When Agility Robotics chose Fremont, California, for its new Digit training center, it didn’t just plant a flag in Silicon Valley. It scored a point for a specific computational philosophy: training complex autonomous agents — humanoids navigating warehouses and unstructured spaces — demands a compute stack under direct control, on-premise, not offloadable to a remote data center.

The news is plain: the company, already known for putting its robots to work with Amazon and GXO, will build a facility to test Digit in simulated and real-world scenarios. But the location — a stone’s throw from Tesla’s factory and its Optimus robots — and the very nature of the operation say much more about where the industry is heading.

Latency and the sim-to-real loop

Training a bipedal robot to walk, grasp objects, and avoid obstacles isn’t just a brute compute problem. It’s a latency issue: every extra millisecond in the feedback loop between sensors, motors, and policy model updates degrades convergence. Simulating thousands of parallel scenarios with realistic physics and then transferring learned skills to real hardware (the so-called sim-to-real process) requires GPUs and CPUs physically close to the rack where the robot is tested. It’s no coincidence that Agility invested in its own facility, not a generic cloud cluster: when deployment is edge by definition — a robot is literally a mobile node — the pre-deployment training phase greatly benefits from being on-premise, inside the company’s perimeter.

This approach resonates with those in the Large Language Model world evaluating self-hosted architectures for fine-tuning on sensitive data. The principle is identical: the tighter the loop between experiment, measurement, and model update, the more it pays to keep the infrastructure locked down. And if for an LLM the cost of cloud fine-tuning can be acceptable, for a robot learning physical tasks the competitive variable is iteration speed. Moving terabytes of sensor data and 3D video over the internet creates bottlenecks that slow down the entire development team.

Proprietary data and sovereignty

Then there’s the data question. Digit robots collect continuous streams of environmental information: shelf layouts, human movements, warehouse floorplans. For a company providing logistics automation, that data is a strategic asset. Uploading it to the public cloud means exposing it to unauthorized access risks, compliance headaches with enterprise clients, and, more subtly, giving up a competitive advantage. Major robotics players — from Boston Dynamics to Tesla — are progressively shifting intensive training to internal clusters, often based on high-end NVIDIA GPUs (A100, H100) interconnected with low-latency networking.

Agility hasn’t disclosed the hardware specs for the new center, but the industry trend shows that on-premise for robotic simulation workloads isn’t a whim: it’s an operational necessity. Multi-GPU workstations linked via NVLink or InfiniBand can run reinforcement learning environments where thousands of Digit instances learn in parallel not to fall, to stack boxes, to navigate moving obstacles. And all this happens inside the company’s local network, without depending on external latencies or cloud provider policies that can shift with little notice.

The Fremont effect: a hub for on-premise robotics

Proximity to Tesla isn’t merely symbolic. It creates a geographic axis where talent, sensor suppliers, and system integrators cluster, pushing hardware vendors to consider solutions tailored for robotics workloads. Already, manufacturers like Supermicro or ASUS offer compact GPU servers optimized for edge computing that could populate training centers like Agility’s. And as control models become more complex — integrating visual perception, language models for voice commands, planning — the demand for VRAM and memory bandwidth will grow, favoring those who can deliver scalable, locally deployed architectures.

Agility’s move highlights an uncomfortable truth for pure cloud advocates: the more AI becomes embodied in physical hardware, the closer the development cycle moves to that hardware. It’s no accident that Total Cost of Ownership (TCO) discussions for on-premise training infrastructure, once confined to hyperscalers, now involve robotics and advanced manufacturing companies. For those evaluating on-premise deployment, trade-offs exist: capital investment, in-house expertise, vertical scalability. But when the final product is a robot that can’t afford to phone a cloud API for every step, the alternative is no longer a luxury. It’s the premise of the product itself.