Written by: Agata Wojtas, Chief Commercial Officer, Digital Colliers
Meta's Q2 filing carried a number worth staring at. Around $279B in future AI data-centre lease commitments, up 53% quarter over quarter, sitting off the balance sheet as contractual obligations. Not debt, not capex on the face of the accounts, but a hard cash claim on future operating income. If you're doing credit work on hyperscalers, their suppliers, or anyone downstream in the AI build-out, this is the shape of the next data gap.
The commitment stack nobody's aggregating
AI infrastructure spending doesn't show up where a classical credit model looks. The audited balance sheet catches drawn debt and finance leases. It doesn't catch the three commitment types now driving hyperscaler cash claims:
- Long-dated data-centre operating leases, often 15 to 20 years, disclosed as contractual obligations in the notes.
- GPU purchase agreements and reserved-capacity contracts with the chip vendors and ODMs.
- Power purchase agreements with utilities and independent generators, some running two decades, some with take-or-pay clauses.
Each one is legally binding future cash out. None of them roll neatly into the leverage ratio your credit committee reviews. And the same pattern is now showing up at the smaller AI-native tenants, the neoclouds, and the colo operators borrowing against the hyperscaler contracts. If you lend into that chain, your counterparty's counterparty risk is now your problem.
Why the classical credit model misses it
Most mid-market credit shops still consume quarterly financials on an 8 to 10 day lag after the borrower's own close, and the borrower's own close is usually a spreadsheet exercise pulling numbers across systems by hand. By the time the ratios land in your model, the commitment picture is already two quarters stale. Even the best-in-class finance teams closing in under 5 days aren't restating footnote disclosures at that cadence.
So the model sees leverage that looks fine. The reality is a borrower who has signed 15 years of fixed payments against a revenue stream that depends on AI workloads clearing at expected utilisation. When roughly 95% of enterprise AI projects don't reach production or ROI, the demand assumption underneath those commitments deserves a second look. You don't need to believe the AI thesis is wrong. You just need to price the dispersion.
What a lender's data model needs to capture
If you're rebuilding the credit view for names with heavy AI exposure, the input set has to widen. The pattern the sharper credit shops are moving toward:
- Footnote extraction as a first-class data feed. Contractual obligations tables, purchase commitments, and lease maturities pulled from every 10-Q and 10-K, normalised into a maturity ladder per counterparty.
- Power and grid interconnect data. Queue positions, interconnect agreements, and PPA terms are public in most US ISOs and increasingly in Europe. They tell you which announced capacity is real and which is a press release.
- GPU allocation signal. Vendor concentration, delivery slippage, and secondary-market pricing on H100 and B200 capacity are all leading indicators for the tenants underneath.
- Utilisation proxies. Colo power draw, hyperscaler capex guidance versus depreciation schedules, and hiring patterns on the AI engineering side. If the workloads aren't landing, the commitments don't get covered.
None of this is exotic data. It's public. What's missing at most lenders is the pipeline to ingest, normalise, and hold it against the borrower record so the credit officer sees one view.
The regulatory angle nobody's talking about
For European lenders, DORA has been in force since January 2025, which means your operational resilience obligations already extend to the ICT third parties propping up your credit and risk models. If your PD model quietly starts depending on scraped hyperscaler disclosures and a vendor's GPU pricing feed, that's in scope. And if you're automating credit decisions off any of it, the SCHUFA ruling from December 2023 already put automated credit scoring squarely inside GDPR Article 22 territory. The data gap is real, but the fix has to be built with the audit trail on from day one.
The left-behind risk here isn't missing the AI story. It's carrying counterparties on a credit view that stopped describing them two years ago. The operators building the wider input pipeline now are the ones who'll price this cycle correctly.