Demand Durability

Demand Durability

The contracted demand that justifies Microsoft's AI build is unusually concentrated in one counterparty. Microsoft's own disclosures show it: commercial remaining performance obligations grew 99% to $627 billion including OpenAI, but only 26% excluding it [1]. That counterparty is loss-making, spending far ahead of revenue, and dependent on serial external financing. Whether the capex earns a return turns partly on a customer whose solvency the filings do not vouch for.

How much of the backlog is one customer

Microsoft frames nearly every contracted-demand figure twice — with OpenAI and without — which makes the concentration measurable rather than a matter of inference. In the March 2026 quarter, commercial bookings fell 4% including OpenAI's Azure commitments but rose 7% excluding them; the RPO backlog grew 99% including OpenAI and 26% without [2]. Two quarters earlier, bookings had jumped 112% — "driven by Azure commitments from OpenAI" — and even that surge, Microsoft noted, excluded "the incremental $250 billion Azure commitments from OpenAI announced yesterday" [3].

Sources: Q3 FY2026 earnings call [4]; Q1 FY2026 earnings call [5].

The gap between the two bars is the load a single customer carries.

Source: Q3 FY2026 earnings call, CFO remarks [6].

An analyst put the concern to management directly. On the October 2025 call, JPMorgan's Mark Murphy described "a new era where the contractual commitments from a small number of AI natives are just incredibly large… contracts worth hundreds of billions of dollars that are 20 times their current revenue scale," and asked how Microsoft evaluates the ability of those companies to follow through and "placing guardrails on customer concentration for any single entity" [7]. The $250 billion Azure commitment against OpenAI's revenue base is roughly that multiple.

In the December 2025 quarter investments in OpenAI added $10.0 billion to other income and $1.02 to diluted EPS, lifting reported EPS to $5.16 (+60%) against adjusted EPS of $4.14 (+24%), while in the March 2026 quarter that same OpenAI relationship carried Microsoft's commercial RPO to +99% growth ($627 billion) versus only +26% excluding OpenAI and turned commercial bookings to -4% including OpenAI versus +7% without it — and Microsoft's equity-method loss from OpenAI was about $4.1 billion in the September 2025 quarter alone, nearly the entire $4.7 billion 'Other, net' drag recorded across all of fiscal 2025, even as Microsoft surrendered its right of first refusal to be OpenAI's compute provider as OpenAI's Azure commitment rose by an incremental $250 billion [8] [9] [10].

What the counterparty's economics look like

Microsoft does not report OpenAI's financials, but its own equity-method accounting is a window onto them. Microsoft holds an approximately 27% as-converted interest in OpenAI, accounted for under the hypothetical-liquidation-at-book-value (HLBV) method, and has committed $13 billion of funding, of which $11.8 billion was funded by March 2026 [11]. Across all of fiscal 2025, the "Other, net" line — which Microsoft says "primarily reflects net recognized losses on equity method investments, including OpenAI" — was a $4.7 billion drag [12].

Then the pace jumped. In the September 2025 quarter alone, the OpenAI investment sat in other income at roughly $4.1 billion of loss — nearly the entire prior full year in three months — and CFO Amy Hood confirmed "that increased loss was all due to our percentage of losses in OpenAI due to the equity method" [13]. HLBV recognition does not map linearly to OpenAI's total losses — Microsoft absorbs a share weighted by its liquidation priority, not a flat 27% — so the figure understates as much as it reveals. But the direction is unambiguous: the losses Microsoft records from OpenAI are large and accelerating.

External reporting fills in the counterparty's own scale. OpenAI was valued near $500 billion in its late-2025 recapitalization, with reported annualized revenue around $20 billion at the end of 2025 and cash losses on the order of $10 billion or more a year; its own circulated projections show losses continuing to build through 2028 before a targeted turn to positive cash flow near the end of the decade. Against that revenue base, OpenAI has signed roughly $1.1 trillion of multi-year compute commitments across a handful of suppliers.

Source: external reporting on OpenAI's disclosed infrastructure agreements, 2025–2026; figures are company announcements and press accounts, not Microsoft filings, and are estimates.

The arithmetic is stark: about $1.1 trillion of contracted spending against roughly $20 billion of revenue, funded by a plan that depends on both scaling and continued fundraising. Microsoft's $250 billion is the third-largest single line in that stack — and it is the RPO that Microsoft's investors are, in part, capitalizing when they underwrite the AI build.

The loop, and the backstop Microsoft gave up

The relationship is also circular in a way the filings make explicit. Microsoft funds and part-owns OpenAI, and OpenAI's Azure purchases flow back into the RPO backlog that underwrites Microsoft's capex — the same dollars appearing as Microsoft's investment on one side and Microsoft's contracted revenue on the other. The wider version of this loop, in which chip vendors, cloud providers, and model labs finance each other's purchases, has drawn open comparison to late-1990s vendor financing.

The October 2025 agreement sharpens the demand-durability question in a way that belongs to this chapter's lens. As OpenAI's dollar commitments rose, Microsoft gave up a protection: it "will no longer have a right of first refusal to be OpenAI's compute provider" [14]. The commitments grew even as exclusivity loosened. The same recapitalization also produced a one-time equity-method gain and reset the carrying value of the stake, but those accounting mechanics bear on reported earnings rather than demand durability and are taken up in The OpenAI Stake.

What makes the risk manageable — and what would change the read

Management's answer to the concentration question is the strongest counter-case, and it is a substantive one. Nadella argued that the "risk of concentrating on a single customer can be mitigated" because Microsoft builds "a highly fungible fleet" usable for first-party products, enterprises, and other digital natives alike [15]. Hood added the practical point that capacity commitments convert to cost on a delivery schedule: the "CPU and GPU and the storage gear doesn't come into play until the contracts start happening," giving Microsoft "a lot of lead time" to reassess as each tranche approaches [16]. If OpenAI faltered, the datacenters it would have filled could in principle serve the broader Azure book — where demand still "exceeds available capacity" and the AI run-rate reached $37 billion, up 123% [17]. Fungibility is the reason a counterparty shock need not become a stranded-asset event.

The bounding facts on the other side: RPO growth outside OpenAI is a healthy-but-ordinary 26%, so the 99% headline that most visibly "backs" the capex is not repeatable without the same customer; the right-of-first-refusal Microsoft surrendered was exactly the mechanism that would have redirected OpenAI's compute back to Azure; and re-tenanting purpose-built AI capacity at the same price and utilization is an assertion, not yet a demonstrated fact. The read that fits the evidence: the demand is real and partly fungible, but its growth is concentrated in a counterparty that is years from self-funding, so the return on the AI build is more exposed to one customer's trajectory than the consolidated backlog implies.

What would move that read, in Microsoft's own disclosures:

• RPO growth excluding OpenAI — if the ex-OpenAI line re-accelerates toward the headline rate, the concentration is diversifying away; if it stalls near or below the low-20s while the headline stays inflated, the reliance is deepening [18].
• The equity-method loss run-rate — a quarterly OpenAI drag that keeps climbing from the ~$4.1 billion September-quarter mark signals the counterparty's burn is widening faster than its revenue [19].
• Azure ex-AI utilization — Microsoft's fungibility defense only holds if non-OpenAI, non-AI demand can absorb capacity; continued "exceeds available capacity" commentary supports it, a swing to excess capacity would not [20].