Louwrentius

All major AI services like ChatGPT or Claude are heavily subsidised by investors. It's not just the initial investment — it's the running costs of these major AI services that are also astronomical.¹

At some point the money will run out. Assuming that Ed Zitron is right about his assessment of the AI market (and he's not alone in his assessment), the current big AI services are not sustainable. And there seems to be no realistic path to profitability.

How are the big AI providers going to make $2 Trillion in the next 4 years?

So what happens when money runs out? When the music stops, will there be any chairs at all?

I see two major scenarios:

1. The AI services are just shutdown overnight
2. The AI services are priced according to actual cost + profit

In the first scenario, those who have build their entire workflow on AI support have a serious problem. AI servies as a massive single-point-of-failure. Quite a few very optimistic roadmaps have to be adjusted, I guess.

The second scenario may be effectively identical to the first, as AI services may become so expensive that their usage can't be justified. For no reason I can think of, the company "VMWare" comes to mind, strange isn't it?

As there are open-source models available, it may be possible for some smaller players to jump into this market vacuum, but they'll face exactly the same problems: operating AI at scale is stupidly expensive. They can't be as 'cheap' as current AI services. As they can't afford to subsidise their AI service, they may still be too costly. I can see that smaller players can provide 'one-trick pony' AI services, like 'just for programming', as opposed to the generic AI services provided by the large players. But due to market demand they may become priced at: "if you have to ask, you can't afford it."

Self-hosting an AI with your own hardware is probably just as cost-prohibitive, even if you don't value your time. In part because a ton of people will get this idea at the same time, impacting hardware prices even more. And the operating costs of AI seem significant. Would it even be possible to setup your own AI and achieve the same productivity level?

Given current circumstances, I can see a scenario where the USA government will step in and subsidise these services. If the AI investors buy some specific crypto coins obviously.

Quid pro quo, Clarice.²

Hacker News thread

Context

I'm running a virtualisation lab environment, with four Intel-based CPUs (i5-6500) and one AMD Ryzen 3 PRO 2200GE.

When migrating virtual machines from one of the Intel hosts to the AMD host, the migration would fail with the following error:

error: operation failed: guest CPU doesn't match specification: extra features: cmp_legacy"

The really strange thing is that some virtual machines did migrate successfully from an Intel system to the AMD system without any issues.

So I dumped the XML of the virtual machines and compared them. It turns out that the virtual machines that migrated successfully were all single-core and all failing virtual machines were multi-core.

As the cmp_legacy CPU feature seems to be related to (lack of) support for hyper-threading, the configured CPU topology of the virtual machines was suspected to be the culprit.

I've configured a custom CPU topology on the virtual machines with 1 socket, 1 die and one or more CPU cores. This caused the migration failure.

Solution

If you configure multi-core virtual machines with multiple sockets instead of cores migration between Intel and AMD hosts will succeed.

In the past, migrations between an Intel and AMD host always succeeded, assuming that the appropriate CPU features were disabled. Why this suddenly stopped working is not clear to me, as this configuration never changed. Maybe it's related to a Libvirt update, but I have no proof for this.