Intel Promises 700 TOPS for Humanoid Robots With Razor Lake-AX Chip Due in 2028

Intel has confirmed that a new processor built specifically for the age of humanoid robots will arrive in 2028, promising roughly four times the artificial-intelligence computing power of its current robotics platform. The chip, called Razor Lake-AX, was confirmed by a senior Intel executive at a semiconductor conference in South Korea on Friday, turning months of industry leaks into an official roadmap.

According to Digital Trends, citing a report by the South Korean outlet The Elec, Intel Senior Vice President Cho Min-sung laid out the company’s robotics chip plans during the “Semiconductors Powering Humanoids” seminar held in Seoul’s Samseong-dong district on October 16. The executive said the upcoming Razor Lake-AX processor is expected to launch in 2028 and deliver approximately 700 trillion operations per second — known as TOPS — of AI computing power, aimed squarely at robotics and humanoid machines.

“We plan to provide performance at the 700 TOPS level with the next-generation Nova Lake early next year and Razer Lake AX in 2028,” Cho said, according to Wccftech, which also covered the seminar and cited The Elec’s reporting. The figure represents nearly four times the 180 TOPS of combined AI performance offered by Intel’s current Panther Lake-based robotics platform, which the company sells under its Core Series 3 branding.

A chip designed for robots, not laptops

Razor Lake-AX is not being positioned as a conventional computing processor. According to Wccftech, the chip will resemble AMD’s Ryzen Halo platform — a large system-on-chip combining multiple processor cores with a massive integrated graphics unit and expanded memory capabilities — and will target premium devices alongside what Intel calls “Physical AI” workloads.

Months of leaks had already sketched the rough outline of the processor. Notebookcheck reported that the chip was initially rumoured under the name Nova Lake-AX, as a high-performance accelerated processing unit intended to compete with AMD’s Strix Halo lineup. Leaked specifications — none of which Intel has confirmed — suggest the chip could carry either 16 or 32 Xe3P graphics cores, making it potentially one of the most powerful integrated-graphics solutions Intel has ever built. Technetbooks added that the configuration is designed for real-time machine vision and the intensive AI inference required by autonomous robotics hardware.

Other unconfirmed details come from hardware leaker Jaykihn, cited by Notebookcheck: the AX variant is expected to reuse Nova Lake’s Coyote Cove performance cores and Arctic Wolf efficiency cores, rather than the newer Griffin Cove and Golden Eagle cores rumoured for the rest of the Razor Lake family. The chip is also expected to feature on-package memory — a design choice last seen on Intel’s Lunar Lake “Core Series 2” processors — which would give the large graphics unit the substantial memory bandwidth such designs demand. Wccftech noted that a large, unified memory pool is considered essential for AI workloads, making the return to the on-package-memory approach a logical step.

Nova Lake arrives first, in 2027

Razor Lake-AX is the second act of a two-part robotics roadmap. Intel plans to bring Nova Lake for robotics applications to market in early 2027, with commercial deployment beginning then, according to Technetbooks. Razor Lake-AX will follow in 2028. The roadmap pushes the company’s robotics platform from today’s 180 TOPS toward the 700-TOPS level across the two generations.

Beyond the silicon itself, Intel is building an ecosystem around the robotics push. Digital Trends reported that the chip could power the company’s existing robotics development platform, known as Robinson Bay, alongside software tools including OpenVINO, Physical AI Studio, and the Robotics AI Suite. The additional computing power would allow robots to analyse camera feeds, identify objects, understand their environment, and make decisions with less dependence on remote servers — a significant advantage for humanoid robots, which must coordinate perception, movement, and real-time control, and where cloud latency can be a liability.

Intel has also attached dates to the market it is chasing. The company forecasts mass production of autonomous mobile robots for some customers by 2027, and of humanoids by 2028, according to Digital Trends. Notably, Intel has not confirmed whether Razor Lake-AX will also arrive in consumer PCs — the chip’s home, for now, is robotics and premium devices.

How 700 TOPS stacks up

The headline number invites comparison with the competition. According to Wccftech, Intel stated that Razor Lake-AX’s 700 TOPS would far exceed NVIDIA’s Jetson AGX Orin platform, which offers 275 TOPS — roughly a 2.5-times advantage. But NVIDIA’s newer hardware tells a more nuanced story: the Jetson Thor T2000 offers around 400 teraflops of FP4 AI compute at 40 watts, while the T3000 reaches 865 teraflops at 70 watts.

Intel is not the only chipmaker converging on this space. Wccftech pointed to AMD’s Ryzen AI MAX+ Halo chips and NVIDIA’s recent move into the PC space with its RTX Spark system-on-chips as evidence that “Physical and Agentic AI-specific chips are becoming the new standard.” Intel’s Panther Lake is already described as a potent solution in the physical-AI space, with Nova Lake and Razor Lake intended to take the company further into the growing market.

The numbers come with caveats

As Digital Trends cautioned, the 700-TOPS figure remains a roadmap target rather than a fully fleshed-out performance specification. Intel has not specified which numerical precision the number is based on — a critical detail, since TOPS figures can inflate dramatically at lower precisions — nor how much of the performance will come from the CPU, GPU, and neural processing hardware respectively. Power consumption, memory capacity, and real-world performance are all unknown. The figure, in other words, should not be taken as proof that robots running on the chip will be four times faster than those on today’s hardware.

None of the leaked architectural details — the Xe3P core counts, the on-package memory, the core architectures — has been confirmed by Intel, which so far has disclosed only the product name, the 2028 launch window, and the 700-TOPS AI performance target.

Analysis: Why It Matters

For Intel, the Razor Lake-AX announcement is less about a single chip than about choosing a battlefield. The company has spent the last several years losing ground in the data-center AI race — the market dominated by NVIDIA’s graphics processors — while executing a difficult internal turnaround. Rather than fighting that war on NVIDIA’s terms, Intel is betting that the next computing platform will be physical: machines that see, move, and decide in the real world. That is a market where nobody has yet locked in the equivalent of the CUDA software moat, and where Intel’s long history of platform-building, developer tools, and volume manufacturing might actually count.

The numbers deserve scepticism, however. TOPS has become the megahertz of the AI era: a single figure that compresses a forest of technical assumptions into a marketing-friendly headline. A 700-TOPS claim without a stated precision is, strictly speaking, incomplete information. The same hardware might be rated at a fraction of that figure in the higher precisions that serious robotics workloads often demand. Intel’s own caution-by-omission here is standard industry practice — every chipmaker plays this game — but it means the “four times Panther Lake” comparison is best read as direction rather than measurement. The serious evaluation will come when Intel discloses the precision, the power envelope, and the CPU-GPU-NPU split.

The more interesting technical story is memory. A giant integrated graphics unit with dozens of Xe3P cores is only as fast as the data feeding it, which is why the rumoured return to on-package memory matters. Intel tried this once with Lunar Lake and then stepped back; bringing the design back for Razor Lake-AX suggests the company concluded that for AI-centric system-on-chips, wide and fast memory access matters more than the modular flexibility that on-package memory sacrifices. Wccftech’s observation that a large, unified memory pool is “essential for AI” captures the industry’s consensus: the bottleneck has moved from raw compute to data movement. If the leaks hold, Razor Lake-AX is being engineered around that reality.

Then there is the question of timing. A 2028 launch is an eternity in semiconductor terms — two full product cycles away. By then, the humanoid robot market Intel wants to serve may already have its supply chains settled. Tesla’s Optimus, Figure AI’s machines, and Unitree’s rapidly maturing lineup are all racing toward volume production, and each is likely to bring its own preferred silicon partners. Intel’s forecast — autonomous mobile robots in mass production by 2027, humanoids by 2028 — aligns conveniently with its own chip schedule, which is either a sign of coordinated planning or of a roadmap drawn backwards from the marketing. The Seoul audience, presumably, noticed.

Finally, consider the venue. Intel chose to confirm its most aggressive AI chip roadmap at a humanoids conference in South Korea — not at a US tech event, and not through a press release from Santa Clara. South Korea is home to Samsung’s foundry business, a dense ecosystem of robotics startups, and a government pushing hard into humanoid robotics. Announcing there is a signal to that ecosystem: Intel wants to be the platform underneath the Korean robotics industry’s ambitions. It is also, quietly, a signal to Samsung and the memory industry that on-package, high-bandwidth memory designs are coming — and that Intel intends to be a customer.

The physical-AI chip race is now unmistakably a three-way contest — NVIDIA bringing Jetson and its new PC chips, AMD pushing its Halo-class APUs, and Intel betting on Nova Lake and Razor Lake. All three see the same future: intelligence moving out of the data center and into machines that walk factory floors and, eventually, living rooms. The 700-TOPS figure may be part roadmap and part theatre, but the direction of travel is not in doubt.

What to Watch Next

1. Specification disclosure: the next milestone is Intel putting numbers behind the headline — the TOPS precision, the CPU, GPU and NPU contributions, power draw, and memory configuration. Without those, the 700-TOPS claim cannot be benchmarked against rivals.
2. Early-2027 Nova Lake launch: the first robotics deployment will show whether Intel can ship physical-AI silicon on schedule — a credibility test after years of manufacturing delays.
3. NVIDIA’s response: the Jetson roadmap and any counter-announcement aimed at humanoids will indicate how seriously the market leader takes Intel’s challenge.
4. Consumer PC question: Intel has not said whether Razor Lake-AX will reach ordinary computers; a confirmation would turn a robotics story into a mainstream computing one.
5. Design wins: which robot makers actually commit to the platform — the difference between a roadmap and a business.

Sources

• Digital Trends — “Intel confirms Razor Lake-AX which could deliver nearly four times Panther Lake’s AI performance” (17 October 2026)
• Wccftech — “Intel SVP Confirms Razor Lake-AX Will Hit 700 TOPS & Launch In 2028” (17 October 2026)
• Notebookcheck — “Intel Razor Lake-AX gets official 2028 roadmap confirmation after months of leaks” (17 October 2026)
• Technetbook — “Intel Plans To Release Nova Lake and 700 TOPS Razor Lake AX for Robotics” (17 October 2026)