase borrowed from the telecoms collapse: a glut of dark GPUs, machines installed that nobody switches on because capacity grew faster than demand ^[3].

The model already had two contracts before it had a name

Sharon AI announced on 12 June a six-year collaboration with Nvidia covering seventy-two megawatts in Australia, with up to forty thousand GB300s, under a revenue-sharing and credit-support arrangement ^[4]. Firmus signed its own weeks later, and that is the one that matters for this region. The company will raise a three-hundred-and-sixty-megawatt campus in Batam, Indonesia, with up to a hundred and seventy thousand Nvidia accelerators, in a relationship running to 2034, under the same “revenue-sharing and credit-support model” ^[5]. Behind it stand buyers committed to between twenty-five and thirty billion dollars across the first six years ^[5]. The support has its consideration, because Nvidia collects the price of the equipment and also takes a share of the cloud revenue the supported capacity produces ^[5].

Indonesia did not offer an exceptional tax regime to win this. It got the machines because somebody put a floor under the value of the collateral and because demand was contracted in writing. On 17 August came the large deal, with Nvidia backing up to a hundred and five billion dollars of the financing for OpenAI's Ohio campus ^[6]. The figure comes from a securities filing and arrives cut down: earlier discussions had run to two hundred and fifty billion for ten gigawatts, and what is committed covers 4.25 with an option on 3.75 more ^[6]. A day later, CNBC summarised the shift in a headline the industry repeated all week: Nvidia's moat has moved from chips to capital ^[7].

What Argentina offered

In October 2025, a fortnight before the midterm elections, OpenAI signed a letter of intent to raise a twenty-five-billion-dollar, five-hundred-megawatt data centre in Patagonia ^[8]. The project was announced under RIGI, the most generous investment-incentive regime in the region, carrying thirty years of tax and customs benefits ^[8]. Seven months later nothing had happened. The Ministry of Economy said so without hedging: “Stargate has not been presented. We do not even include it in the ‘announcements' because there has really been no mention of timescale, sums or localities” ^[8]. RIGI's own observatory had no reliable information beyond the campaign event ^[8].

The fact that orders the comparison came from Emiliano Kargieman, co-founder of Sur Energy: the money to build “will not come from OpenAI but from investment funds”, and what OpenAI signed was a commitment to buy the capacity ^[8]. It is the same ingredient that holds Batam up. The difference lies in who answers when the buyer fails, and nobody put that signature down in Patagonia. Argentina today adds up to thirteen data centres and thirty-two installed megawatts; the announcement alone promised five hundred ^[9].

What the guarantee's worth depends on

The whole building rests on one supposition: that the machines will be worth, the day they have to be sold, something close to what the loan assumed. That supposition is in open dispute. The largest buyers stretched the accounting life of their servers to six years, while the real replacement cycle runs at two or three ^[10]. Michael Burry calculates that depreciation declared that way falls short by some hundred and seventy-six billion dollars between 2026 and 2028 ^[10]. A used H100 already sells below half the price of a new one by year three ^[10].

The Bank for International Settlements put numbers on the underlying shift in January ^[11]. The firms driving the wave used to pay for their investments out of the cash they generated and now borrow, “with private credit playing a rapidly increasing role” ^[11]. Loans of that kind to AI companies went from near zero to more than two hundred billion dollars, and the BIS projects three to six hundred billion by 2030 ^[11]. The figure that disarms the calm is a price comparison. Lenders charge an AI company almost the same as they charge anyone else, while the stock market pays for those same companies as though they were going to return extraordinarily ^[11]. Both cannot be true at once, and the BIS says it in as many words: either lenders are underestimating the risk or shareholders are overestimating the earnings ^[11].

The size of the commitment is drawn better by a figure from Morgan Stanley. Global data-centre investment through 2028 adds up to 2.9 trillion dollars, of which the largest companies pay 1.4 out of their own cash ^[12]. The rest — a trillion and a half — has to come from outside, and the bank projects some eight hundred billion of it flowing from private credit, whose end investors are insurance companies and pension funds ^[12]. Nvidia's plan, with its half trillion, is a part of that gap and not the gap itself.

That a software advance can move this price is not hypothetical either. On 27 January 2025, on the sole claim that a model had been trained for 5.6 million dollars, Nvidia lost five hundred and eighty-nine billion in a single session, the largest one-day fall in market history ^[13].

Serving a model is a memory problem. Generating each word, the machine spends most of its time moving a cache of intermediate state between memory and processor, and the arithmetic it performs on top is small. Shrinking that cache is the declared aim of half the industry, and the most cited technique, DeepSeek's latent attention, compresses that state before storing it. The result reads in one line: each machine serves more people and fewer machines are needed for the same work.

The guarantee therefore rests on a bet about the calendar. Nobody is betting that efficiency will stop, because that has never happened; the bet is that demand grows faster than the saving, and that it does so before the loans mature.

The outcome is a short chain: if the saving arrives before the demand, the machines are worth less than the loan assumed; Nvidia covers at most a quarter of that difference and whoever lent loses the rest. A creditor who loses stops lending, so construction halts half-finished and the dark GPUs Sacks fears appear ^[3]. The machines, however, do not evaporate: they get cheaper.

The dress rehearsal is documented: between 1996 and 2001 the telecoms industry buried close to a trillion dollars, almost all of it borrowed, in optical fibre, and less than 5 % was ever lit ^[14]. Global Crossing, WorldCom and Qwest went bankrupt, bandwidth fell by around 90 % and the cable stayed where it was ^[14]. That idle fibre ended up carrying the next decade of the internet, and there sits the lesson nobody tells: the demand arrived, it just arrived late. The ones who went under were not wrong about the future; they financed it with a loan shorter than the wait.

The counterpoint is serious and has to be conceded whole: efficiency can enlarge demand rather than shrink it. “As AI gets more efficient and accessible, we will see its use skyrocket”, wrote Satya Nadella on the very day of the 2025 collapse, invoking the Jevons paradox by name ^[15]. Sacks adds a brake from the opposite side: power, land and permits limit how fast supply can grow, so the glut may never arrive ^[3]. The fibre analogy has a limit besides: nobody ever promised that bandwidth would manufacture its own demand, and with intelligence that is exactly the promise. The laboratories themselves signed that they believe they are “close to automating AI research” ^[16]. At that point demand stops being a curve and becomes a loop. The argument is not about direction then but about the date, and there sits the asymmetry nobody signs: a debt matures on a written term and a demand curve has none.

The region watches all this from the other side of the counter. Its bottleneck was never the weights but everything underneath ^[17], and that layer is financed today by whoever sells it. It does not receive the machines, does not sign the guarantees and buys the finished product, so the event that would ruin the creditors is the same one that would cheapen its input. The honest caveat is that regional retirement savings do travel on the equity side: Chile's Fund A returned 14.89 % real in 2025 and its own regulator attributes that to the technology boom and the adoption of AI ^[18]. A correction is felt here, once, against a decade of cheap input.

THE PLAY

1. Your signature is worth more than your location. What is scarce here is not electricity or land but the buyer who commits for years, and that is a paper a company in this region can sign. If your consumption is too small to sign alone, sign it aggregated: a chamber, or five firms in the same sector, negotiate as one contracted customer rather than as five walk-ins.
2. Do not tie your cost to today's tariff. Short contracts, revision clauses, and an architecture that can change provider and model without being rewritten. Half the outcomes end with the input far cheaper, and being locked to the current price is the one sure way to miss that fall.
3. Keep the shopping list ready. Write down today the work you are not running because it costs too much, and price it at a third. The winners from buried fibre did not predict the bankruptcies: they knew what to do with cheap bandwidth the day it appeared.