November 9, 2026 — today in AI.
1. Nvidia Cracks Down on Overseas Cloud GPU Rentals for China, Imposes 51% Non-Chinese Quota on Partners
Nvidia is restricting overseas cloud providers' GPU rentals to Chinese customers by enforcing a 51% non-Chinese quota on partners, while black-market B300 rack prices reportedly reach $2.8 million. Cloud operators and GPU resellers should expect tighter KYC and end-user checks, and anyone sourcing B300 capacity through intermediaries faces compliance and supply risk.
2. Google Cloud's TPU 8tx Pairs AMD EPYC Hosts for Hardware-Enforced Confidential AI Computing
AMD's 5th Gen EPYC processors will serve as host CPUs for Google Cloud's TPU 8tx, which Google describes as the first cloud TPU system with hardware-enforced confidential computing, displacing Arm-based host processors. For teams running sensitive training or inference, this is a new option for confidential AI workloads, though Google has not disclosed pricing or general availability dates.
3. Anthropic Secures 1.7 GW of Gas-Fired Generation From OnePWR Solutions for About $3.4 Billion
According to the report, Anthropic committed in late September 2026 to 1.7 gigawatts of heavy gas-fired generators from San Antonio-based OnePWR Solutions for roughly $3.4 billion. Power access, not just accelerators, is becoming the binding constraint on frontier-lab capacity, which will affect future API rate limits and regional availability.
4. China's Galaxy AI Campus Runs 800V DC Power Architecture in Commercial Operation in Ulanqab
Envision reports an integrated 800-volt DC power architecture in commercial operation at the Galaxy AI campus in Inner Mongolia, which is planned as an 800 MW site delivering its first 120 MW within a year at about one-fifth the reported cost of comparable builds. If the efficiency and cost claims hold, 800V DC distribution could become a reference design for rack-scale AI power, so operators tracking power conversion losses should watch the published metrics.
5. Nvidia Plans Mixed Data Centers Pairing Its GPUs With Rebellions NPUs for Inference
Nvidia is building a framework that runs its GPUs and Korean startup Rebellions' NPUs inside a single data center, with the NPUs handling inference workloads. Heterogeneous fleets are a signal that inference is being split onto cheaper, specialized silicon, so teams planning serving stacks should prioritize portable runtimes across GPU and NPU targets.