The next useful AI device may begin with a brilliant idea and a very small chip-design team. It may also stall when that team confronts the cost of licensing processor technology before it knows whether the product will work. Arm is trying to remove that early obstacle by bringing its Armv9 edge AI platform to startups through Flexible Access.
The programme works like a “try before you buy” model for semiconductor design. Participants receive low-cost upfront access to Arm technology, tools, and resources, while qualifying startups can enter at no cost. They can experiment and revise a design, then pay licence fees only for the technology included in the finished product.
Arm says Flexible Access has contributed to roughly 400 successful tape-outs—the point at which a completed chip design is sent for manufacturing—during the past five years. Raspberry Pi, Hailo, and SiMa.ai are among the companies already associated with the programme.
A Billion-Parameter Model Can Stay on the Device
The new platform combines the energy-efficient Cortex-A320 processor with the Ethos-U85 neural processing unit. Together, they can run AI models containing more than one billion parameters locally, without relying on a cloud connection for every inference.
That capacity expands what a relatively small device can understand and do. A camera can interpret a scene rather than simply record it. A home product can adapt to a user’s habits. A robot can respond through vision, voice, and gesture. The intelligence moves closer to the sensor, reducing the distance between an event and the machine’s response.
Paul Williamson, who leads Arm’s Internet of Things business, expects the next wave of AI innovation to happen at this edge: inside the devices, interfaces, and systems located where information is created. That is a different design centre from the enormous clusters that train frontier models, but it may be where many people encounter AI most frequently.
Flexible Access Changes the Startup Calculation
Chip development demands decisions long before a product earns revenue. A startup must choose an architecture, build prototypes, validate performance, and prepare a design for manufacturing. Licensing a wide set of technologies at the beginning can make exploration expensive and lock a team into choices before the evidence is available.
Flexible Access lets designers evaluate the Armv9 components while they iterate. Payment follows the final selection rather than every possibility considered along the way. The structure does not eliminate the cost or risk of producing hardware, but it allows scarce capital to remain focused on proving the product.
Timing is staggered. Cortex-A320 is scheduled to enter the programme in November 2025, with Ethos-U85 following in early 2026. Developers can therefore begin planning around the processor before gaining programme access to the dedicated AI accelerator.
Local Processing Strengthens Privacy and Security
On-device inference can keep speech, images, and behavioural data from travelling to a remote server merely to be interpreted. It can also support devices where connectivity is unreliable or a rapid response matters. Privacy is not automatic, however; sensitive information still needs protection inside the product.
Armv9 includes Pointer Authentication Code and Memory Tagging Extension. These security features are intended to make it harder for software flaws to corrupt control flow or misuse memory, strengthening the environment in which local data and AI workloads operate.
VDC Research predicts AI will become the dominant technology across IoT projects by 2028. If that forecast holds, competition will depend not only on the largest models but on who can turn capable inference into efficient, secure, affordable devices.
Arm’s move is therefore an infrastructure decision aimed at widening the field of inventors. Giving startups earlier access to powerful edge components creates more opportunities for specialised cameras, household devices, and robots to reach a working chip. The most important result may not be a single flagship processor, but the variety of products small teams can now afford to test.
