“Build or surrender: for now, the choice is still ours.” Good piece by @arthurmensch, and I'm with him on building (a lot!). But the thing is that Europe isn’t going to out-spend the US on GPUs. The reality is that we have less capital, more expensive power, and fewer data centers. So building looks different in Europe. Every GPU and every watt we have needs to do more work than it does anywhere else. If we get more work out of every watt, we don’t have to wait years for a grid connection to get there. We should still invest in power, obviously. But the reality is that new generation and grid capacity takes years, and we can’t afford to sit around and wait for them. So this is where I’d put the effort now, and this is also where startups can shine. Build different, do more with less 🇪🇺.

The timeline tunes fast here

Coinbase made training a specialist model one command in their harness. My sense is that serving is where the work goes next. If every company ends up with a handful of models like this, each one has its own latency target and its own scorer. A fraud model that has to answer in milliseconds and an overnight code-review model want different batching, and often different GPUs. This means that a company's fleet ends up running many models on mixed hardware, with perhaps a frontier API kept for the cases nothing else passes. Serving all of that well is an interesting problem space.

Look at those two lines together📈: the token index halved since May while B300 rent went up 38%. Whoever serves the model pays the GPU rent and gets paid per token. The only way through is more tokens out of every GPU-hour. In Europe, where adding GPUs takes years, that matters most...

$36B of neocloud backstops and $20B of DC leases to hand off to neoclouds. Nvidia now carries part of its customers’ demand risk. The question if that ever costs them anything depends on how busy those GPUs are.

Agree on demand. The margin is the fun part though. Same weights, same GPUs, and the cost per finished task still swings a lot depending on the serving stack. Free weights turn it into an inference engineering problem imho.

Inference is Everything?

In one quarter, open-weight models went from about 30% to a record 78% of the tokens on Vercel's AI Gateway. That's just the one gateway, on a record day, and of course these tokens aren't 1:1 revenue. But the direction of travel is hard to miss. When you can just download the model, the margins move to the infra and inference layer. A token starts to cost what it costs to serve, and the value moves to whoever runs the model best: fastest, cheapest, at the quality the job needs. For Europe, this changes the equation. As long as the best AI was a closed API, "European AI" meant renting capability from a few US frontier labs, served from their infrastructure. Open weights change that. The models doing most of the work (the "Work Horse") can now run on hardware Europe controls, under European rules. That doesn't directly make Europe competitive on its own. Where the models come from still matters, and unfortunately Europe has very little compute to run them on (more on that at a later time). But the question has moved from "who builds the model" to "who runs it best."

Extremely proud to have been backing this team from the very beginning with @nphardvc and this is a great time to join them on their mission! 🧱🤖🏘️

Claude sure likes “the gap”.

Turns out codex doesn't care about getting a sandwich.