Oracle Wasn’t Weird Enough
In September, Oracle looked like an outlier: a company trying to finance the future with the performance of demand. Six months later, it looks more like the least disguised version of a sector-wide operating model. [1][2]
Back in September, I wrote about Oracle as if it were a strange object: a software company wandering into project finance by way of AI spectacle. The immediate question was whether Oracle was uniquely absurd or merely early. That distinction matters more now than it did then. At the time, it was still possible to tell a comforting story in which Oracle was just Larry Ellison doing Larry Ellison things at industrial scale. It is harder to tell that story now. [1]
Because March did not bring the clean rebuttal. It brought the uglier outcome. Oracle posted a strong quarter: $17.2 billion in Q3 revenue, remaining performance obligations up 325% year over year to $553 billion, and a raised fiscal 2027 revenue target of $90 billion. Investors, being romantics when sufficiently oxygen-deprived, sent the stock up more than 8% after hours. On the surface this looked like Oracle beating the criticism. In practice it looked more like the criticism had matured into a business model. [2]
That is the key point. The quarter did not invalidate the September diagnosis. It clarified the mechanism. Oracle’s own November 2025 10-Q had already shown the shape of the problem: $523.3 billion in remaining performance obligations, with only about 10% expected to be recognized over the next twelve months; $248 billion in additional lease commitments, mostly for data centers and cloud capacity, expected to commence between Q3 fiscal 2026 and fiscal 2028 and run for fifteen to nineteen years; and roughly $106.1 billion in senior notes and other long-term borrowings outstanding by late November. That is not just growth. That is duration, leverage, and obligation arriving in heavy material form. [3]
So I think the September pieces and the March update are describing the same phenomenon from different altitudes.
“Latent capitalism” named the regime in which projected demand, strategic positioning, and contractual theater could function as provisional capital. It described a system in which commitments did not merely point to future revenue; they could be treated as financing substrate in the present. “The Oracle Papers” pushed that logic further, arguing that Oracle’s bookings were not just a business metric but a kind of collateral performance, a closed loop in which the appearance of future capacity helped finance the construction of the capacity that would later ratify the appearance. [1]
What March adds is not contradiction but hardening.
Latent capitalism is the grammar. Zombie infrastructure is the building.
One describes the valuation regime: a world in which commitments, narratives, and strategic aura can be booked socially before they are fully validated economically. The other describes what happens when that regime congeals into leases, debt, power arrangements, physical campuses, hardware reservations, and contractual lock-in too large to unwind gracefully. The old point was never that Oracle was simply faking demand. The point was that performance had become entangled with financing so tightly that the line between signal and substrate was beginning to disappear.
The bridge between those two levels is temporal. The mechanism works because the commitment horizon outruns the verification horizon. By the time the real utilization picture arrives, the leases are signed, the debt is placed, the sites are underway, the counterparties are committed, and the strategic story has already been normalized. Future necessity is imported into present solvency. Verification comes later, if at all. The lag is not a bug in the story. The lag is the story.
That is why Oracle’s latest quarter should be read less as vindication than as industrialization. The company’s results helped calm investor concerns about whether its multibillion-dollar AI push would generate profits quickly enough. But in the same breath Oracle was describing smaller engineering teams, AI coding tools, and restructuring that lets it build more software with fewer people, while also planning thousands of job cuts amid cash strain from the data-center buildout. This is not a side effect of the model. It is part of the model: labor is compressed so capital can be extended. The regime is eating its own workforce to feed its capex cycle, then calling the result efficiency. [2][4]
This is also where the Microsoft and Amazon comparison matters, because it turns Oracle from a weird case into a diagnostic specimen.
Microsoft is not cleaner in kind. It is cleaner in presentation. Its commercial remaining performance obligation reached $625 billion with a weighted average duration of about two and a half years, and roughly 25% of that is expected to be recognized in the next twelve months. That is a much healthier temporal profile than Oracle’s. It is also still capable of self-funding the distortion: quarterly capex of $37.5 billion, operating cash flow of $35.8 billion, and free cash flow still positive even after the spending surge. But it comes with an asterisk large enough to have its own weather system: about 45% of that commercial RPO is from OpenAI. The distortion is still there. Microsoft can just afford better tailoring. [5]
Amazon sits in a different position again. Its disclosed long-term customer commitments were about $244 billion at year-end, primarily AWS, with a weighted average remaining life of 4.1 years, and Amazon explicitly notes that the timing of revenue recognition depends on customer usage. That makes the backlog less theatrical than Oracle’s and less concentrated than Microsoft’s current OpenAI bulge, but it does not make the larger civilizational move any less obvious. Amazon generated $139.5 billion in operating cash flow in 2025 and spent $128.3 billion in cash capex, then told investors to expect roughly $200 billion in capital expenditure in 2026. Its year-end commitments totaled about $439.7 billion, including $96.4 billion in leases not yet commenced and $84.8 billion in unconditional purchase obligations. This is not an app company having a good run. This is utility formation with better branding. [6]
Seen this way, Oracle is not doing something categorically different from its peers. It is doing the same thing with less camouflage. Microsoft and Amazon make the regime look stable because they have more cash, more ballast, and broader underlying businesses. Oracle makes the regime visible because it has less room for aesthetic softening. The same migration is underway across all three: from software margins toward infrastructure duration, from flexible cloud mythology toward fixed commitments, from “services” language toward something much closer to grid logic. [3][5][6]
That is why I am less interested now in the old temptation to call this another 2008. The mortgage parallel was useful as a shock device, and The Oracle Papers was right to notice the way future claims were being packaged, narrated, and redistributed. But the more precise diagnosis is not “fraudulent assets waiting to be exposed.” It is institutionally sponsored duration laundering. [1][3]
The assets are real. That is the problem.
The facilities will be built. The chips will be installed. The power will be contracted. The debt will be serviced. The obligations will sit there on the timeline like concrete. And because the buildout is increasingly strategic, politically legible, and narratively protected, the system may never be forced into a clean market correction at all. It can remain overextended, under-verified, and economically ambiguous for quite a long time, provided enough actors agree that the future must justify the present eventually. [3][6][7]
That is what makes the sector-level numbers so revealing. The major hyperscalers are expected to pour about $650 billion into AI-related infrastructure in 2026, up from roughly $410 billion in 2025. Capex is rising faster than the comforting stories told about it. What matters is not whether every dollar is irrational. What matters is that the sector is reorganizing itself around a logic in which duration, strategic necessity, and financing access can substitute for near-term proof. That is not a market disciplining itself in real time. It is a strategic overbuild regime trying to make deferred verification look like inevitability. [7]
In September, Oracle looked like a company treating future demand as present collateral. By March, it looked more like the first honest portrait of an industry doing the same thing at scale. [1][2][3]
The pathology stopped looking eccentric and started looking normal.
Oracle wasn’t weird enough.
Notes
Previous pieces: Latent Capitalism and The Oracle Papers. (neutralzone.substack.com)
Oracle’s Q3 FY26 results: revenue of $17.2 billion, RPO of $553 billion, FY27 revenue target raised to $90 billion; Oracle also said large AI contracts were increasingly supported by customer prepayments or customer-supplied GPUs, and that AI code generation was allowing smaller teams to produce more software. Reuters also reported the stock rose more than 8% after hours on the earnings release. (Oracle Investor Relations)
Oracle’s November 30, 2025 10-Q: $523.3 billion of remaining performance obligations, with roughly 10% expected within 12 months; $248 billion of additional lease commitments related mainly to data centers and cloud capacity with 15- to 19-year terms; and $106.1 billion of senior notes and other long-term borrowings outstanding. (SEC)
Reuters on Oracle’s planned job cuts amid cash strain from the AI data-center expansion. (Reuters)
Microsoft FY26 Q2: commercial RPO of $625 billion, weighted average duration of about 2.5 years, roughly 25% recognized within 12 months, about 45% of commercial RPO from OpenAI, quarterly capex of $37.5 billion, operating cash flow of $35.8 billion, and free cash flow of $5.9 billion. (Microsoft)
Amazon 2025 10-K and February 2026 reporting: operating cash flow of $139.5 billion, cash capex of $128.3 billion, about $244 billion of long-term customer commitments with a weighted average remaining life of 4.1 years, total commitments of about $439.7 billion including $96.4 billion of leases not yet commenced and $84.8 billion of unconditional purchase obligations; Reuters reported Amazon expected about $200 billion in 2026 capex. (SEC)
Reuters on AI infrastructure spending: Bridgewater estimated Alphabet, Amazon, Meta, and Microsoft would collectively invest about $650 billion in AI-related infrastructure in 2026, up from about $410 billion in 2025. Reuters also reported investor concern that hyperscalers are moving from more asset-light models toward materially more capital-intensive ones. (Reuters)