T.J. Rylander is a general partner at N47, a global firm with more than $2 billion under management. He backs technical founders working on real machinery, infrastructure, and security problems. Before N47 he spent over a decade at In-Q-Tel, the intelligence community's venture arm, where he helped national security agencies work with companies like FireEye, Cloudera, and Pure Storage from early product to IPO.

At N47 his focus is AI in the physical world. Portfolio names include Skydio, Luminary Cloud, and Harmonic Security. He has a phrase for the category: near tech, not deep tech. The science mostly exists already. The hard part is turning customer pain into product.

This one lands for anyone building in dirty, dusty, and dangerous industries. TJ is a skeptic on the consumer humanoid craze and blunt about pure play defense tech. His read is that both get sold on a story the market wants to hear instead of the problem in front of them.

The conversation

Jay: When you say physical AI, what do you want people to imagine?

TJ: Ultimately this is AI applied to everything that isn't just a pure software experience. Imagine the next car you buy is more beautiful, more innovative, and more performant than anything you could have pictured, on the back of applying AI to physical product design. That's what our company Luminary Cloud is doing. It collapses the design engineering cycle from weeks and months down to near instant feedback for designers. That speeds time to market and yields better products.

Jay: Do you differentiate physical AI from AI for the real world, or is it all a spectrum?

TJ: There's physics AI, which is what Luminary Cloud does, and then there's AI applied in the physical domain. Those can be separate. But anytime it transcends a pure software experience, I'd call it AI in the real world.

Jay: If it's a hardware or robotics product, that's clearly physical AI to me. Then there's the case where the data inputs come from the physical world, concrete or electrical data, but it's manipulated in software and the output is a dashboard some boots on the ground are using. To me it's more a spectrum than a hard vertical.

TJ: That's a great framework. What I'd add is that the IP is almost always still software driven. The way it shows up, and how we interact with it, can take different forms.

Jay: Why this, the interaction between AI and the physical world, as opposed to regular software where plenty of people are making money?

TJ: We still focus on core software investing too. We just think the time is right, and the next horizon is applying these advanced techniques to the physical world. The stat that's well worn is that less than 20% of global GDP comes from the digital world. It's massive, but there's a lot of room left in the rest of the world. You've got great examples of technology brought to bear on dirty jobs.

Jay: I look at it through labor shortages. But the real unlock for me was seeing the divide between the tools available to digital workers and the tools available to the steel-toed-boots crowd get wider and wider. About three years ago I started fishing in this category.

TJ: I could not agree more. When a talented founder takes a clean sheet approach to a known problem, they have an incredible leg up over the last generation of products. And that clean sheet today involves using AI in the most advanced ways, which yields a better product and a big time-to-market advantage.

Jay: What are examples in your portfolio where AI interacting with the physical world solved a problem pure software wouldn't have?

TJ: A great one is Tractian. It's hardware and software together, easy to deploy, and it drives advanced condition monitoring for industrial systems and rotating machinery. It all comes from the team's stark focus on customer needs.

Jay: What do you need to see from a founder in that first pitch, the aha moment where you lean in?

TJ: So much of it comes down to, let's talk about what's special here, and let's start with the product. The product is a manifestation of what's unique about the founders and of how they've internalized the customer's needs. Tell us what's happening at the heart of the product. Let's draw a diagram, map out the elements of the system, and then drill into where it's going.

Jay: With how fast the models are improving, do companies need to be multi-product within the first year?

TJ: The long-term winners have almost always been all about product. These days you can go from zero to first release in nine months and have ten sub-products, just because of the rapidity of development.

Jay: We've seen it. Multitude Insights is AI for law enforcement. How do you make it easier for frontline patrol officers to input crime data as they see it? It's almost an Instagram thing. Take a picture, grab license plate info, dictate it, upload it. The second product was Smart Link. There's a string of home burglaries and it looks like the same person hitting the same kind of house. Cops didn't know because precincts don't share information. He could be doing the same crime in Burlingame, Redwood City, and Palo Alto and those three departments wouldn't connect it. Both products shipped in the first year. The expectation now at Series A is that you're multi-product.

TJ: I like to see it. No hard and fast rule, but we see it more often, and the expectations are moving up.

Jay: Where do founders most often get stuck building product in physical industries?

TJ: We see incredible intellectual horsepower from these teams. The temptation for very smart people is to use first-principles thinking to figure everything out. That breaks down when it's time to go from product definition to go-to-market. You need to lean on standing expertise there, because it's been done so many times before.

Jay: Because of what we do with storytelling, founders come to us and say there's a bigger story here I'm not telling. Sometimes it's on me to say, you think you're building tools for homeowners, but what you're unlocking is way bigger. There's a narrative for investors and a different one your employees need to hear. It's not the same story.

TJ: That's the biggest trap technical founders fall into. They tell the same story to three different audiences. You look and go, how did these guys raise $25 million but the traction isn't there? Sometimes they have to bring in the talent to say, no, it's a different story for customers, and here's how we tell it.

Jay: Have you seen the hype videos on Twitter and LinkedIn? People are dropping a hundred grand to launch a product with a hype video. And I ask, is your job as a marketer to deliver a hype video, or to drive a business outcome? Half the time a founder comes in wanting PR and I ask, do you need new customers? No, we've got a two-year pipeline. Do you need employees? No. Investors? I just closed $10 million. So what are you doing PR for?

TJ: It can be hugely effective if you're clear-headed about who you're talking to and the message. Doing it just to be part of the scene is probably not right.

Jay: Is there a product throughline the best founders share?

TJ: It's customer obsession, and knowing a great product comes from an intimate understanding of the pain point, and that the job never stops. One of our teams defined its vision as, we want to delight our customers and deliver a software product they enjoy using. That was starkly different from legacy data management platforms that had great revenue ramps but customers who admitted big spend and real unhappiness.

Jay: On customer calls, what tips you off that a team has cracked something?

TJ: We ask, what was your life like before, and what is it now that you get to use this product. So much of it is intonation. Reading a transcript doesn't get it done. The real conversation is about emotion and how wed a team is to the experience of using the product.

Jay: How has the barbelling of specialist and generalist funds affected how you invest in physical AI?

TJ: When you compare specialist firms to the big supermarket funds that need one of everything on the shelf, the degree of direct fit between expertise and a company's strategy accrues most to the specialists. That's critical going from zero to one. Later, when you're scaling, there's great reason to bring in larger investors.

Jay: You have a strong view on education as a training ground for founders. Why does it still matter when AI is this great leveling tool?

TJ: Look at all the PhDs driving so much of the innovation in this AI wave. I don't think we've had a startup cycle in a long time so dependent on advanced degrees. Where I disagree is programmatically encouraging people to stop a productive college experience to start a company. What I counsel people coming out of college is to think in terms of forty years. That's about how long you'll be working. In that context, another couple of years of school means something different than if you're only looking at the next five.

Jay: So someone hands your teenager a YC SAFE to go build something. How do you counsel them?

TJ: Side hustle. You've got twenty-four hours in a day. Do both. You can run a side hustle until you see profound product-market fit. You don't have to drop out to get to MVP.

Jay: You once told me the most common failure mode in tech is infinite extrapolation. Explain that.

TJ: I call it the Segway rule. The day before the Segway launched, there was no Segway. The day it launched, the prognosis was the whole world would change, we'd reconfigure cities around it. That extrapolation happens all the time. Same with the dot-com boom. Timing is usually the issue, not whether the prognosis is right. Infinite extrapolation assumes accelerating growth instead of any asymptote. Miss the asymptote and you keep building in a way that misses the adoption cycle.

Jay: Is there a vertical falling into that trap right now?

TJ: The consumer humanoid craze is a great example. If we start believing a new technology is capable of everything, then we haven't done the hard work of saying what exactly it's good for and what it will be dominant at first. Think back to IBM's Watson. It won Jeopardy, dominated, great. But right after, the pitch became, if it can win Jeopardy, what else can it do? It got sold as an expert Q&A system for medical situations. It didn't work. Build for a problem, create a product that's a direct hit for that problem, and don't start thinking that because it can do X, then Y is imminent.

Jay: How does that apply to an auto OEM that's going to make millions of humanoid robots?

TJ: There's a labor shortage reason there, so it's a bit different. When BMW says we're putting this in our factory, they can't hire enough electricians and welders. That's a very clear need. Now, you could argue point solutions already solve that. There are FANUC and KUKA arms you can put in for soldering. But why do I need a human-sized robot in that spot? I don't know why they've all of a sudden decided that humanoid is the answer, as opposed to a bunch of point solutions. I think there will be a stopping place between stationary arms and humanoids that includes wheeled systems.

Jay: Why do you think we moved away from drones for that? Before Amazon acquired Kiva Systems, the natural warehouse extension was to go from ground-based movement to the air.

TJ: So much of a factory or warehouse is about moving goods, so drones are a great source of intelligence, but you need something that can carry and manipulate. On drones though, one of my favorite proof points is Skydio as a first responder. In a jurisdiction that has Skydio, you call for help and the police can have a drone on station in under two minutes, much faster than a patrol vehicle. In one county, two kids were out back firing a handgun. Officers showed up, but from the overwatch they knew these were two eight-year-old kids, so let's not go in hot, let's diffuse it and keep people safe.

Jay: We don't have a term yet for physical AI. Call it near tech. This isn't drilling holes in the Sahara or fusion. This is real, and it's happening in the next five years.

TJ: Deep tech requires additional science. In so many cases we're operating as a pure software investor, or software applied to the real world, but almost always the software is the key lever and the key IP.

Jay: Where do you see unhealthy growth in AI companies today?

TJ: Negative gross margins are dangerous if there isn't a trajectory toward healthy ones. The first principle is to be an enduring company. Nobody should forget that in the dot-com boom, public companies were valued on eyeballs. That didn't work out, and it won't work here if there's no ability to transition into positive gross margin. The companies with a credible trajectory aren't dependent on a substrate of cost. If I'm spending all my money on API calls hitting LLMs, then I'm dependent on their cost profiles over time in a way that's shaky.

Jay: When I raised my first fund I talked to 440 LPs, and less than 20% came in. How has rejection shaped how you invest?

TJ: There's a temptation to portray our trajectories as well planned and linear, and it's almost never the case. I applied to five grad schools and got into one, and it happened to be the only one that didn't interview me live. Every long-term business success story has its own points of crisis. It comes down to grinding through it in a way that preserves your ability to respect yourself and operate with a great reputation. And second, staying in the game, because you have to be in the game to win.

Jay: Let's close by looking ahead ten years. What widely held view will be proven wrong?

TJ: That pure play defense tech investing is a good idea. There are great teams building important products in that domain, and I know the federal government is the worst customer in the world. I don't want anyone to abandon the mission, but I recommend building for dual use or multiple customers whenever possible.

Jay: What did you learn at In-Q-Tel about how the federal government buys software?

TJ: It just takes forever. If you're selling into the federal government, you can count on very talented users, exquisite problems, and an important mission. You can also count on heartache, things taking far longer than they should, and decisions that are pretty opaque. So have multiple markets you're selling into. Palantir is a fantastic defense-focused success story that also has, I believe, the majority of its business coming from non-defense customers.

TJ: The best founders sell a long-term vision that's both exciting and possible. It doesn't have to get into crazy hyperbole to succeed. The circular economy where everyone in the large-cap AI scene funds everyone else does feel like stacked risk, like the collateralized debt obligations before 2008. But even if that's true, there's profound value that's been built and can't be taken back. Worst case, it's like the Panama Canal, half built, someone else completes it, and the whole world benefits from the new technology.

Jay: As an investor, how do you plan knowing the cliff could be years out or six months out?

TJ: We focus on really special products and enduring businesses that will be great regardless of where we are in the capital cycle.

Pull quotes

1. "I don't know why they've all of a sudden decided that humanoid is the answer, as opposed to a bunch of point solutions."

2. "The federal government is the worst customer in the world."

3. "Build for a problem, create a product that's a direct hit for that problem, and don't start thinking that because it can do X, then Y is imminent."

4. "What I counsel people coming out of college is to think in terms of forty years. That's about how long you'll be working."

5. "The people rooting for the failure of the AI labs don't fully appreciate what they're rooting for."

Source

From CLIMB Episode 084 with T.J. Rylander (N47). Transcript cleaned from the published episode. Watch the full episode: https://youtu.be/_1BnZTkLGLM