Microsoft switches off Oracle-rented AI servers as old Nvidia GPUs lose economic edge
rohanpaul_ai · x · 2026-11-07
Per Bloomberg, Microsoft declined to renew a 2023 contract and is switching off AI servers rented from Oracle, because the hardware runs older-generation Nvidia chips that no longer match its specs.
The economics: Microsoft pays per GPU-hour. A Blackwell GPU yields several times more inference output than an H100 at similar power, so each Hopper/Ampere hour buys far less work. Frontier MoE models need lots of memory and fast interconnects — an H100 has 80GB and fast links to only 8 GPUs, while Blackwell racks link 72 GPUs with 192-288GB each; big models either don't fit or run slowly on old boxes.
Oracle says AI demand remains strong, and in September renewed contracts for 4-year-old GPUs at a 20% premium to the original price.
Related event: Microsoft Winds Down Oracle AI Compute Lease, Oracle Shares Fall(7 posts)→
More from Infra
- Ben Bajarin: $9.6T AI capex vs $4.1T revenue — AI pays for itself, but later than you think — BenBajarin · 2026-11-07
- Apple details AFM context sizes; fm serve spins up a local Chat Completions API — rxwei · 2026-11-07
- Vera Rubin NVL72 hits 5x GB300 per-GPU throughput on DeepSeek V4.1 Flash — ArtificialAnlys · 2026-11-07
- DeepSeek V4.1 Flash gets up to 1.5x faster on GB300 NVL72 in one month — ArtificialAnlys · 2026-11-07
- HiSilicon launches first CMOS image sensor: 50MP with 1.2um pixel pitch — pstAsiatech · 2026-11-07
- Open-source tool shows system prompt and tools can eat 98% of API tokens — behrouzbk · 2026-11-07