←back to thread

117 points LordAtlas | 1 comments | | HN request time: 0.191s | source
1. Nevermark ◴[] No.46183585[source]
> training compute looks more like an operating expense with a short payback window than a durable capital asset

Today they are a durable asset functionally, longer than they are economically. So there is no reason in a market with less demand, that their economic payback windows cannot be extended further into their functional lifetimes.

There will be energy cost incentives to replace GPUs. But turnover can respond sensibly to demand as it revives, while older GPUs continue working.

Also, the data centers themselves, and especially any associated increase in power generation, will carry forward as long term functional value.

I doubt any downturn in compute demand lasts long. The underlying trend, aside from AI, was for steady increases in demand. Regardless of bad AI business models, or investment overhangs, a greater focus by more entities on AI product-market fits, along with cheaper compute, will quickly soak up cycles in new and better ways.

The wildflowers will grow fast.