Intel's robot pitch: drop the discrete GPU and run everything on one chip

Intel in robotics: Core Ultra Series 3 chips and more

Oversonic Robotics trains its RoBee humanoids on discrete GPUs. But the robots it ships don't carry one.

Instead, the Italian company's humanoid and "centaur" robots now run exclusively on Intel's Core Ultra Series 3 processors, handling language, vision, reasoning and motion control on a single chip with no cloud dependency. Singapore's Sensory AI made the same switch for its robot barista. Trossen Robotics is testing the chip in its robot arms. Circulus in South Korea runs its humanoid operating system on it.

That's Intel's bet for physical AI, according to John Healy, VP and GM of Intel's Industrial and Robotics Division, who I spoke to today at Web Summit in Lisbon, Portugal: most robots don't need a data center on board. They need one efficient system-on-chip that can do control and AI at the same time.

"Our approach has been less about form factor, more about underlying assets that enable the robots to be built," Healy told me.

Intel isn't a household name in robotics, and it's barely in our HumanoidMegaHub database, which tracks about 1,500 robots including roughly 600 humanoids. Nvidia dominates robot compute, AMD is pushing in, and well-funded startups like D-Robotics are building robot brains too. But Intel has been in robots for decades, Healy says, mostly doing motion control and real-time management, then computer vision for quality inspection and defect detection.

Physical AI changes the workload

The Core Ultra Series 3, formerly codenamed Panther Lake, integrates three compute blocks on one die: a CPU, an integrated GPU, and an NPU. The CPU keeps doing deterministic motion control. The GPU and NPU take on inference and fast processing of sensor data.

"So now you have a consolidation of those two worlds, the controlling world and the AI world in a single platform," Healy says.

The engineering argument is that once a model is trained, a robot mostly needs to run inference, quickly and cheaply. Intel calls this "inference-first." A discrete GPU adds cost, power draw, heat and board space, all of which matter more in a battery-powered humanoid than in a server rack. Intel's Nagesh Puppala says in a company report that robotics can't scale on data-center models, and that the real design question is what runs on the robot versus on-prem or in the cloud.

Intel in modern intelligent robots

Sensory AI makes Ella, a barista robot that's essentially a self-contained retail store in a kiosk. Founder Keith Tan previously paired an Intel CPU with a discrete GPU, and the GPU cost more than the rest of the system combined. Now Ella runs entirely on Core Ultra Series 3, including three AI agents running concurrently on different parts of the chip: one for customer interaction, one that learns store-level business patterns, and one that reasons about system health and recovery. A deterministic orchestrator commands the arm. At Computex in June, Intel said Ella would serve up to 200 drinks an hour.

Intel's involved ...

"... from order entry, which can be through mobile phone, touchscreen, or voice, all the way through to prioritizing the order, and a robot arm physically delivering the order," Healy says.

Ella also tracks line density, customer sentiment and inventory, and reorders supplies itself. It runs around the clock in hotels, train stations and airports.

Oversonic Robotics builds RoBee, which has a humanoid upper body on a wheeled base. Intel has worked with Oversonic on two versions: one for manufacturing, moving goods along a production line, and one in elder-care facilities, where it helps healthcare workers manage patients.

Healy thinks the form factor is a preview of where humanoids are heading.

"It's a sort of a hybrid between humanoid high end or top end, and second half is more mobile base, very stable, and allows for larger payload carrying," he says.

Trossen Robotics, based in Illinois, makes robot arms for food service, manufacturing and research. Intel has worked with Trossen on moving arms from rules-based behavior to adaptive, learned behavior via imitation learning.

Trossen's principal solutions architect Marc Dostie told Intel that the chip's I/O is on par with Nvidia's Jetson systems, calling it a very powerful system to develop on.

Circulus in South Korea builds Pibo, a social robot, plus AI workstations for humanoid platforms. It started with education and is now expanding into manufacturing, Healy says. Circulus prioritizes on-device AI so its humanoids can run entirely offline.

Intel is also showing an autonomous wheelchair from a partner, built on the same underlying stack as an autonomous mobile robot, for moving people around airports and hospitals.

There's post-training software too

The software side matters as much as the silicon.

Intel's stack includes OpenVINO Physical AI, the Robotics AI Suite, and Physical AI Studio for fine-tuning. Healy says Intel builds reference pipelines by robot type: one for AMRs, one for intelligent arms, one for humanoids. Intel isn't chasing foundation model pre-training, which Healy expects to stay in the cloud.

It's focused on the step after.

"The real challenge we've found is in taking that, fine-tuning and optimizing it for the robot, and then deploying it at scale," he says.

Two recent deals extend that. Intel and SAP built a proof of concept that connects SAP Extended Warehouse Management directly to autonomous robots, so warehouse tasks get dispatched to robots and status flows back into SAP. And yesterday Intel and Altera announced FPGA-based real-time sensor bridges for robotics and other edge systems, aimed at the growing number of cameras, lidar units and other sensors robots need to fuse.

Why deployment?

Because Intel's own survey says that's where buyers are stuck.

Robot readiness gap report

The Robotics Readiness Gap, conducted for Intel by Coleman Parkes Research, surveyed 800 leaders at organizations with at least $500 million in annual revenue. Respondents came from the U.S. (240), China (160), Japan (130), Germany (120), the U.K. (95) and South Korea (55), across manufacturing, retail, healthcare, smart cities and defense.

The headline: six in 10 expect to run a robot fleet within five years, and on average they think full-scale deployment would double operational output. They also expect a robot to become cheaper than a human hire in their sector within about three years. But only 40% have a formal human-robot workforce strategy.

"And focused ourselves on saying, how do we take what we have in terms of assets and relationships with our market and start to address that deployment gap," Healy says.

Dig into the report and a few numbers stand out for anyone building or buying robots.

• On latency: 23% of leaders say critical robotics applications need decisions in under 10 milliseconds, and 55% need 10 to 100 milliseconds. Only 18% can tolerate more than 100 milliseconds. That's the core technical argument for on-robot compute, though it applies to Nvidia and AMD edge silicon as much as Intel's. Yet just 42% of organizations have a defined, funded edge AI strategy, and edge computing is the area where respondents report the weakest internal skills.
• On ownership: nobody agrees who owns robotics. 29% say the CTO, 18% the CIO, 16% the COO, and only 12% have a dedicated robotics or automation leader. The confidence-versus-preparedness gap is widest in defense, at 48 points.
• On safety: 48% have already had a safety incident involving a deployed robot, 55% say safety concerns have delayed deployments, and only 37% have standardized global robot safety standards. 42% say cyber or data risk is blocking scale.
• On workforce: leaders say robotics has already freed up or redeployed 14% of human workforce capability, rising to 24% by 2030. 75% say they'll need entirely new roles to manage and maintain robots.

About humanoids?

The news isn't great. Asked which robot types will have the biggest impact in their industry over the next five years, respondents put AMRs first at 20%, fixed industrial arms at 14% and inspection robots at 11%. Humanoids got 6%, tied with quadrupeds. Four in 10 said humanoids will never be viable for mass deployment. Nearly two-thirds said current robot form factors are dictated by technology limits rather than real-world need.

That's a survey of enterprise buyers, not builders, and buyers tend to be conservative about hardware they haven't deployed yet. Also note that the top blockers they cite (capex, integration, technical complexity, maintenance and board-level ROI doubts) apply to every robot category, not just humanoids.

But it does line up with Healy's view that the near-term winners may look more like RoBee than Optimus: humanoid on top, wheels on the bottom, and a single efficient chip inside.

(Which, frankly, is a humanoid to me: a general-purpose robot, as Matter Venture Partners’ Haomiao Huang told me recently.)

Physical AI data is also scarce, Healy says, so much of the learning will happen in real deployments. That's the case for open, adaptable platforms, and it's a not-so-subtle shot at vertically integrated stacks.

"Vertical implementations have their place, but they tend to limit variety and adaptability," he says.

Sources

• Interview with John Healy, VP and GM, Industrial and Robotics Division, Intel (video call, November 2026)
• The Robotics Readiness Gap report (Intel, August 2026)
• Intel Core Ultra Series 3 for edge AI robotics (Intel newsroom, May 2026)
• Intel and SAP physical AI demo (Intel newsroom)
• Intel and Altera sensor integration (Intel newsroom, November 10, 2026)
• Oversonic Robotics on HumanoidMegaHub