Intel is building a robotics chip roadmap that targets 700 TOPS by 20283 while holding separate acquisition talks with Tenstorrent that could value the startup above $5 billion2. ANALYSIS The convergence of these moves, set against Nvidia's own push into edge hardware through RTX Spark, frames a three-way architectural contest over which silicon stack will run the next generation of autonomous machines and AI-native PCs.
Why it matters
Intel unveiled its Robotics AI Suite and Physical AI Studio at an invitation-only workshop in Seoul, positioning Core Ultra Series 3 as a single-chip alternative to the multi-component hardware stacks that currently dominate AI robotics deployments4. Cho Min-sung, vice president at Intel, told a Seoul conference that the company plans to introduce Nova Lake early next year, followed by Razor Lake AX in 2028, "targeting performance of around 700 TOPS". Nvidia's Jetson AGX Orin, the incumbent edge robotics SoC, delivers approximately 275 TOPS.
The big picture
Intel's pitch rests on integration. Core Ultra Series 3 puts a CPU, GPU, and Neural Processing Unit on one die, delivering up to 180 TOPS of total AI compute, with about 120 TOPS coming from the NPU. Intel says the CPU cores handle the real-time control loop on a fixed-interval schedule, the GPU handles vision-language-action model inference at 10 to 30 frames per second, and the NPU handles power-sensitive always-on tasks like sensor signal processing. The company calls this a solution to the "deployment gap": its commissioned research found roughly 60 percent of senior leaders across six countries expect to deploy robot fleets within five years, yet only 40 percent have established a formal strategy for managing a mixed human-and-robot workforce.
Three South Korean robotics firms, Movensys, One AI, and Circulus, demonstrated systems on Intel's platform at the Seoul workshop. Intel says more than 130 design engagements are underway using Core Ultra Series 3 processors for edge AI and robotics, though it characterizes those as qualification-phase numbers, not production deployments.
Meanwhile, Bloomberg reported that Tenstorrent, the AI chip startup run by veteran processor designer Jim Keller, was holding separate talks with Intel and Qualcomm about a potential sale that could value the company above $5 billion. Intel shares jumped as much as 6.4 percent on the report. Earlier reporting from The Information pegged a possible Qualcomm deal at $8 billion to $10 billion; Keller pushed back, telling reporters Tenstorrent had not been in acquisition talks with Qualcomm specifically, though he added that the company talks to "many companies" about strategic deals.
On the Nvidia side, Jensen Huang described the new Surface Laptop Ultra, the first Microsoft device powered by an Nvidia chip, as a product that took "four years to build" with "several thousand engineering years dedicated to it"5. The device runs Nvidia's RTX Spark platform with an N1X processor and RTX Blackwell graphics, supporting up to 128GB unified memory.
ANALYSIS Intel's robotics roadmap and its Tenstorrent courtship address the same vulnerability from different angles. The Razor Lake AX target of 700 TOPS would represent a substantial leap from the current 180-TOPS Core Ultra Series 3, but it arrives in 2028, the same year most South Korean customers are targeting for humanoid production. That timeline gives Nvidia room to iterate on Jetson and gives Tenstorrent, which builds on the open RISC-V architecture, time to scale its own designs.
A Tenstorrent acquisition would pair Jim Keller's design team with Intel's in-house manufacturing. Intel says the integrated Xe2 GPU in Core Ultra Series 3 is built on its 18A process node. ◆ That combination of a proprietary leading-edge foundry process and a RISC-V portfolio could let Intel offer edge customers architectural options that fabless competitors cannot match on their own.
The Surface Laptop Ultra launched on October 16, 2026, putting RTX Spark hardware into developer hands. ◆ Nvidia's entry into PCs through this Microsoft partnership gives its inference software a new install surface beyond the Jetson line already established in robotics.
One detail from Intel's Seoul materials is worth noting: Intel's Physical AI Studio listed Nvidia's GR00T N1 among its supported training targets alongside Pi Zero and SmallVLA. ◆ That cross-platform compatibility suggests Intel is positioning its toolchain to capture developers regardless of which training framework they use, rather than betting on ecosystem lock-in alone.
What's next
Intel's Nova Lake is scheduled for early next year, and general availability of OpenVINO Physical AI is expected in the second half of 2026. Whether Intel or Qualcomm closes a Tenstorrent deal, or whether Keller keeps the company independent, will likely become clearer as the Bloomberg-reported banking process advances. South Korean customers are planning to begin mass production of autonomous mobile robots next year, with most targeting 2028 for humanoids.