Skip to content
Fei-Fei Li & Lisa Su at CoRL 2026: Physical AI, Compute, and Robotics' Future
Industry Insights 27 min read

Fei-Fei Li & Lisa Su at CoRL 2026: Physical AI, Compute, and Robotics' Future

At CoRL 2026, Fei-Fei Li and AMD CEO Lisa Su discussed the $8.2B World Labs acquisition, emphasizing physical AI as a trillion-dollar opportunity, hardware-software co-design for robotics, the compute demands of embodied AI, academia's GPU shortage, and whether PhDs remain essential in the AI era.

Machine Heart
Machine Heart
Machine Heart

Fei-Fei Li and Lisa Su's First On-Stage Dialogue After AMD's $8.2B World Labs Acquisition

On November 12, 2026, at the Conference on Robot Learning (CoRL) in Austin, Texas, Fei-Fei Li and AMD Chair & CEO Lisa Su held their first public conversation since AMD announced its ~$8.2 billion all-stock acquisition of World Labs on September 28. The deal, expected to close by year-end, will see Li become AMD Executive Vice President and Chief Scientist, reporting directly to Su. The dialogue was moderated by Peter Stone, UT Austin CS Department Chair, Texas Robotics Founding Director, Sony AI Chief Scientist, and CoRL 2026 General Chair. Stone noted Li's academic lineage: CoRL co-chair Jie Zhu is her former PhD student, and conference organizer Roberto Martín-Martín is her former postdoc.

Reference: YouTube recording of the dialogue (https://www.youtube.com/watch?v=vzyHMAlvglU).

A "Trillion-Dollar" Opportunity and a Chief Scientist Who Understands Workloads

Stone opened by asking about AMD's biggest opportunity in robotics and physical AI, and where World Labs fits. Su replied that physical AI, especially robotics, is a "huge opportunity" worth "hundreds of billions of dollars" over the next few years, emphasizing that success requires truly understanding physics, models, and the entire environment. She framed the World Labs acquisition around workload understanding: "In our industry, optimizing hardware takes many years. The deeper you understand the workload, the clearer you see the technology trajectory, the better you can execute. The philosophy of hardware-software-system co-optimization applies equally to robotics. Over the next five to ten years, this will be an extremely important domain."

Li highlighted their nearly decade-long relationship, praising Su's transformation of AMD. She then articulated a core thesis: "The relationship between modeling and compute in AI has fundamentally changed. AI is compute, compute is AI." She argued that multimodal models — world models, language models, video generation — are converging, demanding hardware-software co-evolution and co-design. Li also stressed that AI is becoming a tool for scientific discovery, calling this an era of "team science," with the AMD-World Labs combination exemplifying that practice.

Will Foundation Models "End-to-End" Solve Robotics?

Stone referenced CoRL's new late-breaking papers track, which juxtaposes papers submitted six months ago with those submitted one month ago, revealing rapid technical shifts. He asked whether future robots will be driven by modular, layered systems or by end-to-end foundation models.

Li began with "we must be intellectually honest." She acknowledged stunning results from frontier labs and at CoRL, saying we are at an early inflection point with much unexplored territory. She predicted Vision-Language Models (VLMs) and LLMs will play increasingly important roles in robotics, but cautioned that many fundamental problems in physical AI still lack root-cause solutions. Her forecast: "Next year's CoRL will look very different." She cited the BEHAVIOR Challenge results (leaderboard: https://huggingface.co/spaces/behavior-1k/2026-challenge-leaderboard) as evidence of unbelievable progress speed, adding that she does not believe we have seen the endpoint.

Li admitted the tension of betting on research directions: "It's nerve-wracking but exciting because we see signals from multiple directions and have a chance to combine these potential approaches."

Su added an industry perspective: "In many vertical applications, there isn't a single model or a single direction. Facing real-world applications, you need a very diverse set of models." She noted frontier large models will excel in some scenarios, but specialized smaller models may be more cost-effective, power-efficient, and lower-latency in others.

Robotics Will Push Compute to the Next Level

Stone turned to compute, noting the massive training requirements and the strict power constraints for deployed robots performing real-time perception, reasoning, and action. He asked whether robotics is fundamentally a systems problem.

Su responded with the now-famous line: "Who in the audience needs compute?" She then stated: "Compute is the foundation. From where we stand today, in many dimensions it is effectively equivalent to intelligence." Drawing on 30 years in computing, she said never has there been a more important moment. Compute is needed at every layer: cloud for training massive models, enterprise, edge where robots operate, and the physical environment. Her conclusion: models will come in many sizes, compute will come in many sizes, but fundamentally it is a solution problem — hardware, software, and system co-optimized to solve problems as efficiently as possible.

Li added a remark that made Su say "I like that framing": "Since we're at a robotics conference, I want to say — and not just because I'm at CoRL, I truly believe — robotics will push compute to the next level. Robotics progress has lagged behind LLMs because it's a hard problem that must interact with the physical world. But the inflection point we're seeing is real. I can't predict the exact slope, but we are definitely in an upward trend, and it will drive massive compute demand at scale, not just in training but also at the edge."

The Public Sector's "Compute Famine"

Addressing the students and professors in the room, Stone asked how to make frontier robotics research more accessible. Li deferred to Su but then delivered the sharpest critique of the session. She revealed that she and Stone have long advocated for public-sector compute, including pushing the NAIRR (National AI Research Resource) Act since 2019.

"Let me be blunt: failing to provide sufficient resources to the public sector is one of the biggest mistakes in the most important moment in AI history." She explained that the public sector — universities and nonprofits — is the first stage of innovation, driven by curiosity. The U.S. innovation engine starts there, and we are training the next generation of inventors and workforce. Yet the public sector in AI faces a severe famine. This must be solved because we need to invest in the future.

Li linked this to responsible AI: "Whether safety or security, the public sector must play a key role in understanding and evaluating AI, telling the public and government the good, the bad, and the ugly of these models."

Su agreed fully, joking that her earlier quip about asking AMD's Chief Scientist "wasn't entirely a joke." She stated AMD is building "many gigawatts" of compute capacity, but demand far exceeds supply. Allocating a portion for long-term research and universities to ensure the next generation of innovators has access is "absolutely one of our responsibilities as industry leaders." She expressed hope that AMD, and other companies, can turn this promise into reality, noting robotics is an area where returns have not yet been fully demonstrated.

Does AMD Have a Robotics Team?

The top-voted audience question on Slido: "Will AMD build a robotics team?" Su countered: "Are you assuming we don't have one?" She revealed AMD already has a "respectably sized Physical AI and Robotics research team, though still in very early stages." From a compute stack perspective, AMD already supplies embedded components — FPGAs, embedded processors, AI engines — that control many industrial robots today. But for the type of robotics the audience researches, "we are still in very early stages."

Su explicitly stated that part of Li's new role is to heavily invest in robotics research and cultivate ecosystem partnerships. "I firmly believe no single company can own the entire stack. That's why startups are doing so much, why there are industrial labs, universities, and big companies. You want the best of all worlds." Open ecosystems, strong hardware, composable software across layers, and diverse labs each pursuing their own work — that, in her view, accelerates robotics.

When pressed for a yes/no on building a team, Su laughed and said "Yes, that's easy."

PhD in the AI Era: Still the Right Path?

Many students asked Stone whether a PhD remains the correct path to research. He passed the question to Li with a warning: "I might not like your answer."

Li said the three on stage are somewhat envious of the students, born into the AI revolution. Her own answer has evolved: "I don't think a PhD is for everyone. Especially in this AI era, learning tools and software engineering have changed dramatically. The paths to becoming a researcher, an innovator, to having impact are more numerous than before."

However, she emphasized this doesn't devalue the PhD: "I still think it's great; I loved my PhD years." She defined the PhD's purpose: "Fundamentally, a PhD exists for curiosity. It's the most protected time in your golden years to chase the boldest, most curious, most interesting, and hopefully most impactful problems."

She added: "No matter how advanced AI becomes, human intelligence remains incredibly precious, perhaps increasingly scarce. Massive token streams now compete with our brainwaves. But the civilization we're building needs human intelligence, needs our bold curiosity. So if you have that burning desire and curiosity, a PhD is still a great path. But you must have the conviction to tackle problems in the most fearless way."

Su shared her personal story: she didn't want a PhD initially; her parents urged her to continue. Those "truly were the best four years of my life." The specific problem matters less because problems change rapidly. What matters is the opportunity to become among the best in a field through freedom to explore. "Life is long; no one can take away your PhD. You'll always know you solved a problem to the highest level in that field."

Stone quipped to his own students: "Note Lisa said four years, not six or seven." Su agreed: "Exactly, don't drag it out; four years is enough."

On research topics, Su addressed AGI hype: "I hear from peers we're one minute away from AGI. I don't believe it." AI progress is stunning, but its performance depends on what problems we apply it to and how we guide it. "Use the best tools, but you still have your own domain accumulation — intuition. When hiring, you can tell who has real intuition versus 'book smart' or today's 'AI smart.' Find a problem you're deeply interested in, build that intuition, then tool changes don't matter." She noted industry lacks patience for this; universities provide the environment for deep intuition-building.

The 20-Watt Brain vs. Billions of Years of Evolution

Stone posed a sharp research direction: current AI is associated with data centers — massive compute, power, data — while the human brain uses ~20 watts and far less data. Achieving comparable performance with far less energy is an important direction, even if it conflicts with selling more chips.

Su didn't dodge: "This is absolutely critical. Compute has no one-size-fits-all; the physical world isn't one size either." She acknowledged today's large data centers and scaling laws, but argued we ultimately need "ubiquitous compute" — lower power, higher efficiency, longer lifespan, lower latency — all must be optimized.

Li cautioned against conflating two things: "We must be careful not to equate the brain's 20 watts with the massive energy needed to train models. Evolution is expensive. Our brains are the result of billions of years of evolution; I don't even know how to calculate the collective energy consumed by the entire animal kingdom." She distinguished neuroscience as a science — understanding, replicating, treating the brain — from AI as a science that need not merely copy the brain, because the computational problems we tackle (coding, simulation, robotics) already surpass human capabilities in some aspects. These are two orthogonal problems.

Unchanging Values and the Hardest Problems

Near the end, Stone read a "tricky" Slido question: the two are in a "honeymoon phase"; any disagreements? Li joked: "I think the Bay Area is more livable; I love the climate there." Then added "But I also love Austin." They concluded they haven't had their first fight.

In closing remarks, Li addressed students facing turbulence: personal transition, field transition, civilization transition driven by "civilization-level technology." She urged: "In an era of change, look at eternal values. As Lisa said, life is long; many things change — problems, challenges, opportunities. But whether you live the life you want depends less on what changes and more on what doesn't. For me, the eternal values this journey taught are: curiosity, integrity, and the fearless pursuit of truth — intellectually daring and intellectually honest."

Su encouraged students to be "extremely ambitious, extremely optimistic" because they live in the best of times. She recounted her early IBM career: after two years, asked what she wanted to become, she chose the VP track over IBM Fellow, thinking she wasn't smart enough for Fellow. A mentor advised: "Always run toward the hardest problems." She didn't fully grasp it then, but it stayed with her. "As engineers, scientists, researchers, our job is to solve problems — very interesting, very important, very exciting problems. We gain joy, success, and push the world forward. So whatever you do, tackle the hardest problems and believe you can make a massive difference."

Final Takeaways

The article distills three key signals from the dialogue:

Acquisition logic: AMD bought World Labs because "chipmakers must understand workloads earlier." Su repeatedly stressed workload understanding and hardware-software co-design; Li previously wrote that without dedicated hardware investment, AI efficiency and scale are constrained. World models and robotics are among the most uncertain workload shapes with the most demanding edge-compute requirements.

No single-paradigm bet: Both speakers avoided betting on one paradigm. Li acknowledged VLM/LLM-driven robots will grow but said physical AI's fundamental problems lack root solutions. Su explicitly argued large models and specialized small models will coexist long-term. "As many model sizes as there are, there will be that many compute sizes" — a narrative that fits a company spanning data center to FPGA and embedded processors.

Compute commitment for academia: Su pledged to allocate part of AMD's compute for long-term research and universities, but gave no scale or timeline. With demand far exceeding supply, fulfillment remains to be watched after Li formally assumes her role. Li has advocated for public-sector compute for years; now she sits inside a compute supplier's decision layer.

Stone's question about power efficiency potentially conflicting with chip sales went unanswered directly. Li's statement — "robotics will push compute to the next level" — may serve as a response: if true, whether more power-efficient or more powerful, the outcome is demand for more, more diverse compute.

Original Source

Signed-in readers can open the original source through BestHub's protected redirect.

Sign in to view source
Republication Notice

This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactadmin@besthub.devand we will review it promptly.

RoboticsComputeAMDLisa SuFei-Fei LiPhysical AIWorld LabsCoRL 2026
Reader feedback

How this landed with the community

Sign in to like

Rate this article

Was this worth your time?

Sign in to rate
Discussion

0 Comments

Thoughtful readers leave field notes, pushback, and hard-won operational detail here.