Microsoft is turning off the AI computing power it rents from Oracle little by little, and the contract will not be renewed when it expires.
This agreement was signed in 2023 to supplement Bing’s computing power. Three years later, the servers provided by Oracle are still equipped with the previous generation Nvidia chips. In the eyes of Microsoft, they can no longer meet the current technical specifications, and it is not cost-effective to keep them.
This is not an isolated action. Not long ago, Microsoft just tightened its review of cooperation with a group of "new cloud" computing power suppliers, and now it is Oracle's turn. The direction is clear: those rented from outside and whose specifications cannot keep up should be cut first.
From Microsoft's perspective, this account is easy to calculate. The data center I am building uses the latest generation of chips, and the cost per unit of computing power is lower; continuing to rent an old card from three years ago is equivalent to spending money to buy inefficiency. What's more, Bing's business model has also changed in recent years. The computing power signed for it back then is no longer suitable for today's uses and specifications.
The timing is tough for Oracle. It has invested a lot of capital in AI data centers in the past two years. It just canceled a US$7 billion chip lease from Tencent a few days ago, and now it loses Microsoft as an old customer.
The speed of chip generation replacement is turning "computing power leasing" into a very risky business. A contract signed three years ago may no longer want the equipment when it expires.