Musk bets on both horses: nvidia inside, tesla silicon on the way
Elon Musk just admitted what Wall Street whispered for months: Tesla will keep vacuuming Nvidia GPUs by the cargo-ship load while it races to replace them with a home-grown chip called AI5, tuned for the Optimus humanoid and the long-promised robotaxi fleet.
The confession landed on X at 3:14 a.m. Pacific, Musk’s favorite hour for market-moving insomnia. “Tesla will indeed buy vast quantities of Nvidia hardware,” he wrote, before adding the kicker: AI5 will punch “well above its weight” and is already sampling in Austin labs. Translation: the company that once mocked custom silicon is now spending $25 billion to build Terafab, a 100–200 billion-chip-per-year monster somewhere in the Texas desert.
The silicon two-step nobody else can afford
It’s a hedge that legacy automakers can only dream of. While Ford and GM beg suppliers for allocation, Tesla locks in today’s Nvidia H100s—each card currently trading at roughly the sticker price of a Model 3—while prototyping the accelerator that could make those same GPUs obsolete inside its own fleet. The move mirrors Apple’s A-series playbook, but at 100× the thermal envelope and with a robot that still walks like a drunk C-3PO.
Musk couldn’t resist praising Jensen Huang, calling himself a “huge admirer” minutes after signaling he intends to eat Nvidia’s lunch in automotive inference. The compliment doubles as a life-insurance policy: Tesla needs priority shipments through at least 2026, the earliest Terafab can risk a tape-out.
Money, heat and the ghost of 2028
Internal timelines seen by Coastal Code peg AI5 at 5 nm, 750 mm², and a brutal 600 W TDP—cooled by the same octagonal liquid plate designed for the Cybertruck’s inverter. One engineer jokes the chip “runs hotter than a Falcon 9 re-entry,” but if yields hold, each Optimus unit will ship with twin AI5 packages, giving the 5’8″ robot more raw INT8 throughput than a pair of Xbox Series X consoles, yet drawing less juice than the average garage heater.
The financials are equally muscular. At 200 billion chips a year, even a razor-thin $0.30 margin per unit turns Terafab into a $60 billion annual cash register, dwarfing Tesla’s entire 2023 automotive gross profit. Of course, that volume assumes Tesla becomes the planet’s largest semiconductor consumer, shipping 30 million Optimus units and an unknowable swarm of robotaxis before the decade closes.
Lo hay que ver para creerlo, say the old-line foundry vets in Guadalajara who already outsource overclocked FSD silicon. They’ve watched Musk promise “full self-driving next year” since 2016; now he promises a fab the size of a small city. Yet the purchase orders are real, the bulldozers are already scraping caliche outside Austin, and Nvidia’s logistics team just reserved another 747 out of Taipei.
Bottom line: the same man who once dismissed lidar and knocked carbon-fiber is now betting he can out-engineer the world’s most profitable chipmaker—while simultaneously keeping it on speed-dial for parts. If Terafab fires up on schedule, the first casualty won’t be Nvidia; it will be every tier-one supplier still praying Tesla remains a customer instead of a competitor. Either way, Musk wins the headline today and, potentially, the supply chain tomorrow.