Alfred Lin on Physical AI: Why Humanoid Robots Won't Dominate / Deployed Podcast

The Sequoia partner and former Zappos COO reveals in Episode 2 of Deployed how Amazon discovered robotics, what investors look for in physical AI companies, and why the world isn't built for humans.

Guest: Alfred Lin, Partner at Sequoia Capital (#1 on Forbes Midas List 2024)

Host: Brad Porter, CEO of Cobot, former VP of Robotics at Amazon

🎧 Hear the untold Zappos-Amazon story now in Episode 2:

The Zappos Secret That Changed Amazon Forever

During Amazon's due diligence of the Zappos acquisition in 2009, something didn't add up.

Zappos' cost per order was *lower* than Amazon's own fulfillment cost. How was a shoe company out-operating the operational masters of e-commerce?

When Alfred Lin, then COO and CFO of Zappos, showed them the answer, it surprised everyone: Kiva robots moving shelves to workers, plus a custom-built warehouse management system that routed humans and robots together with ruthless efficiency.

Jeff Wilke's initial reaction was dismissive. Amazon's holiday peak was 25-50x normal volume compared to Zappos' 4x spike. "This will never work at Amazon," he reportedly said.

Three years later, Amazon acquired Kiva Systems for $800 million. Read the acquisition announcement → here.

Brad Porter, who scaled that fleet from 50,000 to over 500,000 robots in five years as VP of Robotics at Amazon, sat down with Alfred to discuss what physical AI means for enterprise operations, why humanoid robots might not dominate, and how trust gets built in automated systems.

What Is Physical AI?

Alfred Lin offers a sharp observation about how we define artificial intelligence:

"AI is kind of a misnomer. Things that seem magical for a period of time, that's what AI was. Then it becomes infrastructure. We've been working on AI since the 1950s, and we just keep moving the goalposts."

Speech recognition was the frontier when Alfred and Brad first worked together at TellMe Networks in 2000. Then computer vision. Then natural language processing. Now large language models.

Physical AI represents the next frontier: autonomous systems that operate in the real world. These systems move goods, navigate environments, and manipulate objects. Unlike digital AI that processes information, physical AI must contend with physics, uncertainty, and real-world consequences.

The pattern Alfred identifies matters for investors and enterprise buyers alike: what seems impossibly advanced today becomes invisible infrastructure tomorrow.

Kiva robots went from "this will never work" to powering over one million robots across Amazon's global fulfillment network. Autonomous vehicles went from science fiction to commercial robotaxi fleets navigating San Francisco's streets. [How Waymo operates today →]

Alfred Lin's Criteria for Backing Physical AI Companies

After two decades investing in technology transformations, from Airbnb to DoorDash to the robotics companies in Sequoia's portfolio, Alfred has developed clear criteria for evaluating physical AI investments:

1. Founder-Problem Fit Over Product-Market Fit

"We're looking for founders who have founder-problem fit. You want someone who understands software, understands operations, understands AI, and understands robotics. That rare combination is what we look for."

Most robotics companies fail because they're strong in one domain but weak in others. The founders who win combine software architecture experience for scaling systems, operational expertise for understanding real-world deployment, AI and ML depth for building learning systems, and hardware intuition for knowing physical constraints.

2. Mission-Driven, Not Exit-Driven

"I don't really think about exits as an investor. What you want is a founder with big dreams. Our mission at Sequoia is to help the daring build legendary companies. We're not helping the daring build companies to flip."

Alfred points to WhatsApp, Instagram, YouTube, and Kiva as examples where acquisitions enabled scale the original company couldn't achieve alone. Kiva had approximately $8-11 million in revenue when Amazon acquired it for $800 million, but Amazon scaled it to millions of robots globally.

3. Solving Major Problems, Not Chasing Trends

"My challenge to founders: if physical AI is going to be that powerful, what major problems are we going to solve for society, for humanity? If you solve that, some things will get disrupted, but we'll find other callings."

Why Humanoid Robots May Not Win

There's a line that's gone viral in robotics circles: "Humanoids are going to win because the world is made for humans."

Alfred offers a sharp counterpoint that enterprise buyers should consider:

"I remind people that the world is quite hostile to humans. The ocean isn't built for humans. Space isn't built for humans. Even highways are built for cars, not pedestrians."

The case against humanoid dominance:

Hospitals are designed around wheelchairs and beds, not walking humans. A wheeled robot with manipulation capabilities may be more practical than legs. Warehouses with flat floors and narrow aisles favor wheels over legs for speed and stability. Underwater and space environments require form factors completely unlike human bodies. Industrial throughput often demands specialized mechanisms like suction grippers, conveyors, and shuttle systems that outperform human-like hands.

Brad Porter adds technical evidence from Amazon's picking challenge:

"The biggest improvement in pick eligibility, from the low 80s to the low 90s, came from increasing suction power. A stronger vacuum meant you could pick up items that a human-like gripper couldn't grasp effectively."

The exciting frontier: When robots discover capabilities humans can't perform.

"I'm going to be fascinated when the robot teaches the robots the ways it can do things that a human cannot do. A human wrist doesn't turn 360 degrees. What's the new form factor that makes things interesting?"

For enterprise buyers, this suggests evaluating physical AI solutions based on task fitness, not humanoid aesthetics.

The Trust Principle: How to Deploy Automation Without Destroying Customer Confidence

Both Alfred and Brad trace their understanding of trust back to Tellme Networks, the late-1990s speech recognition company where they first worked together. Tellme’s voice design team discovered a counterintuitive principle:

"Users were more willing to engage with automation when they knew they could easily reach a human agent. Trust came from the safety valve, not from the technology's perfection."

This principle, now 25 years old, applies directly to physical AI deployment:

Trust is built in failure recovery, not just success cases.

"When we see something that doesn't work, as a consumer, we're disappointed. But when we see the company, or the combination of human and technology, fixing the problem quickly, you turn a detractor into an immense promoter."

Practical implications for deploying physical AI:

Design for exceptions first. How does the system escalate when uncertain? At Cobot, robots pause and call for human assistance rather than proceeding incorrectly.

Make the human-in-the-loop visible. At Zappos and Amazon, Kiva robots started slowly alongside humans. Trust built as workers saw the system work reliably, then expanded.

Invest in recovery speed. The metric that matters isn't error rate alone. It's time-to-resolution when errors occur.

Gradual autonomy expansion. At Zappos, they asked: "Can we run this faster in areas where no humans are present?" Trust expanded zone by zone.

The Economic Case for Physical AI

Alfred Lin makes a data-driven case for why technology-driven workforce transformation creates more opportunity than it destroys:

The agricultural transformation:

Two centuries ago, roughly 80% of Americans worked on farms. One century ago, that number was around 40%. Today, 2% of Americans work on farms. [Historical employment data from BLS →] The result: 98% of the workforce freed for other work, not unemployed.

The spreadsheet effect:

"Spreadsheets actually increased the number of finance professionals. You didn't have to be as good at arithmetic, the spreadsheet did that. So it became about logic and thinking, not calculation."

Rather than eliminating accountants, spreadsheets expanded demand for financial analysis. The number of finance professionals grew.

The bowling pin setter:

Automatic pin-setting machines replaced human pin setters but also reduced operating costs, enabled bowling alley proliferation in the 1960s, and created more jobs in operations, construction, maintenance, and service than the machines displaced.

Alfred's father's story:

Alfred's father grew up in a farming village, destined to work the fields. But he became fascinated with the abacus, became a high school champion, and got a job as a bank teller. From there he studied accounting and rose through the bank. He watched calculators replace abacuses and computers replace calculators. He came to the US, got a master's in computer science, returned to revolutionize the accounting department, and eventually became president of the bank.

One day, he pointed at a spreadsheet on his computer and told Alfred:

"Each cell replaced a person who used to sit at a desk doing that calculation."

He wasn't nostalgic about it. Technology didn't replace his ambition. It unlocked it.

Where Physical AI Breakthroughs Will Come From

Alfred and Brad identify several frontiers where physical AI is advancing rapidly:

Autonomous vehicles as the "ChatGPT moment" for robotics:

"I think we have the ChatGPT moment for robotics. It's called the autonomous car. It can navigate flexible environments it hasn't seen before. That's the breakthrough, from hand-coded instructions to learning in new environments."

Dexterous manipulation:

Current approaches focus on teaching robots human-like hand movements. Alfred believes the breakthrough will come when robots discover novel manipulation strategies humans can't perform.

Flexibility as the game-changer:

"Flexibility in robotics is going to be a game changer. Having a robot that can bring bedding, then drugs, then food, then reading material. When you can have that many use cases, it makes sense to deploy more robotics."

The hardware-software co-evolution:

Every major technology transformation combines hardware and software advances. For physical AI, this means better sensors for perception, more efficient compute for on-device inference, novel actuators beyond human-like joints, and software that learns from deployment through continuous improvement.

Key Takeaways

For Investors Evaluating Physical AI Companies:

Look for founder-problem fit. The rare combination of software, operations, AI, and hardware expertise predicts success. Evaluate exception handling, because how a system fails matters more than demo performance. Assess flexibility potential, since single-task robots face ceilings while multi-use platforms compound value. Consider form factor pragmatically. Humanoid isn't automatically optimal. Task fitness matters.

For Enterprise Buyers Deploying Physical AI:

Start with trust architecture. Design human-in-the-loop escalation before deployment. Expand autonomy gradually, zone by zone and use case by use case, to build organizational confidence. Measure time-to-resolution because recovery speed determines user trust more than error rate. Optimize for your environment. Hospitals, warehouses, and offices have different optimal form factors.

For Founders Building Physical AI:

Solve major problems. "If physical AI is that powerful, what will you solve for humanity?" Build for the long arc. Exits happen, but missions attract talent and capital. Combine disciplines because pure robotics or pure AI expertise isn't enough. Learn from deployment. The systems that win will discover capabilities humans can't perform.

Frequently Asked Questions

What is physical AI?

Physical AI refers to autonomous systems enhanced by modern artificial intelligence that operate in the real world. These systems move goods, navigate environments, and manipulate objects. The key distinction from traditional robotics is the ability to generalize to new environments rather than executing hand-coded instructions. Autonomous vehicles represent what Alfred Lin calls the "ChatGPT moment" for physical AI because they can navigate environments they haven't seen before.

Why did Amazon acquire Kiva Systems?

During Zappos due diligence in 2009, Amazon discovered Kiva robots achieved lower cost-per-order than Amazon's own fulfillment operations. Despite initial skepticism that the technology could scale to Amazon's 25-50x holiday peaks, Amazon acquired Kiva for $800 million in 2012 and scaled the fleet from approximately 50,000 to over one million robots globally.

Will humanoid robots dominate physical AI?

Not necessarily. Alfred Lin argues the world is "hostile to humans." Oceans, space, highways, and industrial environments aren't optimized for human form factors. The winning physical AI designs will likely be task-optimized, potentially discovering capabilities humans can't perform like 360-degree wrist rotation and novel grippers. Hospitals favor wheels over legs, and warehouses often benefit from specialized mechanisms over human-like hands.

How do you build trust in automated systems?

Trust comes from failure recovery, not perfection. Design visible human-in-the-loop escalation, expand autonomy gradually as confidence builds, and prioritize time-to-resolution when errors occur. Users trust automation more when they know they can reach a human easily. This principle emerged from TellMe Networks' voice design research in the late 1990s and applies directly to physical AI deployment today.

What does "founder-problem fit" mean for physical AI?

Founder-problem fit describes the rare combination of software architecture, operational expertise, AI and ML depth, and hardware intuition. Physical AI companies fail when founders are strong in one domain but weak in others. Investors look for founders who've worked across these disciplines because building successful physical AI requires understanding how software, operations, machine learning, and hardware constraints interact.

How does physical AI differ from traditional robotics?

Traditional robotics relies on hand-coded instructions for specific tasks in controlled environments. Physical AI uses machine learning to generalize, adapting to environments and situations the system hasn't seen before. This shift from programmed behavior to learned behavior is what enables autonomous vehicles to navigate new cities and warehouse robots to handle novel products.

—--------

About the Speakers

Alfred Lin is a Partner at Sequoia Capital , ranked #1 on the Forbes Midas List 2024. His investments include Airbnb, DoorDash, Houzz, Instacart, and Zipline. Previously, he served as COO and CFO of Zappos through its acquisition by Amazon. Alfred and Brad first worked together at TellMe Networks in 2000.

Brad Porter is CEO and Founder of Cobot, building trusted physical AI for enterprise operations. Previously, he served as VP and Distinguished Engineer at Amazon, where he led Amazon Robotics and scaled the Kiva fleet from 50,000 to 500,000+ robots. His career spans Netscape, TellMe Networks (acquired by Microsoft), and Amazon.

Listen Now: The Full 50-Minute Conversation

Alfred reveals more about Sequoia's physical AI thesis—including the deals they passed on and why. Brad shares Amazon robotics scaling challenges never discussed publicly.

What you'll learn in the full episode:

- The specific metrics Sequoia uses to evaluate physical AI startups

- Why Amazon's first robot deployment almost failed

- Alfred's prediction for when physical AI reaches mass adoption

Never Miss an Episode

New episodes drop every other Tuesday. Join 5,000+ subscribers getting the Deployed newsletter with episode breakdowns and exclusive insights.

Subscribe to Deployed HERE

News

Blog