The race for artificial intelligence is being fought far from graphics processors too. A data centre must convert and distribute electricity precisely; a robot must command several motors without losing responsiveness. STMicroelectronics is targeting this less visible layer with its new STM32G5 microcontroller series, unveiled on 4 November. The company says the components will bring more compute, analogue control and security to digital power supplies, robots and other embedded systems.
Control chips, not AI accelerators
The distinction matters. A microcontroller does not replace a GPU that trains or runs a large model. It more often manages the physical operations around one: reading signals, regulating a power supply, commanding a motor and responding to sensors. In a computing facility, these tasks help determine stability and energy efficiency. In a robot, they contribute to movement precision and coordination across axes.
According to ST, STM32G5 is built around an Arm Cortex-M33 core running at 280 MHz and delivers twice the compute performance of the STM32G4 generation. The company also highlights a 12-channel high-resolution timer, a position encoder, a programmable logic array and analogue peripherals including converters and comparators. These functions are intended in part for control loops that must synchronise fast electrical switching or multiple motors.
The series is in production, with versions offering between 64 and 512 kilobytes of Flash memory. ST lists a starting price of $1.034 for orders of 10,000 units. That catalogue price does not reveal the total cost of a system or the terms offered to each customer; it does provide a reference point for manufacturers evaluating a new design.
The energy challenge in AI data centres
Electricity demand associated with accelerators makes power control more strategic. Supplies must respond to changing loads, limit losses and remain reliable in dense equipment. ST says the G5 family’s timers can help designers adjust power supplies as server consumption changes. The component does not save energy by itself: results depend on the full architecture, power transistors, cooling and control software.
Wide-bandgap semiconductors, including silicon carbide and gallium nitride, enable high-frequency switching in some converters. They can help reduce size and losses, but they impose demanding timing and design requirements. A microcontroller able to generate precise signals can therefore become one part of the system, without being the sole determinant of efficiency. This is why the announcement is aimed at power engineers rather than users of an AI service.
A role in robotics, from arms to hands
In robots, embedded computing connects perception to action. Sensors report position or force; controllers must turn those measurements into commands quickly and safely. ST cites humanoid hands, which can contain many motors, as an example of the need for compact, multi-axis control. STM32G5’s potential role is local actuator management, not general understanding of an instruction given to an AI model.
The term “physical AI” covers systems in which intelligent software interacts with the material world. But the label does not mean every component contains a machine-learning model. Some functions remain deterministic: a command must be applied on time, a fault detected and the machine moved into a safe state. Functional safety, predictable latency and long-term availability can matter as much as raw computing power.
Embedded security and post-quantum cryptography
ST highlights secure boot, key storage, tamper detection and post-quantum cryptography libraries. It presents these as tools to prepare equipment for evolving cybersecurity requirements. The presence of algorithms or hardware blocks is not enough to secure a product, however. Updates, the supply chain, keys and access must all be managed throughout the device’s life.
Data centres and industrial robots are environments where a breach can have physical consequences. A connected control system therefore needs to reduce its attack surface and isolate sensitive functions. Built-in security can simplify some tasks, but responsibility is shared among the chipmaker, board designer, software developer and operator.
A broader platform strategy
The launch coincides with ST’s announcement of a unified STM32 platform intended to make architectures, software and tools more reusable across industrial, automotive and robotics applications. The company also introduced the STM32A family for automotive and safety systems; its first products are sampling, with broader availability planned for 2027. The goal is to reduce transition work across markets while meeting their distinct qualification and safety requirements.
For industrial customers, continuity in tools may shorten development cycles and ease hiring around a familiar architecture. It may also deepen dependence on a particular ecosystem. Developers will need to compare actual code portability, component availability, certification requirements and migration costs rather than rely only on the promise of a common platform.
Integration will provide the proof
STM32G5 shows how AI’s value spreads into components that do not headline demonstrations: power controllers, sensors, interfaces and actuators. The strategy could benefit ST if manufacturers actually build their next generations around these chips. But announced capabilities do not guarantee adoption, and the claimed improvement over STM32G4 will need to be tested in real applications.
The decisive question is not whether a microcontroller “does AI,” but whether the complete system becomes more efficient, responsive and safe. Results will depend on the power supplies tested, the number of controlled axes, thermal constraints, software and integration quality. As computing facilities and robots grow more complex, these modest-looking components take on strategic importance: they connect digital power to the physical world.
That is also why the market is not limited to humanoid robots or hyperscale facilities. The same needs appear in industrial drives, renewable-energy equipment, commercial drones and increasingly sophisticated vehicles. A common development environment could allow engineering teams to transfer expertise between products, but each application will still bring its own operating temperature, safety rules, certification path and maintenance horizon. Reuse is an advantage only when it does not obscure those differences.
For chipmakers, this shift creates a business opportunity in the supporting infrastructure of AI rather than only in its headline processors. For customers, it creates another design decision: which tasks should run on a central accelerator, and which should remain close to sensors and motors? Local control can reduce latency and keep critical functions working even when a network connection is interrupted. The trade-off is more responsibility for embedded software, updates and cybersecurity at the edge.