Wait until they issue collateralised debt obligations, where the collateral is compute earning rent, but that compute ages as new technology is introduced, making the rent inefficient and the CDOs inevitably default.
Compute Debt Obligations.
This page takes that literally. It maps the money reported between the companies building AI capacity, pools their obligations into a deal you can stress until it breaks, and works out what the newest dollar of capital actually earned.
Inspect the loop ↓
Wait until they issue collateralised debt obligations, where the collateral is compute earning rent, but that compute ages as new technology is introduced, making the rent inefficient and the CDOs inevitably default.
Compute Debt Obligations.
Interactive data sculpture showing capital entering the AI infrastructure cycle. Click or press the arrow keys to change the view.
Nine companies and the thirteen legs of money reported between them. Select one for its role and the filing or announcement behind each leg. A leg on this map means a company announced something or filed something. It does not mean anything more than that.
The waterfallholding
Collateral
Stress the assumptions
A collateralised debt obligation pools obligations nobody wants to price one by one, then cuts the pool into layers. The top layer is paid first and gets called safe. The bottom is paid last and gets called equity. The rating agencies rate the layers. Nobody rates the pool underneath them.
This desk does that to compute. Pick the companies whose contracted obligations go in the pool, then move the three sliders and watch the losses attach at the bottom and eat upward. Concentration is how much of your pool a single payer owes. Circular collateral is how much of it is paid by counterparties who are themselves funded from inside the same loop.
Arithmetic on reported headline numbers. The ratings are synthetic, the deal is imaginary, and nothing here is issued, sold, or rated by anyone.
A company's reported return on capital averages every dollar it has ever put to work, including the ones it spent forty years ago. This worksheet throws the average out. It divides one year of extra after-tax profit by the capital that bought it, and shows which filed lines had to be assembled by hand to get there.
The worksheetclears
Microsoft · NASDAQ: MSFT · FY ended 30 June
2026
Everything the company spent this year to buy capacity. Tick a line in or out to see what the answer rests on.
How much more after-tax operating profit the company earned than it did the year before.
Step 2 divided by step 1. No company files this number.
return 21,528 ÷ capital 140,556
15.3%4.7 points clear of the band
The tall tick is what the filing earned. The short one is the figure that gets quoted. The grey band is the cost of capital.
US technology WACC band, Damodaran (NYU Stern), January 2026. Held at one value across all four worksheets so the four can be compared against each other. It is not the right band for each of them taken on its own. The same January 2026 dataset puts restaurant and dining at 7.16%: on that band McDonald's FY2024 increment is 5.3 points short rather than 7.5, and FY2025 clears by 8.7 rather than 5.2. It is held at one date as well: Cisco FY2000 is read against a 2026 band rather than the higher cost of capital that prevailed in 2000, which makes that reading the conservative one. No company publishes its own cost of capital.
Microsoft's newest capital earned 15.3%, against a hurdle band of 9.3–10.7%. The increment clears it by +6,545 to +8,400. The spread is real, and narrow enough that two points in either direction would close it. The 24.9% that gets quoted falls between the two readings these lines can produce: 25.3% over last year's capital, and 24.4% over a denominator near $88B that appears on no statement. Switch the basis to lagged for the first of them.
Working paper4 worksheets
The figure beside each is that worksheet on its default lines, before you strike anything.
Basisconcurrent
This year's change in profit over this year's spend, or over last year's. Both are defensible. At Microsoft they are ten points apart.
Return on invested capital divides profit by every dollar a business has ever put to work, which buries what a company spent this year under twenty years of older and cheaper spending. The marginal return divides the change in after-tax operating profit by the capital deployed to produce it. It covers one year and one increment, and no statement reports it, which is why this worksheet has to assemble it.
Two choices decide the answer and the filing prints neither. The first is what counts as capital. Cash capital expenditure is one line in the cash flow statement; right-of-use assets obtained for finance leases are a different line in the lease note, and at Microsoft that second line was $24.6B. Cisco is the extreme case. Its cash capital expenditure in FY2000 was $1,086M while its invested capital rose $15.8B, because most of the capital arrived as stock issued for acquisitions and never passed through the investing section at all. The second choice is which year. This year's change in profit over this year's spend is the harsher read; over last year's spend, the kinder one. Tick the lines and a quoted number moves ten points without anybody lying.
The hurdle is a band rather than a figure. No company publishes its own cost of capital, and third-party estimates for the same company move by three points. Which side of that band the tick lands on is a fact about capital allocation with a one-year shelf life, and it says nothing about anyone's conduct. One reading below the band is not a crash either. What tends to follow a run of them is a cut to capital spending, and capital spending is what the rest of this page is made of.
Not used here. Three figures travel with this argument and appear as no ledger line on this page, because none of them reconciles to a filing. Cisco's "$16 billion invested against a $426 million loss": the $16 billion reconciles, the loss does not, and FY2000 operating income was a positive $3,235M. Cisco's headline returns of "121% / 37%": no such figures appear in its filings, and they are reproducible only under a cash-adjusted definition nobody states. McDonald's "28% blended ROIC": McDonald's publishes no ROIC at all, third-party estimates cluster at 17.5–19.0%, and that figure appears nowhere here. Where one of the other two is drawn on the bar it is drawn as a quoted mark and badged as one, because it was quoted rather than filed.
FILED lines are read off the filing linked beside them. DERIVED lines are arithmetic performed here on those figures, using the figures as printed so the columns add up. ASSUMED lines would be stated assumptions, and none is used here. THIRD PARTY and AUTHOR figures are drawn only as quoted marks on the bar, never as ledger lines. A marginal return below a hurdle is a fact about capital allocation with a one-year shelf life and says nothing about anyone's conduct. Figures as filed to 9 September 2026.
The hardware may age before the debt does.
A few buyers can carry a large share of contracted demand.
Infrastructure arrives before durable AI revenue is proven.
Compute has no public price. On 5 October, pending regulatory review, CME lists futures on H100 and B200 rental rates. Every valuation resting on an assumed rental rate becomes markable against a screen.