Waymo has developed a custom chip for its robotaxis as Alphabet’s autonomous-driving company seeks greater control over the computing hardware that powers its self-driving system. The purpose-built 5-nanometer application-specific integrated circuit (ASIC) is already operating in Waymo’s latest generation of vehicles and is designed to process raw camera, lidar and radar data before that information reaches the main driving system.
The move represents a broader shift from relying primarily on off-the-shelf components toward co-designing silicon, sensors and software for the specific requirements of autonomous driving. Waymo says its custom ASIC delivers more than 1,000 trillion operations per second (TOPS) of machine-learning performance for front-end processing. However, the company is not eliminating outside chip suppliers: AMD, Nvidia, Micron, Samsung, Sandisk, Socionext and TSMC remain among its partners.
Waymo Develops Custom 5nm Chip For Robotaxis
Waymo describes the computing system as the “brain” of its Waymo Driver, translating large volumes of sensor information into real-time driving commands. Unlike conventional driver-assistance systems, Waymo’s vehicles are designed to operate without a human driver ready to take over, making low latency, reliability and redundancy critical requirements.
The company says it has scaled its computing capability 20 times over the past eight years as its autonomous-driving technology has become more sophisticated. Its latest system processes data from 13 high-resolution cameras alongside lidar and radar sensors.
Key Specifications Of Waymo’s Custom ASIC
| Specification | Waymo Custom ASIC |
|---|---|
| Manufacturing process | 5nm |
| AI performance | More than 1,000 TOPS |
| Primary role | Front-end sensor processing and ML |
| Sensor inputs | Camera, lidar and radar |
| Cameras supported | 13 high-resolution cameras |
| Main workloads | Sensor processing, fusion and neural networks |
| Deployment | Latest-generation Waymo robotaxis |
| Manufacturing partner | TSMC |
| Design approach | Purpose-built for autonomous driving |
The chip is not intended to replace every processor inside a Waymo vehicle. Instead, it handles the enormous volume of raw sensor information before the data reaches other parts of the autonomous-driving compute system.
What The Custom Chip Actually Does
Waymo’s ASIC is designed to extract useful information from raw sensor streams in real time. Its specialized accelerators process camera, lidar and radar data, including temporal denoising designed to improve perception in low-light conditions.
The processed information then feeds into Waymo’s inference engine, which runs machine-learning models responsible for sensor fusion and other elements of perception.
This architecture allows Waymo to optimize the hardware for workloads that are particularly important to its vehicles rather than designing around the broader range of applications supported by a general-purpose processor.
From Raw Sensors To Driving Decisions
| Stage | Function |
|---|---|
| Cameras | Capture high-resolution visual information |
| Lidar | Generate 3D information about surroundings |
| Radar | Provide additional distance and motion information |
| Custom ASIC | Processes and fuses raw sensor information |
| ML inference engine | Runs specialized neural-network workloads |
| Driving system | Builds environmental understanding |
| Vehicle controls | Converts decisions into driving actions |
Waymo says the system is optimized for low-batch workloads, which are particularly relevant to real-time autonomous driving. Rather than waiting to process large batches of data, the vehicle needs to continuously respond to incoming information within milliseconds.
Waymo Is Still Using Nvidia And AMD
Despite the headline shift toward custom silicon, Waymo’s strategy is not a complete departure from Nvidia or other external suppliers.
The company explicitly lists Nvidia and AMD among the industry partners supporting its computing system. Waymo says it combines its own machine-learning technologies with CPUs, GPUs and other accelerators to create a heterogeneous computing architecture.
This means Waymo is effectively pursuing a hybrid approach: custom silicon for workloads where specialized hardware can deliver efficiency and performance gains, alongside commercial processors for other computing requirements.
Waymo’s Current Hardware Ecosystem
| Company | Role In Waymo’s Compute Ecosystem |
|---|---|
| Waymo | System architecture, custom silicon and autonomous-driving stack |
| Nvidia | Computing technology |
| AMD | Computing technology |
| TSMC | Chip manufacturing |
| Samsung | Hardware/component partner |
| Micron | Memory technology |
| Sandisk | Storage technology |
| Socionext | Semiconductor technology |
The approach reduces Waymo’s dependence on any single commercial chip supplier without requiring the company to manufacture every component itself.
Why Custom Silicon Matters For Robotaxis
Autonomous vehicles have unusual computing requirements. A data center can often tolerate some variation in workload and latency, but a robotaxi must make decisions continuously while moving through a physical environment.
Every camera frame, lidar measurement and radar signal can potentially contain information relevant to safety. The computing system therefore needs to process enormous volumes of data while maintaining predictable response times.
Waymo says its architecture is built around three principles: responsiveness, ruggedness and redundancy.
Three Priorities For Waymo’s Compute
| Priority | Why It Matters |
|---|---|
| Responsiveness | Driving decisions must be made within milliseconds |
| Ruggedness | Hardware must operate through vibration, shock and extreme temperatures |
| Redundancy | A backup system must be available if a critical compute component fails |
The company integrates its computing hardware with the vehicle’s liquid-cooling system to maintain performance in conditions ranging from freezing winters to extreme heat.
For a driverless vehicle, redundancy is especially important because there is no human driver who can immediately take control following a computer failure.
Waymo’s Compute Has Increased 20x
The rapid increase in computing requirements illustrates how autonomous driving has evolved from relatively narrow driver-assistance functions toward systems that attempt to handle the entire driving task.
Waymo says its computing capability has increased 20-fold in eight years. The company’s current system can process data from 13 high-resolution cameras simultaneously while also handling lidar and radar inputs.
| Waymo Compute Development | Reported Figure |
|---|---|
| Increase in compute over eight years | 20x |
| Latest custom ASIC performance | >1,000 TOPS |
| High-resolution cameras processed simultaneously | 13 |
| Autonomous experience cited by Waymo | 200+ million miles |
| Robotaxi network | 10+ cities |
Waymo says its autonomous system has accumulated more than 200 million miles of fully autonomous driving experience. The company is using that operational experience to guide hardware and software design decisions.
The Economics Of In-House Chip Design
One of the biggest potential advantages of custom silicon is efficiency.
A chip designed specifically for a company’s workload can eliminate some of the unused functionality found in general-purpose hardware. That can potentially reduce energy consumption, improve performance per watt and lower the amount of computing hardware needed for a particular task.
For robotaxis, energy efficiency is particularly important because computing consumes power that otherwise could be used to move the vehicle or extend its operating range.
Waymo says its highly integrated compute system is designed to deliver substantial processing capability without sacrificing trunk space or battery efficiency.
Potential Benefits Of Custom Silicon
| Benefit | Potential Impact On Robotaxis |
|---|---|
| Workload specialization | Faster processing of sensor data |
| Lower latency | Quicker autonomous-driving decisions |
| Power efficiency | Less energy consumed by computing |
| Hardware integration | Better coordination with sensors and software |
| Supply-chain flexibility | Less dependence on one chip architecture |
| Cost optimization | Potentially lower long-term hardware costs |
| Design control | Greater ability to optimize future generations |
However, developing custom chips is expensive and technically demanding. Waymo must absorb semiconductor design costs, validation requirements and manufacturing complexity that would otherwise be handled by commercial chip suppliers.
Nvidia Still Has A Major Role In Autonomous Vehicles
Waymo’s decision to develop custom silicon comes as Nvidia is expanding its own autonomous-vehicle platform.
Nvidia’s DRIVE ecosystem combines in-vehicle computing, software and sensor technologies for Level 2++ through Level 4 autonomous vehicles. Its DRIVE AGX platform is designed to process autonomous-driving workloads in real time, while DRIVE Hyperion provides a broader production platform for robotaxi developers.
Nvidia has also been building a wider robotaxi ecosystem involving automakers, mobility companies and autonomous-driving developers.
Waymo’s custom chip therefore reflects an important trend in the AI hardware industry: companies building specialized applications increasingly want control over the underlying silicon rather than relying exclusively on general-purpose AI processors.
Custom Silicon Could Become More Common In Physical AI
The trend extends beyond robotaxis.
AI systems are increasingly moving from data centers into physical machines, including robots, autonomous vehicles, industrial equipment and drones. These systems need AI inference to happen locally, often under strict constraints on latency, power consumption and reliability.
This creates a strong incentive to develop specialized chips that are optimized for particular workloads.
Waymo’s approach offers an example of what this could look like: combine proprietary accelerators for the most important AI workloads with commercial CPUs, GPUs and other components for the rest of the system.
The result is not necessarily an attempt to eliminate Nvidia. Instead, it is an attempt to make the overall computing architecture less dependent on any single type of processor.
The Bigger Picture
Waymo’s custom 5nm ASIC marks a significant step in the evolution of robotaxi technology. The company is moving deeper into hardware design because autonomous driving requires highly specialized computing that must operate with low latency, high efficiency and built-in redundancy. The more sophisticated its AI models become, the more important the underlying compute architecture becomes.
The development also highlights a broader shift in AI infrastructure. Companies deploying AI at the edge increasingly have reasons to customize their hardware, particularly when workloads are predictable enough to justify specialized silicon. For Nvidia, that does not necessarily mean losing customers, but it does show that major AI users may increasingly combine Nvidia and other commercial processors with internally designed chips.
Looking Ahead
Waymo’s next challenge will be scaling its custom hardware economically across a much larger robotaxi fleet. The company will need to balance chip-design costs, semiconductor manufacturing capacity, reliability testing and vehicle integration while continuing to improve autonomous-driving performance. Its continued partnerships with Nvidia, AMD, TSMC and other suppliers suggest that Waymo expects commercial hardware to remain part of that equation even as its proprietary silicon expands.
For the broader autonomous-driving industry, Waymo’s strategy could encourage other companies to reconsider how much of their AI computing stack should be built in-house. As robotaxi fleets grow, the cost and energy efficiency of onboard computing will become increasingly important. Specialized chips could eventually become a key competitive advantage, particularly for companies operating millions of autonomous miles where even small improvements in performance, power consumption and hardware cost can compound across a fleet.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
