Musk says Colossus 2 plans to deploy approximately 1200000 units by year-end Blackwell GPU
BIBIBI
AT A GLANCE
Musk says Colossus 2 plans to deploy approximately 1200000 units by year-end Blackwell GPU。
Article
Musk says Colossus 2 will deploy by year-end 1200000 units Blackwell，200000 units dedicated to inference
Beating AI flash: Musk disclosed that the Memphis supercomputing cluster Colossus 2 plans to deploy approximately 1200000 NVIDIA units Blackwell GPU. Two more facilities are expected to come online: Gun Range is expected to launch by November year-end 220000 units GB300, for model training; Spam is being converted from a warehouse that originally stored food such as luncheon meat, and plans to deploy approximately 200000 units GB300, dedicated to AI inference.
Currently, Macrohard is equipped with 110000 units GB200 and 220000 units GB300, while Macroharder and Minihard each have 220000 units GB300. Including the 420000 units planned to be added, the five clusters will total approximately 1190000 units.
Training and inference also have different networking requirements. Macroharder, Minihard, and Gun Range all use 800Gbps NICs, while Spam's inference workloads do not require high-speed interconnects on the same scale. Musk said its network configuration is simpler and requires less than half the fiber of the training facilities.
Original link https://m.theblockbeats.info/flash/374857
Key points
01
Musk disclosed that the Memphis supercomputing cluster Colossus 2 plans to deploy approximately 1200000 NVIDIA units by year-end Blackwell GPU。
02
Gun Range is expected to launch by November year-end 220000 units GB300, for model training.
03
Spam plans to deploy approximately 200000 units GB300, dedicated to AI inference.
04
The article says that the five clusters, including Macrohard, Macroharder, and Minihard, will total approximately 1190000 units GPU。
05
Macroharder, Minihard, and Gun Range use 800Gbps NICs; the article says that Spam requires less than half the fiber of the training facilities.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
GPU are chips used for computing. Training enables a model to learn, while inference is the use of capabilities it has learned. The article describes the GPU deployment plans and network configurations of multiple facilities.
Why it matters to readers
If the plan is implemented, it would involve substantial GPU and data center construction requirements; however, this is currently a relayed deployment plan, and no original statement or delivery evidence has been seen.
This is a report about the scale of AI computing infrastructure, focusing on the GPU number, purpose, and network connectivity.
Risks and unknowns
The report comes from a Telegram account's relay and does not provide Musk's original statement.
The article does not provide a project-owner announcement or equipment delivery evidence.
The article gives two totals: approximately 2 units for Colossus and approximately 1200000 units for the five clusters. 1190000 total figures.
The planned quantities, launch times, and actual deployment status remain to be independently confirmed.
Related Developments
Loading event timeline…
Related concepts
GPU
This term is not in the glossary yet. Browse related concepts in the glossary.
