Oracle's prepay and BYO-hardware terms move the GPU bill onto the customer

The funding decision sits with the buyer, and Oracle's co-CEO spelled out the forms it takes. "Another mechanism is that a customer says, 'Hi, I'd like to pay for the hardware, but use your operational ability and your cloud infrastructure technology assets and your data center to go out and actually turn that into an AI cluster,'" Clay Magouyrk said on the call ^[8]. A third option, he said, is the customer that has raised money and offers to prepay: "'Hi, I would like to pay you upfront as a prepayment,' and in return, that doesn't require you to front the cash to go out and spend your dollars on that capex" ^[9].

A bring-your-own-hardware contract is a hardware purchase plus a services agreement for someone else to operate it. The customer holds the asset and the depreciation schedule. CFO Hilary Maxson said the vast majority of last quarter's new contracts were "prepay or bring your own hardware or similar mechanic, so will not require incremental capital from Oracle" ^[7].

The delivery figures allow a unit check that is usually unavailable. 850MW across more than 300,000 GPUs is about 2.8kW per GPU, counting cooling, networking and everything else drawing power in the building ^[20]. Magouyrk said the quarter's total was "almost three times what we delivered in all of Q4, and 73 percent of the total capacity delivered last fiscal year" ^[3]. Work that second comparison backwards and the whole of last fiscal year came to roughly 1,160MW ^[19].

IaaS revenue of $7.4bn in the quarter annualizes to $29.6bn, so the $664bn of remaining performance obligations is about 22 times the current IaaS run rate ^[21]. Oracle did not break the backlog down by service, so the multiple is a scale marker and not a delivery schedule.

Executives said utilization ran at 97.9 percent, and that GPUs coming up for renewal in the quarter were "renewed or resold at a 20 percent premium to prior contracts," with the majority of that hardware four years or older ^[12][13]. Both numbers come from the company selling the capacity. If they hold up, they are the only public evidence a buyer can put against an internal assumption that a four-year-old accelerator is worth nothing.

What the disclosure withholds matters to anyone signing. There are no customer names, no split between prepay and BYO inside the new bookings, and no dollar figure for the capex Oracle avoided ^[7].

CEO Mike Sicilia added one line for October: "At AI World in October, we will unveil a new agentic AI accelerator poised to redefine how customers deploy Oracle applications faster, simpler, and at a dramatically lower cost" ^[15]. No further details were provided. Oracle has consistently denied to DCD that it plans to develop custom chips ^[16].

Two conditions decide which structure a buyer should accept. The first is whether the buyer can name the workload that fills the cluster for the whole term, because a prepay is paid whether the machines run or not. The second is whether the buyer's cost of capital beats the supplier financing Magouyrk described, in which the vendor pays for capacity as the customer pays for it ^[8]. When both hold, owning the hardware and renting the operations is cheaper than consumption pricing. When neither holds, the premium on an ordinary cloud contract is the price of not carrying the asset. Oracle's guided capex for the year is $90bn to $95bn, against $175bn at Microsoft and roughly $220bn at Amazon including its retail arm ^[10][11].