Anthropic has just taken over Microsoft’s abandoned $45 billion data center lease, with 460 megawatts of computing power, and next year will use NVIDIA’s Vera Rubin chips. Yet on the very same day, Microsoft’s stock rose 0.9% in the US market.
My take: Microsoft’s AI compute strategy is shifting from “heavy-asset buyouts” to “light-asset revenue sharing.” It exits large-scale leasing arrangements while letting others’ models run on Azure.
Look at the three lines together: it pulled out of the West Virginia data centers, which Anthropic took over; it was reported that it negotiated with Moonshot for Kimi K3 on Azure on a revenue-sharing basis; and it partnered with Humain to provide Allam services through Microsoft Foundry.
The question now is: if cloud providers no longer stockpile compute with heavy spending, but instead make money through hosting and revenue splits, can NVIDIA’s chip orders still keep beating expectations? Is this Microsoft being meticulous with cost control—or is it the first crack signaling a peak in AI capital expenditures?
My take: Microsoft’s AI compute strategy is shifting from “heavy-asset buyouts” to “light-asset revenue sharing.” It exits large-scale leasing arrangements while letting others’ models run on Azure.
Look at the three lines together: it pulled out of the West Virginia data centers, which Anthropic took over; it was reported that it negotiated with Moonshot for Kimi K3 on Azure on a revenue-sharing basis; and it partnered with Humain to provide Allam services through Microsoft Foundry.
The question now is: if cloud providers no longer stockpile compute with heavy spending, but instead make money through hosting and revenue splits, can NVIDIA’s chip orders still keep beating expectations? Is this Microsoft being meticulous with cost control—or is it the first crack signaling a peak in AI capital expenditures?