At its Suzhou technology conference on 22 to 23 September, Intel laid out two answers to one shift: AI value is moving from the cloud model race to deployment at the edge and on the factory floor. Dan Rodriguez, vice-president of edge computing, cited a forecast that about 80 per cent of tokens will be consumed on-device by 2030, and argued the real driver is the closed loop of model, simulation, compute, hardware and field data.
For industry AI, the blocker is no longer technical but trust. Firms will let an agent read, but not act, so Intel’s answer is the Agent Box, a soft-hard appliance built on third-generation Core Ultra, ranging from a fanless edge node to Arc Pro B60 or B70 graphics and full-private Xeon configurations, with graded permissions, audit trails and data staying on-site. Its Open Edge Platform has passed 60 supply partners and 125 optimised ISV components.
For physical AI, Intel’s sharper move is putting the brain and cerebellum on one chip. Its third-generation Core Ultra fusion splits work inside a single SoC: the GPU, new Xe architecture, gives up to 120 TOPS INT8 for model inference and planning; the NPU holds a constant 50 TOPS INT8 watch; the CPU delivers sub-10-microsecond scheduling with motion jitter under 10 microseconds. One chip thinks and controls, instead of three.
The developer burden drops with an embodied-intelligence kit, built with CITIC Tech and Digital China, also on Core Ultra, with BSP, BIOS, ROS 2 and model-optimised reference code preloaded, a redesigned thermal path for full load, and a core board that fits humanoid, quadruped and wheeled chests while swapping only the I/O carrier to reach volume in one to two months. On stage, the “Xiao Ying” humanoid used a VLA model to grasp precisely under varied lighting.
The deeper cleverness is reuse. Both solutions share one processor generation and one toolchain, so a tuned experience carries to the next robot body instead of starting over. Edge AI becomes repeatable infrastructure, not a string of one-off projects.
Editor’s note: This is an adapted translation of the original OFweek report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://robot.ofweek.com/2026-09/ART-8321200-8120-30704817.html.