4 ms·

China built hundreds of AI data centers. Now many stand unused

4 points by belter 2y ago

aurareturn 2y ago
I don't think it's a sign that there is less demand for AI compute. I think it's a sign that many AI labs are bowing out and the leading ones are emerging. In the US, OpenAI, Anthropic, Google, Meta are the only ones doing large scale training. Others will no longer compete at training scale. They've bowed out. In China, DeepSeek, Qwen, and maybe 1 or 2 more are doing large scale training. So the expectation that there are 100 AI labs wanting to train SOTA models and use massive data centers is wrong. There are only a handful left and they mostly have their own data centers. Deepseek probably killed a ton of AI labs worldwide thinking they can compete. They can't compete when there is a model so good and free. The moat was always who has the most compute. That's why there was/is a race for data centers. Signs compute is still a severe limiting factor: * My ChatGPT Plus subscription just turned off Deep Research functionality after I used it less than 10 times. I need to wait 7 more days to use it again. * GPT4.5 is really slow and has a very low quota in my ChatGPT Plus subscription * Deepseek's model release paper states their biggest challenge now is more compute * People want faster answers from reasoning models
- belter 2y ago
  > I don't think it's a sign that there is less demand for AI compute. Note also Microsoft declined to take on extra capacity: "Microsoft chose not to exercise $12 billion Coreweave option" - https://www.semafor.com/article/03/20/2025/microsoft-chose-not-to-exercise-12-billion-coreweave-option https://www.semafor.com/article/03/20/2025/microsoft-chose-n...
  - aurareturn 2y ago
    Yes, but Meta and Google promptly picked up Microsoft's abandoned data centers. I think Microsoft investing less into AI compute is directly related to their gradual split from OpenAI. OpenAI seems to want to chart its own path by partnering with Softbank. Edit: You linked to older news. OpenAI directly picked up Microsoft's Coreweave deal.
    - belter 2y ago
      Could be. But these two data points are from today and 2 days ago. "Microsoft ditches more AI data centre leases as spending bubble looms" - https://www.capacitymedia.com/article/microsoft-ditches-more-ai-data-centre-leases-as-spending-bubble-looms https://www.capacitymedia.com/article/microsoft-ditches-more... "Alibaba chair warns of AI data centre bubble amid massive buildouts: report" - https://www.capacitymedia.com/article/alibaba-chair-warns-of-ai-data-centre-bubble-amid-massive-buildouts-report https://www.capacitymedia.com/article/alibaba-chair-warns-of...
      - aurareturn 2y ago
        The Microsoft story - I think it's due to the gradual split between Microsoft and OpenAI. Google and Meta leased the abandoned Microsoft data centers promptly. The Alibaba story - it's one person's opinion. I think he's worried that competitors are too optimistic and he wants them to slow down so Alibaba can have a chance.

4 points by belter 2y ago

I don't think it's a sign that there is less demand for AI compute. I think it's a sign that many AI labs are bowing out and the leading ones are emerging. In the US, OpenAI, Anthropic, Google, Meta are the only ones doing large scale training. Others will no longer compete at training scale. They've bowed out. In China, DeepSeek, Qwen, and maybe 1 or 2 more are doing large scale training. So the expectation that there are 100 AI labs wanting to train SOTA models and use massive data centers is wrong. There are only a handful left and they mostly have their own data centers. Deepseek probably killed a ton of AI labs worldwide thinking they can compete. They can't compete when there is a model so good and free. The moat was always who has the most compute. That's why there was/is a race for data centers. Signs compute is still a severe limiting factor: * My ChatGPT Plus subscription just turned off Deep Research functionality after I used it less than 10 times. I need to wait 7 more days to use it again. * GPT4.5 is really slow and has a very low quota in my ChatGPT Plus subscription * Deepseek's model release paper states their biggest challenge now is more compute * People want faster answers from reasoning models

belter 2y ago
> I don't think it's a sign that there is less demand for AI compute. Note also Microsoft declined to take on extra capacity: "Microsoft chose not to exercise $12 billion Coreweave option" - https://www.semafor.com/article/03/20/2025/microsoft-chose-not-to-exercise-12-billion-coreweave-option https://www.semafor.com/article/03/20/2025/microsoft-chose-n...
- aurareturn 2y ago
  Yes, but Meta and Google promptly picked up Microsoft's abandoned data centers. I think Microsoft investing less into AI compute is directly related to their gradual split from OpenAI. OpenAI seems to want to chart its own path by partnering with Softbank. Edit: You linked to older news. OpenAI directly picked up Microsoft's Coreweave deal.
  - belter 2y ago
    Could be. But these two data points are from today and 2 days ago. "Microsoft ditches more AI data centre leases as spending bubble looms" - https://www.capacitymedia.com/article/microsoft-ditches-more-ai-data-centre-leases-as-spending-bubble-looms https://www.capacitymedia.com/article/microsoft-ditches-more... "Alibaba chair warns of AI data centre bubble amid massive buildouts: report" - https://www.capacitymedia.com/article/alibaba-chair-warns-of-ai-data-centre-bubble-amid-massive-buildouts-report https://www.capacitymedia.com/article/alibaba-chair-warns-of...
    - aurareturn 2y ago
      The Microsoft story - I think it's due to the gradual split between Microsoft and OpenAI. Google and Meta leased the abandoned Microsoft data centers promptly. The Alibaba story - it's one person's opinion. I think he's worried that competitors are too optimistic and he wants them to slow down so Alibaba can have a chance.