A Practitioner's Stress-Test Framework

Everything in this course converges on one practical act: stress-testing an AI-infrastructure structure before you commit capital to it. Stress-testing is not forecasting. You are not trying to predict the future. You are pushing on the structure to find where it breaks and how much of a shock it can absorb before it does. The output is not a point estimate — it is a map of failure paths and the size of the cushion in front of each one.

Work through it in layers, from the asset up to the system.

Start with the asset — the depreciation stress. Take the operator's assumed useful life and residual value for its accelerators and replace them with harsher ones. Shorten the life to match the vendor's actual release cadence. Cut the residual to what used units would fetch in a soft secondary market, not a hot one. Then recompute. Does the collateral still cover the debt in the later years? Do margins survive the higher depreciation charge? If the whole model only works on an optimistic decay curve, you have found the first fault line. The question to hold onto: does the debt amortize faster than the asset economically decays?

Next, the operator — the utilization and rate stress. The business assumes clusters stay full at healthy rental rates. Break that assumption. Drop utilization by a meaningful margin and cut rental rates the way a new chip generation or a demand dip would. Revenue falls; fixed obligations do not. How far can utilization and rates fall before the operator cannot cover debt service and lease payments? That gap between current cash flow and the fixed obligations is the true equity cushion. A thin cushion means a single soft quarter is an existential event.

Then the funding — the refinancing and covenant stress. These structures roll over debt continuously; they must refinance. Assume refinancing gets harder and costlier at exactly the wrong moment, because it always does — credit tightens when the business weakens. Check the covenants: how close is the operator to a breach on utilization or coverage, and what happens if it trips one? A covenant breach can force early repayment or hand control to creditors while the operator still looks solvent on the surface. Map the maturity wall — when large chunks of debt come due — and ask whether the business can refinance them in a cold market.

Then the contracts — the concentration stress. Identify the anchor tenants whose commitments underpin the financing. Remove the largest one. Does the structure still stand, or was that single contract the keystone? Read the contract terms for real bindingness — take-or-pay with genuine penalties is very different from a soft commitment with easy outs. Concentration is only as strong as the weakest exit clause in the biggest contract.

Finally, the system — the circularity stress. Map who funds whom. Flag every case where an investor is also a supplier or customer. Estimate how much of the reported demand is recycled through the loop versus standing on independent end-demand. Then push down the one variable every party shares — AI compute demand — and trace it through every linked node, including the second-order hit where a funder loses both revenue and its investment at once. If a moderate decline cascades into trouble at several nodes simultaneously, the diversification is nominal and the real risk is concentrated.

Layer these and a discipline emerges. Each layer has a cushion, and stress-testing measures each cushion in the same currency: how big a shock can this absorb before it fails? The asset layer's cushion is the gap between the amortization schedule and the real depreciation curve. The operator layer's cushion is the equity slice against fixed obligations. The funding layer's cushion is covenant headroom and refinancing capacity. The contract layer's cushion is diversification and the strength of the exit terms. The system layer's cushion is how much demand is genuinely independent of the loop.

The common failure across the sector is not exotic. It is assuming the layers are independent when they are correlated. In a real downturn they move together: demand softens, so utilization falls, so revenue drops, so covenants trip, so refinancing dries up, so the anchor tenant renegotiates, so the collateral you would seize is worth less exactly when you need to seize it — and if the structure was circular, all of this hits the connected parties at the same time. A stress test that shocks one layer at a time and finds each survivable is not enough. The real test applies the shocks together, because that is how they arrive. Underwrite the correlation, not the average, and you will price these structures for what they are: capital-heavy bets on the durability of AI demand, wrapped in financing that lowers the cost and raises the fragility of the wager.