The Oracle earnings call had, by the count of the business press, three stories to tell: $638 billion in remaining performance obligations, a promise of double-digit growth in the health unit, and a transition to “outcome-based pricing” that the CFO says is “resonating very, very well with our customers.” The part of the earnings call that matters is the interaction among the three — and the interaction, read carefully, is the hyperscaler lock-in play executed without a single dollar of the hyperscaler’s own capital at risk.
It is true, in the narrow sense in which a supply-chain manager evaluates a bill of materials, that Oracle’s procurement management had a defensible operational logic: insulating the company from inflation in component costs, particularly the acute shortage of high-bandwidth memory that we flagged when the AI memory chip supply crunch deepened in early June. The distinction between fixed-price contracts for locked component costs and floating contracts for variable ones is clean enough on paper. The trouble is that the capital expenditure Oracle is managing off its own balance sheet is landing directly on the enterprise customers whose prepaid commitments appear in the same earnings release as a triumph of revenue visibility.
Let me be precise about what is being claimed. Oracle’s remaining performance obligations — the contracted, legally enforceable future revenue pipeline — stood at $638 billion, a figure roughly eleven times the company’s $57.3 billion in fiscal 2026 revenue. The company disclosed, almost in passing, that $75 billion of that total comes from customers who have either prepaid for GPUs or provided their own. “Prepaid and customer supplied hardware portions of our large AI contracts now total $75 billion,” the release reads, with the gloss that this “substantially reduces the amount of capital Oracle must raise to build out our AI datacenters.”
The gloss has it exactly backward. The $75 billion does not reduce the capital Oracle must raise. It replaces it — with the customers’ own capital, committed under contract — while the infrastructure the customers are paying for remains Oracle’s to own, manage, meter, and gate. Oracle keeps the cloud layer that makes the GPU accessible, the database that stores its output, the identity-and-access management that governs who gets to run a workload and under what conditions, and, as the call made clear, the “outcome-based commercial models that align pricing directly to the value derived.” Translate the last item: Oracle will charge by the value the model produces, rather than by the seat, once the customer is running at scale and cannot easily leave.
This is a chokepoint-capitalism maneuver — seizing a gatekeeper position that forces all transactions through Oracle’s meter — financed with the customers’ own balance sheets. The customer — the AI lab, the enterprise scaling a model to production, the sovereign cloud buyer — signs a contract obligating them to provide the compute substrate. The compute substrate is the bottleneck on which the whole model-training-and-inference superstructure depends, and the customer is providing it because Oracle has convinced them that buying the same GPUs and operating their own datacentre is slower, more expensive, or both. At contract-signing, the customer believes they are purchasing speed-to-market and offloading operational complexity. What they are actually doing is handing Oracle a pre-funded, customer-owned hardware layer running inside Oracle’s managed environment, with the metering and pricing relationship between the two still being written — the move to outcome-based pricing, the CFO noted, is “early days.”
Once the customer’s GPU is installed inside Oracle’s data centre, the customer’s negotiating position erodes along two axes simultaneously. The first is the familiar cloud-switching-cost axis: data, model weights, APIs, and access-control configurations all thicken around the provider, and the thickness grows over time. The second axis is the one the earnings call revealed: the customer owns the hardware that would be the hardest asset to duplicate in a migration, but the hardware is sitting in Oracle’s cage. The customer cannot take it back without disrupting production. The customer cannot replicate the managed services wrapping the hardware at equivalent latency without rebuilding the entire operational stack. And the customer, by virtue of having signed the prepaid-GPU contract, has already paid Oracle’s capital-expenditure bill. Every dollar of margin Oracle extracts from the outcome-based pricing layer above the customer-owned hardware is a dollar the customer paid to be in a position to be charged.
The concept of twiddling, as Cory Doctorow names it — the continuous, computer-mediated adjustment of prices, outputs, and visibility — is the operational consequence of this shift. When a vendor moves to “outcome-based” billing for a black-box algorithmic system, the definition of the measured “outcome” becomes the variable that is continuously adjusted against the business customer, making it impossible for the tenant to determine whether the throughput they are renting is actually delivering the value the contract claims. Oracle’s healthcare division promises that a new AI version of the Cerner patient-care management system will “revolutionize” clinical workflows by offloading administrative burdens from doctors and nurses. The political economy of this claim is standard for the platform era: the platform ingests the hospital’s clinical data as a training corpus, captures the resulting efficiencies not as a public-health commons but as proprietary process improvements locked behind renewed licenses, and sells the optimized system back to the medical provider under a continuous outcome-based toll. The enshittification trace is legible in the filing: capture the domain expertise of the clinician, eliminate the competing modular software vendors by integrating every function into a single opaque stack, and twiddle the toll on the resulting diagnostic throughput until the margin satisfies the shareholder.
The comparison to what the telecom industry calls “vendor financing” — the equipment manufacturer lends the operator the money to buy the manufacturer’s own kit, sinking the operator into debt while locking it to the manufacturer’s upgrade cycle — is instructive but incomplete. In the vendor-financing model, the operator at least owns the equipment after the loan is repaid. In Oracle’s model, the customer owns the hardware but operates it inside Oracle’s walled garden, paying Oracle for the privilege of using its own assets. A closer analogy is the farm-implement model John Deere has spent the last decade perfecting: the farmer owns the half-million-dollar tractor, but the tractor cannot be repaired, modified, or operated outside Deere’s authorized service network because the software locks make ownership of the physical asset contingent on ongoing permission from the manufacturer. Oracle’s prepaid-GPU model extends the same logic to the capital layer.
The mechanism, as Doctorow’s four-forces framework makes clear, is the suppression of adversarial interoperability — the legal and technical erasure of the customer’s ability to extract its own hardware, its own data, and its own workloads from the provider’s enclosure. In cloud computing, adversarial interoperability would mean the ability to pull customer-owned hardware, model weights, and workloads out of a provider’s stack without rebuilding the integration layer from scratch. Every cloud provider has spent the last decade making that form of interoperability technically impossible and contractually forbidden. Oracle’s innovation is to make the customer pay for the hardware first, so that the cost of walking away includes the sunk capital locked inside the provider’s cage.
The broader capital markets are beginning to price this shift. Super Micro Computer, a key supplier of AI servers, announced plans to raise $7 billion through new equity and convertible preferred stock, sending its shares down 18 percent on dilution fears. South Korea’s SK Telecom is pursuing the identical structural play, with analysts projecting a 20-trillion-won revenue stream by morphing its gigawatt-scale AI factory partnership into a GPU-as-a-service model. The modern nomenclature obscures an older reality: it functionally resurrects the commercial time-sharing mainframe model of the 1960s, renting high-performance compute cycles over a network at a premium that vastly exceeds the amortization cost of the physical chips. The GPU-as-a-service abstraction is a twiddler’s utopia. Because the hardware is virtualized and accessible only through a network interface, the service provider can dynamically price every single query, shifting the financial risk of idle compute entirely onto the enterprise developer. Meanwhile, WiseTech Global is finding its logistics-software customers actively resisting a transition to a new pricing architecture, requiring greater financial incentives to migrate legacy accounts off the old billing model. The friction is entirely predictable. Enterprise customers tolerate the platform lock-in only until the economic rent extracted by the vendor exceeds the operational friction of leaving it.
Oracle’s disclosure that large AI contracts now total $75 billion in prepaid and customer-supplied hardware is, read alongside the outcome-based pricing announcement, the clearest signal the hyperscaler business model has produced in the current cycle. The market’s logic is this: the buildout cost is too large for any single firm to finance from its own balance sheet, so the cost will be distributed across the customers who will then be locked into the providers who financed the last round. The distribution of cost to the customer is framed as partnership. The lock-in is the business model. The outcome-based pricing that “aligns pricing directly to the value derived” is the mechanism by which the provider captures the surplus the customer’s own hardware makes possible, after the customer has already paid for the hardware.
There is an open rulemaking docket at the Federal Trade Commission regarding unfair methods of competition, and a parallel consultation ongoing at Canada’s Competition Bureau on digital market conduct. The dockets matter because deadlines are the only part of regulatory processes that the regulated actually respect. The contracts are public. The capital shifts are documented in the quarterly 10-K filings. The renewal notices will arrive six months from now. There is a saying my grandfather used, something like the man who pays for the rope does not get to complain about the knot. The work is to read the contract before signing it.