a16z Podcast - Big Ideas 2026: Physical AI and the Industrial Stack

Episode Date: December 25, 2025

AI is moving into the physical economy.In this episode of Big Ideas 2026, we explore what changes when AI leaves the screen and becomes part of factories, construction sites, supply chains, and critical infrastructure. When the product is physical, reliability matters, real-world constraints appear quickly, and the advantage shifts from standalone software to end-to-end systems.You will hear from Erin Price-Wright on factory-first principles, Ryan McEntush on the electro-industrial stack, Zabie Elmgren on physical observability, and Will Bitsky on why data, not compute, determines who wins.Together, these ideas define what physical AI really means: not smarter chat, but deployable systems built for the real world, grounded in new operating models, industrial infrastructure, and defensible data collection. Resources:Follow Ryan McEntush on X: https://x.com/rmcentushFollow Erin Price-Wright on X: https://x.com/espricewrightFollow Zabie Elmgren on X: https://x.com/zabie_eFollow Will Bitsky on X: https://x.com/willbitskyRead more all of our 2026 Big IdeasPart 1: https://a16z.com/newsletter/big-ideas-2026-part-1Part 2: https://a16z.com/newsletter/big-ideas-2026-part-2/Part 3: https://a16z.com/newsletter/big-ideas-2026-part-3/ Stay Updated:If you enjoyed this episode, be sure to like, subscribe, and share with your friends!Find a16z on X: https://twitter.com/a16zFind a16z on LinkedIn: https://www.linkedin.com/company/a16zListen to the a16z Podcast on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYXListen to the a16z Podcast on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711Follow our host: https://x.com/eriktorenbergPlease note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Stay Updated:Find a16z on XFind a16z on LinkedInListen to the a16z Show on SpotifyListen to the a16z Show on Apple PodcastsFollow our host: https://twitter.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

This shift carries genuine risks, the same tools that can detect wildfires or prevent job site accidents could actually enable dystopian nightmares as well. The way that software will affect the physical world is through these sort of embodied, electrified components. We're seeing founders try to reduce these problems into kind of a decomposable set of modular parts such that you can apply the principles of an assembly line to society-scale problems. The problem of messy data is not a new one, and it's at the heart of this broader movement. The winners in this next wave will be those that really earn public trust,

building privacy-preserving, inter-offerable AI-native systems that make society both more legible without making it less free. What will define the next year of building? Our 2026 ideas reflect the themes our investing teams believe will shape how technology evolves next. This episode is built around four big ideas about AI leaving the screen and entering the physical economy. When AI moves it to factories, construction sites, supply chains, and critical infrastructure, the rules change. Reliability matters, real-world constraints show up fast, and the advantage shifts from teams that can build systems, not just software. You're going to hear three perspectives on what enables that shift. A factory-first mindset, an electro-industrial stack, physical observability, and the industrial data frontier.

To start, we need the operating model. Aaron Pricewright argues that we're entering a renaissance of the American factory, not just as a building, but as a set of principles. The idea is to apply assembly line logic to problems like energy, mining, construction, and manufacturing, using modularity, autonomy, and skilled labor to turn complex work into repeatable systems. Here's Aaron. My big idea for 2026 is the renaissance of the American factory. I think next year we'll see companies approach challenges from energy,

to mining, to construction, to manufacturing with a factory-first mindset. The modular deployment of AI and autonomy, alongside skilled labor, will make complex, bespoke processes operate like an assembly line. America's first great century was built on industrial string, but it's no secret that we've lost a lot of that muscle. Some of that has been from offshoring, from the financialization of everything in the 80s, leading to the large-scale offshoring of industrial manufacturing in the 90s. in 2000s. Some of it dates back to regulations. So rules and agencies and processes that were

put in place, usually for very good and specific reasons at the time, have built up over time into a crust that makes it, you know, very hard to do new things and to build new things in America. But here we are, and we have to figure out how to re-instill a culture of building in this country. I'm not just talking about a factory in a literal sense. Like you have a warehouse with an assembly line where you have some mix of humans and machines and at the end of the factory line, there's a widget that pops out. I'm really thinking about the principles of an assembly line full stop and how are those principles getting applied to industries that aren't traditionally industries you think of when you think of a factory. So housing, the construction

of data centers, the construction of mines, the construction of large-scale energy infrastructure and energy projects. We're seeing founders try to reduce these problems into kind of a decomposable set of modular parts such that you can apply the principles of an assembly line to society-scale problems. AI is a really amazing way to do that because you can understand and map out different complexities

in a regulation in a very formulaic and agentic way without having to completely redesign your entire processes from scratch every single time. How do we take technology and bring, bring the factory out into the world. We're building data centers at an unprecedented rate today, and we're creating standard IP and standard designs and putting them up in record time.

It's a great opportunity for us to test where autonomy, AI, robotics, other technologies that are coming to maturity right now can be deployed on these sort of large-scale physical assets because these building projects are moving so fast. As the data center market develops, these technologies spin out and become useful across a broad cross-section of industrial projects, whether that's the construction of new freeways and airports and landing strips

or the construction of mines and mining and refining facilities, which are so desperately needed. How do we take some of the learnings about how quickly we're able to move in data centers and apply them to building new factories, new fabs, new facilities to manufacture goods, whether it's for the defense sector, for the consumer sector or the commercial sector in the United States? How do we build things at scale? How do we create industrial capacity and use our ability to scale as an advantage? If you're a founder or a builder

and you are excited about reinventing what it means to build a factory in the United States, come and talk to us. Aaron frames the goal. Build faster, more repeatably, and act scale by applying factory principles to industrial problems. Now we make the move to what makes that possible inside the machines themselves.

Ryan McIntush lays out the rise, of the electro-industrial stack, the electrified embodied components that power EVs, drones, data centers, and modern manufacturing. He also makes the key point that the hard part is not just technology. It's building the ecosystem

that can produce, supply, and scale it. Here's Ryan. My name is Ryan McIntosh. I'm an investing partner on the American Dynamism team. My big idea for 2026 is that the electro-industrial stack will move the world.

The next industrial evolution won't just happen in factories, but inside the machines that power them. This is the rise of the electro-industrial sack. Combined tech that powers electric vehicles, drones, data centers, and all of modern manufacturing. I think there are common tropes people report on. People talk about China's so far ahead. We can't catch up.

And actually, you know, you go back a couple years ago and people were saying, you know, China's very far behind, and America's incredibly fast. So we've seen sort of like a whiplash, and now it's the opposite. I think the reality is that, you know, the technology that China has America can do. We're very good at engineering. we're very good at doing specific things, and in fact, even like the recent stuff around rare earths, for example, rare is separation and processing.

We know how to do this, we can do this, we can do it incredibly fast. The real challenge is building the ecosystem to do this industrially at scale and doing it at a low cost. Another example, you know, people typically talk about is companies like SpaceX or Enderil, these large businesses that need to move incredibly fast

and thus vertically integrate. In many ways, they're vertically integrating by necessity, not strategy. There just isn't an ecosystem of companies that can scale with them. That is not the case in China. There are tier one, two, three suppliers, components, raw materials that exist in those ecosystems, as well as the institutions and political bodies

that allow them to move incredibly fast. Those are the things that might take years or decades for us to catch up to China. We can do the technology, but everything else seems to grow with it or else we're just moving the bottleneck. So if you want to build the electro-industrial stack or the core components that feed into these technologies

in the United States, you need to blend Silicon Valley software, talent, and culture with industrial veterans. Even companies like SpaceX, they were pulling propulsion talent for people who worked on, you know, shuttle program and various old school contractors. When Shotwell came from aerospace corporation, there is a world where you need this actual expertise. You need to know what's been tried before. There are smart people out there in these other companies, but you need to be able to move a lot faster. There's a lot of advantages

of software today, so you need to be able to get the software talent that may not exist in these companies previously. You also want to co-locate engineering and manufacturing. Concepts like it's designed for manufacturing are something that, you know, when you're tightly integrated on the same footprint or in the same ecosystem, you can move a lot faster. And I think also you need to build prestige around the mission. For a lot of sort of traditional Silicon Valley talent, the smartest people can work on a number of problems. And there are a lot of problems that are worthy of working on. Some of them pay more than others. So you need to attach sort of a prestige or a purpose to what you're working on and use that to attract the top

talent. The way that software will affect the physical world is through these sort of embodied electrified components. And it's not just a humanoid robot or electric vehicle, but it's the batteries, it's power electronics, it's the compute, it's the motors, all these things we're going to need to either re-shore or vertically integrate within the companies who are building the end product. These are very technical. These require a lot of expertise. These are very difficult problems to solve. But the companies who solve it and the countries who have the talent-based, in order to support it, are the ones who are going to win in the 21st century. And as software and artificial intelligence get stronger,

and they start having more of a presence in automation, industrial, military, owning these supply chains has been to become even more important. And I think as we look forward 50, 100 years, owning the supply chains today are going to have a lot of effects of two controls, both the sort of economic and military powers in the future. Ryan shows that scaling physical AI is an ecosystem, system problem, not a single breakthrough. But even if you build the machines, you still need the ability to see and understand what's happening in the real world in real time. Sabby Elmgrim introduces

physical observability, bringing software-style visibility to physical environments using cameras, sensors, and AI. She explains why this is necessary for deploying autonomy safely and why public trust, privacy, and interoperability are not just add-ons, but requirements. Here's Zabby. Hi, I'm Zabby Elmgren. I'm an investing partner here at A16Z on the American Dynamism team. My big idea for 2026 is that the next wave of observability will be physical, not digital. I think over the last decade, software observability transformed how we monitor digital systems, making codebases and servers transparent through things like logs and metrics,

and the same revolution is going to come to the physical world as well. With more than a billion networked cameras and sensors deployed across the U.S., I think physical observability, which is really understanding what happens in these cities or across infrastructure in real time, is becoming both urgent and possible now. This new layer of perception, which enables the next frontier of really, I think robotics and autonomy also to really be successfully deployed, is becoming possible because you have this common fabric that really renders the physical world as observable as code has become in software. This shift carries genuine risks, the same tools that can detect wildfires or prevent job site accidents. could actually enable dystopian nightmares as well.

The winners in this next wave will be those that really earn public trust, building privacy-preserving, interoperable AI-native systems that make society both more legible without making it less free. And whoever builds that trusted fabric will define the next decade of observability in the physical world. When I talk about physical observability, I mean bringing the same kind of real-time visibility we've had in software to the physical world. In software, if something breaks, you usually see it on a dashboard before a user-notice. But in the physical world, you tend to find out when something sparks or is already stolen

or maybe a machine makes a sound that it should absolutely never make. This matters because there's far more attention falling on how we actually think about securing and automating critical infrastructure especially. Whether it's remote mines or data centers, many sites are becoming honestly too important to just operate blind. If you're a mine, you're running around the clock in places where humans have limited oversight and data centers have effectively become national security assets as well. And securing them is not just about locking the server room.

It's about understanding what's happening around the perimeter as well. And so physical observability essentially is about both successfully deploying autonomous systems into the real world and seeing the smoke both figuratively and literally before it turns to fire. I think what's changing now is that cameras are no longer working alone. For years, cameras have been recording a lot around us, but having to be. have understood honestly a lot less than what they've recorded. It's slightly like a well-meaning intern who takes great notes, but you can't really tell what really matters. Now we have thermal

sensors, RF sensors, acoustic sensors, all of these things that kind of capture a different slice of reality as well. When you fuse it together with modern AI, the system actually ends up interpreting what's going on around you and gives you really more context than just a picture or a video of what's happening. And I think companies like Andrel have actually proved this out in defense, but it's shocking how many other industries are pretty far behind. Construction sites are a great example of where physical observability is really lacking right now. A lot of them are remote. All of them are chaotic, and just getting stable power out there as hard, let alone securing a site or having a real understanding of what's happening on a day-to-day.

On top of that, there's a constant flow of likely very expensive materials being moved around, And it's basically becoming a high-value game of musical chairs where things really do go missing straight out from under you in the same way that a chair does in the game. And you think about deploying robotics into these settings as well. And it becomes incredibly difficult when you're flying blind in these different environments that are changing hourly.

There are loads of steel that land in new spots or temporary walls that go up and equipment that shifts. And so I think just thinking about how you make sure that a robot's mental model is really active. with how things are changing is how you're going to be able to bring autonomy to those settings. I think the tension between surveillance and privacy is real. More true, I would say, in the city use cases for physical observability than on a mining site. But the reality is that the same system that can prevent a wildfire or keep a data center secure can also be misused

for things that we absolutely don't want. And the question has really become not whether we can build these systems, but whether we build them in a way that really aligns with the democratic values. I think the winners in this category will really be the companies that treat privacy and trust as fundamental design requirements and not just bolt-on features. And I think in this space, earning trust is just like, it really is not just a nice to have, it's a license to operate. I think winning in physical observability really means becoming the perception layer that everyone relies on, whether it's robots, infrastructure operators, emergency responders, industrial workflows, all of it. It's about building the real-time map of the physical world,

that other systems can plug into and make these really complex environments simple to understand. You've seen this pattern before in the freight industry when some Sasaura showed up. Just having a single dot, honestly, moving on a map

seemed revolutionary. And now that tiny bit of visibility that unlocked huge operational gains is also true in these other, like, very critical industries. And imagine having that sort of step change. But instead of a dot moving on a map, you have a live,

multimodal understanding of an entire environment where assets are, what's changing, what's risky, what needs action, whoever builds that layer becomes the backbone of countless industries. The most obvious things that you'll see in these settings are deep, multimodal sensing functionalities, and again, that level of trust where whether it's the government for securing an asset that's important to national security or the public when it comes to public safety, that confidence in systems being accurate really matters. Sabby makes the case that physical AI depends on observability,

sensors and systems that make the real world legible in real time. But even with observability, there's one constraint that determines who wins, data. Not just clean benchmark data, messy multimodal industrial data that comes from real operations. Will Bitsky argues that the pendulum is swinging from one compute back toward data constraints, and that the most defensible advantage is not just cleaning data, but collecting it at the source. from installed bases, labor forces, and industrial-scale operations that startups can't easily replicate. Here's Will.

My name is Will Bitsky. I'm an investing partner on the American Dynamism team at A16Z. My big idea for 2026 is that critical industry is the next frontier for the crusade and AI data. In 2025, data centers compute and energy dominated the public discourse. In 26, I think the pendulum swings back from compute towards data constraints. I think critical industry is the next frontier. The problem of messy data is not a new one,

and it's at the heart of this broader movement. It's how do we take a bunch of messy data from different modalities? It's the big problem, and everybody is correct to be focused on it. But I think a lot of the underlying problems here are not necessarily new or unique to AI itself. A lazy first order answer on how we get past it is that scale and quantity tend to fix things over time. This is the whole bitter lesson problem

that I think our infra team is better equipped to apply here. So scale and quantity help, If you have industrial-scale businesses, they have industrial-scale data supply, right? So that's the short answer, maybe the lazy answer. You can borrow on lessons learned just building out the modern data stack. And some of those issues are not new to the last three years. It's a matter of leveraging models to an infrastructure and allowing them to work together. It's complying with consistent data ontologies and vocabularies and labeling data at the source where you can.

These are problems that aren't new. and as industrial incumbents get better at leveraging the modern data stack, at leveraging existing generalized models, we'll be able to get through the whole messy data problem. So the industries that have valuable data modes are anything within industrial supply chains, so across manufacturing disciplines, across defense, aerospace, commercial aviation, energy, mining, et cetera, there's so many opportunities to pull extremely large quantities of data. And so the frontier models will be able to use all sorts of different data types

from these industrial incumbents. If you listen to the World Labs team on one of the A16Z podcasts recently, these are fundamentally multimodal problems. These models are going to have to work together. So for the industrial incumbents, they can leverage language from existing software platforms. They can leverage spatial inputs

than anything sensor-based. So proprioceptive, you know, any feedback from tactile grippers. As they build this infrastructure, they're going to be able to just increase their access to data. The supply is going to increase. Some will be higher.

Some will be lower, But these foundation models are going to need a comparative advantage, a lower marginal cost way to acquire this data, and they'll increasingly come from these industrial incumbents. As I think about the hardest layer to build and what's most proprietary or what's most defensible, where the most value is, I think the answer today is very different from the one that I give in, say, 12 months. Today, everybody's scrambling to build the actual data infrastructure. So they're focusing on cleaning messy data, they're thinking about RL environments, they're thinking about data pipelines. But longer term, I truly think collection, thinking about where the data inputs are at the top of the funnel, that's where the most value accrues. I analogize this to walled gardens in the consumer world, any of these industrial companies that have install bases, that have existing labor forces, that have industrial scale operations, they have a lower marginal cost to collection because they can pull from their operations that already exist. It's impossible to disintermediate these companies from their existing operations.

Startups are trying to hack together their own data collection operations, but they're paying a steep marginal cost. They're building robotic arm farms. They're selling consumer products that are teleoperated. If they have an alternative path to collecting data by leveraging these industrial companies, then they'll be able to lower their marginal costs. I don't see it as sustainable or scalable to see these startups paying for each marginal unit of data being collected. Here's the through line across these four big ideas. Aaron argues that scaling the physical economy requires a factory-first mindset, turning bespoke industrial work into repeatable systems.

Ryan explains the electro-industrial stack and the ecosystem challenge behind building and supplying the components that power modern machines. Zabby shows that physical AI only works with observability, a trusted sensing layer that makes real environments legible in real-time. Will closes with a constraint. The bottleneck is swinging from compute back to data, and the durable advantage will belong to the companies that can collect mess see multimodal industrial data from real operations at scale.

Put together, this is what physical AI actually means. Not smarter chat, but systems you can deploy in the real world, built on new operating models, new industrial infrastructure, and defensible data collection. Thanks for listening to this episode of the A16Z podcast. If you like this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube,

Apple Podcasts and Spotify, follow us on X, A16Z, and subscribe to our Substack at A16Z.com. Thanks again for listening, and I'll see you in the next episode. As a reminder, the content here is for informational purposes only. Should not be taken as legal business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the company's discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures.