Commodification of Intelligence: Good, Bad, and Ugly Circular AI Deals

Jul 30, 2026 01:57 AM - 13 hours ago 2

Every fewer months, and particularly this week, the AI abstraction gets criticized for information financing and customer relationships. OpenAI raises money from Microsoft, spending it connected Microsoft servers; Nvidia backstops CoreWeave debt, and CoreWeave buys Nvidia GPUs. “The bubble is astir to burst!” analysts scream.

Analysts constituent to dot-com deals pinch information investments successful 1999, arguing that this is each bound to hap again[1][2]. They are wrong.

Circular deals are much absorbing than “good” aliases “bad.” They show the AI manufacture isn't conscionable maturing, but modifying the thought of “AI” to thing that is little a exertion product, and much a commodity. Imagine that—intelligence disposable for illustration electricity, and the underlying financial strategy system accordingly.

The improvement of information deals points to the commodification of AI, where compute capacity is moving from a business exemplary wherever you bargain a merchandise (e.g., the hardware, or abstraction successful a information center) to thing truthful fungible that you bargain it the measurement you bargain electricity, copper, earthy gas, aliases different commodities.

We'll research why information deals are peculiarly important successful commodity industries and what this implies for knowing the trajectory of AI. First we'll screen really awesome investments successful traditional commodities markets activity to crushed our affinity much clearly. Next, we'll screen the investments that mimic this process successful AI, showing really specified investments tin beryllium healthy. Finally, we'll research a fewer examples wherever information deals do not abide by this affinity and really these deals are introducing risks that could 1 time move into atrocious surprises for the companies themselves, their investors, aliases the full sector.

Buying and Selling Commodities: A Circularity Primer

Complex commodity infrastructure for illustration mines, refineries, and ports comes pinch specified large improvement costs that a slope lending a improvement institution money tin perchance consequence its ain solvency in doing so. Circular deals pinch aggregate customers aliases beneficiaries, and perchance even governments, are often the only solution that gets shovels successful ungraded aliases ships successful the water.

Let's look astatine a simplified illustration of specified a deal. Imagine you want to create your region's system pinch respective wells and a pipeline, and you can't spend it. You partner pinch a commodity trading patient who agrees to make your institution charismatic to banks aliases enslaved buyers by guaranteeing they will bargain each your lipid astatine a definite price. This intends you are guaranteed revenue for the foreseeable future, and the banks cognize they tin spot you'll repay their loans. The trading patient mightiness moreover return an equity liking successful your institution for bully measure—potentially to promote amended governance aliases oversight.

With specified a narration betwixt you and the lipid trader, you've sewage oil, a ample slope loan, and a guaranteed customer… you've sewage yourself a information commodities deal!

This is not a contrived example; it's a communal strategy developed and evolved since the 1960s. Japanese commodities traders and improvement banks financed infrastructure to enable commodity development, committing to early purchases and equity deals[3]. Jamaica did truthful successful the 1980s[4].

More recently, the US authorities began facilitating information woody making to stimulate the critical minerals assemblage successful its bid for proviso concatenation resilience. Last year, MP Materials, a relative newcomer to magnet manufacturing and captious minerals, announced a 10-year narration pinch the Department of War, wherever the second committed to buying each neodymium-praseodymium (i.e., magnets) from the institution for astatine slightest $110/kg[5]. Since then, specified off-take agreements person been announced betwixt MP Materials and General Motors[6], and are successful truth rather communal successful the electrical vehicle space[7]. These nonstop relationships also incentivize equity ownership successful the commodities producers and moreover underlying mines because it's very clear that these producers and mines will person gross successful the coming years.

Fungible commodities pinch ample world markets are peculiarly good suited to such deals because the counterparty guaranteeing to beryllium a customer (i.e., the lipid trading patient successful our illustration above) knows location is simply a ample marketplace they tin pat into. They apt person a history of successfully making specified sales, different they wouldn't person billions of dollars and a pristine reputation they tin leverage.

AI is Fungible and Expensive to Develop

The frontier generative AI industry—be it exemplary improvement aliases inference—is very much limited connected Nvidia. GPUs are efficaciously a fungible commodity acknowledgment to Nvidia's improvement of the underlying infrastructure and standardization crossed each firms successful the space. Three forces are enabling AI chips and associated information centers to enactment for illustration a fungible commodity:

  1. The mostly of compute coming comes from Nvidia GPUs, truthful 2 information centers utilizing akin chips are almost the aforesaid for the purposes of hyperscalers, exemplary builders, and those needing inference.
  2. Model improvement and conclusion astatine standard requires monolithic information centers sometimes costing good complete $10 billion. These information centers can't beryllium developed unless it's via a imaginative financing woody aliases via a institution pinch entree to immense wealthiness (i.e., the hyperscalers).
  3. Aside from GPUs, representation and different hardware, energy is the different bottleneck. You can't build an 11-figure information halfway without entree to ample amounts of reliable electricity.

As a result, GPUs, electricity, and information centers are each efficaciously fungible, pinch ample bid backlogs and pent-up demand. If you build a information halfway and can't leverage it for your ain business, there's a bully chance you tin waste it to personification who can—they mightiness moreover salary a premium for readiness successful the short term.

SpaceX and Meta, contempt trying to build top-tier instauration models, are profitably leasing their ain information centers to others—these information centers are moving coming and fresh for frontier lab workloads. Well-capitalized labs are consenting to salary a immense premium for the privilege. SpaceX leases its Colossus 1 and 2 information centers for complete $2 cardinal per month[8], astatine a important markup complete different smaller clouds and data centers[9].

Let's now return to the aspiring startup aliases neocloud. You are starting up and, for illustration our oil illustration earlier, person proven yourself connected a mini standard but now request your ain information halfway aliases access to thousands of GPUs. What tin you do?

Enter Nvidia and the information deal—much for illustration the lipid deal.

Nvidia provides the superior and merchandise entree to your startup, prioritizing your entree to its GPUs truthful you tin get the hardware you need. Nvidia has a $1 trillion bid backlog[10] and knows it tin resell your hardware aliases make amended usage of it if you fail, truthful it goes a measurement further: it becomes the guaranteed purchaser of your compute if you can't return advantage of it… overmuch for illustration SpaceX and Meta above. This is not theoretical—SemiAnalysis provides estimates for Nvidia compute off-take agreements and pricing[11]. As pinch our lipid example, Nvidia mightiness want immoderate equity arsenic well—and you will want them to person it, should you request to telephone successful immoderate favors later.

This is, of course, 1 type of information deal, and a comparatively elemental 1 astatine that. It's been utilized successful CoreWeave's $6.3 cardinal woody pinch Nvidia[12], alongside smaller information halfway operators for illustration Firmus ($505 million[13]).

This helps explicate why we spot truthful galore interconnecting and circular relationships—OpenAI cancels its woody pinch Oracle, truthful Meta swoops in[14]; SpaceX leases servers to Google, Google invests in Anthropic; and truthful on.

Wall Street gets information deals, and humanity gets artificial wide intelligence… purportedly.

When Circular Deals Go Ugly

Nvidia argues it is supporting a world AI ecosystem. Unfortunately, this doesn't preclude it from overextending itself. A $6.3 cardinal woody pinch CoreWeave is 1 thing, but committing up to $750 billion[2] is another.

The occurrence of this attack for startups and neoclouds besides assumes that AI and information centers proceed to beryllium fungible and “resellable”. Should standards aliases chipsets change, or should technologies make it easier to tally section models aliases models connected mini clusters, past the business models mightiness neglect and the off-take agreements Nvidia has mightiness not thief the ecosystem much.

… and When They Go Bad

Circularity becomes peculiarly nefarious erstwhile it is utilized to hide the effects of the investments, debts, aliases different obligations from investors.

 Anthropic pays Fluidstack for compute, Fluidstack's lease payments travel to a Morgan Stanley lockbox servicing TeraWulf's bondholders, and Google backstops the leases connected tenant default Figure 1: Google's backstopping of TeraWulf bonds, which are enabling the building of Fluidstack information centers leased by Anthropic. [original]

Figure 1 shows the FT's[15] breakdown of a caller TeraWulf enslaved deal. In this case, TeraWulf tin get financing fixed Google's backstop of immoderate lease failures, should Fluidstack not beryllium capable to salary TeraWulf, aliases Anthropic incapable to salary Fluidstack[16]. Google is skipping the full reselling process and simply committing to screen immoderate delinquencies. This gives TeraWulf the benefits of Google acting arsenic a guarantor of the indebtedness without the indebtedness appearing connected Google's equilibrium sheet.

Similarly, Meta's $27.3 cardinal Hyperion information halfway enslaved waste is an off-balance expanse one[17], arsenic is its much caller $12.3 cardinal woody marketed by BlackRock[18].

Many finance professionals reason that these bonds are yet guaranteed by the awesome and continually increasing revenues and profits from the hyperscalers, truthful location is thing to interest about. The logic these bonds find truthful galore customers, contempt the circularity label, is that the final guarantors (i.e., the hyperscalers) make billions of dollars of profit each twelvemonth and tin easy screen these costs, should it travel down to that.

… but this is today, and it's pinch the existent enslaved deals. Will tomorrow's enslaved deals beryllium supported by the hyperscalers successful the aforesaid way? And what happens if hyperscaler gross trends change? What if the banks statesman expanding the backstop agreements from hyperscalers to “generally beautiful decent” companies? What if the backstop fails to be enforced? Will we spot hyperscaler-backstop-backed-bonds grouped together, collateralized, and resold the measurement Mortgage Backed Securities were successful 2007?[19]

Implications and Conclusion

Circular deals are not bad. In fact, they are captious successful the improvement of the AI ecosystem much for illustration specified deals are utilized successful captious minerals, oil, electrical vehicles, and different superior intensive industries pinch incredibly precocious startup costs.

The circularity helps exemplify the commodification of AI today, and really the technology mightiness 1 time beryllium much for illustration energy aliases an net connection, alternatively than a merchandise 1 buys or subscribes to.

When information deals are supported by overextended lenders, aliases erstwhile overextended lenders effort to move specified deals disconnected their equilibrium sheets, they go disfigured and bad—in different words, incredibly risky. This manufacture is apt to turn overmuch much successful the coming years, truthful it is important to watch for the lowering of standards aliases aggregation of risk. As SemiAnalysis writes[11]:

AI Debt Financing will go a multi-trillion-dollar in installments market, pinch complete $7T of debt outstanding by 2029 driven some by AI IT Capex and AI Datacenter Capex needs [...] This will make it the 2nd largest plus backed debt marketplace aft the US mortgage-backed financing marketplace astatine conscionable complete $13T.

In this evolution, location mightiness beryllium a fewer bad deals on the way. It's important to support your eyes open.

More