Musk said that Memphis Colossus 2 plans to deploy approximately 1200000 NVIDIA units Blackwell GPU, of which approximately 200000 units will be used for inference. Other facilities, including Gun Range and Spam, are planned to come online or expand.
DeepSeek Reportedly open-sourced DeepInfer and disclosed cluster processing capacity; based on its API off-peak pricing, theoretical billing revenue per GPU per hour is approximately 38 USD, but this does not equal actual revenue or profit.
BlockBeats reported that the head of research at SemiAnalysis believes that currently GPU is easier to obtain than data center power, and warned that 2028 data centers could face oversupply.
Citing PANews,NVIDIA the official blog stated that, in a October 21 GTC Berlin keynote, Jensen Huang said NVIDIA is shifting to providing AI factories and announced that the Vera Rubin platform has entered full production.
PANews reports that Ant Group and the vLLM community have jointly open-sourced Humming, a general-purpose library of quantized inference operators, claiming that it supports multiple categories of GPU, quantization formats, and Dense and MoE models.
NVIDIA said it used Germany's JUPITER supercomputer's 8192 GPU to simulate a 104-qubit fault-tolerant quantum computing system and validate that “magic-state cultivation” technology could reduce the required physical qubits by approximately 90%.
Plans to expand GPU compute pool and is in discussions with Switch and Volta AI The report says several Silicon Valley VCs are securing a16z compute capacity to support portfolio 20000 companies;GPU plans to expand its approximately
According to reports, the two SpaceXAI Colossus data centers currently have approximately 780000 units GPU, with another 220000 units GB300 planned to become operational next week; Musk says its models may be in a leading competitive position in months.