The AI Vendor Due Diligence Questions Nobody Asks: Does Your Supplier Own the Metal?
A sublet market in AI compute is forming, and the landlord is often the tenant's competitor. If your supplier is renting the silicon that serves you, the scarcity story you were sold at signature looks very different at renewal.
The most informative sentence in AI infrastructure this year was not about a model. In an interview published on 9 July 2026, Mark Zuckerberg said the offers coming in for Meta's compute were high enough that in some cases renting capacity out rather than using it internally could make sense. Read that as an operator rather than a headline. A company that has spent two years explaining why it needs every chip it can get is being asked what it would take to hand some over, and it has an answer. That single fact should reorder your AI vendor due diligence questions.
Eight days later, Reuters reported that Meta and Anthropic were in early talks over a compute arrangement potentially worth up to $10 billion, discussions that the reporting was careful to say might produce nothing. Treat the specifics as provisional, because early-stage talks fail more often than they close. The structure is the story. A model developer renting silicon from a company that competes with it in models, on a term long enough to outlive several strategy cycles, is not a variation on cloud procurement. It is a different asset class wearing a familiar contract.
Why would a company rent out compute it says it needs?
Because the price got interesting. Meta's first-quarter Form 10-Q tells investors it anticipates approximately $125 billion to $145 billion of 2026 capital expenditure, a range whose upper bound gets quoted as though it were a forecast. Spending at that rate produces a portfolio of hardware with wildly uneven marginal returns. The first tranche serves the products that pay for it. The last tranche serves whatever the research roadmap turns out to want, which is exactly the part nobody can underwrite.
So a lease is a revealed preference. An organisation that rents capacity out has run the comparison between what it expects to earn from that hardware in its own hands and what someone else will pay for it, and the tenant won. That is a rational trade and I would make it too. It is also information the market has not priced properly, because the same firms have been selling urgency to enterprise buyers. "Take capacity while there is capacity" is a much weaker argument coming from a party that is quietly shopping its own surplus.
What AI vendor due diligence questions actually matter in 2026?
Four, and no procurement function I have seen asks them. Do you own or rent the compute serving our workloads? If you rent, from whom, and is that counterparty a competitor of yours or of ours? What is the remaining term, and how does it sit against the term you are asking us to sign? And what happens to our service if that agreement is not renewed on comparable economics?
The answers are rarely commercially sensitive in any real sense, which makes refusal to answer the finding. What buyers are actually testing here is whether the continuity risk they think they bought is the continuity risk they hold. A three-year enterprise commitment sitting on top of a compute agreement with an eighteen-month tail is not a supply chain, it is a mismatch, and the exposure lands on whoever is furthest from the hardware. That is the customer. This is the same discipline we argue for in assessing readiness before committing to a build, applied one layer further down the stack than most assessments bother to go.
The competitor-as-landlord wrinkle deserves its own line in the risk register. In normal cloud procurement the provider's incentive is renewal, because that is the whole business. A landlord whose primary business is competing with the tenant has a second incentive that only shows up under stress: prioritisation. When capacity gets tight, the party allocating it is choosing between its own roadmap and a customer it is trying to beat. Nobody needs to behave badly for that to matter. The tenant simply cannot verify how the decision was made.
The landlord's new profit centre is its largest credit risk
Here is the second-order effect that the coverage keeps missing. Leasing converts a hardware position into a book of counterparty exposures. Multi-year compute agreements with AI developers are, in balance sheet terms, long-dated receivables from firms whose revenue is growing fast and whose funding conditions can change quickly. The lessor has swapped an asset it controls for a promise it does not. If a downturn in AI capital markets ever arrives, it will hit the landlord's rental income and the resale value of the underlying hardware at the same time, which is the definition of correlated risk.
Now the contrarian test, because agreeing with the alarm in full would be lazy. Anthropic is not a fragile tenant scrambling for scraps. It publicly describes a strategic Microsoft and NVIDIA partnership including $30 billion of Azure capacity, and a five-gigawatt agreement with Google and Broadcom. It confidentially submitted a draft Form S-1 on 1 June 2026, and its Series H raised $65 billion at a $965 billion post-money valuation. Against that, adding a further supplier reads as deliberate diversification by a company with the balance sheet to pay for optionality. Concentration risk is the thing that kills you, and this reduces it. The exposure I have described belongs to the enterprise three layers down the chain, not to the model developer.
What would change my mind. If compute leases start being disclosed with counterparty and term in vendor security and continuity documentation, most of this concern evaporates into ordinary supplier management. If spot pricing for frontier-class capacity falls sustainably, the negotiating asymmetry disappears with it. And if the reported talks collapse without a comparable deal appearing elsewhere within a couple of quarters, then this was a rumour about one relationship rather than the start of a market, and I will have overweighted a single data point.
The asymmetry nobody has priced is on the buyer's side, and it is temporary. Vendors are still negotiating as though capacity were the binding constraint while their own suppliers advertise a willingness to rent it out. Reuters noted Meta shares closed down more than 2% on the day of the report amid a broader technology selloff, which tells you the market has not settled on whether leasing reads as strength or as doubt. Buyers do not need that question resolved. They need one clause, requiring disclosure of material changes in underlying compute supply, written into contracts being signed this quarter. That is a cheap thing to ask for now and an expensive thing to want later, which is roughly the definition of good technical strategy.
Questions people ask
How can we tell if our AI vendor is sub-leasing compute rather than owning it?
You usually cannot tell from the outside, which is the point. Public cloud commitments get announced; sub-leases between AI firms often do not. The practical route is contractual: ask for the ownership status, counterparty and remaining term of the compute serving your workloads as part of onboarding, and add a notification obligation for material changes. A vendor that owns its capacity answers immediately, so hesitation is itself a data point.
Should compute supply chain risk go in an AI vendor risk assessment questionnaire?
Yes, and it belongs under continuity rather than security, where most questionnaires currently have no home for it. The relevant questions are term mismatch (does their supply agreement outlast our commitment?), counterparty identity (is their landlord our competitor or theirs?) and failover (what capacity moves our workloads if the arrangement lapses?). Existing third-party risk frameworks handle sub-processors well and hardware tenancy chains almost not at all.
Does compute leasing between AI companies mean the shortage was never real?
No, and claiming that would overstate the evidence. Frontier training capacity can be genuinely tight while individual holders still find that renting out marginal capacity beats using it. What the leasing behaviour undermines is the universality of the scarcity claim, and with it the urgency argument used to justify long lock-ins and premium pricing. Scarcity here is a price, and prices move.
Related
- AI Deskilling Is the Workforce Risk Your Board Isn't Pricing
- The Memory Oligopoly Behind AI's Cost Inflation
- Least privilege for AI agents is a 1975 idea your pilot is skipping
- AI & Automation
Written by an AI editorial persona of Abyshire's proprietary editorial system and reviewed by our team.