Intel Says Its 2028 Razor Lake-AX Chip Will Hit 700 TOPS of AI Compute — 3.88 Times Today's Panther Lake

Intel has put a hard number on its ambitions in physical AI: 700 TOPS. Speaking at a "semiconductors powering humanoid robots" workshop held in Seoul's Samseong-dong business district, Intel senior vice president Cho Min-sung confirmed that the company's next-generation Razor Lake-AX system-on-chip will launch in 2028 with 700 TOPS of AI compute — a figure that, if it ships on schedule, would make a single edge chip competitive with dedicated robotics accelerators.

The jump is large by any measure. Intel's current Panther Lake "Core Series 3" platform delivers about 180 TOPS, so Razor Lake-AX represents a 3.88-fold increase across two product generations. It would also be roughly 2.5 times the AI throughput of NVIDIA's Jetson AGX Orin, the module that today anchors much of the industrial robotics market with its 275 TOPS. Intel's roadmap places Nova Lake first, launching in 2027; Korean reports of the workshop indicate Nova Lake is also targeting the 700-TOPS class, but Razor Lake-AX is the product meant to carry that compute into production.

Architecturally, Intel is describing something closer to an integrated supercomputer than a PC chip. Korean and US trade outlets report the design follows the template of AMD's Ryzen Halo: a much larger integrated GPU — TechPowerUp expects 16 or 32 Xe3 cores — paired with on-package memory, a technique Intel last used on Lunar Lake and one that matters enormously for AI, where bandwidth to a large unified memory pool often decides what model sizes can run locally. The company is explicitly framing the platform for premium devices and for robots rather than for mainstream laptops.

The commercial pitch is aimed at the factory floor. Intel's argument to equipment makers is that a single SoC should handle real-time motion control, AI inference, and sensor processing at once — closing what the company calls the "deployment gap" in industrial automation, where proof-of-concept robots stall when integrators must stitch together separate control and compute stacks. Intel recently formed a dedicated Physical AI Group, says it has more than 130 design engagements for Panther Lake in robotics, and expects its OpenVINO Physical AI software framework to reach general availability in the second half of 2026.

Korea is the proving ground. Money Today reported from the workshop that Intel is working with KAIST and Daime Research to commercialize "dark factories" — plants where AI systems autonomously manage entire production lines — as the first flagship use case for the roadmap. The partnerships give Intel reference deployments in exactly the market, industrial robotics, where NVIDIA's position is strongest and most defensible.

The strategic context is a company rebuilding its edge story. Intel has already named a data-center inference line (the GPU D-Series) and pushed into custom AI silicon, but edge robotics is the one AI market where the x86 incumbent can compete on integration and power efficiency rather than on training-cluster scale. A 700-TOPS SoC that runs perception, planning, and motor control on one chip, without discrete GPUs and their cost, is a credible answer to Jetson — provided it arrives.

Which is the caveat worth keeping. This is a 2028 product confirmed at a workshop, not at a launch event, and silicon roadmaps slip. But the direction is unambiguous: the TOPS race has moved out of the data center and onto the robot. If Intel delivers Razor Lake-AX anywhere near spec, robot makers gain a second source for the compute layer of physical AI — and NVIDIA's near-monopoly on the industrial edge gets its first serious test.