Tensor says its Robocar will go on sale in the U.S., the EU, and the Middle East in 2026. (Tensor)

In late February, Tensor and Arm announced a multi-year strategic collaboration to develop a new foundational compute architecture that would create the “world’s first agentic AI personal robocar.” Tensor first announced in August 2025 that it was developing its L4-capable Robocar.

What is an agentic robocar, and how much compute will it require? Tensor COO Jewel Li said to be considered agentic, a vehicle will have to be able to complete tasks on its own, the way agentic software-only agents can (at least theoretically) do in the virtual space today. Li said, in a video discussion with Drew Henry, the EVP of physical AI at Arm, that the car’s SAE Level 4 central autonomy stack (with NVIDIA-accelerated AI processing) is about 8,000 TOPS, but autonomy is just one of the compute platforms in the vehicle. Tensor integrated 433 Arm-based cores in each of its test vehicles, which the companies said was enough for their AI-first approach to L4 autonomy and is, unsurprisingly, also the highest concentration of Arm technology in a consumer vehicle to date.

The overall idea behind the AI-first design of the Tensor Robocar is that, since AI is shifting to the physical world by becoming embodied in machines, new autonomous driving systems will be defined by their ability to sense, decide and act, so Tensor is defining its vehicle around intelligence instead of trying to put autonomous driving tech onto a legacy OEM platform.

The Tensor Robocar uses 37 cameras, five lidars, 11 radars, 22 microphones, 10 ultrasonic sensors, three IMUs, GNSS, 16 collision detectors, eight water-level detectors, four tire-pressure sensors, a smoke detector, and triple-channel 5G connectivity. Of the 433 cores, Tensor went with the Arm variety pack to power its server-class compute. It will use Neoverse AE for high-throughput AI processing, Cortex-X for agentic AI cabin and peak performance system control, Cortex-A for drive-by-wire, lidars, redundancy, and general compute, Cortex-R for real-time safety-critical systems, and Cortex-M for low-power subsystem management. The distributed cores will also each have their own distributed safety-capable intelligence throughout the vehicle, “from the onboard supercomputer to the smallest sensors,” the companies said.

“This car, it's a four-wheel robot,” Li said. “It's not a humanoid robot, so there are things that it cannot do. But the definition of an agentic robot, an agent, is that it can complete tasks from beginning to end without your intervention, without the need of a human operator. And in the future, this is the car that will complete tasks on its own.”

Tensor claims it will have the Robocar available for sale in the U.S., the EU, and the Middle East in 2026. The Silicon Valley start-up is also working with Autoliv, ZF, Continental, NVIDIA, AMD, Qualcomm, Samsung, and Oracle in the development of the Tensor Robocar. In early February, Tensor announced that it had received approval from the UAE Government’s Regulations Lab to test it’s Level 4 technology in the United Arab Emirates.

“Delivering personal autonomous vehicles at scale requires more than breakthrough AI and autonomy, it demands advanced engineering for safety, redundancy, reliability, and power efficiency,” Li said in a statement. “Our collaboration with Arm leverages their deep, decades-long expertise and leadership in AI-capable compute, which alongside our broader ecosystem of strategic partners ensures that the Tensor Robocar moves seamlessly from advanced technology to real-world roads, safely and reliably.”