Nvidia, the dominant supplier of the chips and systems that run AI, has assembled a consortium to mobilise more than $500 billion of outside capital for AI infrastructure.
Its partners are Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, a roll-call of the largest names in private capital.
The plan treats processing power, known in the industry as compute, as an investable asset that throws off long-term, usage-linked revenue.
Nvidia calls compute an investable asset that delivers cheap output, strong revenue and a long working life.
Days earlier, Google paid $10 million at a bankruptcy auction for the internal data of Spirit Airlines, the collapsed low-cost carrier.
The package runs to about 100 million emails and 500 million Microsoft Teams messages, plus operational and financial records, bought to help train Google's AI models.
Between them, the two deals put a price on AI's two essential raw materials, the hardware that does the computing and the data that trains the models.
The pitch, increasingly, is that both are becoming commodities, traded and hedged like oil, copper or natural gas.
That comparison is worth testing, because it is doing a lot of work to make the trade sound solid.
What makes a commodity
A commodity market needs a few things that oil and metals have and most assets do not.
The product must be fungible, meaning one unit is interchangeable with another, and standardised enough to price without inspecting each batch.
It helps if the product is durable and storable, so it can be held, delivered later and pledged as collateral.
And it needs transparent price discovery, usually through deep spot and futures markets where buyers and sellers agree a public price.
On the first of AI's inputs, compute, the machinery of a commodity market is genuinely being built.
Compute is starting to trade
Exchanges are racing to turn raw processing power into a tradable contract.
Intercontinental Exchange, the owner of the New York Stock Exchange, plans compute futures tied to an index of live spot prices for Nvidia chips such as the H100 and H200.
The derivatives exchange CME Group is working on a rival set of contracts with its own benchmark, both pending approval from US regulators.
Prediction markets already let users bet on where GPU rental prices go next.
Rental rates are volatile enough to justify the effort, with one-year prices for the H100 chip rising more than 50% in five months earlier this year.
In that narrow sense the oil comparison is becoming literal, and compute is joining electricity, gas and shipping freight as something the financial system prices and hedges.
But the Nvidia financing deal is not really about this flow of rented compute.
It is about the hardware underneath, financed as durable, yield-bearing infrastructure.
Special vehicles own the chips and lease out their capacity, much as investors own toll roads or office blocks.
Where the oil analogy breaks
This is where the commodity framing starts to strain.
A barrel of oil does not become obsolete while it sits in a tank, but a graphics chip loses value fast as newer, faster models arrive.
Jeff Gundlach, the DoubleLine bond investor, questioned why hardware of uncertain lifespan should serve as collateral for long-term debt, and predicted the structure would not age well.
The short-seller Jim Chanos has made a similar point, arguing that data centre operators depreciate chips over about six years when the real economic life may be far shorter.
Tellingly, Nvidia has offered to backstop up to 25% of the residual value in some deals, an admission that the collateral may not hold its worth.
Compute, in other words, behaves less like storable oil and more like perishable electricity, which is workable for a spot market but awkward as security for long-dated loans.
Data is the least like oil
If compute is a stretch, data is a bigger one, despite the well-worn line that it is the new oil.
Oil is fungible and finite, but data is neither, since it can be copied endlessly and used by many buyers at once without being used up.
Nor is it standardised, because Spirit's emails are not interchangeable with any other company's, and their worth to a model is uncertain until the data is cleaned and tested.
The Google deal looked less like buying a commodity and more like a one-off purchase of a bespoke asset, closer to acquiring a brand or a patent library.
It also carried frictions a commodity never does, from court approval and a mandatory scrub of personal information to an objection from Spirit's flight attendants' union.
Bespoke assets sold at distressed prices under legal supervision do not add up to a liquid market in a standardised good.
'Financialisation', not commoditisation
The better description of this month is 'financialisation', the wrapping of AI's inputs in the instruments of capital markets, rather than true commoditisation.
For the thin slice of rented GPU time, a real commodity market is forming, and it may end up resembling power markets more than oil.
For the hardware behind the $500 billion, and for corporate data archives, the commodity label flatters the trade more than it describes it.
That matters, because dressing fast-depreciating, idiosyncratic and legally encumbered assets as though they were durable, fungible commodities is a well-worn way of obscuring risk.
Gundlach's warning that this looks like a market top is really a warning about that mismatch.
The plumbing of commodity trading is being bolted onto artificial intelligence at speed, but compute perishes and data is unique, and neither will ever trade quite like a barrel of oil.
