Waymo Opens the Trunk: Custom 5nm ASIC Pushes Robotaxi Brain Past 1,000 TOPS

For a company famous for saying almost nothing about its hardware, Waymo just gave the industry an unusually candid look at what sits in the trunk of its robotaxis. In a blog post titled “A look under our trunk: what’s in our compute,” published August 20, 2026, the Alphabet-owned autonomous driving company detailed its onboard computing architecture, introduced a purpose-built 5nm ASIC of its own design, and — remarkably — published the names of the seven suppliers it builds the system with.

The disclosure is significant both technically and strategically. It arrives as competition in the robotaxi market intensifies, as Washington tightens rules around Chinese vehicle hardware, and as Waymo’s parent company Alphabet doubles down on custom silicon across everything from data centers to consumer devices.

A Front-End Powerhouse, Not a Full Brain

Precision matters here, because early coverage blurred an important distinction. The new chip is an application-specific integrated circuit (ASIC) built on a 5-nanometer process, and Waymo describes its job as handling “the massive influx of raw data before it reaches our core ML brain.” Its specialized accelerators extract critical information from raw lidar, radar, and camera streams — including temporal denoising that improves perception in low light — and feed the results into a purpose-built inference engine that runs sensor-fusion models.

In other words, it is a front end. It prepares what the driving system sees; it does not single-handedly decide what the car does. Waymo says the ASICs “alone deliver over 1,000 TOPS of ML performance dedicated to front-end processing and ML models” — a figure Bloomberg noted would put that one component in the same throughput class as Nvidia’s latest autonomous-driving platforms. That comparison, as The Next Web carefully pointed out, measures one component doing one job against complete driving platforms — a claim Waymo itself did not make.

Just as important: Waymo has not dropped its merchant silicon suppliers. AMD, Micron, NVIDIA, Samsung, Sandisk, Socionext, and TSMC all appear on the newly published partner list. The company describes a “balanced, heterogeneous system” that pairs its own ML silicon with what it calls the best CPUs, GPUs, and accelerators available. The custom chip adds a capability; it does not replace the parts around it. And Waymo hinted that more is coming, calling the ASIC “just one of several exciting custom components we’re developing.”

Three Non-Negotiables: Responsive, Ruggedized, Redundant

Waymo’s post, written jointly by VP of Engineering Satish Jeyachandran and Compute Lead Daniel Rosenband, frames the compute design around three requirements — and they explain why a robotaxi computer looks nothing like a data-center card.

Responsive. The system operates entirely onboard, “constantly processing decisions within milliseconds.” Waymo has coined the phrase “pixels-to-actuation” for the delay between a photon hitting a sensor and the vehicle acting, and says raw compute has scaled 20x in eight years to shrink that latency. The company optimizes especially for low-batch regimes — the real-time, small-workload conditions that benchmark numbers rarely capture.

Ruggedized. The computer lives in a car, not a server rack: constant vibration, shock, and temperature swings from freezing Midwest winters to blistering Phoenix summers. It integrates directly with the vehicle’s liquid cooling system, runs silently, preserves battery efficiency, and still leaves room for luggage.

Redundant. This is the most distinctive design decision. “Our compute is designed like two independent engines,” the post explains. They normally run as one unit executing full parallel workloads — but if one develops a fault, the other takes over seamlessly. There is no human backup driver to fall back on, so the redundancy has to live in the hardware itself.

The latest system processes high-fidelity data from 13 high-resolution cameras simultaneously and in real time — a count that matches the Zeekr-built Ojai, Waymo’s purpose-built four-seater that also carries four lidar sensors. Bloomberg reports the Ojai is now being equipped with the custom chip. A day before the chip announcement, Waymo opened the Ojai to all riders in Los Angeles, Phoenix, and San Francisco, ending its invitation-only period.

Why Now: Scale, Supply Chains, and Politics

The timing of the disclosure is hard to separate from Waymo’s market position. The company now runs roughly 4,000 vehicles across more than 10 US cities (11 by its own website count), completes about 500,000 paid trips per week, and draws on more than 200 million miles of fully autonomous driving experience to inform its designs. It plans to expand to nearly 20 more cities, plus London and Tokyo. In February, it raised $16 billion from Dragoneer, DST Global, Sequoia, and Alphabet at a $126 billion valuation.

Publishing a supplier list is not normal corporate behavior for a company this secretive, and the context matters. Washington has spent the year tightening rules on Chinese vehicle hardware, and Waymo’s newest vehicle body is built by Zeekr, an arm of Chinese automaker Geely. Designing the brain in-house while a Chinese partner builds the body is one way to reduce that exposure — though neither company framed it that way. The supplier list also quietly signals to regulators and customers that the silicon story runs through TSMC, Micron, and other US-aligned vendors. Socionext — a Japanese firm that designs custom ASICs for others rather than selling its own chips — is the name to watch: it’s the kind of partner a company turns to when it wants proprietary silicon without building a full chip division from scratch.

There’s also the Alphabet angle. Google has been designing its own silicon for years to control AI infrastructure costs — TPUs for the data center, Tensor chips for Pixel phones — and just this week Marvell handed Google a $12.2 billion share option in a custom-chip arrangement. Waymo’s ASIC extends that playbook to the road.

Context and Caveats

Two sober notes belong in any honest account. First, the node number — 5nm — is a mature process, not the leading edge. BYD has built a 4nm driving chip for a car that costs $10,000. Node size is not capability, and the two parts do different jobs, but the ranking isn’t what most readers would assume. Second, Waymo has had a rough year on the safety front: a sixth recall in June pulled nearly 3,900 robotaxis over software that let vehicles enter closed freeway construction zones (following 13 such incidents and a temporary freeway service suspension in May), on top of a May recall of ~3,800 vehicles over flooded-roadway behavior. Faster front-end processing does not fix a decision made downstream of it, and Waymo does not claim otherwise.

All of Waymo’s performance and safety figures are self-reported; no outside body has audited them. Engineers will present more detail at the Hot Chips symposium, where specifications get questioned by people equipped to question them.

What It Means

The robotaxi industry is converging on a familiar pattern from cloud computing: at sufficient scale, you stop buying general-purpose hardware and start designing your own. Tesla built its own inference chip years ago; Waymo has now shown its hand. When your fleet runs 500,000 paid trips a week, every watt and every millisecond of latency compounds into real money and real safety margins. A chip purpose-built for temporal denoising and low-batch inference doesn’t need to top any benchmark — it just needs to do exactly one job, deterministically, millions of times a day.

The lesson for the broader AI industry: as models move from the cloud into physical products — cars, robots, drones — the winning architecture won’t be the biggest GPU. It will be heterogeneous systems where custom silicon handles the sensor firehose, merchant chips handle the general compute, and redundancy is designed in from the start, because there’s no human in the driver’s seat to catch the failure.