Tarun Chitra is the founder of Gauntlet and managing partner at Robot Ventures, which just led Nous Research’s $90 million round. He joins Tommy on The Delphi Podcast to explain why he backed Nous and Hermes Agent after months of diligence on OpenRouter data and GitHub forks. Tarun traces how crypto’s first run at AI went wrong, from GPU networks to token-launched agents, and why the crypto people who dropped the token have done well. He breaks down why OpenClaw is the worst developer experience he’s used, why Hermes is closer to Ubuntu than to Android, and how robotics makers built on Hermes without Nous knowing. The conversation closes on his market structure thesis: AI inference looks like crypto, OpenRouter looks like a DEX, some providers game cache pricing, and the money will flow to the harnesses on top and the inference providers underneath. Hermes, in his words, is like investing in Phantom or FOMO.

🔗 Connect with Delphi

🌐 Portal: https://delphidigital.io/
🐦 X: https://x.com/delphi_digital
📺Youtube: https://www.youtube.com/@Delphi_Digital

🎧 Listen on all podcast platforms: https://thedelphipodcast.buzzsprout.com

[00:00:00] Tarun Chitra: There’s a sense in which the rise of open source and open weight models in the last six months, like the immense rise, if you look at the fingerprints, people from crypto are all over that. OpenClaw is like kind of the worst developer experience of like anything I’ve used personally. I think that kind of thing is like the difference between product engineering and like backend engineering.

I would find these people who are like, “I’m some random Chinese hardware robot manufacturer, and the interface to my like little robotics embedded hardware thing was like for Hermes.” One thing I was kind of looking for was like, are there people who are lying about how much they’re giving you in cache token?

Like I ran these experiments- Interesting … a- and like tried to make sure I went through all the different providers and, and basically found there were some providers doing this. I think that market structure looks like crypto. All the money goes to the top and the bottom. What is the top? The harnesses.

What’s the bottom? Inference providers. Hermes is like investing in Phantom or Fomo, right?

[00:01:03] Tommy Shaughnessy: Hey everyone, I just hosted Tarun, who’s the managing partner of Robot Ventures and the founder of Gauntlet, on why he led Nous Research’s latest round and his thesis on Hermes Agent. He’s an early investor in Nous and spent three to four months digging into usage and the developer community and got extremely bullish on the opportunity.

We’re three-time investors at Delphi, huge supporters of the team, and I’m a vocal Hermes power user. Tarun opens with the history of crypto and AI. We trace how early Nous has been. Hermes models with tens of millions of downloads, yarn for longer context windows, synthetic model evaluation, and distributed training through Distro.

Now, Hermes Agent. His point is that those years of research and infrastructure were in a detour. That’s why the team understands the hard problems underneath the product That’s why his OpenClaw comparison goes deeper than the interface. His analogy is that OpenClaw hands you the parts and asks you to build the car.

You’re choosing the models, managing the contacts, and figuring out how much reasoning to pay for. Hermes was designed around a different philosophy. Learn your preferences, generate skills automatically from how you use it, and make it easier to share context and route work across models. The it just works experience comes from research depth meeting real product taste, with hundreds of decisions and optimizations compounding from thousands of users and those contributing to the repo.

His bigger thesis is that Hermes could become a front end for AI, like Phantom or FOMO in crypto. You don’t need to understand the infrastructure to use it, and as models improve, you shouldn’t have to rebuild your agent or leave the context behind. The evidence that really stood out during his diligence was he found robotics manufacturers building on Hermes without Nous even knowing it.

They were just trying an assistant. They were choosing it as the foundation for their own products. And the enterprise opportunity takes that a lot further. Imagine businesses having their own version of Hermes connected to their data sets and workflows with their own permissions, privacy, and compliance requirements.

They don’t have to build everything from scratch or accept a one-size-fits-all assistant. Tarun explains how Nous is actively pursuing that enterprise market alongside its consumer community, and discusses Microsoft and Nvidia joining the round, and how they can help lead enterprises to Hermes. That’s the bigger opportunity when packed, not just an agent you use, but an AI foundation your business can make its own.

The models underneath can change while the knowledge, context, and workflows you’ve built remain yours. It’s why Tarun sees Nous as potentially the most important open source challenger to the centralized AI companies. My personal thesis is that Hermes allows businesses and people to own their entire AI stack as an asset, which doubles as providing them unique outcomes versus using the same models as everyone else and getting IP farmed.

Satya Nadella from Microsoft and Alex Karp from Palantir have voiced this thesis vocally. I hope you enjoy the episode. Hey, everyone. Welcome back to the Delphi podcast. I’m your host Tommy Shaughnessy, and today I’m thrilled to have Tarun on. Tarun is the founder of Gauntlet, managing partner, Robot Ventures, and has way better glasses than I do.

Tarun, how’s it going?

[00:04:13] Tarun Chitra: I appreciate it. Uh, g-great to be on and, uh, you know, I think, uh, everyone has to find their… go through their own glasses journey. You know, like, uh, I, I, I will say it– the first time I ever really started looking into crazy glasses was I happened to randomly be in France, uh, like 20 years ago, and there was this like

I, I randomly walked into this conference that was like the eyewear conference, and then that- … changed my life. So like I said, you can, you, you just, you never know when it’ll hit you. You might suddenly, next time, next time Tom’s on, he’s gonna just have something crazy. He’s gonna have- There- … some steampunk glasses Were you there for the

[00:04:50] Tommy Shaughnessy: eyewear, like, conference, or was it just random?

[00:04:53] Tarun Chitra: No. Random, random. Uh, uh, the thing is like, I was like just out of college, and I was like, “Ooh, I see free food.” And- It sounded very stupid … I didn’t know if you’d be into the glasses. And then I was like, oh, okay, like, like, like this conference, I think like didn’t really have that many people, and they over-ordered food.

So they were just like, “Here’s, here’s some free food.” I have these, uh- You know, sometimes, sometimes having a party, you know, you never know what, what it’ll create, what monsters it creates. No, it works.

[00:05:22] Tommy Shaughnessy: I have these expensive Avulux glasses. They’re like migraine glasses, but they look really weird when you wear them, so I don’t usually do it, but they’re amazing.

Yeah, you know, you know, I will say,

[00:05:32] Tarun Chitra: and, and this is a, this is a very controversial opinion, ’cause a lot of people- Um- … get quite offended when I say this, but, um, I think it, I never would wear Transitions, because you look like a serial killer when you wear Transitions. Yeah. But there are a lot of people who wear Transitions, and they’re always like very self-conscious about it.

But like, okay, if you are in New York and you take the subway, or anywhere in the world, you take the subway- Every time you come on the subway, you look like you just came out of one of those, like, spy movies and you killed someone- Yes … in the subway. So it’s like, I don’t know. I can’t do that.

[00:06:06] Tommy Shaughnessy: It’s… I feel like it’s hard to trust somebody who wears Transition lenses.

Like, your life is that dynamic, you can’t buy two? I, I don’t know.

[00:06:12] Tarun Chitra: Yeah, exactly. Exactly. Like, sunglasses, totally fine. Not sunglasses, totally fine. But like, the middle thing, I don’t know, something about it is like, I’m either a card shark or a serial killer. I’m not sure which one.

[00:06:23] Tommy Shaughnessy: Well, will founders know if they pitch you or I to not wear Transitions lenses or progressive lenses?

There you go. Auto red flag.

[00:06:32] Tarun Chitra: Auto red… Yeah, exactly. Exactly.

[00:06:34] Tommy Shaughnessy: Well, Tarun, I’m really excited to talk with you. Robot Ventures just led Nous Research’s $90 million new round. Uh, so huge congratulations. Um, maybe we could start quickly there, and then, uh, we- we’ll gonna dive into the eras of, of crypto and AI.

[00:06:50] Tarun Chitra: Yeah, definitely. Uh, definitely appreciate… I mean, you, you guys as well as us, we’ve been there for the whole journey, and Nous has sort of been through a lot. You know, I think a lot of people… The interesting thing, uh… So there’s oftentimes this type of advice I always hear in, in, uh, you know, that changes in venture capital over time.

You know, I think in the 2010s the advice was like, you want the company to be much more famous than the founder. Mm. You want, you know, you know, you want the brand of the company to be known much better than, than the founder because that means, like, the founder is not just, like, wasting all their time on social media.

Th- this is like the early social media era. And then in the pandemic everything changed into the, like, the founder is the only marketing agency you have. So like, you need them to always be on, uh, social media and have a, a public presence and image and whatever, right? You know, like founder first marketing.

Um, and that, that sort of flipped the script where you, like, saw a lot of people who ran companies where they’re more famous than their company. And then, uh, fast-forward to now, we’re kind of in this, like, interesting thing where sometimes there are companies who have products where people don’t even know who makes…

You know, they have so many sub-products that no one even knows the company, let alone the founders. And I think Nous is a great example of a company like that where, uh, tons of people know what Hermes is. Um, Hermes Agent, you know, the very popular harness. I mean, their social media is extremely popular be- uh, and their social media is popular not initially because of Hermes, but because of their amazing design skills in the beginning.

Um- Mm … but then obviously they made some amazing products over time, but they’ve done a lot of different things. And I think the interesting thing about Nous is- You know, the, their evolution over time really reflects how crypto kind of had to learn what its place is in the new AI world. And a lot of that has changed over time.

You know, what people thought crypto and AI would look like in 2023, the, basically, like, post-ChatGPT-

[00:08:58] Tommy Shaughnessy: Feels

[00:08:58] Tarun Chitra: like forever ago … ha- has, it has evolved so much, right. Exactly. And, like, in general, it’s kind of funny because- Obviously everyone in AI is the, “I don’t think about you at all” meme about crypto. Um, but there’s a sense in which the rise of open source and open weight models in the last six months, like the immense rise, if, if you look at the fingerprints, people from crypto are all over that.

And I think that comes from really understanding the lessons from crypto and, like, applying them elsewhere, versus just, like, shoving a token into an AI thing and hoping it works, which is, I think, how, you know, a lot of the crypto AI stuff started.

[00:09:37] Tommy Shaughnessy: So Tarun, you, you mentioned, like, a lot of early crypto infra projects back in 2023, especially, like, the rented GPU ones.

The idea that if you’re running, like, a big Bitcoin mining network, you could kind of, like, flip a switch and, and like, you know, serve AI inference, host models, things like that. Um, i- in hindsight, it is, it does seem really naive, right? Uh, because the models, the open source models weren’t even around to serve, right?

We’ve only, like, recently, the last year or two, gotten really good open source models from China. So it just seemed, um … Yeah, it definitely seems a bit naive in hindsight.

[00:10:13] Tarun Chitra: Yeah, like I said, I think a lot of people saw the success of existing miners converting their operations, and then they also saw, to some extent, stable diffusion as an example of someone who was running a mining operation converting their existing GPUs into making a model.

I think what people didn’t realize at that time was that language models would just have a lot of idiosyncrasies that, like, image models didn’t. And, like, no one really wanted a decentralized image model, to be honest. Like, no one seemed to care about that that much. Um, and I also just think it’s like, it came from, like, crypto as an open marketplace, and so people are inherently lazy and they’re, like

I- in the sense that they’re just like, “Well, hey, I could launch a token for saying that I’m doing this thing that, like, probably won’t work, but, like, the token will still have some valuation.” So I think in the beginning there were a lot of, like, disingenuous projects that were, like, just, you know, meme coins with more steps, as I like to call, uh, certain types of projects.

I like meme

[00:11:21] Tommy Shaughnessy: coins with more

[00:11:21] Tarun Chitra: steps. I

[00:11:23] Tommy Shaughnessy: don’t.

[00:11:24] Tarun Chitra: Uh, and, uh, but i- i- in, underneath all of that, you know, I, as someone who had, um, you know, worked in sort of centralized hardware stuff 15, 16 years ago and then, you know, at the beginning, I was always like, “Oh, the decentralized things are all scams.” And then after, like, four or five years, I, like, figured, finally understood sort of a little more of, like, why do blockchains work?

Like, what, what is different about them? Why do all these kind of things that you were taught in school about distributed systems in, like, grad school classes, like, why are they violated but still seem to work? And there’s a sense in which because I was in crypto for a while, but also have been on the other side, I was like, you know, there is this sense in which there’s, like, a 1% chance of, like, something, some architecture, some, some real kind of technical advancement being able to get over some of these problems for, for decentralized training.

Maybe I was a little less optimistic on inference, I was much more optimistic about. Um- And I think the thing that stood out to me about Nous was, like, they were one of the only people who was, who were straddling the, straddling the line between wanting to try to do, you know, in 2023, this kind of, like, decentralized, uh, training, but also taking a more scientific approach to it.

So they, like, they wrote a bunch of research and did a bunch of experiments, uh, some with, um, you know, one of the co-founders of OpenAI in the early days on how to do low bandwidth training. And at that time, I think, you know, realistically, there was only OpenAI and Anthropic that had real language models.

There were a lot of, like, really shitty language models, like Llama probably just came out around then. Um, and- I

[00:13:18] Tommy Shaughnessy: have the, uh, I have the… I actually have it in my backpack. Oh, wow. I’ve never sorted it out. Just the, the Distro preliminary report- Yes … like, on the- Which is- … distributed training they were doing.

Exactly. It, it was, uh… I think, I think one of the things you’re getting at too, and, and kind of one of the things we saw early on was the team was just really forward-looking, right? Like, they were doing, like, synthetic evaluation of models. Like, everybody tries to, like, game these benchmarks, but they were doing that in a synthetic way early on.

I think Geoffrey did, uh, yarn pretty early, which was, like, to increase the context length back when you couldn’t put that much into an AI model. Then they did this distributed training run. I forgot if it was 40 billion parameters or something like that. Um, so I, I think to your point, like, they definitely, they never launched a token, they never did, like, anything infra-wise on, on the token side.

Mm-hmm. It definitely seems like they kind of delved into it a little bit, but also understood kind of the opportunity- Right … and where they could, could thrive there.

[00:14:16] Tarun Chitra: And I think they were being rigorous at a time where, like, everyone who was, like, talking about the decentralized side was a little bit of a scammer or, like, like I said, meme coin with more steps.

Like- … scam and meme coin with more steps are, like, it’s like, it’s, it’s a very fine line between the two. ‘Cause, like, you could be well-meaning but just don’t have a good idea, but people still give you money for it, right? And that, that’s a meme coin with more steps.

[00:14:40] Tommy Shaughnessy: Well, well, what got you over the line early on?

‘Cause, like, you’re, you’re kind of a long-term investor here too. Right. Like, did you… Like, you’re, you’re seeing, like, you come from a rigorous technical background. You’re seeing a lot of distributed stuff in AI that doesn’t really make sense. But this sort of, this sort of did to you early on. Well, well before for me.

So, like, in this era- I, I s- Yeah.

[00:14:59] Tarun Chitra: Yeah. I think it’s, like, Distro… Okay, so they, Yarn as well as sort of some of their early research on, like, different types of embeddings for better compression Just sort of stood out to me versus, like, the normal crypto AI companies at that time who … Because, like, they were, like, focused on very specific well-defined problems versus, like, promising the moon of like, “Hey, this system, this complicated system that has a million moving pieces will get perfectly right on the first shot.”

That’s kind of like, you know, right? Any distributed system has, like, 100 times or a million times more moving parts than, like, the centralized version oftentimes. Um, and you know, most of the crypto AI people are like, “Oh, we’re just gonna like, you know, take an b- existing blockchain and, like- somehow just, like, force the node op- node operators to run- Mm

the, the kind of the software. And that’s like, well, none of them can actually do that at, at sufficiently high bandwidth or at a level that, uh, was useful in the sense of like, okay, if I’m doing it at like 0.001 token per second, who cares? Um, and the other thing was like, there were a lot of things like Bittensor that were, like, kind of in this vein where they, like, really jumped too fast to the token and decentralization piece, but then the, like, utility of the actual projects was kind of weak.

In Bittensor’s case, I think they found a medium ground where they kind of became this, like, funding platform, like Launchpad effectively, and like so, you know, arguably there was some residual value from that, but it didn’t really feel like that was, like, a technical innovation. Nous were one of the only people who, you know, Nous and Prime Intellect realistically were the only people doing real research and, like, focusing on well-defined specific problems that would potentially, if, you know…

Like, every successful distributed system, um, is success by like a thousand little optimizations. It’s not like- Mm … there’s one big idea that’s like, oh, okay, this big idea automatically solves everything, right? Like, if you look at even just Bitcoin, going back to the original paper, like, there’s a bunch of different concepts that are, need to work together in order for the thing to, to work, right?

Like, first of all, you need the hash functions to have very high collision resistance. Um, you need this difficulty adjustment to, like, not be gameable. Like, the idea that I can, based on how many people join and leave the network, it can, like, adapt to them, so, like, you can dynamically scale. You need the inflation curve to be, uh, not too steep and not too inflationary also, right?

Like, you know, it needs to have some balance between those. And you need the node software to have, like, sufficiently low bandwidth, right? That’s why Bitcoin blocks are so slow. Um- Mm … and like each of these is like a separate technical problem, oftentimes in, you know, from a different type of specialty.

Like, a person with a cert specialty knows each one of those. And usually to build these dec- decentralized and distributed systems, you have to bring those all together. Mm. So you need kind of a, like, a wider breadth of knowledge rather than depth. And so I think the thing that Nous and, in a lot of ways Prime Intellect- At that time had were these, where they were like the only ones kind of doing, doing that.

I think the difference to me, for me for Nous was, I think Prime Intellect maybe wasn’t focused on the training side as much to- I mean, I know they did do a training run, but like, you know, they kind of hedged their bet and were like, “We’re also gonna be like a GPU reseller at the same time.” Whereas Nous was really focused on this research piece and like that was sort of what stood out to me, um, as, as like, you know, i- in order to ever have decentralized training or decentralized inference be a viable concept, um, you just need these 10, you, you need these 10,000 tricks that when you- To get

[00:19:11] Tommy Shaughnessy: there

[00:19:11] Tarun Chitra: to get them all together comp- they compound and give you the f- final product, versus like a one-shot thing that like… ‘Cause like the latter is like real needle in the haystack and often never exists in science and engineering. Yeah. Um, which is wh- what I think the early crypto AI stuff is like there’s a lot of hopium.

It, it, it, it, it, it, in some ways it’s a little bit, it reminded me of the 2017 ICO boom, right? Whereas like staple a token and c- take something that exists already and say it’s decentralized, like decentralize Uber. Like, you know.

[00:19:43] Tommy Shaughnessy: Well,

[00:19:43] Tarun Chitra: that- And so I thought, I think that’s the thing about Nous that stood out, was like they were actually trying to do the real thing by doing the hard work and not launching a token and not making- Yes

promises that were crazy.

[00:19:53] Tommy Shaughnessy: Yeah. No, I, I’m, I’m with you. I mean, it is kind of funny like mixing some hindsight with, with what we’re seeing today in terms of like the token side, the distributed network side, I’m pretty sure for both Nous and Prime is sort of abandoned, right? Like we’ve kind of from first principles figured out like, hey, that, that doesn’t work, right?

And that’s fine, right? They went through it, they tried it, and now it seems like there isn’t a fo- focus on, uh, decentralization or, or distribute- like distributed networks, but there’s a huge focus on the open source side. Um, and that seems like it’s going really well. I mean, Hermes has, you know, millions, maybe tens of millions, I’m not sure, of downloads around the world.

So I’m, I’m kinda curious like how you link the two worlds together, ’cause like you saw distributed centralized networks and then they were solving really hard problems, but now they’re just sort of kind of the champion for open source.

[00:20:51] Tarun Chitra: Yeah, so I think, you know, Nous obviously, one interesting thing to think about, right, is like in 2023 Everyone who’s building open weight models didn’t exist or existed the way like Nous did, which was like a GitHub and a group of people and like maybe a little bit of funding, not that much, right?

Um And arguably, like, the decentralized training people were competing with the Chinese model creation companies. And I think, you know, if you look at DeepSeek, you look at Minimax, you look at, at, uh, GLM, Z.ai, if you read the papers of, like, the model development, again, all of these things boil down to, like, there’s 10,000 possible tricks you can do, and you have…

You only– You can only pick 100 in your model. How do you pick the best 100 little tricks, like, little optimizations that you can compound to, to make your model faster or better or whatever? And I think the Chinese companies rightfully focused on data collection and, like, basically making their own Mercor-style stuff, except for much cheaper than in the US.

[00:22:03] Tommy Shaughnessy: Hmm.

[00:22:03] Tarun Chitra: And that allowed them to just, like, get better economies of scale, while also doing a lot of these optimizations. But their optimizations weren’t focused on, like, lower bandwidth or lower power requirement or make it decentralized. Their requirements were focused on, like, if we have shittier chips, how do we make it run faster?

And so it’s, it’s still- Yeah … a resource constraint, but it’s a different resource constraint. I mean, some of it’s

[00:22:24] Tommy Shaughnessy: pretty innovative, though. Like, the DeepSeek model one paper- Oh, no, for sure … was nuts. Like, mixture of expert- No, no, no, no … second rate hardware. Yeah.

[00:22:30] Tarun Chitra: For sure. No, no, no. I, I’m, I’m not at all knocking that.

I’m actually just saying, like, if we look back in history, when Nous and Prime Intellect and other people are doing decentralized training, was the same time these Chinese model companies were also doing, “Hey, we’re gonna build our own model.” And they both started with the resource constraints. Good. The decentralized training people started the resource constraint of, well, it’s decentralized, we don’t know how much comp- computational resource we’re getting, uh, and we have to pay for it, meter it, and attract it somehow.

So that’s, that’s, like, one type of resource constraint on training. And the, the, the Asian companies were like, “Well, we actually only get shittier chips and, and, like, resources and worse data centers and worse data. So given those constraints, what can we do?” Right? Like, and, and so the… If you look, both sides may have a different bag of tricks.

But out of the, you know, 10,000 possible optimizations, they both chose kind of, like, different subsets of 100, and the Chinese ones, their 100 was just better, right? Yeah. And, like, that’s why they got to, to building, you know, world-class models faster. But as we’ve learned from the kind of agentic world and the kind of recursive improvement stuff, once you have a good enough model, you just start increasing a lot faster and it becomes much harder for people to catch up to you.

And so I think what you saw was, like, a lot of the early infrastructure people, a lot of the decentralized training people, by 2024, 2025, uh- kind of starting to pivot out, right? Yeah. And that was also when we start- started to see the first versions of, like, trading agents and early AI agents. Now, those things, again, were a little bit just like meme coins with a lot

[00:24:22] Tommy Shaughnessy: more steps and a lot more replies on Twitter

[00:24:25] Tarun Chitra: Right, exactly.

Yeah. Uh, and, and, and, and in some ways, those probably set people back because, like, they were correct that agents were going to take over, but they were, like, the shittiest possible agents, right? ‘Cause, like, you fast-forward a year from when aixbt and ai16z were at the peak, that’s when Claude Code, Codex, et cetera, were just, like, having the huge growth curve, not just in users, but just, like, in capabilities and, like, stuff they could do.

Mm-hmm. Um, and so I think, again, crypto kind of right idea. Like, like in the same way decentralized training was the, like, right idea of, like, having resource constraints might make you make a better model or, like, a competitive model. Um, but you have to have the right ones. In the same way it’s like the AI agent thing, you have to kind of have the right thing and, like, you don’t need the token piece- Mm-hmm

in all cases. Well, it,

[00:25:16] Tommy Shaughnessy: it’s kind of awesome to kind of go through memory lane before we get to where we are today, ’cause it sounds like what you’re saying is, like, in crypto land, we had people really approach AI with well-meaning intentions. Like, they wanted, people wanted to train models, like Nous and Prime, but the technical choices they made just, like, weren’t competitive with the Chinese research at the time.

And then, like, , 2024 to 2025, like, we had AI agents, but they weren’t managing our lives or trading our portfolios. They were just, like, replying to us on Twitter. So it kind of seems like we’re getting close each time, but we’re not exactly on target with kind of what the, the big picture is. Yeah,

[00:25:58] Tarun Chitra: and I think the interesting thing is, like, people in crypto who, uh, pivoted, um, have e- into, like, non-token AI have done really well.

So, like, a great example of this being, you know, Nous, to some extent OpenRouter, Together. Together AI actually was gonna be, like, a decentralized inference-

[00:26:25] Tommy Shaughnessy: I

[00:26:25] Tarun Chitra: mean, CoreWeave’s- … uh, network when they started. CoreWeave- Well, one of, one of the ones- Yeah, but CoreWeave’s a little bit different in that they weren’t, like, trying to be a network, you know?

Like, they weren’t trying to launch a token. Like, like, like, the people who were trying to launch a token, I think the, the ones who pivoted had this kind of … Now, OpenRouter’s sort of weird. They were definitely not trying to launch a token, but, but in the sense of, like, philosophically, I think they came from this crypto view.

And again, OpenRouter is a great example of finding the right optimization can get you under resource constraints, can get you very good performance in some of these systems. And in some ways, that’s like, that is the same ethos that decentralized systems have, of like, I wanna have the highest capacity while having low resource usage.

Mm. Because the lower resource usage means I can have many more people participating, right? Um, and so I, I think there’s a lot of shared ethos in the Chinese open source model development and the decentralized crypto development. They just, like, kind of went different routes, and I think the, to be honest, the crypto people didn’t have as many

I would say the talent was not as good, and then also, like, they just chose the wrong op- optimizations, right? Now- Yeah … you look at someone like Nous, the interesting thing, you know, first era they were doing decentralized training. They’re gonna launch this, uh, thing on Solana. Um, and then, you know, they were kind of trying to figure out what to do.

They were doing a lot of research. They were doing a lot of private inference stuff, uh, working with kind of optimizing models for running in TEs. Um, which, you know, what, what you actually see, like, Near doing right now, Venice, stuff like that. Um- And sometime in 2026, you know, I think they saw OpenClaw kind of have this moment at the end of last year, early this year.

And one thing about OpenClaw is it’s, like, kind of the worst developer experience of, like, anything I’ve used personally. And, and I don’t mean this in, like, “Oh, really” … I’m not trying to be kind of an asshole ’cause I’m a news investor. I’m trying to basically say, like, OpenClaw feels like an Android because it’s like it makes you have to make too many decisions, like, to use it, right?

And it’s like if you, you look at the comparable centralized products of Codex, of Claude Code, of Muse now, of GrokBot, well, you know, like all, all, all these things. They all have this thing where, like, my grandmother who doesn’t know what AI is could use it, right? And, and OpenClaw was, like, decidedly the opposite.

It was like, tell me what inference agent you want. Tell me whether you want sparse attention or not sparse attention. Tell me like, um-

[00:29:04] Tommy Shaughnessy: Your reasoning level,

[00:29:05] Tarun Chitra: model choice, everything … you know, how much reasoning model choice- Effort … context choice stuff. Yeah. Right? It, it, it, it, it, it had this Android feeling of like you give me 10,000 options and yes, I could choose a better thing than the default, but like you kind of hamstring- Well, the-

who can actually use it.

[00:29:19] Tommy Shaughnessy: I mean, the crazy part… I mean, maybe, like, give you a quick example. I was building an agent on n8n, which is like the, you know, more complicated version of Zapier. Um, and that Zapier is where I built our first, uh, bot in 2024, 2023. Like, you come to Delphi’s site, it’d run through 20 steps, analyze, send us a message.

So I thought, hell, let me make a personal agent on n8n. So I was doing that and building it, and I remember, you know, deciding, you know, using Claude Haiku to decide whether to go to R1 reasoning on DeepSeek or a basic version to answer my emails, things like that. And OpenClaw came out the same weekend.

And I stopped using n8n, which is basically where you have to draw arrows between services and, and build, like, this manual process. Used OpenClaw and hated it and literally went back to n8n. It was just so clunky and hard to use that I just couldn’t, I couldn’t figure it out. I couldn’t, like, get excited about it.

[00:30:13] Tarun Chitra: Yeah. I, I think, like, the interesting thing is, like, OpenClaw in some ways is, like, the right idea for a product, but, like, the wrong execution on it. And, like, there is a sense in which someone had this tweet the other day that I kind of stuck in my head a little bit, which is, um-

Like in AI, it seems like second mover advantage is actually like there’s like a real thing to second mover advantage in that you learn a lot from the mistakes of the first mover. Um, and like you could argue that I would say like Claude getting a high uptake, you know, late 2025, early ’26, right? Where it like kind of eclipsed OpenAI.

Obviously, OpenAI has come back with a vengeance recently, um, was because it was just easier to use and it like had more human-like touches so that like the average person who d- is not technical could kind of make a product with it, right? Like, you know, there were like reporters and journalists making web apps and dashboards, right?

Without having to like know a lot of the technical language of like what to say to make the web app, right? Like you really had this no-code thing, which I think obviously OpenAI came back and caught up. But I think, you know, why did Anthropic have this opportunity is w- was like they learned from the lessons of OpenAI.

I think arguably you could say the same is true for the open weight models. I think ZAI and Kimi at like Minimax just like learned a lot from DeepSeek, where like DeepSeek did a really good job of like super optimizing like inference time, making the model small. Um, but where ZAI and Kimi kind of, um, did a better job was like changing how they’re thinking about reasoning, changing how they’re like doing resource usage- Mm

changing a lot of the kind of like attention, uh, uh, topology, stuff like that. And I think the second mover thing is kind of very real here in that OpenClaw had clearly the right market, right? Everyone was excited for this agent concept A- and again, crypto tried it a year earlier, but because they shoved the token in it- Mm-hmm

it, like, no one was able to use it.

[00:32:27] Tommy Shaughnessy: Yeah. The o- the other thing too about, I mean, like, I think OpenClaw deserves a lot of credit for, like, creating the category, right? Like- For sure … and, uh- For sure. I mean,

[00:32:36] Tarun Chitra: no doubt.

[00:32:36] Tommy Shaughnessy: Yeah, definitely, like, a lot of credit. I think the hard thing though is, like, you- you’d get, like, you get people in, like, this weird AI psychosis with it, where, like, they build stuff and they’d spend, like, 10 hours fixing it, and then they finally fix

[00:32:48] Tarun Chitra: it.

But no one else could use it.

[00:32:49] Tommy Shaughnessy: Exactly.

[00:32:50] Tarun Chitra: It couldn’t train. Yeah. That’s, like, the problem. Yeah. Like, they made something that was great for them, but then they, uh, then they, but they couldn’t ever get anyone else to use it. And so then no one else could be like, “Oh yes, this is awesome.” I remember when OpenClaw came out, I wanna say it was, like, December or January, uh, like end of ’25, early ’26, where I went to this meetup.

That’s right. Which is, like, an OpenClaw, like- Yeah … hackathon type thing. And it felt a little bit like going to, like, a crypto hackathon, where it was, like, people building their weird little project of, like, how do I, like, get my cat to, like-

[00:33:29] Tommy Shaughnessy: Love me …

[00:33:30] Tarun Chitra: you know, wh- when it eats food out of the cat dispenser-

it, like, sends me a, a USDC payment type. You know? Like, some, some, some weird-ass project that makes no sense and you’re like, “Does this have a market size of one?” Yes. It, it- And so the thing was, like, all these people who did these O- OpenClaw hacks, and they were, like, presenting them and whatever. Hm. But, like, almost all of them were the worst demos in history.

Like, none of them, their demos worked, first off. Secondly, everyone who was like, “Oh, I thought that person’s was cool,” like, “Can I, like, download the setup?” And it was like, “No. Actually, it’s-” “… like, too hard to fucking transfer it to you.” And so I think, like, the irony is, like, going to the OpenClaw events showed you a lot of the reasons, like, uh, well, what was, like, wrong with it from that perspective.

Um, and so I think Nous, you know, now, now coming back to Nous again, you know, kind of, they, especially Teknium in particular, ’cause it was just, like, his side project to start, uh- It’s nuts that it was his side project.

[00:34:31] Tommy Shaughnessy: Just insane …

[00:34:31] Tarun Chitra: he just, he was just like, “I love OpenClaw as a concept, but, like, it’s unusable.”

And because Nous had spent all this time building all this infrastructure for doing decentralized training, doing, uh, they’ve trained a bunch of models of different sizes. They understood, like, the difficulties in that. On the data side especially, right? ‘Cause, like, that’s the hardest- They’re

[00:34:56] Tommy Shaughnessy: models have tens of millions of downloads.

Yeah, it’s-

[00:34:58] Tarun Chitra: Right …

[00:34:58] Tommy Shaughnessy: no joke. Yeah.

[00:34:59] Tarun Chitra: But they, they, they have problems with, like, the how do they get more data, right? Like, decentralized data collection has always been much harder than centralized, and like, so for training, that’s, you know, one problem that inference doesn’t have. Okay. So then they were doing inference.

In the process, they learned a lot about the infrastructure for scaffolding around an inference engine, right? So it’s like- Yeah … I have an inference engine, how do I pass context between it? How do I share context? And then I think, you know, around the time OpenClaw came out was also when, like, Codex and Claude Code were really coming out.

And, you know, I think the thing people really got used to when those came out was this idea that I could have the same context shared between multiple models.

[00:35:39] Tommy Shaughnessy: Hmm.

[00:35:39] Tarun Chitra: And I think in the early days, everyone thought that was gonna end up being a form of lock-in for OpenAI and Anthropic, because they were like, “Oh, y- our models are gonna have, like, a certain format for context, and you can move between our models, but, like, we’re gonna lock you into that.

If you wanna move to an open model, it’s gonna be hard.” And as of yesterday, that’s completely obliterated, right? Right? Like, OpenAI now supports open models in the same context. So context sharing, all the stuff that, like, Nous, uh, and OpenClaw to some extent really- Mm-hmm … focused on of, like, having the same, the same context be able to move freely between models across- Yeah

multiple agents, across multiple models. Um, that stuff, uh, I think is just, like, interoperability is here, right? Like, uh, to put in, in blockchain-y language. Um, and so I think they, they were just, like, in the right position at the right time to be like, “We’ve built all these things. We know why they’re hard.

We know why training’s hard, data. We know why inference is hard, orchestration and figuring out how to, like, do context management.”

[00:36:48] Tommy Shaughnessy: Yeah.

[00:36:48] Tarun Chitra: And I think Taki was like, “The problem with OpenClaw is they don’t realize that because, like, they don’t do that much of their own inference. They’re not like… Like, it’s, like, built around giving the user all the options to figure out, to go figure it out yourself.”

Yeah. It’s like, “Here’s all the parts to this car. Go build the car.”

[00:37:05] Tommy Shaughnessy: True. Could you, um… I mean, uh, I, I’m glad you’re getting into it, ’cause I think people might on the surface say, “Hey, Hermes agent is just, you know, a similar kind of better version of OpenClaw.” But what you’re describing to me is, like, there’s real architectural kind of genius behind it that makes it a much better experience without, like, an easy trick or two.

Is, is that sort of the case? I think it’s

[00:37:30] Tarun Chitra: a- Or not? It’s, it’s just like an Apple versus Google thing, but, like- Yeah … in the inference era, right? Because- You know, a lot of the most successful, uh, you know, uh, the interesting thing is, like, all the crypto ideas that didn’t work have analogs in no-token land, like no crypto token land- Mm-hmm

that are huge companies and have worked a lot, right? So, like the rented GPU stuff, all these inference providers have been raising at, you know, decacorn valuations plus because they’re getting a ton of demand. But the problem is what, what all… So, like, the Maslow’s hierarchy of compute needs as a, as a, as I’ll call it is-

[00:38:12] Tommy Shaughnessy: I like where

[00:38:12] Tarun Chitra: this is going

you know, everywhere in the beginning was like, “Okay, rent a GPU.” Okay. Well, now I can rent the GPU, but then I have to go figure out how to set up the environment, I have to go set up the models, I have to go make sure I, like, manage memory. Oh, I want it to run inference really fast on ultra-fast mode. I have to go set up a ton of DevOps.

And so now it’s like, okay, I have to be a developer to start the GPU. And then there are companies like Together, Modal, Base10, Fireworks who are like, “We’ll give you a served endpoint for a particular model. We’ll, we’ll do all the optimization. We’ll, like, optimize the inference engine, all the sampling.”

And so it’s almost like they make an RPC for you. So the crypto version of this is like, one, I run my own node, that’s run a GPU. Two is I use a node that someone gives me an RPC endpoint to, like an Alchemy, right? That’s what Together and Base10, Modal… You know, you can view them as doing that. And then, but the problem is that’s still not that easy.

So, like compare the experience in crypto between using, like, FOMO or Phantom to, like, going to Etherscan and submitting a transaction on your own, right? Like, that’s more like what the RPC- It’s insane … right?

[00:39:25] Tommy Shaughnessy: It’s a good, good comparison. Yeah.

[00:39:26] Tarun Chitra: And, and, and so, like, clearly that, like, the TAM of the number of people who can use that, while bigger than the one that’s, like, run your own node, is, like, still not that big.

And the harnesses are truly, like, these, like, kind of wallets that manage your, a lot more state for you. They manage your history, they manage your preferences, they manage private keys and stuff. They, like, try to give you a ton of defaults so you don’t shoot yourself in the foot. And so I sort of think of, like, OpenClaw a little more like MetaMask or MyEtherWallet.

Like, it, it, it clearly has the right stuff, but it’s like im- it’s annoying to set up and it doesn’t give you the right defaults. Like, it still forces you to do a ton of stuff on your own. And I think the Hermes model, Claude Code, et cetera, is all about, like, giving you the right defaults, learning your behavior, right?

Like generating skills on, based on how you’re using it, and, and learning about you as a way of, like, having some memory and, like, preferences. But then also figuring out how to manage all the infrastructure of, like, which model should I use? How much inference should I spend on… How much should I spend on this model for tokens?

How much to spend on this model for tokens? And I think it, it sort of was built with the idea that a model router would exist. Hmm. So I think one problem for OpenClaw, it’s built under the assumption that the user will do their own routing. The user will choose their own model. The user will choose, like, what context they wanna share and not share.

Most

[00:40:56] Tommy Shaughnessy: people don’t know the difference, though, especially at scale. Like, people don’t know the difference between reasoning, not reasoning at the time. Th- yeah.

[00:41:01] Tarun Chitra: But they do, do th- know the, they do know the difference when the answer they get looks shitty, right? Like- For sure. Yeah … so,

[00:41:06] Tommy Shaughnessy: like- But then they just think OpenClaw is broken, I think, for, like, mass retail.

[00:41:09] Tarun Chitra: Yes. Exactly. Exactly. So s- but I think that’s why, you know, hey, the MetaMask thing is the same thing. Yeah. Um, uh, i- i- it’s, realistically, right? Like, the MetaMask thing is, like, the same, you know, it’s like, is it broken for retail? Why is it very universal, too? There’s a ton of FOMO users- Yeah … who could never use MetaMask, right?

That’s like, that’s just a fact. Um, and so I think that’s sort of, that’s sort of the way you should be viewing these. And so I think Nous was in the right place at the right time because they understood all of these problems one level below. Whereas OpenClaw was like, “We’re gonna shove it to the user.

That’s their problem.” And that difference in philosophy, while simple sounding and obvious, changes the architecture of how you design the harness, right? It changes, like, what decisions you unearth to the user, which things need to be their own sort of like models that learn something about the user, which things need to be asked to the user.

You know, like when you use any of these, um, bots, whether they’re the consumer ones like the, the Muse, um, or Grok Bot or the, or the, the productivity ones like, like Hermes, Claude Code, et cetera. Um, they all ask you questions, right? They’re like, “Hey, do you think this is the correct interpretation of this?”

Whatever, right? And it’s like learning something about a preference of yours. OpenClaw didn’t do that. It would just be like, “You should’ve written that in the config file. If you didn’t, fuck you,” right? Like- And I think, I think that kind of thing is, like, the difference between product engineering and, like- backend engineer, right?

And, like, the problem is OpenClaw, like, treats its users as backend engineers.

[00:42:49] Tommy Shaughnessy: That’s interesting. Okay. And then it, it is really cool to talk to you, ’cause you’re so deep technically, and you understand- Yeah … like, how things are actually built, and then a lot of the design decisions. I think a lot of people try Hermes and they say, “Hey, this just works.”

But from what you’re describing, the, “Hey, it just works,” is because of kind of 100 correct decisions on the product side and the backend side.

[00:43:14] Tarun Chitra: Yeah. I, I, I think, like, the interesting thing on Nous, right, is, like, they have this mix of, like, these technical people, but also this kind of, like, design, product side stuff.

And, like, you know, I think a lot of people on the internet made fun of them in the beginning for, like, oh, is this company- For being a merch

[00:43:28] Tommy Shaughnessy: company?

[00:43:29] Tarun Chitra: Yeah, exactly. Like, oh, is this company just, like, a very expensive merch company? They have great merch.

[00:43:33] Tommy Shaughnessy: Yeah.

[00:43:34] Tarun Chitra: But I think that also teaches you a lot about, like, aesthetics about your user base.

Like, you, you care about the aesthetics of your user base, and, like, what they’re, what they have to do.

[00:43:42] Tommy Shaughnessy: Right.

[00:43:44] Tarun Chitra: And I think, like, that’s something OpenClaw doesn’t care about, right? Like, or… A- and not just OpenClaw. I think, like, a lot of the harnesses ha- that, that were built at that era, and still a lot of the ones you see now, right?

Like, there’s this whole- Mm-hmm … DIY harness thing, right? Of, like, people trying to build their own, like, optimize it for, you know, my partic- It reminds me of people who, like, you know, spend two years customizing their text editor because, like, they need, like, every shortcut so that, like, when they’re coding it’s really fast.

Like, 10 years ago. You could have built something- Obviously now… Yeah. Yeah. A- and there’s, like, a DIY community for this harness stuff, right? And Pi has clearly taken over that community. Yeah. Everyone forks Pi and starts their own. But, um, I think what Hermes went for is, like, there’s gonna be people who want DIY.

There’s gonna be people who want enterprise. But then there’s people in the middle who are, like, prosumers, and, like, over time, more consumer. And, like, that philosophy is what I really liked about what Nous was doing, is, like, they started with prosumers, but they made it easy to use, and then they kind of widened the, the, the space.

And-

[00:44:45] Tommy Shaughnessy: For sure.

[00:44:46] Tarun Chitra: And- Around the time they were growing, sorry, I, I, was when OpenRouter started really growing. Um, like, the two of them kind of grew synergistically. Um, I, I, I, I, I sort of think of it as, like, the way Binance and Tron grew synergistically in some ways. But, um, but basically- The beauty of it was now there was public data, right?

You, you had a leaderboard, you could see which apps were generating the most tokens, you could see which models were being used. Uh, you had like almost like blockchain crypto data for the stuff that you did not have prior to Nous and OpenRouter growing. And that was when it was like, okay, this company is doing something really special.

And like, yes, I’ve invested before, but like clearly this is like, you know, probably gonna be the most important, you know, competitor to the centralized companies. Um, and, and, and in, in a sort of crypto-like way of like being the most credible competitor to normal finance.

[00:45:49] Tommy Shaughnessy: Yeah, no, I, I couldn’t agree more.

I, I mean, Tarun, I wanna, I wanna like double-click on where you’re, where you’re taking this, right? Like you mentioned like the consumer agents, there’s a lot of vendor-owned consumer agents, like Meta owns Muse, OpenAI owns OpenClaw, Instinct I guess is, is its own. There’s Grok X… You know, SpaceX, xAI owns GrokBot.

Nous is the only one that’s open source and not lab-owned. Um, do you think that that has any implications in terms of like what people will use it for, or the restrictions or, or like what they’ll build or like second-order effects of being open source? Because it’s the only one on that spectrum that is open source, not vendor-owned.

[00:46:33] Tarun Chitra: Yeah, I mean, there are, like I said, there is this whole DIY community, but again, it’s like, it’s a little more like OpenClaw-ish, where it’s like you have to pick up, pick everything yourself. They, they’re the only one that’s like open source, but packaged. So I, I think of it a little more like Linux distributions.

You know, like there’s the Linux kernel, there’s Red Hat, and then there’s people who build distributions on top. There’s Red Hat, Ubuntu, Gentoo, whatever, right? And realistically, I’m probably gonna get flamed for saying this by old heads, but Ubuntu is probably the only real Linux distribution that ever got like non- Non-develop, like technical user usage, right?

Realistically. ‘Cause it’s like, it, it, it is, and it’s because they did all this defaults, packaging, making it easy to update stuff, figuring out how to deal with, like, non-open source code interacting with open source code. So like, Hermes like is, is a good example of that, right? Where it’s like I can use Claude and OpenAI with an API key within Hermes.

Whereas a lot of the pure open source stuff is, like, extremely, like, militant of like, “No, no, no. We, we, we don’t wanna touch any of the closed…” You know, Ubuntu had the same philosophy. So I c- sort of view it a little bit like Linux distributions. However, I think one really cool thing, another thing prior to us doing this investment that we’re doing when we’re doing diligence, obviously the OpenRouter thing made it be like, “Oh, I’ve turned this into a blockchain data thing.”

And so, like, for me, as my background, that’s like, that made everything make sense. Yeah. But the, the second thing was, uh, like- I started just looking at GitHubs of people who fork Hermes. Like, who, if it’s open source, right, one way of measuring the community strength or people like the product is forks, right?

And to, to some of the other competitors’ credits, like I think Pi is probably the most successful from a pure DIY standpoint. Like, it completely eclipsed OpenClaw, like, like from the D- DIY community, um, because it’s very lightweight. I don’t think a lot of people know that, by

[00:48:34] Tommy Shaughnessy: the way.

[00:48:35] Tarun Chitra: Yeah, yeah, yeah. It’s, it…

I mean, Pi, Pi, you look at the, the usage forks commits, it’s like, it’s like s- surpassed. I think OpenClaw, the, the whole like OpenAI sale, pseudo sale thing, like that was like the death of them.

[00:48:46] Tommy Shaughnessy: It kind of feels like TBPN. Like, I don’t listen to TBPN since OpenAI bought them. But,

[00:48:50] Tarun Chitra: yeah. It, it, it, it’s somehow like they like sold themsel- sold too early, you know?

They- But Pi, Pi is really community-based, and I think the interesting thing about Pi is Pi probably has the most forks, um, because it’s like, you know, I… In fact, my, my, my brother’s a CS professor, and he was saying like in one of the like intro CS classes now they like fork Pi and modify it, and it’s like that’s like how they teach you how the harness model works, because Pi is like very simple.

It, it, it, it… But it, it, it has this problem of like, again, you have to build all the adapters to everything that you might use. Um, so I, I was like, okay, is Hermes open source stuff still sufficiently competitive for people who are DIY builders? You know, like, ’cause like I think that’s actually still important.

Like, their ethos should still be that. It shouldn’t just be, um, you know, “Hey, we’re an open source harness. We have a lot of users,” but then we’re actually like no one uses our open source code, because that sort of means you, something is wrong usually. So I started indexing a bunch of forks, looking through forks, and I started finding some really crazy and interesting forks, um, which ma- And, and these are not things Nous or Hermes will monetize, right?

But these are things of like brand recognition. People really care about how it’s been built, and they like how it’s been built. This is a way of measuring that, right? It’s almost like measuring, doing sentiment analysis about a brand via people who forked it versus like- That’s

[00:50:20] Tommy Shaughnessy: super interesting, by the

[00:50:21] Tarun Chitra: way

versus revenue. Yeah. Right? Like, I’m not, I’m not trying to assess like monetization on this. And I would find these people who are like, “I’m some random Chinese hardware robot manufacturer, and the interface to my like little robotics embedded hardware thing was like fork Hermes Right? So it’s like, I, I…

It wasn’t like a Unitree humanoid. They were like small robots, but like they don’t have money, right? Like they’re try- they have to be cheap. Like the robotics stuff is like they have to spend all the money on the hardware. Like they, the, they need to lower their inference cost to zero. That’s why they love open source models.

But for robotics, it’s, you know, e- especially, uh, you’ve seen a lot, in a lot of the, um, uh, new evals, it’s actually quite hard and competitive to have a single model be really good. In fact, one of the reasons I think the open source world is better is that I think the routers are much, and like the ability to move between multiple different models based on like specific tasks.

For robots, it seems very true that it’s like, okay, I need to use one model for planning. Like I have a robot that has a task: walk to the laundry room, open the laundry machine, and like fold the clothes. Open the dryer and fold some clothes, right? Like the folding clothes thing is always used as like the hard test for a humanoid robot, ’cause like no- we, we still don’t really have one that can do it generically, as crazy as that sounds.

Mm-hmm. Um, so you might have one model that’s actually really good for the like moving my feet, walking stuff. You might have one model that’s really good for planning of like, okay, what’s the layout of the space? Uh, which direction do I need to go? What path do I need to take? Uh, you might have another one that’s like for dexterity, for like, okay, to open the door I need like a specific amount of force.

And interesting thing is like these have different amounts of reasoning versus like amounts of like tool calls or like external actions needed. And a lot of the evals have shown that like segmenting to different models for different sub-tasks and then putting it together is like oftentimes a lot, giving you much better performance than a single model.

And- So the interesting thing is, like, how do they manage that? How does, how does the average user who’s gonna use this robot… You, you know, how do they figure out how to do it? And, like, these guys basically, a couple of these Chinese companies, like forked Hermes, basically had some modifications for, like, their…

You know, they, they obviously modified it to, like, connect to their device. They made tool calls for, like, move hand, move head, whatever, you know. But that to me is a sign of product, brand market, product market fit of your brand, right? Is like people want to use it. So for that type of stuff-

[00:53:02] Tommy Shaughnessy: And they’re not going to sell this robot comp- They’re not doing, like, door-to-door sales or

[00:53:05] Tarun Chitra: marketing.

No, no, no. It… I don’t e- I remember I told them, and they’re like, “We didn’t even tell this, this person for it.” Uh, so, like, like, my point though is, like, for open source projects, this, like, community sentiment, community… It, it’s something like a brand of, like, how much people use your stuff is actually very important.

Like, you know, I think oftentimes, especially in the current AI market, where first of all, half the AI numbers are fake, everyone is just so focused on, like, rushing to tell you monetization stories about companies. But there’s a sense in which retention and stickiness is much more, will be much more important, say, if there’s, like, a, a crash or downturn.

And this is, this is one of the things that as we’re looking at Nous is like, do they have this type of like really good brand with their user? And that, that was something we, we found a lot of and, and I think that kinda cemented the, the deal for us of like, you know, it’s like if not now, then when? You know, like it feels like this is like the, the rocket ship.

[00:54:09] Tommy Shaughnessy: No, I’m, I’m with you. I mean, Drew, maybe just to recap, like you’ve kinda walked us through why you like Hermes so much more than, than the competitors, and then like its design and the people using it, and these really interesting companies like basically proving PMF and, and social buy-in and like brand longevity, right?

But I guess maybe my more like pointed question for you is like looking to the next year or two, like where does it go, right? Like you have Muse, you have Instinct, you have a lot of these consumer agents. I personally don’t feel Nous is, is really competing with them or, or vice versa. I view Nous as like, or I view Hermes as like kinda what you described, like for pro builders building applications for the, the long term and really wanting to own it.

And maybe like my second order of thought is I really feel like people wanna use Hermes for two reasons, and that’s because if you own the architecture around Hermes, everything you build, the data, the memory, where you’re hosting it, the interactions, everything, that’s obviously an asset for you or your business.

But also you get like actually unique outcomes, like in terms of like when you ask the model something. So like I feel like there’s a lot more to the open source side than people are giving it credit for, but I’m not sure if you, you agree given what we’ve seen. Yeah. Uh, so I’ll tell you

[00:55:29] Tarun Chitra: what, you know, Nous, uh, you know, has been working towards based on s- pub- stuff they’ve said publicly so far at least.

Um, and then I’ll tell you what I think, where I see, uh, the dream kind of going and also where I think, uh, the crypto, you know, kind of we can- kind of a You know, uh, come back to crypto. You know, so far we’ve- Yeah … been like crypto pe- people have fucked up. They’re wrong every time on the Sei stuff, and, like, everything they’ve done has just like-

[00:56:06] Tommy Shaughnessy: We can also leave crypto behind on this one.

I mean, there, there is a-

[00:56:10] Tarun Chitra: We can. Yeah But I think, I think there’s actually a very real reason to not do it. Um-

[00:56:14] Tommy Shaughnessy: Interesting. Okay, cool. I’m eager to get to it.

[00:56:17] Tarun Chitra: Um, and so the, uh … You know, I think Nous has obviously been focusing a lot on, um- increasing sort of, uh, their enterprise usage. So like, you know, like the nice thing about Hermes being open source is like enterprises can clone it, customize it, use it internally, like add their own privacy layer, add their own compliance layer, add in permissioning, um, connect it to their databases.

And Nous has really been focusing on that. I think they have a bunch of big news coming up on that soon. Um, the– And, and that’s been going really well, I think, uh, you know, in this- And, and true, sorry to- Yeah, go ahead …

[00:57:03] Tommy Shaughnessy: sorry to interrupt you, but also, like, when you have millions of users, you have advocates internally at every company who want to bring this to their whole enterprise, right?

Yes. So you already, you already have like the GTM sort of in the bag.

[00:57:18] Tarun Chitra: You know, nothing’s in the bag until the contract’s signed, but yeah.

[00:57:22] Tommy Shaughnessy: Fair. Fair.

[00:57:23] Tarun Chitra: Fair.

[00:57:23] Tommy Shaughnessy: A true

[00:57:24] Tarun Chitra: OG. But, uh- Yeah. But, uh, but, uh, like, yeah, I think, I think that there’s certainly that. I mean, they have like… Realistically, they’re in the, like, tens of millions of users, so like they’re– In, in terms of like…

I think, if I remember correctly, it’s like tens of millions of unique GitHub accounts access them daily. So if you use GitHub as your civil resistance out, it’s not perfect. Like, obviously people can have tons of GitHub accounts, but GitHub has made it expensive to have multiple GitHub accounts versus most social media accounts.

So they have a ton of users coming, um, from everywhere. I think obviously, A, their subscription model with Portal has been growing, um, as well as their OpenRouter usage. I think they just passed 2 trillion, um, as a number on apps there. Uh, they are also, um, you know, have been working really closely with NVIDIA.

Like in this round, NVIDIA and Microsoft both participated. Um, I think enterprises are going to access Nous. Nous is gonna be like one of the main open source, um, harnesses you can use, um, at Microsoft. Um, and NVIDIA has installed it on their hardware by default for DGX Sparks machines. Um, so they’ve been actually kind of trying to play both parts of the market, like the consumer open source closer to the, the ground model, which they’ve obviously excelled at, as well as expand enterprise.

Now, um- The interesting thing, uh, a- and so like the, you know, the, the, the thesis we had, you know, kind of for investing beyond the fact that they’re growing and like They clearly have this kind of, like, developer market, brand market fit. Um, is one thing I got from scraping all this OpenRouter data in the process of due diligence for, like, three or four months, um, was that I started really analyzing how the OpenRouter market itself started to work, and it really looked

And I, I wrote this post about this, uh, like a month ago. But it was sort of like a culmination of work since April or May. Um, and basically I’d analyzed and found, you know, the way OpenRouter works is, like, there’s applications. The applications say, “I want access to a model.” There are many providers, almost like blockchain operators or intent operators in blockchains, who, who quote a price that is like, “Here’s how much I’m gonna charge to run this model for a million tokens on input tokens, cache tokens, and output tokens.”

And then the router matches you to a provider. So it’s like node operators basically say, “Here’s how much I’ll sell you access to, to run this model.” And then, um, you know, OpenRouter does the matching. So you can kind of think of this as, like, OpenRouter’s a little bit like a dex router or crypto exchange.

The harness or application is like a wallet. It stores the user’s private data, it stores the user’s API keys in, in their tool calls. It stores potentially private secret keys, like if they’re trading. Um, and then you can view the model as a little bit like a protocol. Like, it’s like a open source thing that has a license.

In fact, a lot of the open source models now have, like, BSL style licenses where, like, if you make more than X revenue, you have to, like, give them a licensing fee. Kimi has

[01:00:55] Tommy Shaughnessy: that, I think.

[01:00:56] Tarun Chitra: Which is- Yeah. Yeah, exactly. No, I think all of them. GLM as well. They all do? The- they’re … GLM also added something similar.

Not, not the same amount. Kimi’s is much lower. Like, it’s, like, a very low cap of how much you can make. Um, but this is very much like the crypto licenses of, like, if you fork, if you fork me, you have to pay some fee. Um, and then the inference provider is a little bit like a node operator or validator or a liquidity provider.

They kind of are doing both functions. And if we take this analogy a little more seriously and, you know, the data kind of shows a lot of things that look very similar, uh, what phenomena can you expect? Well, one phenomena you can expect is MEV in crypto. Like, people trying to front-run, people trying to manipulate the price, uh, and most of my post is about finding some examples of that and just to show that, like, hey, a lot of this, in any open marketplace, whether it’s decentralized or not, it just has to be open, like open entry.

Um-

[01:01:57] Tommy Shaughnessy: Tarun, sorry to interrupt, but what is MEV with, when you’re asking for AI inference? Because, like, in crypto you’re, you’re competing to, to make money, solve a block faster, fill a transaction, but In AI, like if I beat you, yeah, sorry, just want to understand that.

[01:02:09] Tarun Chitra: No, no. Yeah, for sure. So basically, um, remember I said there are three prices that, uh, an infrastructure quotes you?

There’s a input token price, there’s a output token price, and there’s a cache token price. Um, let’s just pretend the input and output are the same. They oftentimes differ, but like, uh, there, it’s a marginal difference usually. Um, but the cache token is usually hundreds or thousands of times cheaper. Why?

Because, like, the infrastructure is using way less compute, ’cause, like, cache tokens are, like, saved, so they don’t have to redo a whole calculation. And so an interesting question is, like, how does Open Router allocate you to a node provider? They don’t just give you the lowest price one. What they do is they take, they give you a, a, they, they, they give a probability of each provider being chosen based on the price that they offer, and then they assign you t- based on that probability.

Now, that probability also includes some type of like, uh, uh, quality score. Like, have they had good uptime? Have they, you know, delivered more than X tokens per second? Like, so that part is like Open Router grades them. That’s almost like a credit score for each infrastructure. But then they have a price.

So it’s like… And so one question you might, one thing I was kind of looking for was, like, are there people who are lying about how much they’re giving you in cache tokens? So they, they tell you there’s a really low price, and then they get a bun- so that increases their probability. But then they say they never cache tokens, ’cause there’s no cryptographic verification style thing.

I can’t verify, like, they just say, “This is how many tokens I cached.”

[01:03:48] Tommy Shaughnessy: And, and true, and sorry, if they’re not providing you the, the amount of cache tokens they’re quoting you, your level of service goes down in terms of, of how? Like, the memory or, or your response or- Uh, your

[01:03:58] Tarun Chitra: level of service goes down in, in the price of it.

So I’ll give you just a, a similar… Let’s suppose I have an agentic workflow that, like, involves fi- the age- you know, you ask a prompt, the agent asks another prompt, agent asks another prompt, agent asks another prompt. So there’s five prompts. So as an example, in GLM 5.2, there was ZAI, let’s say that they were charging $1 per million tokens, and the cache tokens are like .001 cent.

Um, there was a, a provi- and that’s done by ZAI. Like, the model creator usually runs an infrastructure. And, and you can treat the model creators, one, as sort of the gold standard. Like, they’re not gonna lie about their caching because then everyone’s just gonna be like, “Your model’s shitty.” Like, it, it’s a brand thing for them.

Um- So, um, basically, uh, you say, okay, the true optimal or, like, close to it caching rate is, like, whatever the model provider tells you. Okay, now suppose there’s someone who’s saying, “I’m gonna offer you five cents per million tokens,” so a 95% discount, right?

[01:05:03] Tommy Shaughnessy: Okay.

[01:05:04] Tarun Chitra: Uh, a- a- and actually imagine there’s now, like, 21 queries, so like, you know, I make a prompt, agent asks another prompt, agent asks another prompt, and that happens 21 times.

On the ZAI model, it’ll charge me… Imagine that the, the, the caching is perfect. It’ll charge me $1 for the first query, and then the agent’s repeated queries are, like, .001 cent. Very small. So it’ll cost me basically a dollar. Like, under a dollar and 1 cent. On the other hand, if the person who is charging five cents says five cents, but then they lie about caching, then I get charged five cents on each query.

So now I get charged- Okay … five cents 21 times, so I get charged a dollar and five cents. So, so the idea is, like, by lying about how much I cache, I can charge five cents on each query. Yeah. Whereas if I, the person who’s caching, they charge me a dollar at first, but then basically zero on the repeated one- rounds.

Yeah. And so-

[01:05:57] Tommy Shaughnessy: Okay …

[01:05:57] Tarun Chitra: the question is, like- Interesting … but why, why would I put five cents as my price? To get a higher probability of being chosen.

[01:06:05] Tommy Shaughnessy: Yeah.

[01:06:07] Tarun Chitra: Yeah. So it, it’s, it’s sort- I understand … it’s sort of like a, it’s not exactly analogous- Yeah … to, like, a front-running or j- Right … like, making too much of the analogy to MEV is, is wrong, but it is a manipulation.

Like, they’re ma-

[01:06:18] Tommy Shaughnessy: No, no, there, there’s a way you can game it. There’s a way you could

[01:06:20] Tarun Chitra: attract more- Yes, exactly …

[01:06:21] Tommy Shaughnessy: inference to provide a bad service, for sure. Your way

[01:06:23] Tarun Chitra: of providing bad service- Yeah, and there’s a way to prove it really, yeah … is like, you’re probably caching, but you’re not passing on the price benefits of caching.

That’s the way of looking at it. Yes. Like, it’s like- Yes, yes, yes. And the idea is, like, you’ll lose money on the first query, but then if the agent asks a ton of sub- subsequent queries, then, like, you, that’s where you make money. And so I found statistical evidence on it. Like, I ran these experiments- Interesting

a- and, like, tried to make sure I went through all the different providers. Um, and, and basically found there were some providers doing this. So how does this relate back, though, to, like,

[01:06:56] Tommy Shaughnessy: I guess using an

[01:06:58] Tarun Chitra: agent or, or the inference side? Yeah, so, so, so what this made me realize was like- Like,

[01:07:01] Tommy Shaughnessy: back to the Hermes side?

[01:07:02] Tarun Chitra: if th- this analogy to this crypto thing of, like, wallet, front end harness, dex router, kind of like exchange/dex aggregator, model protocol, inference provider, liquidity provider. If I look at crypto from 2021 to now, all the money went to either the top, the wallet or the harness, or to the validator/LP, right?

Like, everyone in the middle- … kinda got squeezed, right? Yeah. Like, the dex aggregators got squeezed. The models can’t collect as much fees, right? It’s like FOMO makes more fees than Uniswap, right? Like, why, right? Like, it, it, it’s like the front end is, is able to capture more than the protocol, right? And so my point with this is, like-

If you think this looks like crypto, and I think any open market where people can, like, join and leave, like these infrastructure can join. I mean, OpenRouter is a quality control mechanism. Like, they, they can, like, not let some people in, but they’re … So in that sense, it’s not decentralized. Like, I can’t just join with my own node.

But it- once I’m on there- Yeah … it’s like an open market competition, right? And, and that, that aspect- Yeah … is very different than- For sure … the centralized, like, OpenAI or Claude, where you don’t even know who the provider is. Like, they might be using CoreWeave for some of your stuff. They might be using-

[01:08:26] Tommy Shaughnessy: People just wake up and they say, “The model’s been nerfed this week,” or they don’t.

Like, they just dynamically could turn off-

[01:08:31] Tarun Chitra: So-

[01:08:31] Tommy Shaughnessy: Yeah.

[01:08:32] Tarun Chitra: Scale up, scale down … I think this open source stuff, especially in the routing model where there’s, like, multiple models used as well as multiple providers, and potentially multiple apps, where when I say app, I mean, like, a Hermes or an Instinct or a whatever.

I think that market structure looks like crypto. All the money goes to the top and the bottom. Mm-hmm. What is the top? The harnesses. What’s the bottom? Inference providers. And, like, it’s kind of borne out in the last six months where, like, they are- Interesting. Yeah, I think, like, OpenRouter really sold the top because, like, I do think the margins are gonna get squeezed there, um, ’cause, like, the router is, like, a very competitive market.

Like, rever- like, every SaaS company is making a router right now, right? Vercel has a marketplace, Ramp has a market- So, my point is it’s gonna be … The router is, like, hard to kind of like … Like, it, you can be a s- a successful business, you just, like, it, it’s, there’s low barriers to entry. Yeah. Right? So it’s like, um

So I guess my, my long shot point here is from that model, Hermes is, like, investing in Phantom or Fomo, right? That’s sort of, that’s sort of the, the game plan.

[01:09:39] Tommy Shaughnessy: I’ve never heard someone say it like that.

[01:09:41] Tarun Chitra: And- It’s

[01:09:41] Tommy Shaughnessy: interesting. Yeah.

[01:09:42] Tarun Chitra: And- It’s a good way to think about it … I think a, a difference between AI and crypto is in, in, to some extent, i- in crypto, the wallet is, like, the final end state.

But actually, the interesting thing about something like Hermes is you could have people build Instinct on top of Hermes, and then inherit a lot of the security properties of Hermes. And my guess is actually people are going to, like, view the app harness layer as, like, a security and brand, like, context management layer.

And so people will start building apps like an Instinct, you know, like, like different types of bots that exist on top of a harness that provide you the consumer experience, but then the back end development is done via these harnesses. And so in that sense, it’s not like the end point like a wallet, right?

The wallet is a end point ’cause it goes to the final end user, whereas, like, I actually … You’re already starting to see people, like, build stuff on top of these. Um, and so I think that is gonna be a very interesting game to see. The other thing that’s really interesting is the harnesses, the inference providers are also already doing kind of, like, payment for order flow type of stuff, where it’s like, if you guarantee me X billion tokens a day, I’ll give you a rebate or I’ll pay you back some of the fee.

And so in some sense this token, being able to aggregate token order flow is like an extremely valuable market position, just like in the crypto front end market.

[01:11:11] Tommy Shaughnessy: For sure. No, that is, that is interesting. So, I, I mean, Hermes though, it, like Nous isn’t directing that usage though, right? It’s each individual person kind of making their own decision where they wanna send the

[01:11:21] Tarun Chitra: inference to, or- When it goes on Open Router, but when they’re using Portal, when they’re using another

Like, the point is, like, you can use many different … You could just pick your own provider, right? Like, you could just be like have a contract with Together or something and then send, send directly. Like, the, the point is that these open models allow you to unbundle Codex or, or Claude Code, right? Like where, when you use Claude Code, you’re getting all of these features.

Like implicitly they’re using some router type of thing, or at least model selector. Um, implicitly they’re using different providers, right? Like, you don’t know, you know, they’re doing all these long-term data center contracts. You’re getting run on one of them, but you don’t know which one you’re getting run on, and I’m sure you’re getting migrated.

It’s not like why would they lock you to one particular provider? It doesn’t make sense for them either.

[01:12:07] Tommy Shaughnessy: Tarun, I know we have a hard stop in, in two minutes. Yeah. So I just wanna ask you in closing, like, a, a hard question, but, like, what do you think the world looks like with Hermes in a year? You mentioned enterprise, you mentioned retail.

It’s a lot of technical stuff. Like, I don’t know. Like, how do you envision, like, what the world’s like, how people are interacting with it, how they’re using it kind of a year from now?

[01:12:29] Tarun Chitra: There are lot of people who are building, who have existing apps and are adding in agentic functionality, like Robinhood a- You know, you’ve tweeted a lot about the Robinhood agentic experience not being as good.

Maybe it, maybe it, maybe it’s changed after yesterday or whatever. Well, use it anymore. Yeah. My belief is that every SaaS company either has to have a harness or be a router. Mm-hmm. Because they need to get into the token order flow. Otherwise, they’re like f- they’re, they’re kind of like, they’re kind of fucked if they’re not- Interesting

in the token order flow game, because, like, that’s, that’s the easiest thing to monetize. And in that world, Hermes is the white label harness for anyone, right? Like, every company that … No one wants to start from zero. You have a security and, and brand risk thing. Like, if your harness has been hacked, no one’s gonna use it.

A lot like a wallet. And so I think, like, being able to have this really strong brand means other people will want to, to use you as their backend beyond the prosumer. Mm. And I think that’s, that’s, that’s kind of where I, I see the, the growth. And, like, the enterprise usage is one. The other might be, like, consumer usage.

The other might be this kind of, like, apps that build on top of Hermes because, like, they don’t wanna start from scratch.

[01:13:44] Tommy Shaughnessy: So true, and if every company becomes a harness, and Hermes is the best harness to build on, w- why does every company wanna become a harness, though? Or, or

[01:13:51] Tarun Chitra: yeah, why do they all need one?

I mean, they bring users and data, they get a rev share from someone who’s managing the token order flow for them. It, it, it, it, it’s really like a fintech and a bank, right? Yes. A fintech … Most fintechs don’t have banking licenses, so they just do rev shares with the bank underneath.

[01:14:08] Tommy Shaughnessy: Love it. It, it, Tarun, you took us down such a, a unique and, like, differentiated view of Hermes today. Like- I could go, I, I think this is where it,

[01:14:15] Tarun Chitra: like, looks

[01:14:15] Tommy Shaughnessy: like crypto. A lot of people just kinda, just, yeah. It’s like… And, and, and another thing- No, it’s cool. No, it’s interesting, ’cause a lot of people just talk about, like, what they’re

[01:14:21] Tarun Chitra: building,

[01:14:22] Tommy Shaughnessy: but-

[01:14:22] Tarun Chitra: The last thing I’ll end on, which is why I think this crypto thing is true, Muse is going to monetize in a very crypto way.

It’s charging transaction fees on stuff you buy in the, uh, app. Mm-hmm. And it’s basically doing that by doing rev shares with Stripe and Shopify and whatever. How is that any different than this world I just presented? Not really, right? So, so I think a lot of the ways order flow management is done in crypto, in any open system where there’s competing counterparts, will end up looking very similar to what crypto has.

And by trying to take that, the lessons you’ve learned from there, you can make some pretty hopefully useful hypotheses.

[01:15:00] Tommy Shaughnessy: I love it. Tarun, thank you for, for coming on. I mean, huge congrats on leading this news round. We’re, we’re co-investors with you. We’re so excited. Uh, I’m also just excited for humanity, ’cause this is, like, open source and where I want the world to, to go and to be.

Um, and your, your comments were just so unique. Uh, people come on and they talk about, like, what they’re building or, or what they wanna do. Like, you brought us down the eras of crypto, like, why this is so interesting, like the differentiated build on the dev side, the app side, like the relationship to crypto and, like, how it could be related.

So this was really unique. Definitely. Thanks so much. So I appreciate your time.