No Buyer for a $40,000 GPU: Microsoft Winds Down Oracle and Puts a Rental Rate on 'Good Enough' AI Compute

Bloomberg reported on Nov 6 that Microsoft is winding down its 2023 arrangement to rent AI computing from Oracle. The deal was not renewed, Microsoft has been turning off the rented capacity in recent months, and the hardware no longer fit its specifications; the size of the contract was never disclosed. The arrangement dated back to the 2023 tie-up in which Oracle's superclusters, scaled to 32,768 A100 or 16,384 H100 GPUs, were used by Microsoft for workloads including Bing conversational search. Oracle's shares fell 4% on the report.

The resale channel that exists is a brokerage channel, and its price band is wide enough to tell you there is no clearing price underneath it. New 80GB H100 units list at $25,000 to $40,000 in 2026, refurbished units at $21,000 to $34,000, and used units at $15,000 to $28,000. In practice, a GPU in a powered rack is a location, not a product: it cannot be sold without unbolting it from a liquid-cooled rack, rewiring it and re-commissioning it, so the market that matters is the market for hours.

The failure-rate spread is what kills the institutional buyer. A rigorous refurbisher runs at roughly 1.5% to 2% annual failures while a broker marketplace runs at 4% to 5%, and a buyer with no way to verify a fleet's history cannot underwrite the spread. The practical consequence shows up in the discount: a 16% sticker saving on a refurbished card becomes a 9% effective advantage over 24 months once warranty gaps and failure rates are counted, roughly $156,000 on a $1.8M, 64-card cluster. The 2026 pricing guide's own conclusion is that this resale channel functions as a timing tool rather than a savings tool, and it expects resale prices to fall another 10% to 20% once Blackwell supply ramps.

Oracle spent $75.7B on property and equipment in the twelve months to Aug 31, 2026 against $46.9B of operating cash flow, and funded the gap with $37.2B of net new debt and $20.1B of common equity, including a $20B at-the-market program in the September quarter alone. Net property and equipment now stands at $127.8B, of which $77.8B is computers, network and machinery: servers and networking equipment that the company depreciates over six years. Another $48.5B sits in construction in progress, not yet earning anything.

The consequence of Microsoft's exit is not yet an impairment, and that is the point. Oracle's own filings warn that a change in customer demand on long-term cloud arrangements could require an evaluation of the recoverability of related long-lived assets, yet the depreciation charge has already doubled, to $3.2B in the September quarter from $1.4B a year earlier. SoOracle has to refill released racks at a price it chooses, and the only price it has published is the one Bloomberg reported in September: contracts on four-year-old GPUs renewing at a 20% premium to the original deal price.

The same shock lands on one balance sheet and misses another, and the dividing line is depreciation intensity. When more than half of revenue is a non-cash charge on hardware bought at today's prices but rented at yesterday's prices, a fall in the rental rate passes through to operating income almost one for one. The same price reset is existential in the top rows of the table below and a rounding error in the bottom three.

Set against the neocloud sell-off, the parts of the chain that sell space, power and wafers are not showing a glut. Equinix reported record annualized gross bookings of $533M in the September quarter, up 35% year over year, on what the company called strengthening pricing, and is delivering nearly 40% more cabinets this year than it originally planned. Digital Realty signed roughly $855M of total leasing bookings in the same quarter, its booked-not-billed backlog rose30% sequentially to $2.5B at 100% share, and renewal rents on leases signed in the quarter rose 11.8% on a cash basis and 17.0% on a GAAP basis. Wafer demand does not care which generation of accelerator runs; it cares how much compute gets sold.

That is an inference rather than a reported fact, and it is the load-bearing one: a discount on 'good enough' compute cannot go deep while memory prices are producing a 79.7% operating margin on the stacks those accelerators ride on. The practical version is that the floor under AI compute prices is rising even as the ceiling on yesterday's silicon falls, which turns the release of old capacity into a squeeze on operators rather than a deflation event for the stack.

Fact: Microsoft walked away from a 2023 Oracle lease built on older-generation NVIDIA silicon and is trimming a neocloud roster of roughly half a dozen partners; the hardware stays inside Oracle's data centers. Fact: Oracle carries $664B of backlog with only 13% converting within a year, $127.8B of property and equipment on a six-year life, and $34.2B of unconditional purchase obligations for components and power. Inference: because no hardware channel exists to arbitrage released capacity, the marginal price of 'good enough' compute is set by rental hours and utilization. Speculation: neocloud equity keeps de-rating until renewals are repriced, and the first hard data point arrives with IREN's January handover.