Waymo opens the trunk: a 5nm ASIC, a quadrillion ops, and a supplier list rivals can price

Waymo has published, for the first time, key details of the compute "brain" it carries in the trunks of its robotaxis: chip architecture, processor specs, internal component details, and a list of the hardware suppliers it uses to build the machines ^[1]. According to The Verge, the disclosure came in a company blog post ^[2], which matters because it converts the most-watched autonomy program in the industry from a software narrative into a bill of materials that vendors and competitors can benchmark against.

The peak figure is up to one quadrillion operations per second, sitting behind the rear seats ^[21]. At the front of the pipeline is a custom 5-nanometer ASIC that, rather than passing raw sensor data straight to the main compute, cleans it up, combines sensor inputs, and runs fast AI checks as data arrives ^[3][4]. Around it, Waymo describes what it calls a machine-learning-primary architecture paired with CPUs, GPUs and accelerators to handle orchestration, data movement and logging: "a balanced, heterogeneous system," in the words of the post ^[5].

Three constraints are described as non-negotiable by Satish Jeyachandran, VP of Engineering, and Daniel Rosenband, Compute Lead: responsive, ruggedized, redundant ^[6][7]. Redundancy here means two systems running in parallel so that one can fail ^[8]. Ruggedized means surviving road vibration and temperature extremes ^[9]. The pair also write that they are building something "that would be considered impressive for a data center, with the added complexity of an in-vehicle operating domain and real-time requirements" ^[17]. Two further constraints are product ones rather than engineering ones: the stack cannot occupy the whole trunk, because passengers have luggage, and it has to run nearly silent ^[12].

The growth number is the one to sit with. Waymo says it has scaled compute power 20 times in the past eight years ^[13], which works out to roughly 45 percent compound annual growth in onboard compute ^[18]. That is the cost curve every would-be competitor has to fund, and it runs against a fleet of about 4,000 vehicles in more than 10 cities doing roughly 500,000 paid trips a week ^[10], or about 125 paid trips per vehicle per week ^[16]. Waymo builds the stack itself but depends on third-party suppliers for the components it does not make ^[11], and it claims its technology significantly reduces crashes and injuries relative to human drivers ^[14].

The timing is not neutral. Tesla CEO Elon Musk has called lidar a "crutch" and "a fool's errand" and argues that fusing camera, radar and lidar data creates "sensor contention," with disagreement between sensors producing dangerous ambiguity ^[19][20]. Waymo's counter is that redundancy is precisely what permits deployment at scale ^[15], and the compute disclosure is the supporting exhibit: sensor fusion is expensive in silicon, and it is showing the silicon.

What to watch: whether the named suppliers begin referencing Waymo in their own materials, since a public customer list changes how those firms are valued and how quickly rivals can approach the same parts. Watch also for what Waymo did not put in the post, particularly per-vehicle compute cost, power draw, and who fabricates the 5nm part ^[3]. And watch whether the 20x figure is repeated as a forward commitment ^[13]; a company that has to keep multiplying trunk compute at that rate has a unit-economics problem it will eventually have to describe in public.