GPUs in the Wine Cellar: Why Techies Are Hoarding AI Compute in Their Homes

AI Summary

Former Microsoft executive Gary Flake built a home AI computer equipped with seven Nvidia GPUs to train custom models on brain-wave data. He says owning a machine assembled from discounted parts was cheaper than renting cloud compute, despite additional electrical and cooling upgrades.

The basement of Gary Flake’s contemporary home in a quiet neighborhood in Bellevue, Wash., is furnished with everything you’d expect in a multipurpose man cave. There’s a cozy-looking couch situated between a large screen and a digital projector, a Peloton treadmill and a rowing machine. The real showpiece of the space, however, is a custom computer that resembles an oversized erector set, one with seven large Nvidia graphics processing units dangling from its frame. Flake, a former Microsoft executive and startup founder, built the computer to train custom AI models on terabytes of data from electroencephalograms, the tests that measure electrical activity in the brain—part of a startup project he has since moved on from. To train his AI models, Flake could have paid a cloud service to train them on the industrial-strength GPUs inside their data centers, but he said it was far cheaper to assemble his own machine out of discounted parts (total cost: under $6,000) than to fork over fees to Amazon Web Services or Microsoft every month. That’s true even when including the $10,000 to $20,000 he spent for electrical upgrades and a new HVAC system he bought to keep his basement cool while his AI computer throws off a lot of heat. “I want to own,” said Flake, who plans to begin using the computer again for a new AI project in the coming months. “I don’t want to rent.”