20 of 720 GPUs went offline without warning. The remedy was a discount at the end of the contract.
That detail sits in a Wisevoter report (21 October 2026, drawing on TokenPost) on US power limits and AI capacity. To get around capacity shortages, companies now spread deployments across 20 to 30 specialised GPU cloud operators. Those operators ask for 20–30% upfront, and large platforms require three-year minimum contracts. Reliability, the report notes, remains a challenge.
20 / 720
GPUs offline in one deployment (≈2.8%)
20–30%
Paid upfront to specialised GPU clouds
3 years
Minimum contract on large platforms
Source: Wisevoter, 21 October 2026. The 2.8% figure is 20 ÷ 720.
A Discount Is Not an Explanation
Read the 720-GPU line again. Roughly 2.8% of a deployment went dark, and the response was settled after the fact: a discount, not a cause.
I keep coming back to this gap in my work with Vitali de Mille. A GPU rarely vanishes without a reason. Before it drops, it usually leaves traces.
What often comes before “without warning”
- Remapped memory rows and uncorrectable ECC errors
- An Xid in the kernel log, sometimes days earlier
- A PCIe or NVLink link that degrades or retrains at lower width
- A temperature or power reading that slowly drifts
“Without warning” often means “without anyone reading the warning.”
When capacity is scarce and contracts run three years, a GPU’s health record stops being an internal metric. It becomes part of what you are buying.
A question for operators and buyers: does your GPU contract give you the health history of the hardware you rent, or only a discount when it fails?