From his latest article:

One of my sources has come forward and brought me a story that will possibly burst the AI bubble. The reason they brought this to me is that I’ve shown — and will continue to show — that I actually give a shit about this industry and the people in it. If you’re wondering what the story is, know that it’s the information I’ve wanted for years, delivered as I have always wanted it, and I will treat it with the reverence it deserves. Imagine what the worst possible thing for me to get would be and you’re probably close.

Link: https://www.wheresyoured.at/ai-is-slowing-down/

I must say I’m pretty excited.

  • humanspiral@lemmy.ca
    21·
    6 days ago

    If we take Sightline Climate’s data from February at face value, there are 190GW of data centers planned. If we take NVIDIA CEO Jensen Huang’s statement that data centers will cost $80 billion to $100 billion a gigawatt at face value, this means that said data centers will cost anywhere from $9.5 trillion to $15 trillion. Bloomberg incorrectly states that this is a “$3 trillion” buildout.

    There is a lot of oudated misinformation in this setup, and these numbers must stop getting repeated.

    2026 active construction datacenters is 5gw, 2027 is 8gw. These are max numbers. Globally. Just today, Crusoe announced that MS/Oracle/OpenAI datacenter of 1.9gw was paused at request of customer, though it is unclear if that was taken out of the active 26/27 pipeline. It is impossible for 50gw to come online by 2030. The power queue is too long.

    The average cost this year for liquid cooled nvidia nvl72 racks is $20b/gw for building and $32b for GPUs. You can get to $100b by adding opex costs over 5 years, or imagining more expensive future GPUs. But the buildout is indeed in $3T ballpark. The important figure though is $20B/year in depreciation and expenses (including 10% interest/ROI for GPU, and a bit for building) per gw capacity, for operators to not go bankrupt.

    That is simple basis (can add 3rd party margins on top for extra effect) for making Ed’s “general doom” more tangible/grounded. OpenAI’s revenue is $11B/gw they control. While Anthropic’s latest claim is $27B/gw they control, it ignores $15B rental deal with Anthropic because, it is something they are likely to cancel after SpaceX IPO. It’s still an overall loss (about $8b) on just compute costs.

    Sure both companies have huge commitments for more compute (OpenAI double $121B vs $55B in 2028), and how fast they can grow to pay for it matters a lot, but a bigger factor is that software advances are delivering more with less compute, Gemini flash, composer 2.5, other fast models by frontier labs, and of course China’s models provide much higher value. But OpenAI is in much more trouble, unlikely to both scale revenue, and scale it by a much greater factor to just pay for compute bill.

    The other factor is that compute capacity is not at all constrained now, nor after what is 10gw planned by the 2 main labs. Leaving 40gw by 2030 for everyone else outside of China.