How advances in robotics, foundation models, and vertical AI are converging — and what it means for founders

Over the past three years, our team at nvp capital has made 20+ investments behind the thesis that foundational models will enable digital automation across a range of under-digitized industries. Many of these industries remain under-digitized because the previous generation of SaaS products functioned as systems of record and data repositories, but didn’t address the repetitive, operational work required day-to-day. That gap remains a significant opportunity — and we believe it’s exactly where the next generation of category-defining companies will be built.

As exciting as the future is for applied AI in digital use cases, I’ve been thinking more about how this same pattern of AI-driven automation will take hold in the physical world — and where we are in that journey. While many industries remain under-digitized or reliant on legacy software, there are also large sectors where physical tasks still dominate the value chain. In these environments, work may be process-driven, but it hasn’t been digitized because doing so requires robotic automation.

Jensen Huang of Nvidia has been saying that the ChatGPT moment for physical AI is coming — and he doubled down on that view at CES this month. I agree. I see strong parallels between where physical AI applications are today and where digital AI stood in late 2022, when the launch of ChatGPT unlocked a wave of early, high-impact use cases.

Why This Time Might Actually Be Different

Robotic automation isn’t a new idea. The promise has existed for decades, long before foundation models. There are real examples of robotics deployed at scale — Amazon’s use of Kiva Systems (venture backed acquisition in 2012!) for warehouse operations is one, Waymo and robo taxis are another — but, by and large, the broader robotics thesis has underdelivered.

There are many reasons for that underperformance, but three stand out: robotic quality, intelligence, and the cost of deployment relative to edge use cases. In theory, robotic automation is compelling. In practice, deployment breaks down when robots are placed in non-standardized environments — where lighting changes, parts arrive unpredictably, and edge cases vastly outnumber “normal” cases.

What’s different today is that progress in robotics and intelligence is finally happening in lockstep.

On the robotics side, the cost curve continues to decline. Actuators, batteries, sensors, and onboard (edge) compute are all getting cheaper and better, making “off-the-shelf” hardware far more viable than it has ever been.  If you’re interested in going super deep on how we’ve gotten here Packy McCormick’s Not Boring piece on the Electric Slide is a great read. 

On the intelligence side, foundation models are making genuinely agentic systems possible. In robotics, that shows up first through vision-language-action models (VLAs) that map perception and intent directly to motor commands. Over time, these action models will be augmented by the emerging world models that provide richer spatial and temporal context — allowing robots not just to act, but to reason about how their actions will change the environment.

The Data Problem

Even if advances in inference and robotics hardware are no longer the primary constraint, that doesn’t mean we’re suddenly seeing humanoids or task-specific robots deployed productively at scale. The next major bottleneck is training data.

Simulation helps, but real-world robot training data is still scarce, expensive, and slow to collect. Every mistake in production costs real money. Every near-miss can require stopping a line, pausing equipment or not moving the boxes.  As one founder with a pilot in market said, “95% right doesn’t work in industrial robot deployments like it does in much of enterprise software.”

I believe the data problem is addressable, but it will be solved incrementally. We’re already seeing progress through a combination of real-world data collection by foundation model providers, training data companies and deployed robotic companies – plus simulated data.  The use of teleoperations, where humans guide robots through tasks to generate initial training datasets, is the bridge that can help get deployments in the field faster and actually accelerate data collection at the same time since models will learn from human actions.

So the founder challenge becomes: you need data to deploy robots, but you need deployed robots to collect data.

The strongest founding teams we’re meeting aren’t waiting for perfect VLAs. They’re picking a narrow enough problem that they can get initial deployments with 85% automation with humans handling the long tail. Those early deployments generate the data needed to move from 85% toward 99% — and that initial wedge is what makes everything else possible.

What Does This Mean for Company Building + Founders

If the stars are finally aligning across robot capability, cost, inference, and access to data, that creates a platform for founders to build robotic, industry-specific applications — vertical AI in the physical world.

The real acceleration will come from founders with an earned view of a specific industry and its workflows. Building physical AI isn’t about generic automation; it requires understanding which problems are actually worth solving, how environments differ, and where failure modes and safety requirements change. A robot that folds laundry can’t transfer-learn its way into operating a CNC machine. Physics and context matter.

This is why vertical specialization will win. The strongest teams will be led by people who understand both the industry and the technical constraints of automation — or who have the ability to assemble that expertise. I’m looking for founders who have spent time watching manual processes eat up time and capital, who know exactly which 20% of tasks drive 80% of the pain, and who now see that advances in VLAs and robotics finally make automation viable for their specific problem.

This moment will also unfortunately produce many products driven by what’s technically possible rather than what is needed in the market. The tension between capability and need  is where being founder-driven pays off for us: we may have a high-level thesis, but we rely on mission-oriented founders to show us where the real opportunity lies.

Why Start-ups Can Compete With Big Tech

The staggering amount of money it’s taken to reach this point in self-driving cars, somewhere well north of $100B, makes the promise of robotic autonomy seem out of reach for regular start-ups. Even a platform company like Figure requires so much capital to build their humanoids that it seems like a no-go.

But I don’t think the first wave of large outcomes in physical AI will come from humanoids or generalized robots. Instead, I expect them to come from the vertical software equivalent of robotics: non-humanoid systems designed to automate a specific, high-value manual workflow — and then expand over time if they work.

Think ship cleaning, metal cutting and drilling, crop inspection, picking and packing, unloading trucks, or other narrow but painful tasks that I haven’t ever thought of, but are critical. Early deployments need to focus on workflows that are both valuable enough to justify automation and structured enough to avoid endless edge cases.

And just like the biggest ERPs and horizontal software players didn’t dominate every vertical – Tesla isn’t going to build a robot specifically for commercial HVAC installation. Amazon’s warehouse robots won’t work for hospital logistics. Figure won’t optimize for the specific workflows of chemical plant maintenance.

The nerd in me is super excited about horizontal robotics platforms because that feels like living in the future, but the investor in me is more excited about founders who say “I’m building the robot that automates this specific thing in this specific industry, and once we nail that, we’ll expand adjacently.”

Where the First Big Opportunities Will Emerge

The best opportunities are in verticals where:

• Labor shortages are severe enough that customers will pay for 85% solutions
• Environments are controlled enough for robots to perform most tasks reliably
• Big Tech lacks existing operations that create a data advantage
• Start-ups can own the customer relationship and build a data flywheel over time

The best part about being founder-first is I don’t have to nail exactly where the opportunity will happen, just meet great founders, but FWIW I think the first wave of venture-scale outcomes will be in manufacturing and logistics, but not the parts people expect. Not automotive assembly lines, which are capital-intensive and deeply entrenched. Instead, I’m looking at:

• Mid-market manufacturers who can’t afford custom automation but need it to stay competitive
• 3PLs and regional warehouses facing the same labor shortages as Amazon, without Amazon-scale capital
• Food production and processing, where margins are thin but labor costs are killing profitability

These markets are big enough to build billion-dollar companies but overlooked enough that start-ups won’t be competing with Tesla and Amazon directly. The need is also urgent enough that they’ll accept 85% solutions with human oversight rather than waiting for perfect automation.

What Excites Me Most

What excites me most is the founder this moment unlocks: someone with an earned view of an under-digitized market and the ability to connect cutting-edge AI to real-world problems. That overlap is rare.

For founders and investors in vertical AI, the leap from digital to physical automation is the frontier. But it’s a frontier where the founder brings domain expertise, data collection strategy and execution and leverages the 10s billions of investment that has already gone into robotics and model architecture. The winners will be determined by who can deploy imperfect systems and learn the fastest.

If we really are entering a ChatGPT moment for physical AI, then somewhere right now a founder is building the Cursor, Harvey, or Abridge of robotic automation….And I’d love to meet you.