When the offtake is downstream of the funding window

Editions 11 and 12 argued that the binding question was shifting — from whether compute holds its value to whether the paper that financed it can be refinanced. Edition 12 traced the architecture being built for that: the Apollo–Blackstone–Anthropic structure that tested whether compute can serve as financeable collateral — not yet as a confirmed asset class.

This edition stresses that architecture, and Oracle has spent the months since making the question concrete.

In December 2025, the cost of insuring Oracle's debt against default reached its highest level since the 2008–09 financial crisis — its five-year credit default swap spread reportedly widening to around 139 basis points. The headlines called it an AI-bubble scare. The more useful read is narrower: offtake is downstream of the funding window.

What happened next is the actual lesson. In early February 2026, Oracle announced it would fund its buildout with a balanced mix of debt and equity, and the spread eased — one analyst noted that equity financing inhibits the credit downside. Then, by late March 2026, the spread reportedly widened to a new high. The funding-mix lever worked, briefly, and then the same pressure reasserted and pushed further.

The round-trip is the entire edition. The market repriced the credit risk of one company's offtake architecture inside a single quarter, and what moved the price was funding mix — not AI demand, not the viability of the contracts underneath.

Oracle is the cleanest public evidence that long-duration offtake commitments are downstream of capital-market conditions most infrastructure analysts treat as fixed inputs. It is that question — the one Edition 12 closed on, of what happens when the residual backstop holds but the financing market that priced it does not — playing out in public, on a corporate balance sheet rather than inside an SPV.

In plain terms: Oracle has promised to deliver an enormous amount of future cloud capacity, is spending heavily up front to build it, and is funding much of that spend with borrowed money against a backlog concentrated in a few very large customers. That is the profile of an intermediary, not a diversified platform.

From Oracle's own Q3 FY2026 results (the quarter ended 28 February 2026): RPO — Remaining Performance Obligations, or contracted future revenue not yet recognised — of roughly US$553B and climbing (US$638B by the FY2026 close); around US$300B of that reported as tied to OpenAI as a single anchor tenant; total debt of roughly US$134.6B at Q3, up from about US$92.6B nine months earlier; and free cash flow of roughly negative US$24B, confirmed at the full year as negative US$23.7B for FY2026.

Compare that to the genuine hyperscalers. Microsoft, Amazon, and Alphabet each run deep, diversified balance sheets, generate large operating cash flows, and own the demand platform their infrastructure serves. Alphabet's US$84.75B equity raise in mid-2026 is the same lever Oracle reached for — but pulled from a stronger starting point: strategic flexibility rather than visible refinancing stress.

The third obsolescence

Edition 11 showed technical obsolescence and economic obsolescence pulling apart — a chip can still compute long after it has stopped earning its cost of capital. Financing obsolescence extends that same divergence by one more step: an asset can still compute, still hold book value, still be earning, and yet become unfinanceable if the refinancing window closes before its debt rolls.

This is not abstract for Oracle. Its depreciation expense nearly doubled to US$7.62B in FY2026 as the AI hardware base scaled — the asset-life clock running faster on the books at the same time the debt that funded those assets has to be rolled.

A simple underwriting test

A bankable AI-infrastructure contract now has to clear three tests: is the tenant credit strong enough to carry the lease; is the financing tenor long enough for the asset life; and is there enough equity buffer if refinancing spreads widen? Oracle appears to clear the first on paper — a reported OpenAI-linked backlog of roughly US$300B — but strains the second and third: assets funded with debt that has to roll while free cash flow is negative. That is why the credit market, not the demand signal, did the repricing.

A broader market risk sits in the background: the largest US technology companies are now a major share of the S&P 500, so equity-market flows can reinforce funding confidence on the way up and tighten it on the way down. That is a risk vector to watch, not the core thesis.

AI infrastructure is not only constrained by power and silicon. It is increasingly constrained by capital flows. Edition 7 found the binding constraint had moved from generation to deliverability. This edition finds another sitting above both — whether the funding window stays open long enough for the offtake to be delivered at all, and whether a fix that works once keeps working.

If offtake credit must now be financeable before an APAC project breaks ground — and if hyperscale data-centre risk is moving from private refinancing markets into public listed vehicles — which planned regional builds get repriced before they ever break ground?