Tomorrow, money might not be enough

The AI bubble is going to crash, one way or another.

The data centers already built will keep running, but not everything that was promised will get built. Of the 16 GW of data center capacity scheduled for 2026, only about 5 GW was actually under construction.

The question isn’t whether AI is useful. It’s how to make it useful. And until we find the answer, access to it will get scarcer.

The US will make the logical move: manage the shortage, let prices go up, and have the state step in to secure access for its key players. It has already tried twice. In January 2025, its export rule on AI chips put Switzerland in “tier 2”, with a cap on how many chips it could import. In October 2025, the Senate voted to make chipmakers fill American orders first. Neither has taken effect so far.

In May, Anthropic signed for the entire capacity of SpaceX’s Colossus 1 data center, more than 220,000 GPUs. In June, Switzerland’s “most powerful commercial AI supercomputer” went live in Arlesheim, with 192 GPUs. Alps, the public supercomputer in Lugano, has about 10,000.

Swiss companies will end up far behind, with processes that aren’t ready yet and an ever higher cost before they see a real profit. That could choke them.

Nobody can know what a token will really cost in two years. But it’s a safe bet it will become too expensive, even for our heavyweights, if they haven’t completely refactored their processes by then.

My take, for what it’s worth: secure access to sovereign AI now. This is a call to Swiss policymakers, companies and industry to pool resources today and build data centers on our own soil. Sovereign compute, planned, with preferential access for Swiss players.

It’s the only way not to end up facing AI-augmented super-companies while we stand there naked, regretting that we missed the wave.

Today, experimenting costs money.

Tomorrow, money might not be enough.