Counterparty, Not Product
Your AI vendor has its own funding dependencies and its own fragility. No individual risk assessment can see the aggregate. Who owns vendor financial durability inside your organization?
In the first quarter of 2026, Alphabet reported net income of $62.6 billion, up 81% from the year before. The same filing reported a $36.9 billion net gain on equity securities, primarily unrealized gains on non-marketable holdings. Alphabet says that gain added $28.7 billion to net income and $2.35 to diluted earnings per share. The quarter remained strongly profitable. The headline still depended materially on marks on holdings the company had not sold.
There is nothing improper about that. The filing reports the gain through income and spells out its effect. This is not an allegation about the accounting, and it is not a stock call. I want the operating question: what does the financial structure of a critical AI counterparty mean for the organization that has built a dependency on it?
The answer begins with what is visible in filings. No insider access required.
First, follow the financial relationships.
The relationships are not simple. A Federal Trade Commission staff study found that major cloud-and-AI partnerships can combine equity or revenue-sharing rights with commitments requiring developers to spend a large portion of a partner’s investment on that partner’s cloud services. Those commitments concern real cloud services and compute capacity. So investment exposure and commercial cloud commitments can sit inside one relationship. And the investment side can move reported earnings. The public disclosures show the parts, but they do not provide an operating view of the dependency.
Then there is the size of what has been promised. As of the quarter ended June 30, 2026, Microsoft's commercial remaining performance obligations, meaning contracted future revenue not yet recognized, stood at $678 billion, up 84% year over year. Oracle ended its fiscal year on May 31 with $638 billion of remaining performance obligations, up 363%.
The disclosures provide partial attribution, and how much they provide changed this year. Microsoft says its remaining performance obligations (RPO) include OpenAI and that all sequential growth in the June quarter came from customers outside frontier-model companies. In its fiscal 2026 annual report it went further, disclosing for the first time $24.1 billion of revenue from commercial arrangements with OpenAI and $6 billion still owed by that counterparty at year end. Oracle says most of its Q3 and Q4 RPO increase came from large-scale AI contracts. Neither company breaks out the composition of the backlog itself. More than $1.3 trillion of contracted future revenue, one counterparty now named with a number, and no breakdown for the remainder.
Each figure in the last three paragraphs is a dated snapshot. The numbers do not become false after a quarter; they stop describing the current quarter. That matters more than it sounds.
There is a distinction that stock analysts will not make for you.
A liquid public-market position can often be hedged, diversified or sold. An embedded operational dependency cannot. It sits inside approvals, customer interactions and processes that assume the capability will be there. Its exit is a migration project: data, integrations, workflow logic, approvals and time. Same underlying counterparty, entirely different instrument.
What worries me is that we reach for the dramatic analogy and miss the one that actually fits.
The usual comparison is Lehman: sudden, binary, no off-ramp. That is the wrong shape. A better historical pattern is the bond insurers. At the end of 2006, monoline insurers supported $2.5 trillion of insured risk, measured at the par value of the guaranteed securities, including about $800 billion in structured-finance obligations. By January 2008, MBIA and Ambac, the two largest monolines, had a combined $265 billion of guarantees on mortgage securities and other structured products. When their credit strength came into question, the effect traveled through the securities that depended on those guarantees. The analogy is not identity. It is shared dependence on a small set of counterparties and the knock-on effects that individual exposures do not reveal.
Enterprise AI presents a related structural concern. The FTC study examined partnerships among three of the largest cloud providers and prominent AI developers, including contractual and technical switching-cost risks. That does not prove imminent failure. It does mean each organization can review its own vendor and still miss aggregate exposure that appears only across customers, developers and infrastructure providers. No individual risk assessment can see that aggregate.
I am not predicting a failure. The plausible bad outcomes here are duller than a collapse and more likely: pricing moving under funding pressure, service levels drifting, a capability roadmap redirected toward a more attractive market, terms renegotiated after an acquisition. None of those make headlines. All of them land on whoever owns the dependency.
So, who does own it?
Do not assume the answer is already embedded in security review, architecture review or an AI policy. Those controls may address data handling, output risk and technical fit without evaluating the vendor as a counterparty with its own funding dependencies and fragility. Vendor financial durability can fall between procurement, IT, legal and finance. If no one can name the accountable owner and the review cadence, that is your answer.
Three things worth putting in place, none of which require a new function or a consultant.
Name an owner for vendor financial durability, as a risk category distinct from vendor performance and vendor security. One person, identified, with a standing review cadence rather than an annual gesture. If you cannot name that person today, that is the finding.
Write down what breaks first. Take your two or three most embedded AI dependencies and document, concretely, what stops working if that provider becomes unavailable or materially changes terms in ninety days. Then have somebody who did not write it try to break it.
Know your switching cost before you need it. Data portability, fine-tuning, embedded workflow logic, the realistic timeline. A migration cost you have never estimated is not a plan, it is a hope with a spreadsheet attached.
Now back to the shelf life, because it is the actual lesson.
Every current-quarter number in this piece will stop being current within a quarter. The filings will refresh, the backlogs will move, and the marks will be revised in one direction or another. The historical figures will still be true; they will no longer describe the present. And it is not only the numbers that move. Microsoft’s OpenAI revenue figure did not exist as a public number a year ago. What you can know changes on the same schedule as what you know. If your read on vendor durability is a set of figures somebody pulled once, it decays at exactly that speed, and nothing announces the change. This draft itself carried a superseded Microsoft figure until the final source check.
So put the question on somebody’s calendar. Quarterly.
Nobody has to be at fault for that to cost you.
Your AI vendor is a counterparty, not a product.
Sources and the verification ledger
The material factual claims below are checkable in primary or official sources. The release test is not merely whether a number once appeared in a filing, but whether the wording matches what the source actually supports.
Confirmed and used - re-verified August 12, 2026
Alphabet, First Quarter 2026 results (SEC-filed exhibit)
https://www.sec.gov/Archives/edgar/data/1652044/000165204426000043/googexhibit991q12026.htm
Net income was $62.578 billion, up 81%. Net gain on equity securities was $36.915 billion. Alphabet states that the gain increased net income by $28.7 billion and diluted EPS by $2.35. The prior $2.62 adjusted-EPS / consensus-miss claim was not supported by the filing and has been removed.
Microsoft, FY2026 Q4 earnings release (quarter ended June 30, 2026)
https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast
Commercial remaining performance obligations were $678 billion, up 84% year over year. The earnings call says RPO includes OpenAI, all sequential Q4 growth came from customers outside frontier-model companies, and RPO grew 25% excluding OpenAI. This is partial attribution, not composition of the backlog. The widely circulated figure putting OpenAI at roughly 32% of the backlog is analyst arithmetic derived from the 84%/25% pair, not a Microsoft disclosure, and is not repeated in the piece.
Oracle, FY2026 fourth quarter and full year results
RPO was $638 billion at May 31, 2026, up 363% year over year. Oracle says most of the Q3 and Q4 increase came from large-scale AI contracts. This replaced the earlier $552.6 billion Q3 figure.
Microsoft, FY2026 Q4 earnings call (counterparty attribution boundary)
https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4
Microsoft states that RPO includes OpenAI, all sequential Q4 growth came from outside frontier-model companies, and RPO grew 25% excluding OpenAI. No counterparty-level composition is provided.
Microsoft, Form 10-K for fiscal year 2026 (period ended June 30, 2026)
https://www.sec.gov/Archives/edgar/data/0000789019/000119312526323660/msft-20260630.htm
Discloses, for the first time, $24.1 billion of FY2026 revenue from commercial arrangements with OpenAI including revenue sharing, and approximately $6 billion of accounts receivable from OpenAI at June 30, 2026. Also discloses $13 billion committed to OpenAI with $11.9 billion funded. This is a named-counterparty revenue disclosure. It is not composition of the RPO backlog, which is why the article distinguishes the two. Confirmed 2026-08-13.
Federal Trade Commission, AI Partnerships & Investments 6(b) staff report
Supports the description of partnerships combining equity or revenue-sharing rights with cloud-spend commitments and identifies potential contractual and technical switching costs. Findings reflect information through January 2025.
Financial Crisis Inquiry Commission, final report, Chapter 14 (section: Monoline insurers)
https://fcic-static.law.stanford.edu/cdn_media/fcic-reports/fcic_final_report_chapter14.pdf
Chapter 14, section “Monoline insurers: We never expected losses.” Confirmed: “By January 2008, as a result of the stress in the mortgage market, S&P had downgraded 3,389 tranches of residential mortgage-backed securities and 1,383 tranches from 420 CDOs. MBIA and Ambac, the two largest monoline insurers, had taken on a combined $265 billion of guarantees on mortgage securities and other structured products. Downgrades on the products that they insured brought the financial strength of these companies into question.” PROVENANCE: this PDF encodes its numerals to the Unicode Private Use Area (U+F643 through U+F64C for digits 0 through 9, U+F653 for the dollar sign), an artifact of oldstyle figures in the embedded font instance MinionPro-Regular-OV-RKYGCA. Copying or extracting text therefore returns the prose with every digit dropped, silently and without error. The figure was first read visually from the rendered page, then confirmed on 2026-08-13 by decoding the Private Use Area mapping, which recovered all 360 numeric strings in the chapter with no residual unmapped characters. Note for any future reviewer: a text search for “265” in this document returns zero hits and is not evidence against the figure.
International Monetary Fund, April 2008 Global Financial Stability Report
Chapter 2, Figure 2.6 and accompanying notes. States verbatim: at end-2006, monoline insurers supported $2.5 trillion of insured risk (securities at par value), including about $800 billion in structured finance obligations. The par-value qualifier is now carried in the article text. The same passage explains the cross-sector effect of a major monoline downgrade. Confirmed against the document 2026-08-13.
