Kodiak AI announced on October 1, 2026 that it is using Amazon Web Services as its primary cloud provider for developing AI models and running safety simulations for its autonomous driving system. The announcement comes ahead of a driverless trucking service on public roads in Texas, planned for later in 2026. Preparing for the service is not an announcement that it has already begun operating in that form.
According to the publication in the AWS press center, the infrastructure supports model training, data storage, operations center software and compute-intensive testing. One of the central tools is BreakPoint, an internal Kodiak system that searches in simulation for rare situations that could lead to a crash. The announcement concerns development and testing infrastructure; it is not independent certification of the vehicle's safety.
The search is specifically for the trip that does not end well
In a technical article published in June, Kodiak described how BreakPoint introduces realistic, variable errors into the driving system's signals and deliberately searches for conditions in which it fails. Instead of only repeating familiar scenarios, the system is intended to expose combinations the team has not yet encountered on the road.
Kodiak reported at the time on a scenario in its industrial environment in which the perception system could incorrectly estimate the speed of stranded vehicles. According to the company, discovering this in simulation made it possible to address the problem before encountering it in the real world. This is a case the company describes from its development process, not the result of a test we performed.
The article connects BreakPoint to probabilistic risk assessment, PRA. This framework examines how common a particular scenario is, the likelihood of a crash within it and how severe that crash would be. The model is intended to update as evidence accumulates and to represent uncertainty too. These tools predate the new cloud announcement and therefore should not be presented as an October launch.
Here lies the difference between accumulating mileage and searching specifically for failure. Many journeys can provide useful information, but they do not guarantee encountering every dangerous combination. Simulation allows conditions to be selected and repeated after a software change. The next question is how well the simulated conditions represent what the vehicle will actually encounter.
Finding and addressing a scenario does not end testing either. One fix can change the response in another situation. A serious assessment therefore needs to examine both the original failure and the behaviors on which the system already relied. More computing power may expand the number of experiments, but does not by itself determine whether the right experiments were chosen.
The Texas target was defined before the cloud provider was chosen
In a September 25 announcement, Kodiak identified the Dallas-Houston route as the first lane for its driverless long-haul trucking service. The company described daily preparation trips along Interstate 45 and driverless testing on closed courses. It said trips on the launch lane consistently finish without intervention, while noting the presence of a safety observer on the journeys described.
That last detail changes the meaning of the statement. A trip on which a person did not touch the wheel is not the same as one with nobody in the cab. The former can be part of the evidence leading toward the latter, but the intermediate stage must not be erased when describing progress.
Kodiak reported that its ARM metric stood at 93% at the end of August. According to the company's definition, the metric describes the share of claims and evidence in the safety case that are substantively complete. It is not a 93% safety score, not the share of successful trips and not a 7% probability of a crash.
An internal progress metric can explain how much of the verification process has been completed, but it does not replace examining the quality of the evidence. This distinction is useful beyond autonomous driving too: completing items on a checklist and demonstrating real-world performance are related questions, but they are not identical.
When comparing future safety data, it will also be important to know what is included in the denominator. Trip counts, total distance and the number of complex situations are different measures. Fewer interventions on a fixed route can represent progress, but without details of driving conditions it is difficult to extrapolate to another route. This is a limitation of the comparison, not a claim that the reported improvement does not exist.
The same principle applies when moving from an industrial environment to a public road: the target conditions themselves must be assessed. Previous experience is relevant background, but does not remove the need to present evidence about the new operating environment.
The model is only one part of the system on the road
Kodiak's technology page describes a combination of AI software and modular hardware. SensorPods units include LiDAR, radar and cameras, and the system has redundant power, steering and braking components. The company presents this as a design intended to maintain control even when a component fails, subject to compatibility with platform requirements.
That description explains why a cloud service announcement does not describe the whole driving process. Sensors need to supply information, software needs to interpret it and the vehicle needs to act. The testing process must examine both component failures and the relationships between components, because a correct answer at one stage does not guarantee a correct response from the system as a whole.
The use of AWS therefore does not imply that the truck waits for a remote server's decision every time it brakes. The announcement details development, simulation and operations work in the cloud; it does not describe transferring every real-time driving decision to the cloud. The distinction is between where the system is improved and tested and how it operates while driving.
What the agreement demonstrates, and what remains to be demonstrated
The announcement confirms a choice of computing infrastructure and presents the role the companies intend for it. It does not disclose a contract value, a new vehicle count or an independent finding about the expected accident rate. Operations on public roads in Texas and future expansion in the United States remain plans the company seeks to carry out.
To follow progress, it will be useful to look in subsequent announcements for a clear distinction between closed-course testing, driving with an observer and commercial service with nobody in the cab. Details of operating conditions matter too: the route, the kinds of situations it includes and the boundaries within which the system may operate.
For readers in Israel, the announcement does not introduce a local service or a timetable for Kodiak trucks to enter the country's roads. Its relevance lies in the way AI is used to test a physical system: not only to produce an action, but to search in advance for conditions in which that action could be wrong.
A cloud connection can expand the scale of that work. Kodiak's test will be its ability to turn the additional computation into better evidence, address the failures found and define the service's boundaries precisely before expanding it. That is the practical measure between an infrastructure partnership announcement and a truck driving without a driver on a public road.