Waymo designs a custom robo-taxi chip to stay ahead of Tesla

Waymo revealed in a blog post published Thursday, 20 August 2026, that it has designed its own machine-learning accelerator chip, the company's first custom silicon. Autonomous vehicles have only milliseconds to process streaming data from more than a dozen cameras, and Waymo says the new ASIC, built on Taiwanese foundry TSMC's 5 nm process, is designed to convert that raw sensor data into driving responses as quickly as possible. It replaces the Intel FPGAs Waymo had previously used for this job. FPGAs suit low-latency work (the same reason they are common in high-frequency trading) but are notoriously hard to program and offer less compute density than purpose-built silicon.

According to Waymo, the chip's design draws on more than 200 million miles of the company's autonomous-driving data and is tuned to maximize responsiveness, reliability and redundancy. It is built to run both older-style convolutional neural networks and the transformer models used in AI chatbots and image generators, and it performs temporal noise reduction to improve visibility in low light in real time. 'Within those critical milliseconds, advanced ML models build a high-fidelity understanding of the environment to evaluate the safest path forward,' the company said, without naming an individual spokesperson. Waymo claims the chip can deliver more than 1,000 TOPS of AI performance, though it has not disclosed the precision or power level the figure is measured at, which The Register's Tobias Mann says makes it hard to compare directly with existing autonomous-vehicle and robotics platforms. Mann writes that since Waymo is specifically advertising a TOPS figure, the chip is probably running at INT8 precision, putting it roughly in the same range as Nvidia's Drive AGX Thor platform; that comparison is his own inference, not a figure Waymo has confirmed, and The Register says it contacted Waymo for clarification but had not heard back by publication.

Reliability was a design priority: the chips are liquid-cooled using the same coolant system as the vehicle itself, to keep them at steady temperatures despite vibration, shock and weather extremes that datacenter hardware never faces. Each vehicle carries a pair of the ASICs; under normal conditions they operate as a single unit, but if one fails or produces an erroneous result, the other takes over. The Register notes that aircraft and spacecraft often add a third, tie-breaking system for this kind of redundancy, calling Waymo's dual setup unsurprising for a vehicle sharing roads with pedestrians and other traffic. The custom chip does not run everything on the vehicle, though: Waymo says non-machine-learning tasks such as orchestration, data movement and logging are still handled by components from partners AMD, Micron, Samsung, Sandisk and Nvidia, and that it is working on several further custom chips and systems.

Waymo is not the only autonomous-vehicle maker building its own silicon: the article notes that Tesla, which has developed custom chips for its vehicles for years, launched a limited Robotaxi service in Austin after years of missed deadlines. Waymo says it will share more technical detail on the new ML accelerators next week at the Hot Chips conference at Stanford.

Key facts

Waymo's first custom silicon is a machine-learning accelerator ASIC built on TSMC's 5 nm process, replacing the Intel FPGAs it previously used to process vehicle sensor data.
Waymo claims the chip delivers more than 1,000 TOPS of AI performance but has not disclosed the precision or power level the figure is measured at.
The chip's design draws on more than 200 million miles of Waymo's autonomous-driving data and runs both convolutional neural networks and transformer models.
Each vehicle carries a pair of the ASICs for redundancy: they act as one unit under normal conditions, and either can take over if the other fails or produces an erroneous result.
Non-machine-learning tasks like orchestration, data movement and logging still run on components from partners AMD, Micron, Samsung, Sandisk and Nvidia; Waymo will detail the chip further next week at the Hot Chips conference at Stanford.

Why it matters

Waymo moving from off-the-shelf FPGAs to a purpose-built ASIC is a meaningful hardware bet: it signals that the company sees custom silicon, not just software, as a competitive edge in the split-second decisions autonomous vehicles have to make. That puts Waymo in the same camp as Tesla, which has built its own vehicle chips for years and now runs a limited Robotaxi service in Austin. Waymo frames the move around latency and reliability rather than raw throughput alone: the chip is designed to compress the time between sensing and acting, and its design was shaped by more than 200 million miles of the company's own driving data.

Who it affects

Most directly, it affects Waymo's own robo-taxi fleet and the passengers riding in it, since the chip underpins the real-time perception and decision systems the vehicles depend on for safety. It also affects Waymo's hardware suppliers: TSMC now manufactures the core ASIC, while AMD, Micron, Samsung, Sandisk and Nvidia continue to supply the non-machine-learning components (orchestration, data movement and logging) that the new chip does not replace. For the wider autonomous-vehicle industry, it adds another data point in a trend where vendors including Tesla move toward silicon designed specifically for their own vehicles rather than general-purpose chips.

How to use it

There is nothing here for outside developers or buyers: the ASIC is proprietary to Waymo's own vehicles, not a product or platform sold or licensed to others, and no price or contract value has been disclosed. The Register reports that Waymo has begun rolling out the chip, though no start or completion date has been given for that rollout. Anyone wanting the engineering detail Waymo has withheld so far, such as the chip's precision and power draw, will have to wait for the company's presentation next week at the Hot Chips conference at Stanford, where Waymo says it will share more.

How solid is it

The account rests on a single source: Waymo's own blog post, relayed by The Register with no independent testing or corroboration. The headline performance figure, more than 1,000 TOPS, is Waymo's claim alone, and The Register's Tobias Mann says that without knowing the precision and power level behind that number, it is hard to compare against other autonomous-vehicle or robotics platforms. His suggestion that the chip likely runs at INT8 precision, in the same range as Nvidia's Drive AGX Thor, is his own inference, not a figure Waymo confirmed. The Register says it asked Waymo for clarification and had not received a reply by publication, and the piece quotes no named individual at Waymo, only 'the company.'

Risks and caveats

Waymo has not disclosed the precision or power consumption the more than 1,000 TOPS figure is measured at, so the claim cannot be checked against rival hardware on equal terms. There is no timeline for when the chip will be fully deployed across the fleet, no cost or contract figures for the work with TSMC or its component partners, and no performance data for Tesla's own custom chips, so the comparison with Tesla rests on Waymo's hardware claims alone rather than a head-to-head test. The redundancy design, a pair of ASICs that can fail over to each other, has not been demonstrated publicly, and Waymo has not said how long it used Intel FPGAs before switching.

“Within those critical milliseconds, advanced ML models build a high-fidelity understanding of the environment to evaluate the safest path forward”

— Waymo