A16z · EXTRACTED

Why Physical AI Is the Next Frontier | Applied Intuition with a16z

Physical AI, sovereign autonomy, and why the machines that move the world will matter more than the machines that move data. The company putting intelligence on a billion physical things explains how it actually works.

▶24K views on YouTube

"In this intelligence revolution, the companies that impact the physical world might actually be bigger than the companies that impact the digital world." — Kazer, Applied Intuition
Kazer & Peter, Applied Intuition, on the episode

This is a conversation between two co-founders of Applied Intuition and their early investors on The Z Podcast. Applied Intuition builds the software and AI that makes physical machines autonomous: trucks, tanks, drones, mining equipment, port systems, agricultural vehicles. The pop framing of the company is self-driving cars, but automotive is already below 30% of their business and shrinking as a share. The operating system underneath is closer to a chipmaker than a car company: deep design wins, long relationships, proprietary data, and a platform that deploys across every vertical where machines move. This conversation covers where physical AI actually is today, why the cost curve is about to break, and what their new agentic development platform, Dana, means for who gets to build autonomous systems.

Tactic 01

The Cost Cliff: When Sub-$1,000 Driver Assistance Makes Safety Standard Equipment

The reason most cars are not self-driving today is not the technology. It is cost. The Applied Intuition founders point to Chinese EV manufacturers running L2++ driver-assistance systems at sub-$1,000 per vehicle as the leading indicator of where the curve is heading. Once the package of chips, sensors, and software hits roughly $500, automotive OEMs will not sell it as an option. They will absorb the cost and include it in every vehicle, the same way navigation systems went from a $3,500 add-on to a default feature that nobody talks about anymore. The founders put a specific production date on this: 2028 to 2029 for start-of-production on vehicles with standard driver assistance, with early-2030s as the window when it becomes cheap enough to be default across the market. Every major OEM is already running an FSD competitor program. The question is not whether it ships; it is which price envelope breaks first. At $500 per unit across millions of vehicles, the OEM subsidizes it. Below that, it becomes invisible infrastructure, the same way Bluetooth and GPS did.

The play

If you are evaluating any business that depends on human drivers in personally owned vehicles, build two scenarios: one where L2++ becomes standard equipment by 2029 and one where it is universal by 2033. Run your unit economics and competitive assumptions through both. The technology risk is resolved. The remaining variable is price timing, and the founders' estimate is 28 to 30 for broad availability, early 30s for routine ubiquity. Model accordingly.

Tactic 02

The Data Flywheel: Why Hundreds of Petabytes of Proprietary Physical Data Is the Real Moat

In digital AI, foundation models train on the internet. In physical AI, the internet is not enough. A model that needs to operate a mining haul truck in an underground tunnel in South Korea, or navigate a port in the Middle East, or drive a commercial truck through mountain fog, needs data that does not exist in any public corpus. Applied Intuition has spent nearly a decade collecting it. The founders put a specific number on this: the company has built one of the largest private data collection fleets on the planet and has accumulated hundreds of petabytes of proprietary physical-world data. That data, combined with their own synthetic data tooling called Neural Sim, creates a closed loop. Real-world data trains the models. Synthetic data fills the gaps and edge cases at scale. The models improve. The fleet collects more real-world data. The founders note that synthetic data has been a core investment for over five years, and that the end-to-end reinforcement learning approach now dominant in the field depends on exactly this kind of closed-loop data infrastructure to keep improving. The key structural point: because physical AI data cannot be scraped from the internet and because collecting it requires physical presence in sovereign countries with government approval, the companies that have already built the data infrastructure have a lead that is not easily replicated by writing a check.

The play

If you are building or evaluating a physical AI system, audit your data collection operation before your model architecture. The founders' framework is explicit: proprietary data sets plus scaling laws works the same as in digital AI, but the diffusion is radically different because you cannot access physical AI models through a phone. Map every geography and operating environment your system needs to work in. Identify which require in-country government approval to collect data. Start those conversations now. The companies that skip this step and rely on public data or partners' data will hit a ceiling that is invisible until the model ships into production.

Tactic 03

Dana: The Agentic Platform That Lets a High Schooler Build an Autonomous System

Applied Intuition's core launch is a platform called Dana, named after the street their headquarters is on. The founders describe it as an agentic development environment for physical AI that packages everything the company has built in nearly a decade into a single interface. The benchmark they set is specific: a high school student who can build iPhone apps should be able to use Dana to build an autonomous system. The workflow they walk through is the full pipeline from scratch: define requirements for the environment the robot needs to operate in, generate scenarios from a satellite image of that space, pull training data from publicly available sources, deploy the model onto the actual hardware, observe where it fails, close the loop, and iterate. Workflows that previously took days or weeks now run in minutes. The company already uses Dana internally to develop their own autonomy systems across all their verticals. The founders draw a direct parallel to what coding assistants did to software development: the IDE changed, the surface area of who could build things expanded dramatically, and the applications that emerged were not predictable in advance. Their expectation is the same thing happens in physical AI. The barrier between having an idea for an autonomous system and having a working prototype collapses, and the creative surface area expands to people who are not autonomy engineers today.

The play

Before adopting any new tool, take one task you currently find slow or painful, run it through the tool start to finish, and time it against your normal process. Only commit if the real number, on your own work, actually beats what you do today.

Tactic 04

Solve The Hardware Before You Trust The Software

The public conversation about physical AI focuses almost entirely on models, data, and algorithms. The actual bottleneck in production is messier and less glamorous. Peter Guest said it plainly: "Whenever you're dealing with physical systems, you inevitably hit a lot of gnarly hardware problems. It could be anything from overheating to a sensor being slightly miscalibrated or a funny issue we saw yesterday was basically a fogging sensor, fog impacting a sensor." Those examples are not edge cases. They are the everyday surface area where autonomous systems fail in the real world, and no amount of model quality covers for them. A sensor that reads slightly wrong feeds bad data upstream. A compute unit that overheats during a long haul run degrades inference at exactly the moment reliability matters most. The model can be state of the art and still output the wrong decision because the hardware feeding it is off by a small margin. This is why Applied Intuition engineers hardware competency as a first-class discipline alongside software and AI. They have dedicated engineering teams that can deploy models onto more than 50 different hardware platforms, a feat that sounds routine until you realize there is no iOS or Android equivalent in the physical world to abstract the complexity away. Every platform is its own integration problem. The teams that can do that work reliably are a small group globally, and the ones that cannot hit this hardware layer are the ones that get stuck before they ever get to questions of scale.

The play

Before evaluating any physical AI system, whether you are building, buying, or partnering, run a hardware fault audit as a separate workstream from your software validation. List every sensor type in the stack, every compute unit, and every environmental condition the machine will operate in. For each, identify the specific failure mode: overheating threshold, calibration drift rate, occlusion conditions such as fog, dust, or direct sun. Assign a team member or vendor contact responsible for each failure mode before deployment begins. Systems that skip this step reach production and then get pulled back. Systems that treat hardware reliability as a prerequisite, not an afterthought, stay in production.

Tactic 05

Sell With A Calculator, Not A Story

Kazer draws a sharp line between the consumer car market and the trucking market, and the line is emotional versus rational buying. "You buy a car with your heartstrings," he says. "You buy a truck with a calculator." That distinction shapes everything: how you pitch, how you price, how you prove value. A trucking operator does not care that the technology is impressive. They care what it costs per mile and how that compares to what they are paying today. This is why Applied Intuition chose Japan as its primary trucking market rather than the United States. Japan has a massive labor shortage today and what Kazer calls an imploding demographic situation. The demand signal from operators is not latent or speculative. "The people who run those things want it today," he says about quarry and mining operators in the same breath. "We literally can't make the stuff fast enough." That urgency changes the sales motion entirely. You are not convincing a skeptic. You are fulfilling a demand that already exists and has no other resolution in sight. The lesson extends beyond trucking. When Kazer is asked why autonomous systems are not more broadly deployed yet, he does not cite technical readiness. He cites cost envelope. The OEMs are not resistant because they doubt the technology or think consumers do not want it. They are resistant because the price has to fit inside margins that are already razor thin and the system has to work across deployments in more than a hundred countries from day one. The calculator wins every time. If your ROI case is solid, the sale closes. If it is not, no amount of product enthusiasm moves it forward.

The play

Before pitching any physical AI product to an industrial buyer, build the per-unit or per-mile economics first. What does the current operation cost? What does your system save, and over what period? If you cannot express the return in a single number the buyer can verify against their own books, you are not ready to sell. Then look for the market where the labor shortage or safety crisis is most acute, because that is where buyers have already done the math themselves and are waiting for someone to show up. Japan is Applied Intuition's trucking anchor precisely because demographic collapse made the ROI case obvious without any persuasion required.

Tactic 06

Retrofit First, Redesign Second

The instinct when better technology arrives is to assume everyone will buy the new thing. In mining and construction, that assumption breaks immediately. Karim Lakhani makes the point plainly: a haul system in a port or a Komatsu dirt mover in a mine is built to last 20 to 25 years. The buyer may be three years into that cycle, or ten, or eighteen. They are not replacing the machine no matter how much better the alternative is. The ROI math on capital equipment at that scale does not allow it. This forces a split strategy. The first branch is retrofit: take the machines that exist, that are already in production, that are not going anywhere, and make them intelligent. The second branch is redesign for a world without a human in the cab. Lakhani points out that in underground mining specifically, the constraint is the human. The worker needs to breathe, the environment is dangerous, and the entire tunnel geometry is built around keeping a person alive down there. Remove the person and you can build a completely different machine, smaller, shaped differently, running in conditions that would be impossible with a crew. Both branches matter. Skipping the retrofit branch to focus only on clean-sheet autonomous designs means ignoring the overwhelming majority of machines actually operating in the world right now.

The play

If you are building or selling autonomous systems for industrial environments, map your target customer's fleet by age before pitching replacement. For any machine under 15 years old, lead with retrofit and intelligence layering, not new hardware. Reserve the redesign conversation for greenfield deployments or end-of-life planning cycles. The customer who just committed capital to a 20-year asset will not buy a replacement; they will buy an upgrade to what they already own.

Tactic 07

The Nav System Playbook: How Driver Assistance Goes From $1,000 to Free

The reason most cars don't have full self-driving today isn't technical hesitation. It's a price envelope problem. Karim lays it out plainly: the OEMs know their buyers, they know their margins, and they're waiting for the cost of the sensor-compute-software package to hit a number they can absorb. That number is coming. "We anticipate that there's a very aggressive, once you get to like $500, the automotive OEMs will actually subsidize it for free. They'll just give it to you." He points to navigation systems as the model. A factory nav unit cost $3,500 to $4,000 at the dealer window for years. Then it became free. Not because the technology got worse, or because automakers suddenly became generous, but because the fixed cost of developing and certifying the system got amortized across a full production line. The incremental cost of adding it to one more car dropped below the noise floor. Driver assistance is on the same curve. Every OEM without exception, Karim says, is working on a full self-driving competitor. The fixed cost of that development is already being spent. The question is just when the per-unit cost of the hardware crosses the threshold where it makes more sense to put it in every car than to gate it as an option. The assembly line logic makes this almost inevitable. Homologation, the testing and certification regime that clears a feature for production across an entire vehicle family, is expensive to run once and cheap to extend. Once a system clears that process for one trim level, adding it to the base model is mostly a procurement and margin decision. The automaker's incentive flips: a standard-equipment feature protects against liability, differentiates on safety ratings, and removes a line-item customers can negotiate off the sticker.

The play

If you are building or investing in a driver assistance product, price your roadmap against the $500 hardware threshold, not against the current $1,000 ceiling. Map the fixed cost of your certification process across the realistic production volumes of your target OEM partner. The unit that clears homologation at volume is the unit that goes standard. Time your commercial agreements accordingly: the window between OEM adoption as an option and OEM adoption as standard equipment will compress faster than most forecasts assume, because the fixed cost is already sunk.

Tactic 08

The Ninth Grader Test: Build Autonomy From a Satellite Image and Public Data

Kazer uses a delivery robot as the proof of concept for Dana, and the framing is deliberately humbling. The question he asks is not what a senior roboticist at a well-funded lab can build. The question is why a ninth grader cannot build a delivery robot for their high school campus this weekend. That question, held seriously, is what shaped the entire platform. He walks through the actual workflow step by step. The student defines requirements first: this robot needs to navigate around four buildings on a specific campus. From that, scenarios get generated using a satellite image of the school as the environment. The student then pulls training data from publicly available sources, YouTube videos and similar material, enough to get a rudimentary model running. The robot gets deployed onto the physical machine. It runs into a wall. That failure closes the loop, sends a signal back into the system, and the next training iteration addresses it. That cycle, from requirements to scenarios to data to deployment to failure signal to improvement, is the full autonomy development pipeline, compressed into something a teenager can operate. What makes this more than a demo is that Applied Intuition runs its own production autonomy systems through the same platform. As Kazer puts it, "everything that we've built and developed over the past nearly a decade, every tool, every technique, that's available in Dana." The ninth grader and the professional engineering team are working in the same environment. The barrier is not capability. It is access to the toolchain.

The play

Pick one physical task in your environment and sketch the five-layer structure Kazer describes: requirements (what exactly do you want the system to do and where), scenarios (what situations does it need to handle, mapped to the actual space), data sources (what publicly available footage or sensor data covers those scenarios), a baseline model trained on that data, and a deployment loop where real-world failures feed back into the next training pass. You do not need hardware to do the first three layers. Getting specific at the requirements stage, naming the buildings, the path, the edge cases, is where most autonomy projects fail before they start.

Tactic 09

Why Automotive Safety Culture Runs On Obfuscated Language

Three of the top five consumer lawsuits in American history are automotive. That single statistic, which Kazer says was drilled into engineers at GM, explains more about how legacy manufacturers operate than almost anything else. The liability exposure is not theoretical. It shapes language, process, and product decisions at every level of the organization. The most revealing detail Kazer shares is not a policy but a color code. Inside one of these companies, engineers were not allowed to use red, yellow, and green to mark safety system status. The substitutes were colors like purple and magenta. The reason is exactly what it sounds like: if a safety system marked red ships to production, that word becomes evidence in a courtroom. Magenta does not carry the same weight. "You let a safety system that was marked red go to production," Kazer says, recounting the logic. "I was like, no, it was marked magenta." The frustration in that line is not just personal. It describes a company that has reorganized its internal vocabulary around litigation risk rather than engineering clarity. This is not unique to one manufacturer. Kazer's read of the industry, informed by his time at GM and his reading of John DeLorean's account of the company, is that most major manufacturers still operate this way. The institutional weight of quality and safety as twin priorities, reset hard by Japanese competition on the quality side, creates a culture where process compliance and legal exposure management sit above engineering velocity. For any technology company trying to work with these organizations, understanding that culture is not optional. The distribution defines the business, as Kazer puts it. You have to dance the way the partner needs to dance.

The play

Before your next meeting with a legacy manufacturer or safety-critical customer, map their actual internal vocabulary around risk. Ask a direct question: what words do your engineers use internally to flag a system that is not ready for production? If the answer is anything other than a plain-language status, you have found the gap between their public safety posture and their operating reality. Use that gap to position your own tools and processes as the ones that close it, not the ones that add another layer of obfuscation.

Tactic 10

The Chipmaker Model: Design Wins, Deep Trust, Hard To Displace

Most people assume a company selling AI to car manufacturers is playing the same game as Tesla or Waymo, just from a different angle. Applied Intuition's Qasar Younis draws a sharper distinction. The vertical integrators own the whole stack, build the car, operate the fleet, collect the revenue. Applied Intuition is playing a different game entirely. "The way we can always think about it," Younis says, "is we're kind of like a chipmaker." Not a metaphor for scale, but a metaphor for structure: design wins, long-term embedded relationships, and a position that is very hard to unwind once it is established. The Nvidia parallel is worth unpacking. Nvidia does not build the data centers, does not train the models, does not run the applications. It makes the component that everything else runs on, and it knows its customers so well that switching becomes genuinely painful. Younis makes the same point about Jensen Huang: "The things that Jensen knows is he knows his customers. That's why Nvidia does well beyond the fact obviously they make a very complex technology." Applied Intuition sells the intelligence and the tooling. The manufacturer puts their badge on the truck or the combine or the port vehicle. When you look under the hood of a dirt mover or a diesel truck, it might have a Cummins engine. Nobody says that makes Caterpillar a weak company. It is just how complex industrial systems are assembled. The practical consequence of this structure is that once Applied Intuition is integrated into a platform, the customer has deep technical dependency, accumulated institutional knowledge on both sides, and no clean path to replacement. Younis describes running fully autonomous trucks in Japan carrying commercial loads under the Isuzu brand. The customer gets the brand equity. Applied Intuition gets the design win, the relationship, and the data. "That company's been around for almost a hundred years. They know the government, they have test tracks, they know safety, they know their own trucks very well. So when we go and provide them with the intelligence and the integration into their physical machinery, that's a fantastic one-two punch."

The play

Map your customer relationships against the chipmaker model. For each major customer or partnership, ask: is this a transaction or a design win? A design win means your technology is embedded in their product, your team understands their constraints better than any competitor could after a cold start, and replacement requires them to rebuild something, not just switch a vendor. If most of your relationships are transactional, pick one customer where deeper integration is possible and invest disproportionately in knowing their world, their hardware, their regulatory environment, their internal incentives. The goal is the position Younis describes: "Once we're in, it's really hard to take us out." That is not lock-in through contracts. It is lock-in through genuine depth.

Tactic 11

The 20-Cent Mile: How Autonomous Transport Collapses the Cost of Everything

Kazer's argument for physical AI isn't abstract. It runs through a single number. Long-haul trucking today costs several dollars per mile. Autonomous trucking, at scale, could drop that to 20 cents. That gap is not a rounding error. It is the difference between a supply chain that requires a human behind every wheel and one that moves goods across continents at a cost closer to electricity than labor. The downstream implications reach further than the trucks themselves. When transport becomes cheap, the price of everything that moves drops with it. Food grown in one region and consumed in another. Raw materials hauled from a mine to a factory. Components shipped between manufacturing hubs. Kazer puts it plainly: "Imagine for goods being transported, let's say instead of a few dollars a mile, it's 20 cents a mile, and suddenly it's... the unlock is very very very big." He is describing the same productivity shock the internet delivered to information, now applied to atoms. The unlock does not require every machine to be rebuilt from scratch. A trucking fleet running the same diesel rigs, same roads, same loads, but with autonomous systems instead of drivers, captures most of the cost reduction before a single new vehicle rolls off the line. That is the near-term version of the bet. The longer-term version is what happens when designers no longer have to accommodate a human in the cab, and new vehicle architectures become possible. Both versions run on the same economics.

The play

Map one supply chain you touch, buy from, or invest in, and find the transport cost as a percentage of the final price. Then model what happens to margins and end-user price if that transport line drops by 80 to 90 percent. You are not predicting when autonomous trucking arrives; you are identifying which businesses in your orbit are most exposed to that cost cliff, and which ones benefit most when it comes. The businesses that are currently transport-cost-constrained, in agriculture, mining, construction materials, and industrial goods, are the ones where the unlock is largest and earliest.

Tactic 12

Automotive As Minority Business

Most people think of Applied Intuition as a self-driving car company. Karim Lakhani is quick to correct that. 'Even today, automotive is like 30% of our business,' he says. 'So 70% already is non-automotive.' The car framing stuck because cars are what people see and drive, but the actual business has already moved well past it. The industries that are scaling faster are the ones most people don't think about: mining, construction, defense, agriculture, ports, logistics. These sectors share a critical trait with automotive, which is that they run on machines that move, but they often have more immediate pressure to automate than any car company does. The average American farmer is 58 years old. The share of farmers under 35 is below 10 percent. Demand for food and rare earth materials keeps growing while the human labor pool that extracts and moves those things keeps shrinking. Trucking faces the same demographic cliff. 'These demands are only growing but the humans who are the bottleneck are decreasing,' Lakhani says. The long-run picture is that manufacturers, whether automotive or otherwise, become one category of customer among many. Mining operators, ports, departments of defense, and agricultural operators each become direct customers in their own right. Lakhani describes it as a distribution shift: 'Initially the manufacturers were the distribution to that intelligence to consumers, but then you start working in defense and you start working in construction and mining and agriculture and suddenly the manufacturers are important but maybe the mining operator is actually really important.' Automotive is a large and durable market, but it is a shrinking slice of a much larger pie.

The play

Audit how you frame physical AI investments or partnerships. If your mental model is still organized around automotive, redraw it around machine categories: mining, agriculture, ports, defense, logistics. For each category, identify who controls the machines, whether that is the manufacturer, the operator, or the government, because that is the actual buyer. The markets where labor is visibly shrinking and safety records are worst are the ones with the least resistance to paying for autonomy today.

Your action plan

the plays, one checklist.

01
The Cost Cliff: When Sub-$1,000 Driver Assistance Makes Safety Standard Equipment
Build two planning scenarios for any driver-dependent business: L2++ standard by 2029, universal by 2033. Price timing is the remaining variable, not technology. Model both.
02
The Data Flywheel: Why Hundreds of Petabytes of Proprietary Physical Data Is the Real Moat
Before touching model architecture, audit your data collection operation. Map every geography your system needs. Identify which require in-country government approval and start those conversations now. The data ceiling is invisible until you hit it in production.
03
Dana: The Agentic Platform That Lets a High Schooler Build an Autonomous System
Before adopting any new tool, take one task you currently find slow or painful, run it through the tool start to finish, and time it against your normal process. Only commit if the real number, on your own work, actually beats what you do today.
04
Solve The Hardware Before You Trust The Software
Run a hardware fault audit before software validation. List every sensor and compute unit, identify the failure mode for each, and assign ownership. Overheating, calibration drift, and sensor occlusion are the most common reasons a production-ready model still fails in the field.
05
Sell With A Calculator, Not A Story
Build the per-mile or per-unit ROI case before any pitch to an industrial buyer. Then find the market where the labor shortage or safety crisis already exists, because those buyers have done the math themselves and are waiting.
06
Retrofit First, Redesign Second
Map your target customer's fleet by age before pitching. Lead with retrofit for machines under 15 years old. Save the clean-sheet redesign pitch for greenfield deployments or end-of-life cycles. Customers who just committed capital to a 20-year asset will buy an upgrade, not a replacement.
07
The Nav System Playbook: How Driver Assistance Goes From $1,000 to Free
Price your roadmap against the $500 hardware threshold. Map your certification fixed cost across realistic OEM production volumes. The window between optional and standard-equipment will compress fast once homologation is cleared.
08
The Ninth Grader Test: Build Autonomy From a Satellite Image and Public Data
Map the five layers for one physical task you want to automate: requirements, scenarios from your actual environment, available public training data, a baseline model, and a feedback loop from real-world failures. Specificity at the requirements stage is where most projects break down before anything is built.
09
Why Automotive Safety Culture Runs On Obfuscated Language
Ask a legacy manufacturing partner what internal language they use to flag safety issues. If the answer is not plain-language status, you have found the gap between their public safety posture and their operating reality. Position your tools as the ones that close it.
10
The Chipmaker Model: Design Wins, Deep Trust, Hard To Displace
Audit your key customer relationships and identify which ones are design wins versus transactions. For one customer where deeper integration is possible, invest in learning their world, their hardware, and their constraints at a level no competitor can match after a cold start. The goal is the position where replacement requires them to rebuild something, not just switch vendors.
11
The 20-Cent Mile: How Autonomous Transport Collapses the Cost of Everything
Find one business you own, buy from, or invest in where transport is a meaningful cost. Model the margin and price impact of an 80 percent reduction in that line. Identify whether you are on the exposed side or the beneficiary side of autonomous logistics.
12
Automotive As Minority Business
Stop organizing your view of physical AI around cars. Map the actual buyer in each machine category: manufacturer, operator, or government. The markets with shrinking labor pools and high fatality rates are already willing to pay.