Waymo Built Its Own Chip to Stop Depending on Nvidia and AMD

[SortedCougar]

Waymo has revealed it built a custom ASIC chip delivering more than 1,000 TOPS of machine learning performance, manufactured by TSMC on a 5 nanometer process, and it is already running in the company's newest robotaxi generation. The chip is specifically dedicated to processing the flood of raw camera, radar, and lidar data at the front end of the system before that data ever reaches the car's main driving software.

Up until now Waymo has leaned heavily on off the shelf chips from Nvidia and AMD, integrated into computer systems the company designed itself. This move does not replace those suppliers entirely, Waymo says it is still working with AMD, Nvidia, Micron, Samsung, Sandisk, Socionext, and TSMC across its broader compute stack, but it does give Waymo direct control over a critical piece of hardware that determines how fast its cars can actually react to something happening on the road.

The custom chip is going into Waymo's new Ojai robotaxi platform, built together with Chinese manufacturer Zeekr, which is already rolling out to riders in San Francisco, Phoenix, and Los Angeles with expansion planned to Denver, Las Vegas, and San Diego later this year. Waymo says the goal is millisecond level response times even under the vibration and extreme temperature swings a car actually experiences on real streets, conditions a general purpose chip was never specifically optimized for.

This fits a much bigger pattern across the industry right now, where every major AI company seems to be racing toward owning at least part of its own silicon rather than depending entirely on Nvidia. Tesla tried a similar full vertical push with its own Dojo supercomputer before famously disbanding that team last year and pivoting to newer AI5 and AI6 designs instead, so Waymo's more hybrid approach of building one specialized chip while still buying everything else looks like the more cautious version of the same underlying instinct.

Whether owning this one piece of silicon meaningfully improves safety and reaction time in the messy reality of city traffic, rather than just looking good in a spec sheet, is the part that will only get proven out over the next year of actual deployment

Waymo builds its own chip for its robotaxis, cutting its reliance on Nvidiathe-decoder.com

[Natalie99]

The choice to keep using Nvidia and AMD alongside the new custom chip rather than going fully independent is honestly the detail that makes this feel credible to me.
A company that was truly confident in total self sufficiency would not still be naming five different outside suppliers in the same announcement

[SilverRider]

Millisecond level reaction time genuinely matters enormously in dense urban traffic specifically, where a pedestrian or cyclist can appear from behind a parked car with almost no warning. If this chip actually delivers a real world latency improvement rather than just a marketing number, that is a genuine safety win worth caring about

[ScrollGoblin32]

Front end sensor processing being the specific job for this chip rather than the whole vehicle brain is a smart and pretty conservative scope for a first custom silicon effort. Waymo is not trying to replace Nvidia everywhere at once, just the one bottleneck where owning the design actually buys them something concrete

[error.404]

This feels like the exact same playbook every major cloud and AI company has been running for years now, Google's own TPUs being the obvious precedent inside the same corporate family. Vertical integration on the compute layer keeps proving itself as the move once a company has enough scale and enough capital to actually justify the upfront design cost

[Grace]

TSMC's 5 nanometer process being described as not their most advanced node anymore is a fair point that some of the more breathless coverage of this announcement seems to gloss over. Still plenty capable for this specific application, but worth remembering Waymo is not chasing the absolute cutting edge here

[BookerT]

Building custom silicon around the Zeekr partnership specifically raises an interesting question about how much of Waymo's actual hardware roadmap is now shaped by a Chinese manufacturing partner. Tariff exposure on vehicles built that way already seems to be adding real cost pressure according to some of the other coverage of this same rollout

[SignalMage]

Tesla's Dojo collapse looms over this entire story whether Waymo wants to acknowledge it or not. Building your own chips sounds great in a press release until the actual engineering reality of maintaining custom silicon at scale catches up with you a year or two later