# AI Will Hit a Wall in 2026, if nothing changes.

## Метаданные

- **Канал:** Sabine Hossenfelder
- **YouTube:** https://www.youtube.com/watch?v=XA84pSrPHS0
- **Дата:** 12.05.2026
- **Длительность:** 6:42
- **Просмотры:** 201,077
- **Источник:** https://ekstraktznaniy.ru/video/51291

## Описание

Free GenSpark credits if you register here  → http://www.genspark.ai/?utm_source=yt&utm_campaign=SabineHossenfelder #Genspark #WorkWithGenspark

Current AI technology seems to be making decent progress despite concerns about it slowing over time. But while AI is slowly becoming more “intelligent”, the industry is running into another problem: energy supply. Let’s take a look at why energy is quickly becoming a major problem for progress in AI.

Elon Musk at Davos: https://www.weforum.org/stories/2026/01/elon-musk-technology-abundant-future-davos-2026/

👕T-shirts, mugs, posters and more: ➜ https://sabines-store.dashery.com/
💌 Support me on Donorbox ➜ https://donorbox.org/swtg
👉 Transcript with links to references on Patreon ➜ https://www.patreon.com/Sabine
📝 Transcripts and written news on Substack ➜ https://sciencewtg.substack.com/
📩 Free weekly science newsletter  ➜ https://sabinehossenfelder.com/newsletter/
👂 Audio only podcast ➜  https://open.spotify.com/show/0MkNfXlKnMPEUMEeKQYmYC


## Транскрипт

### Segment 1 (00:00 - 05:00) []

I'm generally optimistic about progress in artificial intelligence, but recently I've become somewhat concerned and not for the reasons you might think. It seems that AI progress could hit a wall in 2026. Not because the models suddenly stop getting smarter, but because we might not be able to turn the damn things on. Before we continue, a word from today's sponsor, Gem Spark, an all-in-one AI workspace. They do AI chat, image generation, slides, sheets, video. But the feature I actually care about is their AI agent builder. No coding required. You write one prompt and it builds a standalone tool anyone can use. So I built a science hype detector. Every week there's a headline like shocking dark energy findings challenge the standard model. Sounds dramatic. Then you read the paper and it's a 2. 8 eight sigma hint that requires negative nutrino masses to work. I deal with this constantly. Now there's a tool for it. I fed that scientific American headline. It found the original desi papers noted the data alone is consistent with the standard model called that it's below five sigma and flagged the neutrino mass problem. The article never mentioned hype score 6 out of 10. Links in the description. Sign up for free and you get free credits to start. And right now, Gen Sparks offering unlimited AI chat and image generation for all paid users in 2026. Try it, break some headlines, build your own agent. And now back to the science news. Flashback to 2024 when computer scientist Leopen Brena predicted the super intelligence explosion will come in 2027. Blade training would require 100 gawatt or so per model. Just to give you a sense of scale, that's about twice Germany's entire electricity generation. According to Ashen Brena though, that could be made to work within a year or so by just quickly drilling for more gas. A truly fascinating example for how detached some Silicon Valley guys are from reality. He was right though in that the major problem for AI progress turns out to be energy supply. Not a lack of computer chips, money, and not a lack of faith either, but plain old energy. The International Energy AY's most recent report forecasts that the electricity use of data centers worldwide will roughly double from 2024 to 2030. This isn't quite as dramatic as Ashen Brena's vision, but it means it's growing at a staggering 15% a year, four times faster than overall electricity demand. Now, the problem isn't just where all this energy is supposed to come from, but also how it'll get to the data centers. The investment firm Morgan Stanley is already warning of the consequences. They are generally optimistic about major AI capabilities and write that long-term demand for computing power is likely to justify current investment levels. But they also report that developers expect power constraints by 2027 to 2028 due to underinvestments in grids and potential supply chain disruption. The slowdown is already visible. In February, Axio reported that as many as half of the world's data center projects, which were supposed to go online in 2026, could face delays. Out of more than 16 GW planned, only 5 GW is actually under construction with the rest pending in the announced stage, but no signs of building. What's the problem? The problem is that the US electric grid isn't catching up remotely quickly enough. According to data from Berkeley Labs, there are about 2 terowatts of power plant projects renewable or not waiting for a grid connection. The median weight is 5 years. Some areas like Northern Virginia hit 9 to 12 years. The issue is partly caused by transformer shortages, but mostly it's bad planning. Now, you can of course just build a power source right next to the data center, but that makes it logistically and operationally much more difficult, expensive, and also slower. If you listen closely, those working in the industry are already warning of the problem. Here is, for example, Elon Musk in a recording from January. — Yeah. I think the limiting factor for um AI deployment is fundamentally electrical power. — It's just like it's energy. — Yeah. Yeah, I mean we're seeing the rate of AI chip production increase exponentially, but the rate of electricity being brought online is uh — 3% 4% a year max. — Yeah, it's clear that we we're very soon, maybe even later this year, uh we'll be producing more chips than we can turn on except for China. The lack

### Segment 2 (05:00 - 06:00) [5:00]

of grid connections is of course not just a problem for data centers. It's also going to slow down the energy transition because if we're supposed to all stop heating with oil and gas and drive electric vehicles, then we're going to need a lot more electricity hopefully from carbon neutral sources. And given that the electrical grid simply won't suffice in the present form, neither in North America nor in Europe, that's where things will grind to a halt. Elon Musk is entirely right, of course, that China doesn't seem to have the same problem. And seeing how competitive Americans are, it'll be interesting to see how they'll try to cope. The Trump administration seems to really like the idea of small modular nuclear reactors and those seem like an obvious solution until you remember that all the mini reactors we've seen so far were like a decade behind schedule and a billion or so over budget. Small modular reactor currently means small expectations and modular delays. All this is going to create a very strong incentive for chip producers to reduce energy consumption for AI training. And once that actually makes any significant difference, it's likely that many of the data centers that are currently under construction will never be completed or completed ones will never find customers. When that starts to filter into markets, this is when things will go to hell in the entire tech sector. In summary, the future may be bright, but we've got nowhere to plug it in. Thanks for watching. See you tomorrow.