Microsoft's $5,999 desk box claims a petaflop, counted at FP4

Microsoft held its first major Windows event in two years in San Francisco on 7 October, with Satya Nadella, Surface chief Pavan Davuluri and NVIDIA's Jensen Huang on stage. The Surface Laptop Ultra opens at $2,599 and ships on 16 October. The Surface RTX Spark Dev Box opens at $5,999 with 128GB of unified memory and up to one petaflop of AI compute.

“How local AI will shape the next chapter of the PC”

— Microsoft, Анонс мероприятия, 7 октября 2026

A petaflop is two different numbers

Roadrunner at Los Alamos hit 1.026 petaflops on 25 May 2008 and became the first machine to sustain one on the Linpack benchmark. It cost around $100m. Eighteen years later a desk box asks $5,999, which looks like a factor of 16,669.

The two figures count different arithmetic. Roadrunner's petaflop was FP64, double precision, the kind scientific codes need. Desk-class AI petaflops are quoted at FP4, a four-bit format built for tensor throughput with almost no numerical range. Both claims are accurate and they do not measure the same operation. Anyone converting one into the other is comparing a scientific benchmark with a tensor throughput figure.

The price ladder is the readable part

$5,999 for the dev box against $2,599 for the laptop puts the developer machine at 2.3 times the consumer one. Microsoft positions the Spark as the machine you build local AI on and the laptop as the machine you run it on.

Jensen Huang on a Surface stage is the other signal. The dev box carries NVIDIA silicon under Microsoft's own brand, which is a different arrangement from Microsoft shipping its own NPUs. The Surface brand now covers a machine whose headline number belongs to a supplier.

What Windows actually gets

Nadella put Agent 365, Microsoft IQ and Copilot forward as technologies moving closer to the core of Windows, alongside Windows ML for deploying AI runtimes across CPUs, GPUs and NPUs, agent security, and a redesigned Search. Agent security has a reason to be on the list: we counted a fleet of 14 parallel agents running against a mapping API last week.

No Windows 12. The event ran on the premise that the next chapter of the PC is local inference rather than a new version number, and the two prices above are what that premise costs this year. Cloud inference moved the other way: Google halved its image prices on 6 October.

Informational material, not investment advice. Compute figures are vendor claims at the precision each vendor chose to quote.