“ According to Mr Yotzov’s study, nine in ten executives report no impact of ai on their firm’s productivity over the past three years.”

Brutal stuff.

> A back-of-the-envelope calculation finds that covering aicapex through identifiable ai income requires revenue on the order of $2.5trn per year, more than tech’s entire combined revenue today.

…

> All these complex calculations roughly tally with a much simpler one: adding up the ai revenue of the firms selling most of the ai. Anthropic pulls in perhaps $75bn, annualised; Openai makes tens of billions; Google, via its ai model Gemini, and Microsoft probably get a bit less. SpaceX may have a few billion dollars’ worth of revenue from enterprise aithis year. Meta also makes a few bucks from ai. Add this up and you land at roughly $150bn a year.

You really see the first mover disadvantages in US AI labs. And how second movers (open Chinese labs) can disrupt them. First movers tend to get overcapitalized and way ahead of their skis. You saw a smaller version of this in the vector database market (remember that?) where companies like Pinecone were getting billion+ valuations for capabilities which now seem like a commodity. There's vector DB companies now with much less debt / VC obligations that don't need to clear insane revenue levels to justify investment.

I kind of think the AI boom is the worst for (non-inference-providing) companies. If all companies are using the same "frontier" LLMs, and if they are competitive, what gives one an edge over the other? I think, just the people. Which was the same as before, but now with the additional AI spend that they can't cut, or they become less competitive.

I mean, all one has to do is review https://isaiprofitable.com/ to see just how bad it is. None of this is a profitable venture. It's just a money pit for all players involved, all the while draining our ability to buy electronics and drink water. This has to end at some point.

Attributing to per-worker might be the wrong thing. I.e. how many kWh power does a person buy direct vs. use indirectly through other trade.

In other words usage per regular worker doesn't matter. Revenue overall does.

> Exponential View, a consultancy, counts $175bn of generative-AI revenue, on an annualised basis, in June. In a recent paper Anton Korinek of Anthropic and Patrick McKelvey of the Bank of Canada estimate total “AI services” revenue. Adapting their methodology, we reckon this was $220bn (again annualised) in the first quarter of this year. Ramp’s data imply that 2-3% of business spending now goes on AI, pointing to $170bn a year.

So between $170bn-220bn in annualized AI revenue today. Maybe it doesn't cover trillion-dollar bets but this is a very substantial number.

A couple anecdotes:

I noticed that were a lot of traditional, non-tech companies interviewing for AI engineers in the Feb/March timeframe that have halted hiring in those roles entirely. It seems that if those roles didn't close by mid-April that they didn't close at all. This seems to match the timeframe in which cost suddenly became prominent in the AI zeitgeist.

I don't know _anyone_ outside of SV who has successfully replaced even a single employee completely with the current models. Maybe someone has pulled this off in call centers, but the POCs have all failed.

I'm very much pro-AI, but I just think we're on a false summit. As the article points out, there is absolutely no way to recoup the investment costs unless the models allow companies to start displacing human workers by the millions _and_ recapture a significant fraction of the displaced workers total comp. If either of those aren't true, then the bubble is going to pop... soon.

The thing is, AI is a productivity boost but staff drive all the productivity. The catch 22 is AI is expensive so in order to provide it to everyone they need to reduce staffing. The result is remaining folks are able to produce more but the organization as a whole is not exceeding pre-layoff output. And now things are being dropped on the floor and falling in between the cracks -- which impedes efficiency and velocity. Happening at my company now.

AI inference needs to get cheaper or there will always be this "terminal velocity." I suspect AI labs' incentives to reduce costs is only where there is overlap to free up hardware/utilization (to then provide inference to more paying customers.) I highly doubt they will want to make things cheaper for users -- they have debts to pay.

Open weight models are the way and forward, it is the only way an organization can truly control costs by self hosting or buying cheaper inference. Relying on closed weight models is a business risk. Kimi models are only marginally worse than opus but significantly better than the bleeding edge of yesterday's sonnet.