Will I ever own a zettaflop?

April 9, 2026
System with various wires managing access to centralized resource of server in data center
Photo by Brett Sayles on Pexels

The dream

It has been reported that George Hotz — known online as geohot — posted a short, feverish meditation on his blog about owning a “zettaflop” machine: a system capable of 1e21 FLOPS. He opens with the familiar mix of awe and obsession: he wants to “feel it,” to command that scale the same way he uses his laptop now. He even borrows a moment from Vernor Vinge, quoting a passage about sensory overload to paint what the experience might feel like—intense, bordering on terrifying. Who wouldn’t be tempted? Command a million copies of a powerful model and you’d have 50,000 people’s worth of thinking in two weeks.

The arithmetic

The post lays out rough, crunchy numbers. Comma allegedly has an exaflop already — “just one little exaflop” — and Hotz extrapolates: a gigawatt of power, a million GPUs, 1,000 exaflops to reach a zettaflop. He tosses out jaw-dropping scalars — 1e27 training runs, “100 lifetimes in 1e6 seconds” — and imagines using a zettaflop to search every book, solve hard math, iterate code until it’s perfect. It reads like a wish list and a thought experiment at once: one million instances of a Claude-like model, all aligned and answering as one.

The bottleneck and the plan

Reality, he admits, is power. Hotz points to efficiency limits—FP4 dtypes, B200s getting ~10 TFLOPS/W—and reckons current designs would need on the order of 100 MW; with a plausible 10x efficiency gain that drops to ~10 MW. From there he sketches a DIY approach: 250 acres of solar plus pumped storage, 100,000 chips at 10 PFLOPS each (priced optimistically), and three $10M line items for machine, panels, and land/construction. It has been reported that he concludes, bluntly, “I’ll own this before I die.” Ambitious? Absolutely. Realistic? That’s the big question—today, exascale systems are still largely the province of national labs and hyperscalers, and the leap to zettaflops implies not only brutal hardware scaling but breakthroughs in efficiency, cooling, software orchestration, and massive capital. But if nothing else, the post is a snapshot of a mindset: hacker optimism, stubbornly optimistic math, and a taste of what the next decade of AI infrastructure dreams might look like.

Sources: geohot.github.io, Hacker News