# T-Mobile Kinetic Edge: Figure Robot Inference Tests and What Remains Unproven

T-Mobile is testing where robot intelligence should run, but moving computation into the network creates a different set of engineering demands.

By Paras Madan | 2026-10-27

T-Mobile's October 27 physical-AI update describes tests with Figure to run portions of its Helix S1 and S2 models nearer the network edge. It also says T-Mobile and AWS are building initial Kinetic Edge infrastructure. This is an architecture and partnership update, not a generally available robot-inference service.

## What T-Mobile is testing

In the [CTO's announcement](https://www.t-mobile.com/news/network/physical-ai-is-leaving-the-lab), the Figure relationship progresses from eSIM connectivity to evaluating model execution outside the robot. The stated goal is to preserve responsiveness while shifting some computation closer to the network.

It also describes development work with Serve Robotics and exploration of connected-camera traffic analysis in San Jose.

The announcement supplies no public price, general-availability date or measured end-to-end result for the Figure tests.

## Why the robot's control loop matters

Figure's [Helix 02 technical explanation](https://www.figure.ai/news/helix-02) provides useful background, rather than a specification for this particular network trial. Its hierarchy separates slower semantic reasoning in System 2 from System 1's full-body joint targets at 200 Hz and System 0's balance and coordination at 1 kHz.

Those rates correspond to five-millisecond and one-millisecond update intervals. They help explain why moving a model off a robot is more complicated than moving a chatbot to another server. Sending observations, waiting for inference and returning an action all consume time. A missed deadline matters differently when the output affects a moving machine.

**Different parts of the task can tolerate different delays.** A high-level request to choose the next object and a low-level response that maintains balance do not necessarily belong in the same place. The engineering question is which computation can move without making the complete system less dependable.

T-Mobile separately mentions a pre-6G radio trial with ping latency as low as four milliseconds. That is not a published Helix response time, a guaranteed service level or proof that robot control can be moved wholesale into the network.

## What the AWS description adds

Kinetic Edge was already named in the October 8 [AWS re:Invent telecom guide](https://aws.amazon.com/blogs/industries/your-guide-to-aws-for-telecom-at-reinvent-2026/). Its session description identifies AWS Outposts, Amazon SageMaker and Amazon Bedrock AgentCore in the architecture. October 27 is therefore a new account of the work, not the first public appearance of the name.

The AWS description is conference-program context, not a substitute for deployment documentation. It does not establish which services are used in each Figure test, which sites are operational or what a customer can purchase today.

## What would make the tests convincing?

For a robotics team, the next useful evidence would include response-time distributions under movement and congestion, behavior during connectivity loss, the amount of onboard computation actually removed and the resulting change in battery use. An average latency number alone would leave several of those questions unanswered.

Shared computation could reduce the hardware each robot carries, but it also creates dependencies on connectivity, scheduling and remote capacity. The valuable result would be a well-defined division of work that continues to behave predictably when the network is imperfect. That is the standard to watch as these evaluations become deployment claims.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain complex AI infrastructure and product updates through clear stories that developers and buyers can understand.

[Book a 30-minute call](https://cal.com/paras-madan/30min).