Interview 1
[Interview] Nikhil Suresh, Silicon Valley VC in His 20s
Opening and Guest Introduction: Nikhil Suresh 0:00
Chester Roh Today, as we are recording, is July 8, 2026. In the immediately preceding episode, I went to Silicon Valley and heard some stories there and organized and shared them with you. We have invited Nikhil directly, one of the main sources of those stories, to have a more vivid conversation. He’s a very interesting person. Today, CEO Kim Min-seok and I will ask him all sorts of questions. Okay, welcome, welcome, Nikhil.
Nikhil Suresh Thank you. Yeah, great to be here.
Chester Roh So actually, we had a lot of talks while we were staying in San Francisco. I really appreciate your time with us. Thank you so much for making time to be with us.
Nikhil Suresh It was amazing.
Chester Roh Yeah, because your career is really, really special, and you’re so young. So we have a lot of questions today. Actually, let’s begin with how you ended up being an investor at such a young age. You graduated from Stanford, and then you were the first engineer at Mercor, and you were a researcher with your own professor at Stanford. So let’s begin with how you ended up in this position. We’d love to hear your story.
Growth Background from an Engineering Family to Stanford 1:09
Nikhil Suresh Yeah, so I think it kind of goes back to my childhood growing up. My parents were both engineers, actually. My dad worked at a company called Early Warning Services— that eventually made the product Zelle. So Zelle is a financial, kind of banking instrument, kind of like Venmo, that all the banks use to transfer money between different companies. It’s a service used for that purpose. And so he was part of the main engineering effort, leading it— to create that product. And so I learned a lot from him on the engineering side. He always kind of pushed me in that sort of direction. So I originally grew up in Arizona and went to high school there, and then graduated and went to college in California at Stanford. So I did my bachelor’s in math and my master’s in computer science at Stanford. I think—my first year at Stanford, I really wanted to be a math professor, actually. I’d always had this dream of teaching people, learning, and doing pure research. I think within the first year of doing math research, I really quickly realized, like, I don’t want to just look at a whiteboard for the rest of my life. And I think—
The Shock of ChatGPT and the Shift from Math to AI 2:20
Nikhil Suresh my freshman year, my first year of college, was when ChatGPT had first come out. I had spent the last four years of high school working on semiconductors, so working on the hardware side. And so I came to college wanting to explore more on the software side as well, and do kind of a mix of both software and hardware. And so, especially when ChatGPT got released, it was kind of this big wow moment: there are so many capabilities that are unlocked, and so many ways you can automate different things in the world and explore different aspects of the world. And being a pure math researcher for the next 20 or 30 years— felt not as important as what this could potentially be. And so—I think I realized that this was what I kind of wanted to focus on. So my first summer, I started to work at a couple of different startups
and was also doing research on hardware chip design, on VLSI circuits— VLSI, VLSI, yeah. for some deep-learning applications, making them very, very specific to whatever applications we had in mind. it was my first exposure to AI, to formal AI research, and I was really excited by what the possibilities were there. And so, over the rest of my Stanford journey, I ended up working at a lot more startups. I think I ended up working at maybe seven or eight startups. I think it was maybe seven or eight startups.
Startup Experience at Stanford, Video Avatar Startup, and Joining Mercor 3:28
Chester Roh Seven or eight startups? Yeah, that’s—I didn’t know that.
Nikhil Suresh Yeah, a lot of startups.
Chester Roh As an intern?
Nikhil Suresh As an intern at most of them, or part-time at some of them. I liked work, and I liked working with other people. And so I would just work 20 or 30 hours a week at some startups. And then, junior year, I was working with some friends on my own startup, on these video avatars, kind of like HeyGen or Tavus. And so, in those early days, one of the first early people that we reached out to was Mercor. So this was back in 2023, before Mercor had even raised their seed round. And we had just seen them in the LinkedIn comment section, and we were like, “These guys would be perfect for us to test with.” We didn’t end up raising money for the company because we thought we could just bootstrap the company and it’d be great. But the only thing that, it turns out, actually mattered was how cheaply and reliably you could serve those video models. And so you actually do need to raise a lot of money to serve them at scale. So we ended up shutting the company down, but then, a couple of months later, I actually ended up joining Mercor. They hired me to help lead a lot of the research engineering efforts. I think it was very early on in the whole data wave, and human data, right? at a very early stage. And so a lot of what we were doing initially was labor markets. But within the first couple of months of me joining, we really shifted our focus to human data: What kind of data is most important for the models to get better at? How can we create these evals and benchmarks that can actually assess what the verifiers look like?
Expansion of the Human Data Business and Transition to Investing 5:13
Nikhil Suresh And not only did the company grow from 15 people to now over 500, but I got to work with the best AI teams at OpenAI, Anthropic, DeepMind, Meta— all the big labs. And so I think that was, honestly, a really amazing experience: to work with great talent and great people, but also get to see what the frontier looks like across all the labs. I think the one thing about Mercor that I really found exciting was having that breadth, having the ability to see the whole field. And so that’s kind of what drew me to investing. And so my partner, Max, started a new firm called
Striker Venture Partners. And so, before he started, he had reached out to me shortly before and said, “Hey, I’m starting a new firm. Would you be interested in joining me?” And so, at the time, I actually wasn’t super interested in joining a venture firm. I actually thought, let me go be a researcher at one of the labs. You get to work with the best people. A lot of my smartest friends were working there, and so I thought there was a lot to learn there. But also, being in a company actually at the frontier, doing the research and building the models, I think there’s just a lot of stuff and material that you can kind of gain from that long-term. And so that was my life plan for a little bit, but I think over a period of time, he eventually
managed to convince me that working as an investor at Striker would be just as fulfilling, if not more fulfilling. You would be able to see a lot of the frontier, work with the best teams across a bunch of different spaces, and be able to control how deep and how broad you want to go. And so I thought, to me, it was very exciting because it felt very much like a startup, where we were starting with nothing. We had no investments, no track record, no name for ourselves. And it was kind of just like: How do we build our brand? How do we build our investments? How do we find the best people and make sure we’re partnering with them? actually, we have a lot of questions about Striker,
Chester Roh but before then, let us ask some more questions about your trajectory at Stanford and Mercor. Actually, after listening to your story, I really envy your position because you were at the right place at the right time: at Stanford at the beginning of ChatGPT. and you were exposed to the proper coursework, proper social networks, and a lot of good friends. So at such a young age you are in really core positions at really renowned companies.
Reexamining the Value of College Education in the AI Era 7:48
Chester Roh My question is about Stanford, because Stanford is such a prestigious school, but some people are saying that undergraduate school is no longer an important step in the life of the younger generation. So, as a Stanford graduate, how has being at Stanford affected you?
Nikhil Suresh I think, especially among the best AI talent and the best young people out there, this feels like an increasingly popular consensus opinion: the world is changing so fast. If I just stay in school for the next four years, am I going to lose a lot of that experience? Am I going to fall behind? I think that’s a super-fair point. I think that’s kind of one of the reasons why I worked so many different internships: to get that experience, to be able to see outside of just what we were learning in college. I will say, though, that I think college is honestly a super-valuable experience, even if that’s not a super-popular opinion. I think, like what you’re saying, the people that you meet, the community that you build, there are a lot of those connections. You know, as you develop, you learn a lot from those people. You get to see what they’re working on five, ten years down the road, and I think that ultimately shapes how you view the world and how you view what problems are the most interesting.
Yeah, and I think the other thing is, especially at Stanford, a lot of our classes were frontier research, right? Our professors were the top people in diffusion, the top people in intelligence, the top people in neural networks. And so, to be able to learn from them in their classes, which they actively update every single quarter with frontier knowledge, was just tremendous. I took a diffusion class with Stefano Ermon,
who is now the CEO of Inception as well, and I will admit that class was very difficult—it was super hard. I think the midterm was open-note, with open internet access and open ChatGPT access, but we got exposed to frontier-level diffusion work that was being done at the time. And so I don’t think a lot of other colleges were really talking about this, but he had a lot of this material. A lot of the frontier people who were working at DeepMind, working at OpenAI, and all these other labs came into our class and talked to us about what the frontier really looked like. And I think having that exposure, especially as you start to build your network and build that trust with other people, having that exposure, both on a personal basis and on a credibility basis, was super helpful. And so I think that’s honestly the value of college: being able to build that credibility, being able to build those networks and connections, and learn a pretty vast breadth of knowledge rather than just being siloed into one specific thing is, I think, super powerful. where you want to be the best is super important.
Semiconductor and Experimental Physics Research in High School 10:38
Matthew Kim I think you had a very interesting interest when you were in high school, because very few people are interested in hardware and semiconductors in high school. What brought you to dig into semiconductors?
Nikhil Suresh So I think for me, initially, very early on in college— I mean, in high school—I was always interested in math, and so I think with that interest in math also came a lot of interest in physics. And so I just wanted to work at a lab that was doing anything related to experimental physics. And so I just cold-emailed maybe 200 to 300 professors, asking, “Can I just work in your lab as a high schooler?” And as you can imagine, most said no. Actually, everybody but one said no, and so I ended up working with that one professor who gave me a chance. I got to work on a lot of cool projects. Semiconductors were probably the biggest one. So I worked on the nanobonding between two different semiconductor materials, looking at surface energies and things like that. I also worked on a lot of materials science related to bio, which I thought was super interesting at the time as well.
One you hear about Theranos, by any chance? So this came out when I was a freshman in high school, And so the whole thing with Theranos, just for a little context for the viewers, is that it claimed you could conduct blood tests with a thousand times less blood than current blood tests required, using the same kind of blood-testing technology. And so there was a huge scandal about transparency and about how the technology actually worked. And so, at the time, we were doing a lot of elemental analysis in our lab for semiconductors and seeing how the bonding process actually worked, how the surface energy bound, and things like that. And so we thought, could we apply this technology to blood analysis? We thought we could. And so I spent a large portion of my time in the lab working on that kind of technology. We developed a polymer coating that could solidify blood in a way that made it analyzable by spectroscopy techniques like X-ray photoelectron spectroscopy or ion beam analysis and things like that.
And so I think, for me, it was just that I always had this interest in real-world things you can touch, like physical-world stuff, especially on the more physical side. I think that was always my interest. Yeah, and so I was just super interested in working on that.
Matthew Kim You sound like a practical physicist.
Chester Roh Do, like, at least, like, a 200-level course. So actually, before we move into the Striker thing, let’s just tap on Mercor because you were definitely also at the right time, right place, right position, with the right people at Mercor. But actually that was a time when frontier models were evolving rapidly. So what did you learn, and what did you experience while working within a frontier lab?
Mercor’s View of Frontier Labs’ Demand for Human Data 13:15
Nikhil Suresh So when I started, we were, I think, around training models like GPT-4, the earliest xAI models, some of the DeepMind models, and things like that. So it was still very early on. Reasoning and chain of thought weren’t really that prominent yet. And so at that time, I think a lot of the model capabilities were still very basic. They couldn’t do very basic math or very basic physics. A lot of our earlier tasks were like, “How do you do middle-school-level math?” That was one of many early tasks. Even that, right?
And so I think a lot of the early problem, especially from Mercor’s side, was: How do we find people whom we can staff on today’s projects, right? It was about building that database, building that expert network, building that muscle out. And so I think I learned a lot on that side: What does the right profile of a person look like for these specific tasks? What do the tasks need to look like? How do you manage those kinds of operations to make sure that the data is pushed out in a very credible way, that the data makes a lot of sense to the customers, who are these big AI labs, and that it’s packaged in a way that is actually valuable to them?
I think the other thing was discovering how we ourselves pitch new projects to the AI labs based on what’s actually valuable to them, right? I think a lot of the value comes from being able to say, “Hey, I think your models are not good at this thing. So what you mean is preparing the pre-training dataset
Chester Roh or the RLHF dataset?
Nikhil Suresh Yeah, a lot of that.
Chester Roh And 2024 was when the OpenAI reasoning model was released. In September 2024, right? Yes. So after that, I guess you guys were asked to create a lot of data about reasoning. So I think, for us, coding data was the biggest spend out of everything.
Nikhil Suresh I think the labs always had a big focus on how to make coding the best, because I think, in terms of economic productivity, coding is probably one of the biggest spends. But then I think after that, it transitioned into more vertical tasks like math, physics, legal, bio, health, and some more niche tasks that they had. And so I think after that, it transitioned
Chester Roh Actually, even though we introduce a lot of technological advancements by frontier labs such as Anthropic, DeepSeek, and Gemini, we introduce a lot of technological advancements by frontier labs we mainly focus on the algorithm part. But the real one is the data. You know, working in a fast-growing startup always means
Matthew Kim addressing lots of problems day to day, like a war. What was the most challenging problem you experienced, in hindsight?
The Challenge of Predicting Model Progress and Mercor’s Strategic Shift 16:27
Nikhil Suresh I think one was that model capabilities, like you said, were developing super quickly. And so the question for us as data vendors was: We needed to be ahead of where the model capabilities were and try to predict not just today, but where they would be in three months, right? I think data, especially, is a matter of trying to be the best at operations, making sure that we can deliver on contracts as well as possible, but also making sure that we understand where it’s going and how we can predict that going forward. So I think that was one of the—I won’t say problems, but one of the challenges that we had: How do we address I think the other thing was, especially in the early days, trying to find our identity. We started off as this labor marketplace company doing full-time staffing, and this human-data business offered a lot of revenue—a lot of revenue—and the potential to grow really, really fast and scale the company. And so how do you, from a strategic point of view, calibrate whether you should focus more of the engineering and operational effort on the short term, with this human data, which is what it seemed like at the time, versus the long term, on labor marketplaces? Now, when we thought about it back then, if you look at the long-horizon, post-AGI world, you assume that these models will be so good that they’ll be able to handle a lot of knowledge work. And they will be able to do that work. And so, in that world, there are a couple of assumptions that you can make, or consequences that you can formulate. One is either that the way humans interact with these models will change significantly and the number of jobs will remain relatively the same, or these models will be so economically productive that the number of jobs in certain sectors will start to decrease, and those are typically the sectors that we would probably be helping to hire for. And so, how do you think about… prioritizing your engineering effort and the tools that you build and the evals that you create for each of these kinds of efforts. And so I think that was one of the big internal conversations that we had for a while.
Because I think at that time, it wasn’t super obvious either that the data spend would continue to ramp up at the scale that it has, right? These models have gotten to trillion-parameter, 10T-parameter models. Back then, they were on the order of 10B, 100B maybe. They were much, much smaller, and so the data spend was also exponentially smaller too. And so it wasn’t as obvious back then to a lot of people that this was the right call. A lot of people thought, “No, this will be a market for maybe a year or two, and then it’ll get saturated after you have enough experts that train it on all the physics data and all the math data,” and But we’re seeing that this is not the case and that the data spend will continue to increase for at least the next couple of years to come.
The Debate Over In-House Data Labeling and Market Outlook 19:16
Matthew Kim I see Twitter debates about data-labeling companies. Some people say that if the frontier labs, OpenAI and Anthropic, are going to be listed, they’ll want to minimize costs, and then they’re going to build internal teams for data labeling. And then the revenue for data-labeling companies is going to get squeezed. That’s one point.
And another point is that we’re not living in an agent era yet. It’s just beginning. If you consider the demand for huge agents, definitely data is a key point for that. So there, what is your perspective on this debate?
Nikhil Suresh I think xAI has been doing expert-level labeling in-house for a while. You’re starting to see other labs like OpenAI starting to transition toward this now. I think that there are kind of two ways that data spend occurs now, right? One is on the model side, for evals, benchmarks, and human data for the actual AI models, and the other is on the enterprise side. So, creating these verifiers, creating these RL environments for these open-source models, to help enterprises reduce their token costs and also increase their performance on certain tasks. That’s the goal. And so I think that second camp is actually a huge market. knowledge work, it’s massive—on the order of trillions of dollars. And so the data spend will also be accordingly proportional, right? And so that’s billions, if not tens or hundreds of billions of dollars, just there alone. And so I think maybe the labs will start to stand up their own operations,
but do they have the resources to stand them up at the scale that Mercor has already been operating, with such a smooth engine up to this point? Where do the priorities lie for OpenAI? Is it more on the model-training side? Is it on the data-creation side? I think the nice thing with data labeling specifically is that you already have a market that’s pretty competitive, right? You already have companies like Snorkel, Mercor, Surge, AfterQuery— And so I think a big part of that is execution:
making sure that customer delivery happens on time, with high-quality data. And making sure it happens on time. And I think Mercor has historically done a very good job with that. And I think that’s something to also keep in mind: you are paying for that expertise; you create a ton of benchmarks across different sets of tasks. And I think that’s the argument for why companies like Mercor have such a strong position. They have a strong position for this reason.
Chester Roh Actually, these guys are, you know, creating the demand by themselves. We met the guys from AfterQuery. They’re creating something in a niche, highly specialized area, and then they go to the frontier labs and say, “Your model is not good at this. You should buy it.” Exactly, yeah. They go to the next one and say, “Okay, this one will excel your model, so you should buy it too.”
Matthew Kim Yeah, exactly. So, as you mentioned, enterprises haven’t hired digital workers yet. What if they leverage their traces and their data? Definitely, that would be a huge market outside of frontier labs. That would be a good point. Exactly, yeah.
Nikhil Suresh I think capitalizing on context and enterprise data is a big theme for this year. models are getting bigger and bigger. And the size of the data is also getting bigger and bigger.
Chester Roh So I think the market outlook is pretty nice. To shift gears to the next topic, you recently created a venture capital firm, Striker.
Founding Striker to Win Through Highly Concentrated Investments 22:45
Chester Roh from what I don’t like about the Korean venture capital industry, because when there is one deal— let’s say it’s a $10 million deal— 10 venture capital firms each pay just $1 million. It’s like spray and pray to alleviate their risk exposure, and I think that’s kind of a dumb approach. So once you figure out the really, really nice firm, you should monopolize the opportunity. you guys are doing exactly. So could you please introduce your investment thesis and your portfolio companies? investment thesis is honestly super simple.
Nikhil Suresh We are aiming for about 10 investments for the full fund. So, just like you said, instead of doing this spray-and-pray strategy— which I think can generate great economic returns— if you want to really get the highest upside possible, and if you really want to leverage your ability to change the outcome of companies and truly work with the companies that you invest in, I think you need a super-high-conviction strategy. And so this means investing in relatively few companies, taking a board seat, and really helping the companies with anything they need, whether it’s hiring, strategy, or ops. Any sort of funding or compute problems— I think that’s what our model is meant for. We’re trying to find the 10 best companies in categories that we very strongly believe will become $100 billion or trillion-dollar categories, and find the best teams in those categories. I think our model is obviously super risky, right? For us, having only 10 bets,
if any of them don’t work out, our fund goes to zero super fast. But I think that also has another side: it forces us to keep the bar very high. in categories that we very strongly believe in, and it creates very strong alignment between the founders and us, right? We’re all in with them, and they’re all in with us. And so it’s in both of our best interests to make sure that the company succeeds, no matter the cost. And so So you’re on your first fund right now? Yes.
Fund Size and Portfolio Management Principles 24:47
Chester Roh How big is your fund, and how much have you already deployed across how many companies?
Nikhil Suresh our first fund is $17 million. We have deployed into around seven companies at this point, ranging from AI for chip design and inference to world models for robotics, video visual-reasoning models, and a couple of cybersecurity companies. Those are our main public focuses right now. So we probably have five or six more investments to make, but that’s the state of the fund right now. I don’t know how much I can publicly share, but we have 10 LPs for the fund.
It’s premier capital, premier investments. There’s no riffraff; there’s no nonsense going on with us. We want to make sure that everything involved is very clean, very blue-chip. I think one of our big personal theses is that, with companies, the number one thing you want to make sure you can— maybe not the number one, but one of the main things you need to ensure— is that the company continues, right? If you can help a company continue to survive
and reduce its mortality risk at any point, then a company with an amazing team and the best people can find a way. And so what does this look like? It looks like helping the company continue to raise money and prove itself out, whether that’s building the product or finding new go-to-market angles. And then the other thing is: can you continue to hire the best people? Can you build a great team? Right? If you have unlimited money, a great team, and you’re directionally in the right category, you can build a very, very great product. a very, very great product. And so the question is:
how do you get to that point? And I think a lot of our focus is geared toward that perspective. And I think How do we help our companies fundraise? How do we help our companies make sure that they’re hiring the very best people, the very best talent? In order for you to identify the opportunities, how do you do that?
Deal Sourcing Strategy and the Ricursive Investment Case 26:44
Chester Roh How do you source prestigious companies? We have P-1, Eve, Ricursive, and Elorian. All of the founders’ backgrounds are really prestigious and second to none. So how do you source those deals? How do you source those deals?
Nikhil Suresh I think it starts with first understanding the category. We want to understand where we think the biggest markets are going to be. I think hardware is obviously a really important theme this year, and I think it has been cyclical for the past however many years. But especially this year, with the chip shortage, it’s been a very prominent theme. And I think the models are starting to get good enough that you can do AI for chip design at a very, very high level, with great accuracy. If you have the ability to reduce the time for improving chips from a year and a half to just a couple of weeks, that’s a huge unlock. And so it’s about thinking through that, talking to researchers, and figuring out what the major categories are; I think figuring out the categories is the first thing that we try to do at Striker. And then the second thing is the founders: who are the founders? Who is the founding team?
and what is their background related to the problem at hand? Why are they best suited for this problem? How long have they been thinking about it? What is their angle? What is their edge? If you can answer those questions, then I think we gain more conviction about why you are the team that we should back and why you’re the team that we think will win in that category. That gives us much more conviction. For example, the Ricursive team was started by Anna and Azalia. Azalia was actually my research professor in college. I was doing an inference project under her at Stanford.
Matthew Kim What a small world. What a small world, indeed. She’s an amazing professor and an even more amazing researcher.
Nikhil Suresh She was one of the co-creators of Mixture of Experts, which is very widely used in many AI models today. She was one of its co-creators. And she was the co-creator with the other Ricursive founder, Anna Goldie, of Circuit Training and AlphaChip. And AlphaChip was the first major AI chip-design method used at DeepMind to generate four generations of Google’s TPUs. It was the first major AI chip-design method. That paper initially got a lot of backlash internally within the Google community: “How useful is this actually, and how accurate is this really?” But they managed to prove all those claims wrong, fight against the backlash, and prove themselves correct. This team is clearly an N-of-1 category leader with top-tier experience in chip design. So when we had the opportunity to meet them and invest, we did not even hesitate. Right after the first meeting, we said, “We want to invest in you.” “How will you take our money?” Let me slightly change the question.
Striker’s Edge in Backing Top Founders 29:41
Chester Roh You just answered in terms of your company. But your professor could have had many different opportunities, I imagine. because many other prestigious funds would approach her and say, “Take our money.” So how could you win that competition? She really did have many opportunities. To give a little background,
Nikhil Suresh after she was at DeepMind working on AlphaChip, she and Anna were actually the 25th and 26th employees at Anthropic. They were there for a year and built much of its infrastructure and post-training stack, then left to return to DeepMind and Stanford. I think it was a mix of factors. We obviously had a personal relationship with them.
My partners are also on Reflection’s board, so they have frontier knowledge of the compute landscape and the hiring landscape. I think not many VCs really have access to this, frankly. Reflection is a U.S. open-source lab,
and outside Anthropic and OpenAI, there are not many places like it. And most VCs are not on the boards of these companies. Honestly, I am very lucky just to have both of them as partners, and to have both of them on Reflection’s board, is an enormous advantage.
Reflection was founded by members of the AlphaGo team and includes Joe Spisak from the PyTorch team, as well as A.C. from DeepMind’s pre-training team. That access gave us a network of amazing AI researchers, and let us say, “We can help you hire staff- and senior-staff-level researchers,” while also giving us that connection. She has worked with many of those founders, and my partners have had a very strong track record with those founders. In the end, it was a mix of that personal relationship and exceptional founder references, plus a frontier-level view of the state of compute and models, and where we think the world is heading. That frontier-level view was an important part of it.
Diagnosing the AI Value Chain and Infrastructure Bottlenecks 31:39
Chester Roh You pointed out that understanding market structure and timing is extremely important. So my next question follows from that first point, because we cannot copy your social network. Let us ask about your view of the current market landscape in AI. There is a clear value chain, from sourcing land and building a data center, to chips, NVIDIA, and many startups, software orchestration, and finally models and services. That is the AI value chain. The timing is different at every stage. Could you share your view of the market landscape and timing for each layer of the stack? Yes, definitely.
Nikhil Suresh I think we can start at a very fundamental level. In America, you really have to start with power, because power is the most constrained resource. In America, the situation is the reverse: power is extremely constrained, and there is also substantial local-government backlash. So bringing new data centers online takes a long time. Combined with chip shortages, photonics, lithography, and all these other factors, this delays data-center build-outs. As a result, companies have to take out leases on chips for use two or three years down the line. At this exact time last year, chips were available. Frontier chip prices were also quite reasonable, around 1x. But now, if you want a large cluster, you are paying 2x or 3x for these chips. So if you are building a very compute-intensive company, you need to think about which resources and inputs you actually need to build it, day to day. You need to think carefully about those needs. You need to monitor them far more closely than six months ago.
On the model side, there are many neoclouds now. We all see these neoclouds raising money in very flashy ways, But the scarcity of models and compute is being underestimated. Maybe six months ago, you could find ten neoclouds and all of them would have access to compute, but now it is extremely difficult. Another question is where value actually accrues. From an economic perspective, does much of the value accrue on the chip side? Does it accrue on the land-and-power side? Does it accrue on the application side? Does it accrue on the infrastructure-layer side? I think that is also an important question to assess. With better open-source models that are easily post-trainable now, and growing post-training infrastructure, we are seeing companies, especially large enterprises like Uber and Coinbase, become smarter about token economics. They are starting to understand that you cannot just throw Fable Max at every problem and hope it works. You do need to be mindful of how and where your token costs are being spent.
I think margins are also becoming a serious issue, because last year and the year before, many application-layer companies were operating at very significant negative margins. “Here is a business with great revenue but very bad negative margins,” and it is an application-layer company that does not own the model; it only owns the wrapper. We are now seeing many people move away from that model, because we are realizing, “This model does not make as much sense as we thought,” because if you do not own the model, the chips, or the core infrastructure beneath it, it is very difficult to make the structure economically viable unless you can truly use open-source and very cheap closed-source models effectively. That is why I think a major theme for next year will be:
Open-Source Post-Training and the Economics of Evals 35:23
Nikhil Suresh how do you post-train open-source models for your specific task? Companies like Mercor are very well positioned, because these verifiers cannot be generated autonomously. You need humans to create these verifiers for specific tasks. So that is where I think much of the value will ultimately accrue: in evals and verifiers. For example, if you take a vertical such as accounting firms, create evals for every task across different firms, train that into a model, and use that model again, you can create a great deal of value. So, probably two years ago,
Matthew Kim if you said, “Every enterprise needs its own eval system and environment,” researchers and engineers would have said, “That is not possible. That is too expensive.” But times have changed. So do you think that this year, or in the near future, enterprises will build their own eval environments, and build multiple agents using open source, minimizing token costs, and maximizing token efficiency? Yes, definitely.
Nikhil Suresh There is a metric I really like, which came from the Hazy Research and Scaling Intelligence groups at Stanford: intelligence per watt. If you have a particular task, what is the cost in watt-hours to accomplish that task? I honestly think that is a very important metric, because it captures not only token count, but cost per watt and cost per token. It encompasses all those factors. We are starting to see this number fall in six-month increments. It is falling fairly rapidly. This is a pretty astounding metric, because it shows that we are moving toward commercial viability for these larger multi-agent frameworks in enterprises.
Chester Roh I would like to hear your view of the market timing. I will give you just 100 points. I wonder how you would distribute those points across each layer of the AI stack. Where would you put your points?
Physical AI and Robotics at Their GPT Moment 37:40
Nikhil Suresh Personally, if we are talking about right now, I think robotics and physical AI foundation models are a very crucial area. I honestly think it is even bigger, but it is like the GPT moment for physical AI. Companies like Skild, Physical Intelligence, and Generalist are building these foundation models for robotics, and they will eventually be used in a form analogous to vertical SaaS companies. But for robotics. I am already seeing construction companies use fine-tuned π models, and excavators and drills use fine-tuned π models as well. These cases are beginning to emerge. So I think I would put 40 or 50 points into physical AI right now. I think some areas of physical AI are still a little too early and need more time to develop.
Robotics Scaling Laws and the Data Bottleneck 38:32
Matthew Kim What do you think about where we are compared to the progress of LLMs? Because if— If you see the history of AI, very few people figured out a pre-training scaling law for GPT-1, and then a couple of people figured out RL and test-time compute figured out after GPT-3. So where are we for physical AI? Do we know a scaling law for pre-training?
Nikhil Suresh Yes, I think the architecture has been evolving pretty dramatically. We started with VLAs and VLMs and are slowly moving into world action models. But I think we are starting to see scaling laws emerge in robotics, as we did in LLMs. But I think one of the main differences between LLMs and robotics is that you had the whole internet to train on. You had this very easily accessible data source that you could just take and say, “Hey, let me just give this to the model.” With robotics, it’s a lot harder, right? A lot of it needs to be handcrafted, formulated, and given sensors for specific applications, then fed into the models. And so I think that scaling up—how do we increase the data, whether that’s in the real world or in simulations?— is what’s going on right now. A lot of new world model companies are emerging,
and a lot of development is being made in larger companies, with Genie and Dreamer, or GR00T even, and Cosmos from NVIDIA. And so, yeah, I just think there’s still a lot more to be done, but I think we’re nearing that point of having that big breakthrough. So I would say we’re probably closer to the GPT-2 moment. I think we’re starting to see, especially with a lot of the general foundation models,
that they have solid revenue from doing these deployments at larger scales in manufacturing and assembly-line tasks. But how that translates to a wide variety of things is the question. Now,
Chester Roh and you just mentioned that you’re highly interested in the physical layer, that might be for one of the reasons you mentioned: there is some exclusivity around the data, and for the environment, you do need some type of physical environment to run tests. That’s why the frontier labs aren’t easily entering that market. From that perspective, I think there might be other areas. For example, maybe AI for science. And you introduced us to a lot of people in the biotechnology area, so there might be some next area of interest for you. What would that be?
AI for Biology and Science as the Next Frontier 41:00
Nikhil Suresh frontier labs are starting to move into robotics now. OpenAI has moved a lot of its Sora team into robotics at this point, right? But I agree. If you ask me where I would put the other 60 points, a good 30 or 40 of that is going into bio. I think bio is one of those big next frontiers. And I think you can learn a lot from where the labs put their focus, right? If you look at Anthropic, if you look at OpenAI, they’re building out these life sciences teams. They’re building out these drug discovery policies because they also believe that biology is one of the biggest frontiers in human civilization. It’s one of the biggest problems, and economically, it’s also one of the most important problems to solve. I think longevity specifically is one of those problems. And so you’re starting to see companies like Isomorphic, Chai Discovery, Lila Sciences, and Edison Scientific start to emerge and become well-capitalized to try to tackle these problems, not only from the wet-lab perspective—
how can we attach Codex and agents to wet labs to do drug discovery and create actual foundation models for bio?— but also on the AI-for-science platform side that you’re seeing from Google, Anthropic, and Edison Scientific: and How do we take all of the information that’s out there online, in all these different research papers, and formulate some sort of automated research to find new findings? Whether that’s with an actual lab or just with a bunch of agents, I think that area is super exciting to me.
Chester Roh then after bio, what would be next? After bio, just to pick something.
Nikhil Suresh the reason bio is super important is because you’re able to solve a lot of key problems with life, right? A lot of the problems involving suffering, in terms of cancer or mental health or things like that. I think you’re able to solve those with a lot of these foundation models for biology, drug discovery, and things like that. Once a lot of those problems start to be resolved in the long term, and we have the ability to live longer and more fruitful lives, I think something that becomes actually really important in the very long-term, post-AGI case is: How do you actually build those communities of people? How do you make sure that now that we have longer lives to do more things and explore more things, people still have the vehicles and those third spaces to be able to interact with each other in a meaningful way? We already have AGI, presumptively, to solve a lot of our core problems, to build out the software, and to build out the infrastructure, hardware, and everything else that we need. We can presume that capability. And so now the focus is on how you foster that human-to-human connection and how you make life really worth living. That’s what I envision as the very long-term future.
Human Connection and Community in the Post-AGI Era 43:13
Sovereign AI Potential in Korea as a Hardware Powerhouse 44:09
Chester Roh Let me ask a little bit about the Korean perspective. You just mentioned the AI landscape, and from the Korean perspective, what are you specifically expecting from Korea, per se? I think Korea is really interesting because it’s one of the few countries
Nikhil Suresh that has amazing AI talent, amazing manufacturing, and amazing hardware. If you look at a lot of countries, most countries don’t have that capability, right? There are also only a few countries in the world that actually even have the ability to maybe do a sovereign lab, right? think Korea could be that next country because of the talent density that’s here, both on the software and hardware sides. And so I could definitely see it becoming one of the next bigger powers on the AI front, developing those capabilities. You have amazing companies like Samsung, with its foundry. That’s a great TSMC competitor. You have SK hynix. You have just a ton of amazing companies here, and so I think there’s a lot that can be done, honestly.
Chips and Hardware as Key Areas for Startup Investment in Korea 45:15
Chester Roh So, Nikhil, if I ask you, you know, if you were to fund one of the prestigious Korean startups right now, what area would it be? For me, if I were looking to fund a Korean startup right now,
Nikhil Suresh I think hardware is probably the first thing I would go toward. I think Korean hardware has historically been very, very strong. Electronics and hardware have a very strong place in Korean history, and I think it’s just something that they’re among the best in the world at, right? Having that knowledge and expertise—better than almost anybody else in the world— is just a no-brainer. You
Chester Roh Hardware you just mentioned is mainly about robotics and physical AI?
Nikhil Suresh Physical AI, but I was actually thinking more on the chip side— more on semiconductors and memory, right? Building specific inference chips for specific architectures and things like that.
Matthew Kim lot of components are required for building data centers. Exactly, exactly. Actually,
Chester Roh we’re going to have a party with SemiAnalysis this evening, and you’re going to meet a lot of talented chip startups in Korea. What actually led you to build an agent at the time?
Experience Building Agent Startups and the Reason for Switching to Investing 46:30
Matthew Kim What did you do, and what did you learn from that experience? Why did you decide to quit and then jump into investing? Yeah, totally. 2025 was an interesting period
Nikhil Suresh where enterprises still weren’t super keen on using AI at that point, especially the more legacy enterprises. They didn’t really want to use AI and didn’t see the economic value. And so convincing them that AI was the answer to a lot of their automation problems was a really hard sell, which was the number one issue. I think number two was that, at the time,
we were working on very bespoke, custom deployments. And I think that just doesn’t scale very well.
So if you want to create a lasting business that is actually super meaningful, and that you can envision yourself spending the next 10 or 15 years doing, a lasting business I just felt that, for me, that wasn’t the business I wanted to be doing. I think there were more interesting problems that could help more people. And so that’s why I ultimately ended up going into investing. I think you have the ability to influence that—to fund ideas that can actually impact a lot of people in a very positive way
Where the Edge Lies in an Increasingly Saturated Agent Market 47:44
Chester Roh we didn’t mention the many agent But we’re seeing a lot of startups in San Francisco and Seoul. They’re all agent startups. So what is your take on those startups? How does the market timing seem? Market timing? I mean, I think it’s great. This is kind of the year of agents,
Nikhil Suresh when the models are becoming capable and enterprises finally want to start using AI a little bit more, which is why it makes sense that now is the time when everybody is going to try to capitalize on this opportunity. You have the whole digital labor market that wants to be uplifted and can be advanced with agents. And so having that insight into how you actually build agents in production systems— the people who started very early, last year in 2024 or at the start of this year, have that knowledge and expertise that they’re going to deploy. The thing to keep in mind, though, is that building agents has a very quick ramp-up. You know, you can easily, quickly scope up on that. And so I think that’s what’s also causing a lot of people to start these companies, because they realize the market’s big, there’s a lot of greenfield space, and there’s a lot of money to be made as well in building these agents for different enterprise tasks. And so I think that’s what makes the market very competitive and very saturated. It’s about figuring out: How do you fit into the market?
Where are you positioned best to really compete? you came to us saying that you wanted to build another agent company, and we said, “What are you building for?” and you said, “I’m going to build for enterprises,” and we said, “Why do you think you’re going to win?” and your answer is, “Oh, I think I’m going to move fast,” that doesn’t mean anything. Everybody thinks they’re going to move fast, but why are you built for this field? Why do you have that mandate? And so I think that’s the question that every founder should be asking themselves. I think Yes, that’s right. This is my last question. You came to Korea three or four days ago,
Benchmark-Centric Research and Model Interpretability at ICML 49:35
Chester Roh so you met a lot of people at ICML. You had your own philosophy before coming to Korea, and then you met many people in Korea and talented people from frontier labs. Has a conversation with someone changed your perspective? What was it?
Nikhil Suresh I definitely changed my perspective on a couple of things. Honestly, I think there is a lot of really interesting research being done at ICML. I think many of the papers we are seeing are about benchmarks and very specific topics. I find that interesting because, if you look back a couple of years, it was like, “What are some cool new model architectures? What new experiments can we run?” But now it is like, “Why are models failing on this benchmark and this specific task?” That is the question being asked. It was not a big change, but it shifted my thinking toward how to make models better at the specific things they are not currently good at. How do we make models perform better in these specific areas, and how do we formulate these benchmarks? In that sense, my perspective shifted a little. It is something that is top of mind for many researchers right now. Another thing is that, to me, many researchers
seem deeply invested in the idea of a singularity, with one super-powerful model that can do everything. Anthropic recently published very interesting work on how a model thinks and on mechanistic interpretability, and I think it is increasingly becoming a larger field in its own right. It started slowly, but it has been accelerating faster and faster, and many researchers at frontier labs are starting to think about what actually happens inside a model. How are a model’s thoughts and memories actually perceived? Anthropic released a paper called “The J-Space” yesterday, with findings on what the internal framing of these models really looks like in these models.
I met a founder yesterday who left his company as CTO to work at Anthropic because he realized there is so much interesting work. He left his role as CTO at the company he founded to go work at Anthropic. He had originally been working on black-box observability and black-box models, but after looking at Anthropic’s findings on the interpretability and transparency of models, he realized this could not really be true. He came to think that his premise could not be true, and he no longer had the conviction that this approach was feasible. You are starting to see many founders and researchers gather at these talent centers and talent nodes. The question then becomes how to draw that talent toward equally important projects outside the lab, such as chip design and biotech, which carry more diverse perspectives than the work the labs may be doing. How do you pull talent toward those projects? That becomes the important question. This is my last question. You know, the three of us are really into AI.
Faster Drug Discovery and the Outlook for Autonomous Companies in Five Years 52:34
Matthew Kim So what will be the biggest change in AI in five years?
Nikhil Suresh Well, honestly, I do not know. I mean, I could say something, but I do not even know if it would be right. I feel like the world is changing much faster.
Matthew Kim It is hard to predict even a year ahead. I cannot.
Nikhil Suresh Honestly, I feel like I cannot predict at all anymore. There are so many amazing people working on so many amazing projects. That is why. At this point, my hope is that in five years, the time to discovery in new drug development the time it takes to identify candidates for new drugs will become very, very short. Not only should drug discovery become very quick, but clinical-side experimentation should also become faster. This is very important because regulation matters greatly here. That is the really important point. We need to test it on actual people and make sure that it works without side effects. Within the next five years, I hope we also make progress on the clinical side and in clinical testing. I hope there will be progress. That is on the bio side.
On the agent-infrastructure side, we are seeing many people move into autonomous-company development, multi-agent frameworks, and auto-research. I think this is a major focus not only for labs, but also for many infrastructure- and application-layer companies. It is a very important focus. Many new startups are also starting to work on autonomous-company building. They are entering this space. So in five years, I think we will probably start to see many more companies of this kind. It may be something like a holding-company model, where one person runs a top-level company, and that company’s agents build multiple different companies. Those companies are all run completely autonomously and connected to each other only through agents. It will be interesting to see how interconnected the world becomes then. That will be fascinating to watch.
If you think of the Industrial Revolution, when steam engines and ships that could travel the globe made the world highly connected, the world became enormously interconnected. This is like the digital version of that phenomenon. Now we have digital workers that can travel and communicate autonomously, and build trust graphs and relationships with one another. I use “relationships” in a somewhat loose sense, but they are relationships nonetheless. They can share that context and act on it. I think it will be very interesting to see how this plays out.
Final Advice for Korean Founders and AI Talent 54:58
Chester Roh May I ask for a final word to Korean entrepreneurs, AI engineers, and people working in AI as a final message?
Nikhil Suresh If I had one thing to say, it would be that there are many very interesting problems out there. Korea has an amazing talent network in both hardware and software, and I think now is the time to truly capitalize on it. Try to identify the biggest and hardest problems that need solving, gather the smartest people you can, and seriously tackle the problems you believe can truly make a difference in people’s lives. Thank you. Yes, thank you, everyone.
Chester Roh Thank you for having me. Yes, thank you so much for your time. Thank you.