Post by @ArtHarg@mastodon.nl

2026-09-30 10:50 UTC

@david_chisnall@infosec.exchange @troberts@theblower.au @luciedigitalni@aus.social My gut feeling for what’s going to happen: the pioneers of the AI boom are going to crumple under the weight of their investments. Creditors will sell off IP and physical assets on one side. Maybe government will pick up some of the losses, to distribute among the masses. Asset vultures swoop in to buy IP and data centres that have already been built. Their investments would be substantially lower than those of the pioneers, because they won’t pay for promised profits that will never materialise. They may even make sense from a valuation point of view. There may be a sensible business case then. I agree that inference is going to become increasingly important. And especially: how do you tie in LLMs with other forms of machine learning and other forms of systems control. That’s going to be a job for humans, I reckon, because it requires understanding and creative thought.

Replies (1)

@david_chisnall@infosec.exchange 2026-09-30 10:55

@ArtHarg@mastodon.nl @troberts@theblower.au @luciedigitalni@aus.social Asset vultures swoop in to buy IP and data centres that have already been built The IP is pretty small. The shape of the models is quite simple, the value is in the weights. These are trained on data that the LLM vendors don't own. Just acquiring the trained weights is not that useful because they age quickly. The situation for datacentres is similar. They're designed around the ludicrous power requirements of NVIDIA GPUs. Those GPUs have an operational lifespan of three years on average (they run hot and burn out quickly). Buying them isn't going to be that worthwhile because they're not that useful for anything else and they'll be much less reliable than new ones. And that would make sense only in the low-opex outcomes where you could use the same datacentres to offer a profitable service. But that's where the limited lifespan of LLMs comes in. LLMs must be constantly trained on new information (current events, changes to APIs and languages for programming, and so on). Offering inference on a five-year-old model is not that useful. So it makes sense to buy the datacentres only if capex also drops. So your bet is on low-capex, low-opex and commodification of the LLM market. That's possible, but so far the trend has been in the opposite direction.

Open ##5171460