Microsoft signed for computing from Oracle in 2023 because it could not get enough of it fast enough. It is walking away for something close to the opposite reason.
Bloomberg reported on Nov 6 that Microsoft is winding down the AI compute lease it signed with Oracle three years ago, on the grounds that capacity rented from an outside provider has fallen behind on chip generations. Oracle shares fell about 4% and closed at a two-year low.
The logic of a lease is that it fixes a price for a period. The logic of a chip is that each generation makes the last one cheaper to run per unit of work. Put the two together and a long contract on older silicon becomes a liability the moment a faster part ships, because the tenant is paying a fixed rate for capacity that produces less per dollar than the equipment it could install itself.
That is the mechanism the report describes, and it is the reason the lease was signed in the first place. In 2023 the binding constraint was access. Microsoft needed accelerators and could not wait for its own buildings, so it rented floor space from a competitor and paid for the privilege. Three years later the constraint is different: the company has its own campuses, and the rented ones are no longer the fastest place to put a workload.
The move is part of a pattern rather than a one-off. The day before, Microsoft abandoned a planned expansion of a lease with Nscale in Texas and cut its capital expenditure guidance by $5 billion. Two retreats in two days, in the same direction, from the same budget line.
For Oracle the reaction was immediate and out of proportion to any revenue that has actually disappeared. A lease that is being phased out still runs until it ends, and the capacity does not vanish. What the market repriced is the assumption underneath Oracle’s cloud business: that a small number of very large AI contracts would keep growing, and that the customers who signed them had no cheaper alternative.
Concentration cuts both ways. A landlord with three enormous tenants has a backlog that looks like growth and a renewal book that looks like risk, and the risk arrives years before the revenue does. Microsoft’s decision tells other tenants that the exit exists, which changes how the next lease gets priced.
The deeper problem is depreciation, and it applies to every company in the buildout. A data center shell lasts decades. The accelerators inside it are economically obsolete in a handful of years, and whoever owns them absorbs that loss. A company that owns its own fleet can time the refresh against its own product cycle. A tenant on someone else’s schedule cannot, and it pays for the difference in the rate.
This is why the numbers that get reported as AI capex are misleading as a measure of capacity. Money spent on buildings and money spent on chips have different lifetimes, and a company that shifts from renting to owning is not necessarily spending more; it is taking on the depreciation instead of paying a margin for someone else to carry it.
The bullish reading is that the capacity does not go away, it gets re-let. Older accelerators are still useful for inference on smaller models, for training runs that do not need the newest part, or for customers who care about price rather than speed. A lease that Microsoft does not want may be exactly what a mid-sized enterprise buyer needs, and Oracle can sell the same floor twice at a lower rate.
That reading has a limit. The rate at which older silicon gets re-let depends on how fast inference demand grows, and inference demand is the part of the market with the least visibility. Everyone can see the training runs. Nobody can forecast how many tokens a company will generate next year, because that number depends on products that have not shipped.
There is a less comfortable interpretation of Microsoft’s two announcements. A $5 billion cut to capex guidance, arriving days after the midterms, can be a sourcing decision or a demand signal. The company has not broken out which. If its own AI products are selling more slowly than planned, the cheapest lever is to stop paying for capacity it does not need, and the Oracle lease is the easiest contract to let lapse.
Either way, the terms of the next round of leases will look different. Tenants who watched a three-year contract get unwound will ask for shorter terms and for refresh clauses that tie the price to the silicon, not to the calendar. Landlords will ask for higher rates to compensate for the risk that their best customer leaves early.
The companies caught in the middle are the neoclouds, whose business rests on long contracts with large tenants and whose debt is priced against those contracts. CoreWeave’s bonds traded at yields above 16% this week, and the pressure in that market predates Microsoft’s decision. A lease is only as good as the tenant’s willingness to keep paying for capacity it can no longer use, and that willingness is now a question analysts will ask about every contract in the sector.
Oracle’s answer will come in its next set of results, when it reports whether the capacity is re-let and at what rate. Microsoft’s will come with its own, in a capex number that either holds the lower guidance or slips again.
