Generative Bionics AI Data Factory in Italy: Robots, AMD Helios and the Missing Megawatts

What Generative Bionics announced

Generative Bionics, a robotics company founded to commercialise physical AI research from the Istituto Italiano di Tecnologia, announced the plan at CES Unveiled Milan on 21 October 2026. Its press release describes a network of three facilities. The first is a 5,000 square metre hub in Italy, followed by a second site elsewhere in Europe and a third serving global markets.

DCD added detail on 26 October 2026. The first hub will sit at an undisclosed location in Liguria, where the company is based, and could be partially operational by the fourth quarter of 2027. It will run on AMD Helios, which DCD describes as a double-width rack-scale system. The company has raised €70 million, according to both DCD and the press release, with AMD Ventures among the backers.

The purpose is training data. The press release points out that language models learn from text online, while humanoid robots have no comparable record of how to act in the physical world. Fleets of Gene.01 robots will perform industrial tasks continuously in reproduced factory environments. Each environment and robot configuration is then rebuilt as a digital twin, so the same task can be repeated across thousands of variations in layout, objects, materials and failure modes. Models trained in simulation go back to the robots for validation, and data from the floor feeds the next training cycle.

What a physical AI data factory asks of a data center

This site puts a working robot floor, a compute hall and the people who run both in one building. The press release says robotics, AI and data engineering teams will be co-located with the infrastructure to manage the pipeline from raw robot data to datasets, models and fleet deployment.

That changes the load profile. A training cluster for language models reads a prepared dataset. Here the dataset grows every shift. The press release lists vision, motion, contact, force, proximity and temperature signals, plus readings from the robot's full-body multimodal skin. All of it has to be stored, cleaned and turned into training sets before the Helios racks can use it. The release says AMD will supply compute and storage for physical and synthetic data.

The compute side runs two kinds of work at once. Model training is long and steady. Digital-twin simulation, which the release says runs thousands of scenario variations in parallel, comes in bursts that follow the test plan. Both share the same racks, power and cooling, so the people scheduling jobs and the people running the facility need the same view of what is running where.

What is still undisclosed

The 5,000 square metres cover the whole facility, robot floor included, so they say little about compute density. DCD states that IT capacity has not been disclosed. Neither source names the town, the grid connection or the cooling design. Rack-scale AI systems draw far more power per rack than general enterprise IT, so the cooling choice will shape the building more than its floor area does.

The press release says operational by the fourth quarter of 2027. DCD says it could be partially operational by then. For anyone supplying power, cooling or services to the site, the first concrete signals will be a grid request, building permits and the cooling specification.

What the operator watches in Sensaka

A site like this needs one record that links the compute hall to the work it serves. Sensaka for AI Data Centers keeps power, cooling, accelerators and jobs in one inventory. Its CMDB relationships run from rack to server, accelerator, job and tenant, so a training run and a simulation batch can each be traced to the racks they use.

AI workload operations shows jobs from Slurm and Kubernetes with their queue reasons, failure reasons, checkpoints and training time lost. When a burst of simulation work pushes training jobs back in the queue, the operator sees why. GPU pool operations shows pools, idle GPUs and quotas by project, which matters when two teams share the same Helios racks.

On the facility side, Sensaka reads rack power and, where liquid cooling is installed, CDU status, supply and return temperatures, flow, pressure and leak events, next to PUE and WUE. This is monitoring only; control of the cooling plant stays with the site's own systems. In the AI Cockpit the operator can ask which GPUs are idle or which racks are near their rated power, get a board for that question, and render it as a 30-second data video for the weekly review. Faults raise tickets that sync with ServiceNow or Jira. Sensaka works with NVIDIA, AMD and other accelerators.

Takeaway

The Generative Bionics hub is an AI factory whose raw material comes from robots in the same building. Until the company publishes a power figure, the useful facts are the floor area, the AMD Helios platform and the fourth quarter of 2027 target. Operators planning similar sites should expect data to arrive every shift, and two kinds of compute work to compete for the same racks.

Frequently asked questions