Summer themes: heatwaves and AI

August is a slow month in Europe and I’ve been using it to strategize. So this won’t be a big analysis. But interesting things are happening that are shaping my plans for the next months.

We’ve had a few heatwaves this summer, obviously.

The first one was in June, with the usual impacts: higher temperature → higher power demand → higher prices. Plus trouble starting for the temperature-sensitive French nuclear fleet.

Then a second round of heatwaves hit mid-July. But the most surprising part came last. The heat coupled with drought called for some extreme measures. Hungary started sinking barges in the Danube river to raise water levels and keep the Paks nuclear plant running. Romania blasted rock to redirect cooling water to their Cernavodă nuclear facility. Poland, despite strengthening the cooling of coal plants after power shortages in 2015, still had to reduce production in two major coal power stations due to low water levels in the Vistula river.

We all know heatwaves will become more frequent and severe. But governments and utilities seem to always be surprised, prices always spike, potential blackouts are always discussed. I think it’s high time to treat it as new normal.

So my takeaway: get solar, a battery, and an offline inverter. Plus air conditioning. If you’re the government or grid operator, do the same - keep building solar (the profile matches AC demand), add as much battery storage as possible, unlock demand response, set procedures for black starts and test them. Seems much better than blowing up rivers.

A few news items dropped recently that are worth reflecting on.

One: Anthropic announced that Claude will now be watermarking AI-generated text. Meanwhile, LinkedIn now has an “AI slop” button that you can use to scrutinize AI-sounding posts. Substack is doing something similar. This has pros and cons. Of course, like everyone, I hate AI slop, and all social media would be better without it. But what prevents you from tagging your competitor’s posts as “AI slop” even if it isn’t? Unless the detection is really good, this might just be a massive disruption for genuine content creators. I write my newsletters and posts myself, but I am actually worried someone will tag them as “AI slop” and all the hard work will go down the drain…

Two: The European AI sovereignty debate is accelerating. Mistral announced plans to “to build up to 1 GW of capacity by 2030” in Europe. But the deal is roughly this: companies and countries will buy “European Compute Units”, Mistral will aggregate their compute demand and build the infrastructure needed. It’s a cool idea to finance the buildout, and also keeps the supply aligned with the demand. They’re also opening the doors for other models, so you could in theory host Chinese open models on Mistral’s infrastructure in Europe. Interesting combo. But I wonder what that means for Mistral’s own models and European LLMs in general.

Three: In the US, Cleanview reported on Amazon’s plans to build the largest gas plant in the country’s history - 7.65 GW, powering the company’s first off-grid data center. It’s baseload gas running 24/7. Which means around 11 bcm of gas consumption a year. That’s more than the whole of Hungary or Czechia - countries with ~10 million populations. I find that surprising. Actually, insane.

Next week I’ll have some news of my own to share. See you then!