이미지: METAL LAB 생성
Summary
- Waymo has developed its own chip for use in robotaxis and has already deployed it in its latest-generation vehicles, according to Bloomberg.
- The chip is manufactured on TSMC's 5-nanometer process and delivers more than 1,000 TOPS of performance, putting it on par with NVIDIA's self-driving systems.
- Waymo had long relied on chips from NVIDIA and AMD, but the new chip will power its next-generation robotaxis, built in partnership with China's Zeekr.
- 개발 주체
- 웨이모(알파벳 자율주행 자회사)
- 보도
- 블룸버그, 2026년 8월 20일
- 제조
- TSMC 5나노미터 공정
- 성능
- 1000 TOPS(초당 1000조 회 연산) 이상
- 기존 공급망
- 엔비디아·AMD 칩
- 탑재 차종
- 지커(Zeekr) 협업 신형 로보택시
- 발표 목적
- 비용 절감
The Robotaxi's Brain Is Now Homegrown
Alphabet's self-driving unit Waymo has designed its own chip for use in robotaxis. According to a Bloomberg report, the chip is already running in Waymo's latest-generation vehicles. Until now, the core computing work in these robotaxis has been handled by chips from NVIDIA and AMD, so this announcement marks Waymo's first real step toward cutting that dependence.
Why Build Your Own Chip
Robotaxis have to process a flood of sensor data from cameras, lidar, and radar in real time, then run that data through AI models trained to read the surrounding environment and decide what to do next. Both tasks demand enormous computing power, and up to now that work has been left to hardware from general-purpose chipmakers. Waymo says its new chip both speeds up sensor data processing and runs the self-driving AI models directly. The company points to cost savings as the main reason for moving to its own design.
Performance and Manufacturing
According to figures Waymo has released, the chip delivers more than 1,000 TOPS of processing power and is manufactured by Taiwan's TSMC on a 5-nanometer process. By the company's own account, that puts it roughly on par with NVIDIA's current self-driving systems.
| Category | Before | This Announcement |
|---|---|---|
| Chip supplier | NVIDIA, AMD | Waymo's own design, alongside NVIDIA and AMD |
| Manufacturing | - | TSMC 5nm process |
| Processing performance | - | 1,000+ TOPS |
| Vehicles used in | Previous-generation robotaxis | New robotaxis built with Zeekr |
The new chip will go into Waymo's next-generation robotaxis, developed jointly with Chinese automaker Zeekr. In effect, Waymo is now bringing both the vehicle and the chip inside its own supply chain.
AI Companies Are Building Their Own Chips, Too
Waymo isn't the only one moving this way. As METAL LAB reported on August 8, Anthropic confirmed it has formed its own silicon team to design chips specifically for Claude models. Anthropic, however, said it plans to keep a multi-chip strategy, using its own hardware alongside chips from other vendors rather than going all in. Big AI companies designing their own chips isn't new — Google has its TPUs, Amazon has Trainium — and Waymo's move fits squarely into that pattern.
NVIDIA isn't jumping into this race as a competitor; it's the company selling the chips that power AI to everyone else. Whether it's self-driving cars or chatbots, watching major customers gradually shift to their own custom chips isn't exactly welcome news for NVIDIA.
Editor's Take
It's worth reading Waymo's move as a cost decision above all else. Sensor and compute costs per robotaxi have been a major factor in whether the whole business turns a profit, so swapping general-purpose chips for hardware built to exactly what Waymo needs can meaningfully cut costs. The specs — TSMC's 5nm process, 1,000 TOPS of performance — don't look like an attempt to copy NVIDIA's lineup so much as the result of stripping out anything unnecessary for the narrow task of self-driving.
For anyone who's watched this play out before, the pattern is familiar. Google built TPUs optimized for search and advertising; Amazon built Trainium for its own cloud; now Waymo has built a chip tailored to the single task of running robotaxis. The math keeps repeating itself: once a company's usage hits a certain scale, designing your own chip ends up cheaper than paying the markup on general-purpose silicon.
For Korean self-driving and robotics companies, this news cuts two ways. Very few have the resources to follow Waymo or Anthropic into custom chip design — most will still depend on general-purpose chips from NVIDIA or AMD. But as bigger customers peel away, competition for the remaining customers on price and performance could actually intensify, which means now might be a good time to revisit existing contract terms.
Expect other robotaxi and self-driving companies to make similar announcements in the coming months. How the self-driving segment's share of NVIDIA's revenue moves in upcoming earnings reports will be an early signal of how fast this shift is happening.
Comments