Tesla plans to sell modular AI data center hardware called 'Megapod'

Tesla filed a trademark this month for something called Megapod, and the paperwork is oddly specific for a filing that's really just claiming a name nobody's shipped yet. The application, filed under serial number 99893717, describes a self-contained AI computing system bundling servers, networking gear, power distribution units, and cooling into one enclosure — plus software to manage the whole thing. Translation: Tesla wants to sell the rack-scale building block that AI data centers are built from, not just a chip or a battery.

That's a crowded room to walk into. Nvidia's GB200 NVL72 already defines this category — 72 Blackwell GPUs and 36 Grace CPUs, liquid-cooled, wired together to act like one giant processor — and Dell and Supermicro both build licensed versions of it. Nvidia's own DGX SuperPOD scales those racks past 9,000 GPUs. There's also a smaller, almost comic wrinkle: a company called Submer already sells an immersion-cooled data center enclosure literally named MegaPod, complete with its own trademark, though in a different registration class than Tesla's filing.

What makes this stranger is that Tesla isn't a compute-hardware company right now — it's a customer. Its own Cortex training cluster in Texas runs on roughly 67,000 Nvidia H100-equivalent GPUs, not homegrown silicon. Tesla did try to build its own AI chips: Dojo got shut down in August 2025 after Elon Musk called its second-generation design an