Hyperbolic

Jasper Verified account @zjasper · 13h

A neocloud correction is coming.

Not because demand is slowing. It’s the opposite. OpenRouter alone is now processing close to 200 trillion tokens a week, more than 10 quadrillion a year.

Strong demand won't save neoclouds with bad operating economics.

Illustrative math: take a 1,024-GPU B300 cluster renting at $4.50 an hour, with an all-in cost of about $3 an hour.

→ 60% utilization: −11% gross margin
→ 80% utilization: 17%
→ 100% utilization: 33%

Same chips. Same price. Utilization is the difference between losing money and running a real business.

The winners won't just own the most GPUs. They'll keep every GPU-hour busy, by connecting their capacity to demand wherever it comes from.

That's what I'll talk about this Friday at the SGLang Summit:

⚡ "Building the Grid for AI Compute"
📍 Decode room, Fort Mason Center, SF
🕜 Fri, Nov 13, 1:30 PM PT

Come find me after the talk or at the Hyperbolic booth.

What utilization do you think most neoclouds actually run at? 👇

View quoted post on X