Written by: Jakub Pietroszek, Partnership Manager, Digital Colliers
Tesla just printed negative $1.1B in free cash flow, its first negative quarter in over two years. The driver wasn't demand or margin, it was AI and robotics capex landing all at once. Google reported its first negative cash-flow quarter on the same theme. If you're a credit officer looking at an AI-heavy borrower, the base rate you've been trained on doesn't apply anymore.
The assumption inside your credit model just broke
Most mid-market credit models assume capex is smooth. You underwrite off trailing FCF, apply a haircut, stress the interest coverage, and move on. That works when the borrower is buying trucks, warehouses, or line equipment on a predictable replacement cycle.
AI capex doesn't behave that way. It arrives in step-function jumps tied to GPU allocations, data-center leases, and multi-year power contracts signed before the revenue exists. A borrower can look fine on a rolling 12-month view, then swing hard negative for two or three quarters as a training cluster comes online. Tesla and Google can absorb it. A mid-market borrower on a covenant-heavy revolver cannot.
The risk you carry, if you don't update, is the classic left-behind pattern. You keep underwriting to a smoother world while your peers build a view of the lumpier one. The borrower who tripped a covenant six months ago at your bank raises debt cheaper somewhere else next year.
The feeds a credit team actually needs now
If you're rebuilding the underwriting file for an AI-heavy borrower, the useful signal sits outside the standard financial pack. The pattern I keep seeing in credit teams that are getting this right:
- Capex disclosure feeds pulled from 10-Q and 10-K filings, broken out by AI-specific line items rather than aggregate PP&E.
- Contracted commitments, meaning the off-balance-sheet cloud, GPU, and colocation deals disclosed in the notes. These are the real forward capex.
- Long-dated energy contracts. A 15-year PPA is a fixed cost the covenant math needs to price in.
- Depreciation schedule assumptions. GPUs on a six-year straight line versus a three-year accelerated schedule move reported earnings by material amounts.
- Vendor concentration on the compute side. A borrower whose training runs sit on one hyperscaler carries the same tail risk as a factory with one supplier.
None of this is exotic data. It's mostly public. The problem is that pulling it, normalizing it, and getting it into the credit file on a monthly cadence is manual work most teams aren't set up for. The majority of mid-market finance teams still close the month in spreadsheets, and month-end close typically runs 8 to 10 days. If your borrower's own team is that slow, your view of them is stale by design.
The borrower profile you're now underwriting
The mid-market AI-heavy borrower looks different from what your credit policy was written for. Expect:
- Negative FCF quarters that are planned, not distressed. You need to distinguish the two.
- Revenue growth that lags capex by 12 to 24 months. Coverage ratios computed off trailing numbers will mislead.
- Balance sheets carrying prepaid compute as an asset. Ask how it's amortizing.
- Working capital swings tied to enterprise pilots that may or may not convert. Around 95% of enterprise AI projects don't reach production, so pipeline-based projections need a heavy discount.
The borrowers worth banking here aren't the ones with the cleanest FCF. They're the ones whose management can walk you through the capex ramp, the contracted revenue against it, and what happens to covenants if the ramp slips by two quarters.
What the winning credit shops are doing
The credit teams staying ahead of this are doing two boring things well. First, they're closing their own view of the borrower faster. Best-in-class finance teams close in under 5 days, and credit teams are borrowing the same discipline, pulling capex and commitment data monthly rather than quarterly. Second, they're being explicit about model risk. Automated credit scoring already carries real legal exposure in the EU under the SCHUFA ruling, so any updated model needs a human review layer documented in the file.
If you're at a mid-market bank, the honest question is whether your credit committee has seen an AI-heavy borrower memo yet. If not, the first one is coming, and the framing you set on it becomes the template for the next twenty.