Why OpenAI Should Be Scared of Open Source
- Robot Ventures led Nous Research’s $90 million round on a thesis that Hermes can own AI’s user relationship. Tarun Chitra compares the opportunity to investing in Phantom or FOMO: harnesses control preferences, context and access to users, while inference providers supply execution. If AI develops crypto’s market structure, he expects those two layers to capture value while routers and models face pressure.
- Hermes’s product advantage comes from years of research compounding into better defaults. Tarun argues that Nous’s work on training, inference and context management lets it absorb decisions that OpenClaw leaves to users. His OpenClaw analogy: “Here’s all the parts to this car. Go build the car.” Hermes instead learns preferences, generates skills and supports work across models.
- The useful inheritance from crypto is resource-conscious engineering, with token-first projects supplying the cautionary history. Tarun credits Chinese model companies with choosing better optimizations than decentralized-training teams, prompting pivots in 2024–2025. Early token agents anticipated demand but delivered weak products—often “meme coins with more steps.”
- Unsolicited robotics adoption was a stronger diligence signal than another monetization story. Over three to four months, Tarun examined usage and forks, finding Chinese manufacturers embedding Hermes in small robots without Nous knowing. He treats those forks as evidence of product quality and brand attachment; “retention and stickiness” could matter more in a downturn, even when individual forks generate no revenue.
- Enterprise customization gives Hermes a second distribution path alongside its consumer community. Businesses can add permissions, compliance, privacy and internal data connections. Tarun says Microsoft and NVIDIA participated in the round; he expects Hermes to be among the open-source harnesses available through Microsoft and says NVIDIA installs it by default on DGX Spark machines. Tommy sees an owned AI stack as a business asset, but Tarun checks his enthusiasm about ready-made enterprise demand: “nothing’s in the bag until the contract’s signed.”
- A low advertised inference price can conceal a higher total agent bill. In Tarun’s GLM 5.2 illustration, a nominal 95% discount becomes $1.05 over 21 queries when caching savings disappear, versus just over $1 from the higher-priced provider with perfect caching. He reports statistical evidence of providers gaming this mechanism: “you’re probably caching, but you’re not passing on the price benefits of caching.”
- The longer-term opportunity is to supply the harness underneath other companies’ agents. Tarun calls Hermes the “white label harness for anyone,” with security, context management and brand credibility supporting adoption. His categorical belief is that every SaaS company needs a harness or router to participate in token order flow, where aggregated demand can earn rebates and revenue shares.
1. Crypto’s AI survivors learned to leave tokens behind
Tommy introduces Robot Ventures’ lead in Nous Research’s $90 million round and says Delphi is a co-investor. Tarun’s broader claim is that crypto participants’ “fingerprints” are everywhere in the recent open-source and open-weight surge, particularly among teams applying engineering lessons without attaching a token.
The initial 2023 pitch often assumed mining infrastructure could become an AI network with minimal adaptation. Tarun argues that language models introduced difficulties image models did not, and that no one seemed to want a decentralized image model. Tokens also let projects attract valuations before demonstrating utility. His description—“meme coins with more steps”—includes well-meaning teams whose ideas simply did not work.
Tarun’s own position changed: he once considered decentralized systems scams, then spent four or five years understanding why blockchains worked despite conventional distributed-systems objections. That left room for a roughly “1% chance” that genuine architectural advances could overcome problems in decentralized training. He was more optimistic about inference than training.
Nous stood out through specific research: YaRN for longer context, compression-oriented embeddings, and low-bandwidth training through DisTrO. Tommy adds that Nous did synthetic model evaluation early and recalls a distributed training run of “40 billion parameters or something like that”; the number remains uncertain. Tarun’s criterion was disciplined work on bounded problems, with no token and no promise that one invention would solve everything. He also distinguishes Nous from Prime Intellect, which he says was less focused on training and hedged by becoming a GPU reseller.
Tarun’s broader reasoning is that distributed systems succeed through many separate optimizations, not one breakthrough. He uses Bitcoin as an example: hash security, difficulty adjustment, inflation and bandwidth each pose distinct technical problems that must work together. He says Nous and, in many ways, Prime Intellect were unusual in bringing that breadth of focus to decentralized AI.
2. Chinese model companies chose the more productive constraints
Tarun compares decentralized-training teams with Chinese model developers starting around the same period. Both faced resource constraints, but of different kinds: uncertain participation, bandwidth and payment on one side; weaker chips and infrastructure on the other. Chinese companies also built cheaper, Mercor-style data-collection operations. Out of thousands of possible optimizations, “the Chinese companies’ 100 were just better.”
Tommy pushes back that DeepSeek’s work was genuinely innovative. Tarun agrees: his argument concerns which constraints produced competitive results. Once a model becomes good enough, agentic and recursive improvement can accelerate progress, making catch-up harder. That helps explain why decentralized infrastructure teams began pivoting in 2024–2025.
Early crypto agents repeated the pattern: directionally right about agents, weak on execution. A year after their peak, Claude Code and Codex were expanding in capability and usage. Tarun points to Nous and, philosophically, OpenRouter as examples of crypto’s optimization ethos finding a place in non-token AI. He says Together AI started as a decentralized inference network, while distinguishing Corby’s approach from teams trying to launch a token.
3. OpenClaw demonstrated demand while exposing an execution gap
Tarun calls OpenClaw “the right idea for a product” with the wrong execution. Its Android-like abundance of choices—models, reasoning, context and infrastructure—restricted who could use it. Tommy corroborates the experience: after building a personal agent in n8n, he tried OpenClaw and returned to manually connecting services because it felt so clunky.
Tommy insists OpenClaw deserves credit for creating the category; Tarun agrees. The failure was turning personal experiments into transferable products. At an OpenClaw event around December or January, late 2025 or early 2026, Tarun saw unreliable demos whose configurations others could not readily download. One example was a cat-feeder project that would send a USDC payment. His test was blunt: “does this have a market size of one? Yes.”
That creates room for second movers. Tarun attributes Claude’s late-2025/early-2026 uptake partly to accessibility and human touches that let nontechnical people build applications, while acknowledging OpenAI subsequently “came back with a vengeance.” He sees a parallel in Z.ai, Kimi and MiniMax learning from DeepSeek and changing reasoning, resource usage and attention designs.
4. Hermes’s advantage starts beneath the interface
Hermes began as Teknium’s side project, prompted by liking OpenClaw’s concept while finding it unusable. Tarun argues Nous’s previous work supplied the foundation: training exposed data-collection problems; inference exposed orchestration and context-management problems. Tommy notes that Nous’s models had tens of millions of downloads. Those research years equipped the team to understand failures underneath an agent’s interface.
Context portability is central. Tarun recalls expectations that OpenAI and Anthropic would create lock-in through proprietary context formats. He says OpenAI now supports open models in the same context, undermining that expectation. Nous had already concentrated on sharing context across models and agents: “interoperability is here,” in his blockchain terminology.
Tarun’s explanation is cumulative rather than dependent on one architectural breakthrough. Successful distributed systems require “1,000 little optimizations”; he also describes the need for “10,000 tricks” that compound. Hermes benefits from understanding the layer below the product. OpenClaw’s contrasting philosophy was to hand users the components: “Here are all the parts to this car. Go build the car.”
5. The harness becomes AI’s wallet and preference engine
Tarun’s “Maslow’s hierarchy of compute needs” starts with renting a GPU, which still leaves environment setup, memory management and DevOps. Together, Modal, Baseten and Fireworks move users toward served model endpoints. His crypto mapping is running your own node versus receiving an RPC endpoint: useful abstraction, but still some distance from an accessible consumer product.
A harness is closer to a wallet such as Phantom or FOMO: it manages state, history, preferences and sensitive credentials, with defaults that reduce mistakes. Tarun puts OpenClaw nearer MetaMask or MyEtherWallet because setup still demands substantial user judgment. The distinction determines the reachable audience, just as using a wallet differs from manually submitting a transaction through Etherscan.
Hermes’s approach includes learning behavior, generating skills from usage and managing model and inference-spending decisions. Tarun says it was built around the idea that a model router would exist. Tommy’s objection is practical: most users cannot distinguish reasoning settings or model choices, but recognize a bad answer—and then conclude the agent itself is broken.
The philosophical difference changes architecture: which choices reach the user, which preferences are learned, and when clarification is necessary. Tarun calls this “the difference between product engineering and back-end engineering.” Nous’s design sensibility matters alongside its technical depth; the team once mocked as an expensive merchandise company had already shown attention to its audience’s tastes.
6. Packaged open source expands the reachable market
Tommy describes Hermes as the open-source, non-lab-owned alternative within the consumer-agent set. Tarun qualifies the distinction: DIY competitors exist, but Hermes is “open source but packaged.” His analogy is Ubuntu’s defaults, updates and accommodation of closed-source components. Likewise, Hermes accepts Claude and OpenAI API keys rather than requiring ideological purity about the models underneath.
Tarun credits Pi with taking over the DIY community and says it “probably has the most forks.” Its simplicity makes it useful for teaching harness design—including in an introductory computer-science class his brother described—but builders must supply adapters themselves. Hermes targets prosumers and progressively broader users while retaining extensibility. Tarun still wanted evidence developers used its source: an open-source label without actual reuse would concern him.
7. Robotics forks supplied the strongest diligence signal
During three to four months of diligence, Tarun combined OpenRouter usage data with inspection of Hermes forks. Public application and model leaderboards made analysis resemble blockchain data research. The forks answered another question: did builders value how Hermes was constructed? He treats this as brand sentiment expressed through engineering choices, distinct from revenue or immediate monetization potential.
His strongest example was small Chinese robotics manufacturers using Hermes as an embedded interface. They had to keep costs down because hardware consumed their budgets, and wanted to minimize inference costs. A robot approaching a dryer and folding clothes might need separate models for movement, spatial planning and dexterity, with different reasoning and action requirements. Tarun says evaluations have shown that separating tasks across models often performs better than relying on one model.
Manufacturers adapted Hermes with device-specific tool calls such as moving a hand or head. Nous had not even known some existed: “We didn’t even know this person forked it.” Tarun saw unsolicited adoption as product and brand fit. Amid what he says is a market where “half of the ARR numbers are fake,” this attachment mattered because “retention and stickiness” could become more important during a downturn.
8. Enterprise expansion adds customization and distribution
Tommy’s ownership thesis is that an organization’s architecture, data, memory, hosting and interactions accumulate as an asset, while customization can produce distinctive outcomes. Tarun describes Nous’s enterprise work in concrete terms: businesses can clone Hermes, connect databases, and add privacy, compliance and permissioning layers. They retain a common foundation while adapting it to internal requirements.
Tommy argues a large consumer community supplies advocates inside companies and effectively puts go-to-market “in the bag.” Tarun immediately qualifies that: “nothing’s in the bag until the contract’s signed.” He agrees internal champions help, while maintaining the distinction between widespread individual adoption and closed enterprise business. Nous is pursuing enterprise expansion alongside its consumer and open-source community.
Tarun reports growing Portal subscriptions and OpenRouter usage, saying Nous had passed 2 trillion tokens across apps on OpenRouter. He also describes tens of millions of users and, “if I remember correctly,” tens of millions of unique GitHub accounts accessing them daily. He acknowledges multiple accounts are possible and presents GitHub identity as an imperfect measure.
Tarun says Microsoft and NVIDIA both participated in the round. He expects Hermes to be one of the main open-source harnesses available through Microsoft, and says NVIDIA installs it by default on DGX Spark machines. These relationships support the enterprise push while the community continues supplying adoption from the consumer side.
9. Cheap inference can become expensive when caching is gamed
OpenRouter connects applications requesting models with providers quoting input, output and cached-token prices. Tarun maps the application to a wallet, the router to an exchange or DEX aggregator, the model to a protocol, and the provider to a validator or liquidity provider. Cached tokens can be “hundreds or thousands of times cheaper” because they avoid repeated computation.
Providers are not selected solely by lowest price. Tarun describes probabilities incorporating quoted prices and a quality score covering factors such as uptime and throughput. His question was whether providers could advertise attractive rates to win allocation, then misrepresent how much they cache. Without cryptographic verification of cache accounting, a provider’s report determines how much of that benefit reaches the customer.
In Tarun’s GLM 5.2 illustration, Z.ai quotes $1 per million tokens and cached tokens at 0.001 cents. With perfect caching, 21 queries cost roughly $1, just over $1. A rival advertises 5 cents per million, a 95% discount, but bills every repetition as uncached: 21 × $0.05 = $1.05. Agent follow-ups make the apparent bargain more expensive.
Tommy asks whether the harm is worse memory or responses. Tarun clarifies that it is price: “you’re probably caching, but you’re not passing on the price benefits of caching.” He reports statistical evidence from experiments across providers. A provider may lose money on the first query but earn it back on repeated queries. Tarun also limits the analogy: equating this too closely with crypto front-running or MEV would be wrong, although it is marketplace manipulation.
10. Harnesses and inference providers could capture the economics
Tarun’s investment thesis follows from that market structure: “All the money goes to the top and the bottom.” In crypto, he argues, wallets and validators or liquidity providers gained leverage while aggregators and protocols were squeezed. Applied to AI, the favored positions are harnesses and inference providers. “Hermes is like investing in Phantom or FOMO,” in his formulation.
Models can attempt licensing monetization. Tarun describes BSL-style terms requiring fees above revenue thresholds; Tommy mentions Kimi, and Tarun says GLM added similar provisions with different terms. He says Kimi’s revenue cap is particularly low. But routers face low entry barriers as SaaS companies, including Vercel and Ramp in his examples, introduce marketplaces. They can become successful businesses while still facing margin compression from abundant competition.
OpenRouter’s admission controls mean it is not permissionless or fully decentralized, Tarun emphasizes. Once admitted, providers compete in an open market. That differs from bundled services where users cannot see the infrastructure supplier or allocation decisions. Open models allow users to unbundle the harness, model selection, routing and inference provision that centralized coding products package together.
11. Hermes could become infrastructure beneath other companies’ agents
A wallet analogy has limits: Hermes need not be the final consumer interface. Tarun’s “guess” is that harnesses become a “security and brand” and context-management layer underneath applications, including consumer bots like Instinct. Companies can build differentiated experiences while inheriting shared infrastructure. His robotics examples already show developers using Hermes as a foundation rather than an endpoint.
Token aggregation supplies a possible revenue mechanism. Tarun says inference providers already offer payment-for-order-flow-like rebates: “if you guarantee me X billion tokens a day,” they can return part of the fee. Tommy asks whether individual users control routing. Tarun agrees they can choose providers directly, while Portal and other access paths create different arrangements for managing that demand.
Looking a year ahead, Tarun’s belief is categorical: “every SaaS company either has to have a harness or be a router,” because participation in token order flow creates monetization opportunities. Hermes could be the “white label harness for anyone.” Existing companies bring users and data; they avoid starting from zero and depend on a foundation whose security reputation matters much like a wallet’s.
His closing analogy is a fintech sharing revenue with the bank underneath it. Tarun says Muse is going to monetize in a crypto-like way by charging transaction fees on in-app purchases and sharing revenue with Stripe, Shopify and others. To Tarun, competing counterparties and shared order flow make these systems resemble crypto markets. Enterprise deployments, consumer use and applications built on Hermes are therefore separate potential growth paths for the same underlying harness.
Full transcript
There’s a sense in which the rise of open-source and open-weight models over the last 6 months—the immense rise—has people from crypto all over it, if you look at the fingerprints. Yeah, OpenClaw is the worst developer experience of anything I’ve used personally. I think that kind of thing is the difference between product engineering and back-end engineering. I would find these people who were like, “I’m some random Chinese hardware robot manufacturer, and the interface to my little robotics-embedded hardware thing was, like, fork Hermes.” One thing I was looking for was whether people were lying about how much they were giving you in cache tokens. I ran these—
Interesting.
Experiments and tried to make sure I went through all the different providers, and basically found there were some provider issues. 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?
Tarun is the founder of Gauntlet and managing partner at Robot Ventures, and he has way better glasses than I do. Tarun, how’s it going?
I appreciate it. Great to be on, and I think everyone has to go through their own glasses journey. The first time I really started looking into crazy glasses, I happened to be in France, like 20 years ago, and there was this—I randomly walked into this conference that was the eyewear conference, and that changed my life. You never know when it’ll hit you. Next time Tommy’s on, he might have something crazy. He’s gonna have—
There—
Some steampunk glasses.
Were you there for the eyewear conference, or was it just random?
No, random. The thing is, I was just out of college, and I was like, “Ooh, I see free food.”
And—
It sounded very stupid.
I didn’t know if you did the glasses.
And then I was like, “Oh, okay, this conference, I think, didn’t really have that many people, and they overordered food, so they were just like, ‘Here’s some free food.’” Sometimes—
I have these—
Sometimes having a party, you never know what it’ll create, what monsters it creates.
No, it works. I have these expensive Avulux glasses. They’re migraine glasses, but they look really weird when you wear them, so I don’t usually wear them. But they’re amazing.
Yeah. This is a very controversial opinion, because a lot of people get quite offended when I say this, but I would never wear Transitions because you look like a serial killer when you wear them. There are a lot of people who wear Transitions, and they’re always very self-conscious about it. But if you’re in New York, or anywhere in the world, and you take the subway, every time you come onto the subway, you look like you just came out of one of those spy movies and killed someone in the subway.
Yes.
I don’t know. I can’t do that.
I feel like it’s hard to trust somebody who wears Transitions lenses. Your life is that dynamic? You can’t buy 2? I don’t know.
Yeah, exactly. Sunglasses, totally fine. Not sunglasses, totally fine. But 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.
Well, will founders know not to wear Transitions lenses or progressive lenses if they pitch you or me?
There you go.
Auto red flag.
Auto red—yeah, exactly.
Well, Tarun, I’m really excited to talk with you. Robot Ventures just led Nous Research’s $90 million new round, so huge congratulations. Maybe we could start quickly there, and then we’ll dive into the eras of crypto and AI.
Yeah, definitely. Appreciate it. You guys, as well as us, have been there for the whole journey, and Nous has been through a lot. The interesting thing is, I always hear this type of advice, and it changes in venture capital over time. In the 2010s, the advice was that you wanted the company to be much more famous than the founder.
Yeah.
You wanted the company’s brand to be much better known than the founder, because that meant the founder wasn’t just wasting all their time on social media. This was the early social media era. Then, in the pandemic, everything changed: the founder was the only marketing agency you had, so you needed them to always be on social media and have a public presence and image and whatever, right? Founder-first marketing. That sort of flipped the script, and you saw a lot of people running companies who were more famous than their company. Fast-forward to now, we’re in this interesting situation where there are companies with products whose makers people don’t even know. They have so many sub-products that no one even knows the company, let alone the founders. I think Nous is a great example of a company like that. Tons of people know what Hermes is. Hermes Agent is the very popular harness. Their social media is extremely popular—not initially because of Hermes, but because of their amazing design skills in the beginning. Then, obviously, they made some amazing products over time, but they’ve done a lot of different things. I think the interesting thing about Nous is that its evolution really reflects how crypto had to learn what its place is in the new AI world, and a lot of that has changed over time. What people thought crypto and AI would look like in 2023, basically, post-ChatGPT—
Feels like forever ago.
It has evolved so much. Right, exactly. It’s kind of funny because everyone in AI has the “I don’t think about you at all” meme about crypto. But there’s a sense in which the immense rise of open-source and open-weight models in the last 6 months has crypto fingerprints all over it.
I think that comes from really understanding the lessons from crypto and applying them elsewhere, rather than just shoving a token into an AI thing and hoping it works, which is how a lot of the crypto AI stuff started.
So, Tarun, you mentioned a lot of early crypto infrastructure projects back in 2023, especially the rented GPU ones. The idea was that if you were running a big Bitcoin mining network, you could flip a switch and serve AI inference, host models, and things like that. In hindsight, it does seem really naive, right? The open-source models weren’t even around to serve. We’ve only recently, in the last year or two, gotten really good open-source models from China, so it definitely seems a bit naive in hindsight.
Like I said, I think a lot of people saw the success of existing miners converting their operations. They also saw Stable Diffusion as an example of someone running a mining operation and converting their existing GPUs to make a model. I think what people didn’t realize at the time was that language models would have a lot of idiosyncrasies that image models didn’t.
No one really wanted a decentralized image model, to be honest. No one seemed to care about that very much. I also think it came from crypto as an open marketplace. People are inherently lazy, in the sense that they think, “Well, hey, I could launch a token for saying I’m doing this thing that probably won’t work, but the token will still have some valuation.” So in the beginning, there were a lot of disingenuous projects that were just meme coins with more steps—
Ah.
—as I like to call certain types of projects.
I love “meme coins with more steps.”
Underneath all of that, as someone who had worked in centralized hardware stuff 15 or 16 years ago, I was always like, “Oh, the decentralized things are all scams.” Then, after 4 or 5 years, I finally understood a little more about why blockchains work. What is different about them? Why do all these things you were taught in school about distributed systems, and in grad school classes, get violated but still seem to work?
Because I’d been in crypto for a while, but had also been on the other side, I thought there was a 1% chance that some architecture, some real technical advancement, could get over some of these problems for decentralized training. Maybe I was a little less optimistic about training. I was much more optimistic about inference.
1. Nous Took The Scientific Route
What stood out to me about Nous was that they were one of the only people straddling the line between wanting to try decentralized training in 2023 and taking a more scientific approach to it. They wrote a lot of research and did a lot of experiments, some with one of OpenAI’s co-founders from the early days, on how to do low-bandwidth training.
At that time, realistically, there was only OpenAI and Anthropic with real language models. There were a lot of really bad language models. Llama probably just came out around then, and—
I have it in my backpack. I’ve never sorted it out. Just the DisTrO preliminary report—
Oh, wow.
—on the—
Yes.
—distributed training they were doing.
Which is—
Exactly. I think one of the things you’re getting at, too, and one of the things we saw early on, was that the team was really forward-looking. They were doing synthetic evaluation of models. Everybody tries to game these benchmarks, but they were doing that synthetically early on.
I think Jeffrey did YaRN pretty early, which was 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 forget if it was 40 billion parameters or something like that. To your point, they never launched a token. They never did anything infrastructure-wise on the token side.
Mm-hmm.
It definitely seems like they delved into it a little bit, but also understood the opportunity—
Right.
—and where they could thrive.
I think they were being rigorous at a time when everyone talking about the decentralized side was a little bit of a scammer, or, like I said, a meme coin with more steps. There’s a very fine line between a scam and a meme coin with more steps. You could be well-meaning but just not have a good idea, and people still give you money for it. That’s a meme coin with more steps.
Well, what got you over the line early on? You’re kind of a long-term investor here, too. You’re seeing a lot of distributed stuff in AI that doesn’t really make sense, but this sort of did to you early on—well before it did for me. So, in this era—
Yeah.
I think DisTrO, YaRN, as well as some of their early research on different types of embeddings for better compression, stood out to me compared with the normal crypto AI companies at that time. They were focused on very specific, well-defined problems, rather than promising the moon: “Hey, this complicated system with a million moving pieces will get it perfectly right on the first shot.”
Any distributed system often has 100 times or a million times more moving parts than the centralized version. Most of the crypto AI people were like, “We’re just going to take an existing blockchain and somehow force the node operators to run the software.” But none of them could do that at sufficiently high bandwidth or at a useful level. If I’m doing it at 0.001 tokens per second, who cares?
There were a lot of things like Bittensor that jumped too fast to the token and decentralization piece, but the utility of the actual projects was kind of weak. In Bittensor’s case, I think they found a middle ground where they became a funding platform, effectively a launchpad, so arguably there was some residual value from that. But it didn’t really feel like a technical innovation.
Nous and Prime Intellect were realistically the only people doing real research and focusing on well-defined, specific problems. Every successful distributed system succeeds through 1,000 little optimizations. It’s not that there’s one big idea that automatically solves everything.
If you look at Bitcoin, going back to the original paper, there are a bunch of different concepts that need to work together. First, you need the hash functions to have very high collision resistance. You need the difficulty adjustment not to be gameable. The idea is that, based on how many people join and leave the network, it can adapt to them, so you can dynamically scale.
You need the inflation curve not to be too steep or too inflationary. It needs some balance between those. And you need the node software to have sufficiently low bandwidth. That’s why Bitcoin blocks are so slow.
Each of these is a separate technical problem, often from a different specialty. A person with a certain specialty knows each one of those. Usually, to build these decentralized and distributed systems, you have to bring all those specialties together.
So you need a wider breadth of knowledge rather than depth. I think the thing that Nous and, in a lot of ways, Prime Intellect had at that time was that they were the only ones doing that.
I think the difference for me with Nous was that Prime Intellect maybe wasn’t as focused on the training side. I know they did a training run, but they hedged their bet and were like, “We’re also going to be a GPU reseller at the same time.” Whereas Nous was really focused on this research piece, and that was what stood out to me.
In order to ever have decentralized training or decentralized inference be a viable concept, you need these 10,000 tricks that, when you—
To get there.
—to get them all together, compound and give you the final product, versus a one-shot thing. The latter is a real needle in the haystack and often never exists in science and engineering.
Yeah.
I think a lot of the early crypto AI stuff had a lot of hopium. In some ways, it reminded me of the 2017 ICO boom, where it was like, staple a token onto something that already exists and say it’s decentralized. Like, “Decentralize Uber.”
Yeah.
And so I think that’s what Nous did: they were actually trying to do the real thing by doing the hard work, not launching a token, and not making crazy promises.
Yeah. No, I’m with you. It is kind of funny, mixing some hindsight with what we’re seeing today in terms of the token side and the distributed network side. I’m pretty sure the distributed network side has been sort of abandoned for both Nous and Prime, right? We’ve kind of figured out from first principles, “Hey, that doesn’t work,” right? And that’s fine: they went through it, they tried it, and now it seems like there isn’t a focus on decentralization or distributed networks, but there’s a huge focus on the open-source side.
Hermes has millions, maybe tens of millions—I’m not sure—of downloads around the world. I’m curious how you link the 2 worlds together. You saw decentralized networks solving really hard problems, but now they’re sort of champions for open source.
2. China’s Optimization Advantage
One interesting thing to think about is that in 2023, everyone building open-weight models either didn’t exist or existed the way Nous did: as a GitHub and a group of people, with maybe a little bit of funding—not that much. Arguably, the decentralized training people were competing with the Chinese model creation companies.
If you look at DeepSeek, MiniMax, GLM, and Z.ai, and read the papers on model development, all of these things boil down to there being 10,000 possible tricks, and you can only pick 100 for your model. How do you pick the best 100 little tricks—little optimizations—that you can compound to make your model faster or better? I think the Chinese companies rightly focused on data collection and basically making their own Mercor-style operations, except much cheaper than in the U.S. That let them get better economies of scale while also doing a lot of these optimizations.
Their optimizations weren’t focused on lower bandwidth, lower power requirements, or decentralization. They were focused on, “We have worse chips. How do we make it run faster?” So it’s still a resource constraint, but a different resource constraint.
I mean, some of it is pretty innovative, though. The DeepSeek model paper was nuts: mixture of experts on second-rate hardware.
Oh, no, for sure.
Yeah.
For sure. No, no, no, I’m not at all knocking that. I’m just saying that when we look back at history, Nous, Prime Intellect, and others were doing decentralized training at the same time these Chinese model companies were saying, “Hey, we’re going to build our own model.” They both started with resource constraints.
The decentralized training people started with the constraint that it was decentralized: we didn’t know how much computational resource we were getting, and we had to pay for it, meter it, and track it somehow. That’s one type of resource constraint on training. The Asian companies were saying, “We only get worse chips, fewer resources, worse data centers, and worse data. Given those constraints, what can we do?”
Both sides may have had different bags of tricks, but out of the 10,000 possible optimizations, they chose different subsets of 100. The Chinese companies’ 100 were just better, right?
Hmm.
That’s why they got to building world-class models faster. But as we’ve learned from the agentic world and recursive improvement, once you have a good enough model, you start improving a lot faster, and it becomes much harder for people to catch up. So I think a lot of the early infrastructure people, and a lot of the decentralized training people, started pivoting out by 2024 and 2025.
That was also when we started to see the first versions of trading agents and early AI agents. Those things were again a little bit like meme coins with more steps.
A lot more steps and a lot more replies on Twitter.
Right, exactly.
Yeah.
In some ways, those probably set people back. They were right that agents were going to take over, but they were the shittiest possible agents. Fast-forward a year from when AIXBT and ai16z were at their peak, and that’s when Claude Code, Codex, and others were having a huge growth curve—not just in users, but in capabilities and the things they could do.
Mm-hmm.
I think crypto had the right idea. Decentralized training was the right idea in the sense that resource constraints might make you build a better or more competitive model. But you have to have the right constraints. It’s the same with AI agents: you have to have the right thing, and in a lot of cases, you don’t need the token piece.
It’s kind of awesome to go through memory lane before we get to where we are today. It sounds like what you’re saying is that people in crypto approached AI with well-meaning intentions. People wanted to train models, like Nous and Prime, but the technical choices they made weren’t competitive with the Chinese research at the time.
Then, from 2024 to 2025, we had AI agents, but they weren’t managing our lives or trading our portfolios. They were just replying to us on Twitter. It seems like we’re getting close each time, but we’re not exactly on target with the big picture.
The interesting thing is that people in crypto who pivoted into non-token AI have done really well. A great example is Nous; some would say OpenRouter. Together AI was actually going to be a decentralized inference—
I mean, Corby’s—
—network when they started.
Yeah, but Corby’s a little bit different in that they weren’t trying to be a network. They weren’t trying to launch a token. I think the people who were trying to launch a token and then pivoted had this kind of… Now, OpenRouter is sort of weird: they definitely weren’t trying to launch a token. But philosophically, I think they came from this crypto view.
OpenRouter is a great example of finding the right optimization to get very good performance in some of these systems under resource constraints. In some ways, that’s the same ethos as decentralized systems: I want to have the highest capacity while having low resource usage. Lower resource usage means I can have many more people participating.
I think there’s a lot of shared ethos between Chinese open-source model development and decentralized crypto development. They just went different routes. To be honest, I’d say the talent in crypto wasn’t as good, and they also chose the wrong optimizations.
3. The Harness Wins On Usability
Now, look at someone like Nous. In the first era, they were doing decentralized training and were going to launch this thing on Solana. Then they were trying to figure out what to do. They were doing a lot of research and private inference work, optimizing models to run in TEEs.
Which is what you actually see Nous doing right now, with Venice and stuff like that. Sometime in 2026, I think they saw OpenClaw have this moment at the end of last year, early this year. One thing about OpenClaw is that it’s kind of the worst developer experience of anything I’ve used personally. I don’t mean this in a really—I’m not trying to be an asshole because I’m a Nous investor. I’m trying to say that OpenClaw feels like Android because it makes you have to make too many decisions to use it.
If you look at the comparable centralized products—Codex, Claude Code, Muse now, GrokBot, all these things—they all have this quality where my grandmother, who doesn’t know what AI is, could use them.
Your reasoning level, model choice—
How much reasoning—
Everything.
Model choice, context choice, stuff like that.
Yeah.
Right? It had this Android feeling of, “You give me 10,000 options, and yes, I could choose something better than the default.” But you kind of hamstring—
Well, the—
—who can actually use it.
I mean, the crazy part—maybe I can give you a quick example. I was building an agent on n8n, which is the more complicated version of Zapier. Zapier is where I built our first bot in 2023. You’d come to Delphi’s site, it’d run through 20 steps, analyze, and send us a message.
So I thought, “Let me make a personal agent on n8n.” I was building it, and I remember deciding to use Claude Haiku to choose whether to go to DeepSeek R1 reasoning or a basic version to answer my emails. OpenClaw came out the same weekend, and I stopped using n8n, where you have to draw arrows between services and build a manual process. I used OpenClaw and hated it, and literally went back to n8n. It was so clunky and hard to use that I couldn’t figure it out or get excited about it.
Yeah. I think the interesting thing is that OpenClaw is, in some ways, the right idea for a product, but the wrong execution. There was a tweet the other day that stuck in my head: in AI, second-mover advantage seems like a real thing, because you learn a lot from the first mover’s mistakes.
You could argue that Claude’s high uptake in late 2025 and early 2026, when it kind of eclipsed OpenAI—though OpenAI has come back with a vengeance recently—was because it was easier to use and had more human-like touches. The average person who isn’t technical could make a product with it. Reporters and journalists were making web apps and dashboards without knowing a lot of the technical language or what to say to make them.
It was really a no-code thing. OpenAI came back and caught up, but why did Anthropic have this opportunity? It was because they learned from OpenAI’s lessons. Arguably, the same is true for open-weight models. Z.ai, Kimi, and MiniMax learned a lot from DeepSeek. DeepSeek did a really good job of optimizing inference time and making the model small, but Z.ai, Kimi, and MiniMax did a better job of changing how they think about reasoning and resource usage—
Mm.
—and changing a lot of the attention topology and stuff like that. I think the second-mover advantage is very real here. OpenClaw clearly had the right market: everyone was excited about this agent concept. Crypto tried it a year earlier, but because they shoved the token in it—
Mm.
—no one was able to use it.
Yeah. I think OpenClaw deserves a lot of credit for creating the category.
For sure.
And—
For sure. I mean, no doubt.
Yeah, definitely a lot of credit. I think the hard thing, though, is that people would get into this weird AI psychosis with it. They’d build stuff, spend 10 hours fixing it, and then finally fix it.
But no one else could use it.
Exactly.
It couldn’t train.
Yeah.
That’s the problem.
Yeah.
They made something that was great for them, but they couldn’t get anyone else to use it. So no one else could say, “Oh yes, this is awesome.” I remember when OpenClaw came out—I want to say it was December or January, at the end of 2025 or early 2026—and I went to this meetup. It was an OpenClaw hackathon-type thing.
Yeah.
It felt a little bit like going to a crypto hackathon, where people were building their weird little projects. “How do I get my cat to—”
Love me?
You know, when it eats food out of the cat dispenser—
It sends me a USDC payment or something. Some weird-ass project that makes no sense, and you’re like, “Does this have a market size of one?” Yes.
It’s—
All these people had done these OpenClaw hacks, and they were presenting them and whatever.
Mm.
Almost all of them were the worst demos in history. None of their demos worked, first off. Secondly, if someone thought another person’s project was cool and asked, “Can I download the setup?” the answer was no. It was too hard to transfer it to them.
I think the irony is that going to OpenClaw events showed you a lot of what was wrong with it from that perspective. Now, coming back to Nous, Teknium in particular started it as a side project—
It’s nuts that it was his side project. Just insane.
He was just like, “I love OpenClaw as a concept, but it’s unusable.” Nous had spent all this time building infrastructure for decentralized training. They’d trained a bunch of models of different sizes, and they understood the difficulties, especially on the data side, because that’s the hardest part.
Their models have tens of millions of downloads. It’s no joke.
Right. But they have problems with getting more data. Decentralized data collection has always been much harder than centralized data collection, and that’s one problem with training that inference doesn’t have.
So then they were doing inference. In the process, they learned a lot about the infrastructure for scaffolding around an inference engine. If I have an inference engine, how do I pass context between it? How do I share context? Around the time OpenClaw came out, Codex and Claude Code were also really coming out. I think what people really got used to with those was the idea that they could share the same context between multiple models.
Mm.
In the early days, everyone thought that would become a form of lock-in for OpenAI and Anthropic. They thought, “Our models are going to have a certain format for context. You can move between our models, but we’re going to lock you into that. If you want to move to an open model, it’s going to be hard.”
As of yesterday, that’s completely obliterated. OpenAI now supports open models in the same context. Context sharing, and all the stuff that Nous and OpenClaw focused on to let the same context move freely between models and agents—that stuff is interoperability. It’s here, to put it in blockchain-y language.
I think Nous was in the right position at the right time. They’d built all these things and knew why they were hard. They knew why training was hard: data. They knew why inference was hard: orchestration and figuring out context management.
Good.
I think technically, the problem with OpenClaw is that they don’t realize that because they don’t do that much of their own inference. It’s built around giving the user all the options and telling them to figure it out themselves. It’s like, “Here are all the parts to this car. Go build the car.”
True. Could you—I’m glad you’re getting into it, because I think people might look at the surface and say, “Hey, Hermes Agent is just a similar, better version of OpenClaw.” But what you’re describing to me is that there’s real architectural genius behind it that makes it a much better experience, without an easy trick or two.
Is that sort of the case, or not?
I think it’s just like an Apple-versus-Google thing, but in the inference era, right?
Good.
The interesting thing is that all the crypto ideas that didn’t work have analogs in no-token land—no-crypto-token land—that are huge companies and have worked a lot, right? The rented GPU stuff: all these inference providers have been raising at decacorn-plus valuations because they’re getting a ton of demand.
But the problem is what I call the Maslow’s hierarchy of compute needs.
Yeah.
As I’ll call it, everywhere in the beginning it was like, “Okay, rent a GPU.” Well, now I can rent the GPU, but then I have to figure out how to set up the environment. I have to set up the models. I have to make sure I manage memory. If I want to run inference really fast in ultra-fast mode, I have to set up a ton of DevOps. So now it’s like, okay, I have to be a developer to start the GPU.
Then there are companies like Together, Modal, Baseten, and Fireworks that are like, “We’ll give you a served endpoint for a particular model. We’ll do all the optimization. We’ll optimize the inference engine and all the sampling.” It’s almost like they make an RPC for you. The crypto version of this is: one, I run my own node—that’s running a GPU. Two, I use a node that someone gives me an RPC endpoint to, like Alchemy. That’s what Together, Baseten, and Modal are doing.
But the problem is, that’s still not that easy. Compare the experience in crypto between using FOMO or Phantom and going to Etherscan and submitting a transaction on your own. That’s more like what the RPC is.
It’s insane.
Right?
It’s a good comparison. Yeah.
Clearly, the total addressable market of people who can use that, while bigger than the group that will run their own node, is still not that big. The harnesses are truly like wallets that manage a lot more state for you. They manage your history, your preferences, private keys, and stuff. They try to give you a ton of defaults so you don’t shoot yourself in the foot.
I think of OpenClaw as a little more like MetaMask or MyEtherWallet. It clearly has the right stuff, but it’s annoying to set up, and it doesn’t give you the right defaults. It still forces you to do a ton of stuff on your own.
I think the Hermes model, Claude Code, et cetera, is all about giving you the right defaults and learning your behavior—generating skills based on how you’re using it and learning about you as a way of having some memory and preferences. It’s also about figuring out how to manage all the infrastructure: Which model should I use? How much should I spend on this model in tokens? I think it was built with the idea that a model router would exist.
Hmm.
One problem with OpenClaw is that 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 what context they want to share and not share.
Most people don’t know the difference, though, especially at scale. People don’t know the difference between reasoning and non-reasoning models.
Yeah.
Yeah.
But they do know the difference when the answer they get looks shitty, right?
For sure.
So—
But then I think they just think OpenClaw is broken for mass retail.
Yes.
Yeah.
Exactly. But I think that’s why the MetaMask thing is the same. Realistically, is MetaMask broken for retail?
Why is it very reasonable to—
There are a ton of FOMO users who could never use MetaMask, right? That’s just a fact. I think that’s the way you should be viewing these.
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.” That difference in philosophy, while simple-sounding and obvious, changes the architecture of how you design the harness. It changes what decisions you surface to the user, which things need to be their own models that learn something about the user, and which things need to be asked of the user.
When you use any of these bots, whether they’re the consumer ones like Muse or Grok Bot, or the productivity ones like Hermes and Claude Code, they all ask you questions. They’re like, “Hey, do you think this is the correct interpretation of this?” They’re learning something about your preferences. 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? I think that kind of thing is the difference between product engineering and backend engineering. The problem is OpenClaw treats its users as backend engineers.
That’s interesting. It’s really cool to talk to you because you’re so deep technically, and you understand how things are actually built and a lot of the design decisions. I think a lot of people try Hermes and say, “Hey, this just works.” But from what you’re describing, the “Hey, it just works” is because of a hundred correct decisions on the product side and the backend side.
Yeah. The interesting thing about Nous is that they have this mix of technical people, but also this design and product-side stuff. A lot of people on the internet made fun of them in the beginning, like, “Oh, is this company—”
For being a merch company?
Yeah, exactly. Like, “Oh, is this company just a very expensive merch company?”
They have great merch. Yeah.
But I think that also teaches you a lot about the aesthetics of your user base. You care about the aesthetics of your user base and what they have to do.
Right.
I think that’s something OpenClaw doesn’t care about. And not just OpenClaw—I think a lot of the harnesses built in that era, and still a lot of the ones you see now, are part of this whole DIY harness thing. People are trying to build their own and optimize it for themselves.
It reminds me of people who spend two years customizing their text editor because they need every shortcut so that coding is really fast. That was 10 years ago.
You could have built something.
Obviously, now.
Yeah.
There’s a DIY community for this harness stuff, and Pi has clearly taken over that community.
Yeah.
Everyone forks Pi and starts their own. But I think what Hermes went for is that there are going to be people who want DIY, people who want enterprise, and people in the middle who are prosumers and, over time, more consumer. That philosophy is what I really liked about what Nous was doing: they started with prosumers, but made it easy to use, and then widened the space.
For sure. And—
Around the time they were growing was when OpenRouter started really growing. The two of them kind of grew synergistically. I sort of think of it as the way Binance and Tron grew synergistically in some ways. But the beauty of it was that now there was public data, right?
Mm-hmm.
You had a leaderboard. You could see which apps were generating the most tokens. You could see which models were being used. You had almost blockchain-crypto data for the stuff you didn’t have before Nous and OpenRouter grew.
That was when it was like, okay, this company is doing something really special. I’ve invested before, but clearly this is probably going to be the most important competitor to the centralized companies, in a sort of crypto-like way of being the most credible competitor to traditional finance.
Yeah, no, I couldn’t agree more. Tarun, I want to double-click on where you’re taking this. You mentioned consumer agents. There are a lot of vendor-owned consumer agents. Meta owns Muse, OpenAI owns OpenClaw, and Instinct, I guess, is its own. There’s Grok—SpaceX and xAI own Grok Bot. Nous is the only one that’s open source and not lab-owned.
Do you think that has any implications in terms of what people will use it for, the restrictions, what they’ll build, or the second-order effects of being open source? It’s the only one on that spectrum that’s open source and not vendor-owned.
There’s a whole DIY community, but it’s a little more like OpenClaw, where you have to pick everything up yourself. Hermes is the only one that’s open source but packaged. I think of it a little more like Linux distributions. There’s the Linux kernel, and then there are people who build distributions on top of it: Red Hat, Ubuntu, Gentoo, whatever.
Realistically, I’m probably going to get flamed by old heads for saying this, but Ubuntu is probably the only Linux distribution that ever got non-developer, technical users. It’s because they did all the work on defaults and packaging, made it easy to update things, and figured out how to deal with non-open-source code interacting with open-source code. Hermes is a good example of that: I can use Claude and OpenAI with an API key within Hermes, whereas a lot of the pure open-source stuff is extremely militant about not wanting to touch anything closed-source. Ubuntu had the same philosophy, so I view it a little like Linux distributions.
4. Forks Reveal Community Fit
Another really cool thing came up while we were doing diligence for the investment. The OpenRouter thing made me think, “Oh, I’ve turned this into a blockchain data thing,” and with my background, that made everything make sense.
Yeah.
The second thing was that I started looking at the GitHub accounts of people who fork Hermes. If something is open source, one way to measure community strength or whether people like the product is by looking at forks. To give some credit to the competitors, I think Pi is probably the most successful from a pure DIY standpoint. It completely eclipsed OpenClaw in the DIY community because it’s very lightweight.
I don’t think a lot of people know that, by the way.
Yeah. If you look at Pi’s usage and forks, it has surpassed OpenClaw. I think the whole OpenAI sale, pseudo-sale thing—
It kind of feels like TBPN. I don’t listen to TBPN since OpenAI bought them.
Yeah, somehow they sold themselves too early.
They—
But Pi is really community-based, and I think the interesting thing about Pi is that it probably has the most forks. My brother’s a CS professor, and he was saying that in one of the intro CS classes now, they fork Pi and modify it. That’s how they teach you how the harness model works, because Pi is very simple.
But it has this problem where you have to build all the adapters to everything you might use. So I was asking whether Hermes’ open-source stuff was still sufficiently competitive for DIY builders, because I think that’s important. The ethos should still be there. It shouldn’t just be, “Hey, we’re an open-source harness and we have a lot of users,” while nobody actually uses the open-source code. That usually means something is wrong.
I started indexing and looking through a bunch of forks, and I found some really crazy and interesting ones. These aren’t things Nous or Hermes will monetize, but they show brand recognition. People really care about how it’s been built and like how it’s been built. That’s a way of measuring it, almost like doing sentiment analysis about a brand through people who forked it, rather than looking at revenue.
That’s super interesting, by the way.
Right? I’m not trying to assess monetization with this. I’d find people who were, say, some random Chinese hardware robotics manufacturer, and the interface to their little embedded robotics hardware was a fork of Hermes. They weren’t making a Unitree humanoid; they were making small robots. They don’t have money. They have to keep costs down, because they have to spend all their money on the hardware. They need to lower their inference cost to zero, which is why they love open-source models.
For robotics, as you’ve seen in a lot of the new evaluations, it’s actually quite hard to have a single model that’s really good. One reason I think the open-source world is better is that the routers—and the ability to move between multiple different models based on specific tasks—are much better. That seems especially true for robots. You might need one model for planning: the robot has a task, like walking to the laundry room, opening the dryer, and folding some clothes.
Folding clothes is always used as the hard test for a humanoid robot because we still don’t really have one that can do it generically, as crazy as that sounds. You might have one model that’s really good at moving its feet and walking, another that’s good at planning—figuring out the layout of the space, which direction to go, and what path to take—and another for dexterity, like applying a specific amount of force to open a door.
The interesting thing is that these tasks require different amounts of reasoning and different amounts of tool calls or external actions. A lot of the evaluations have shown that segmenting tasks across different models and putting the results together often gives you much better performance than using a single model. So the question is, how do they manage that? How does the average user of a robot figure out how to do it?
A couple of these Chinese companies basically forked Hermes and made modifications to connect it to their devices. They added tool calls like “move hand” and “move head.” To me, that’s a sign of product-market fit for the brand: people want to use it. So, for that type of thing—
And they’re not going to sell this robot commercially? They’re not doing door-to-door sales or marketing?
No, no. I don’t even think they knew. I remember I told them, and they said, “We didn’t even know this person forked it.” My point is that, for open-source projects, community sentiment—how much people use your stuff—is actually very important.
In the current AI market, everyone is so focused on rushing to tell you monetization stories about companies, and half of the ARR numbers are fake. But retention and stickiness will be much more important if there’s a crash or downturn. As we were looking at Nous, we asked whether they had this kind of really good brand with their users. We found a lot of that, and I think it cemented the deal for us. It felt like, if not now, then when? This feels like the rocket ship.
I’m with you. Tarun, maybe just to recap: you’ve walked us through why you like Hermes so much more than the competitors, its design, the people using it, and these really interesting companies basically proving product-market fit, social buy-in, and brand longevity.
My more pointed question is: looking to the next year or two, where does it go? You have Muse, Instinct, and a lot of these consumer agents. I personally don’t feel Nous is really competing with them, or vice versa. I view Hermes as what you described: something for pro builders building applications for the long term and really wanting to own them.
My second thought is that people want to use Hermes for 2 reasons. If you own the architecture around Hermes, everything you build—the data, the memory, where you’re hosting it, the interactions—is an asset for you or your business. But you also get actually unique outcomes when you ask the model something. 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 agree, given what we’ve seen.
I’ll tell you what Nous has been working toward, based on the stuff they’ve said publicly so far, and then I’ll tell you where I think the dream is going and how crypto could come back into it.
So far, crypto people have fucked up. They’re wrong every time on the AI stuff, and everything they’ve done has just—
We can also leave crypto behind on this one. I mean, there is a—
We can.
Yeah.
But I think there’s actually a very real reason not to do it.
Interesting. Okay, cool. I’m eager to get to it.
I think Nous has obviously been focusing a lot on increasing its enterprise usage. The nice thing about Hermes being open source is that enterprises can clone it, customize it, and use it internally. They can add their own privacy layer, compliance layer, permissioning, and connect it to their databases. Nous has really been focusing on that. I think they have a bunch of big news coming up on that soon, and that’s been going really well.
And, true. Sorry to interrupt you.
Yeah, go ahead.
But 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 have the go-to-market sort of in the bag.
Nothing’s in the bag until the contract’s signed, but yeah.
Fair. Fair. Fair. A true OG.
But yeah, I think there’s certainly that. Realistically, they’re in the tens of millions of users. In terms of, if I remember correctly, tens of millions of unique GitHub accounts access them daily. If you use GitHub as your Sybil resistance, it’s not perfect. Obviously, people can have tons of GitHub accounts, but GitHub has made it expensive to have multiple accounts compared with most social media accounts.
So they have a ton of users coming from everywhere. Their subscription model with Portal has been growing, as has their OpenRouter usage. I think they just passed 2 trillion tokens across apps there. They’ve also been working really closely with NVIDIA. In this round, NVIDIA and Microsoft both participated.
I think enterprises are going to access Nous. Nous is going to be one of the main open-source harnesses you can use at Microsoft, and NVIDIA has installed it on its hardware by default for DGX Spark machines. They’ve been trying to play both parts of the market: the consumer open-source side, closer to the ground, which they’ve obviously excelled at, as well as expanding enterprise.
5. Inference Markets Mirror Crypto
The interesting thing—and the thesis we had for investing, beyond the fact that they’re growing and clearly have developer market and brand fit—is something I got from scraping all this OpenRouter data during due diligence over three or four months. I started really analyzing how the OpenRouter market itself worked. It looked like—and I wrote a post about this a month ago, but it was the culmination of work since April or May—that the way OpenRouter works is that there are applications that want access to a model.
There are many providers, almost like blockchain operators or intent solvers in blockchains, who quote a price: “Here’s how much I’m going to charge to run this model for a million input tokens, cache tokens, and output tokens.” The router matches you to a provider. Node operators basically say, “Here’s how much I’ll sell you access to run this model,” and OpenRouter does the matching.
You can think of OpenRouter as 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, API keys in their tool calls, and potentially private secret keys, if they’re trading. You can view the model as a little bit like a protocol. It’s an open-source thing that has a license. In fact, a lot of the open-source models now have BSL-style licenses, where if you make more than a certain amount of revenue, you have to pay them a licensing fee.
Kimi has that, I think.
I think GLM does as well.
They all do?
GLM also added something similar, but not the same amount. Kimi’s is much lower; it has a very low cap on how much you can make. This is very much like crypto licenses: if you fork me, you have to pay some fee.
The inference provider is a little bit like a node operator, validator, or liquidity provider. They’re doing both functions. If we take this analogy a little more seriously—and the data shows a lot of things that look very similar—what phenomena can you expect? One phenomenon you can expect is MEV in crypto: people trying to front-run and manipulate the price. Most of my post is about finding some examples of that, to show that in any open marketplace, decentralized or not, it just has to be open, with open entry.
Tarun, sorry to interrupt, but what is MEV when you’re asking for AI inference? In crypto, you’re competing to make money, solve a block faster, or fill a transaction. But in AI, if I beat you… Sorry, I just want to understand that.
No, for sure. Remember I said there are 3 prices an inference provider quotes you: an input token price, an output token price, and a cache token price. Let’s pretend the input and output prices are the same. They often differ, but it’s usually a marginal difference. The cache token price is usually hundreds or thousands of times cheaper. Why? Because the inference provider is using way less compute. Cache tokens are saved, so they don’t have to redo the whole calculation.
An interesting question is how OpenRouter allocates you to a node provider. They don’t just give you the lowest-price one. They assign a probability of each provider being chosen based on the price they offer, and then assign you based on that probability. That probability also includes some kind of quality score: have they had good uptime? Have they delivered more than a certain number of tokens per second? OpenRouter grades them, almost like a credit score for each inference provider. But then they have a price.
One thing I was looking for was whether there are people lying about how many cache tokens they’re giving you. They tell you there’s a really low price, which increases their probability of being chosen. But then they say they cache tokens, because there’s no cryptographic verification. I can’t verify it; they just say, “This is how many tokens I cached.”
And, Tarun, if they’re not providing you the amount of cache tokens they’re quoting you, how does your level of service go down? In terms of the memory, or your response, or—
Your level of service goes down in terms of the price. I’ll give you a similar example. Suppose I have an agentic workflow with 5 prompts: you ask a prompt, the agent asks another prompt, then another, and so on. As an example, with GLM 5.2 from Z.ai, let’s say they’re charging $1 per million tokens, and cache tokens are 0.001 cents. The model creator usually runs an inference provider, and you can treat the model creator as the gold standard. They’re not going to lie about their caching, because then everyone is going to say, “Your model’s shitty.” It’s a brand thing for them.
Let’s say the true optimal—or close to optimal—caching rate is whatever the model provider tells you. Now suppose someone says, “I’m going to offer you 5 cents per million tokens,” a 95% discount.
Okay.
Now imagine there are 21 queries. I make a prompt, the agent asks another prompt, and that happens 21 times. On the Z.ai model, imagine the caching is perfect. It’ll charge me $1 for the first query, and the agent’s repeated queries are all 0.001 cents, which is very small. So it’ll cost me basically $1, or just over $1.
On the other hand, if the person charging 5 cents lies about caching, then I get charged 5 cents on each query. So now I get charged—
Okay.
—5 cents 21 times, so I get charged $1.05. The idea is that by lying about how much I cache, I can charge 5 cents on each query.
Yeah.
Whereas if I use the provider that’s caching, they charge me $1 at first, but then basically zero on the repeated rounds.
Yeah.
The question is—
Interesting.
Why would I put 5 cents as my price? To get a higher probability of being chosen.
Yeah. Yeah.
It’s sort of like—it’s not exactly analogous to—
I understand.
—front-running or—
Yeah.
—making too much of the analogy to MEV is wrong, but it is a manipulation. There might—
No, no, there’s a way you can game it. There’s a way you could attract inference—
Yes, exactly.
—to provide a bad service, for sure.
Your way of providing bad service is that you’re probably caching, but you’re not passing on the price benefits of caching. That’s the way of looking at it.
Yeah, and it’s a way to improve it, too.
Yes. The idea is that you’ll lose money on the first query, but if the agent asks a lot of subsequent queries, that’s where you make money. I found statistical evidence of it. I ran these experiments and tried to make sure I went through all the different providers, and basically found that some providers were doing this.
So how does this relate back to using an agent, or the inference side?
What this side made me realize—
Like back to the Hermes side?
If this analogy is to crypto, then it’s like wallet, front-end harness, DEX router, exchange or DEX aggregator, model protocol, inference provider, liquidity provider. If I look at crypto from 2021 to now, all the money went either to the top—the wallet or the harness—or to the validator or LP. Everyone in the middle got squeezed.
Yeah.
The DEX aggregators got squeezed. The models can’t collect as much in fees. Phantom makes more fees than Uniswap. Why? The front end is able to capture more than the protocol. My point is, if you think this looks like crypto—and I think any open market where people can join and leave, like these inference providers, can join—OpenRouter is a quality-control mechanism. They can keep some people out, so in that sense it’s not decentralized. I can’t just join with my own node. But once I’m on there, it’s an open market competition. That aspect—
Yeah.
—is very different from—
For sure.
—the centralized providers like OpenAI or Claude, where you don’t even know who the provider is. They might be using CoreWeave for some of your stuff. They might be using—
People just wake up and say, “The model’s been nerfed this week,” or they don’t. It can just dynamically turn off.
I think this open-source stuff, especially in the routing model where multiple models and multiple providers are used, and potentially multiple apps, looks like crypto. By apps, I mean something like Hermes, Instinct, or whatever. All the money goes to the top and the bottom.
Mm-hmm.
What’s the top? The harnesses. What’s the bottom? The inference providers. It’s kind of borne out in the last 6 months—
Interesting.
I think OpenRouter really sold the top, because I do think the margins are going to get squeezed there. The router is a very competitive market. Every SaaS company is making a router right now. Vercel has a marketplace; Ramp has a market. The router is hard to build a business around. You can be a successful business, but there are low barriers to entry.
Yeah.
So my long-shot point here is that, in that model, Hermes is like investing in Phantom or FOMO. That’s sort of the game plan.
I’ve never heard someone say it like that. It’s interesting. Yeah, it’s a good way to think about it.
I think one difference between AI and crypto is that, to some extent, in crypto the wallet is the final endpoint. But the interesting thing about something like Hermes is that people could build Instinct on top of Hermes and inherit a lot of Hermes’ security properties. My guess is that people will view the app-harness layer as a security, brand, and context-management layer. People will start building apps like Instinct—different kinds of bots—on top of a harness that provides the consumer experience, while the back-end development is done through these harnesses.
So it’s not the endpoint in the same way as a wallet. The wallet is an endpoint because it goes to the final end user, whereas you’re already starting to see people build things on top of these. I think that will be a very interesting game to watch. The other thing that’s really interesting is that the harnesses and inference providers are already doing payment-for-order-flow-type deals. If you guarantee me X billion tokens a day, I’ll give you a rebate or pay back some of the fee. In some sense, being able to aggregate token order flow is an extremely valuable market position, just like in the crypto front-end market.
For sure. That is interesting. Hermes isn’t directing that usage, though, right? Each individual person is making their own decision about where they want to send—
When it goes through OpenRouter—
—the inference to, or—
When they’re using Portal or another service, you can use many different providers. You could pick your own provider, have a contract with them, and send directly. The point is that these open models allow you to unbundle Claude Code. When you use Claude Code, you’re getting all of these features. Implicitly, they’re using some kind of router or at least a model selector. They’re implicitly using different providers. You don’t know; they’re doing all these long-term data-center contracts, and you’re getting run on one of them, but you don’t know which one. I’m sure you’re getting migrated. It doesn’t make sense for them to lock you to one particular provider, either.
Tarun, I know we have a hard stop in 2 minutes. I just want to ask you a hard question in closing: What do you think the world looks like with Hermes in a year? You mentioned enterprise and retail. It’s a lot of technical stuff. How do you envision the world a year from now—how people are interacting with it and using it?
I think there are a lot of people who have existing apps and are adding agentic functionality, like Robinhood. You’ve tweeted a lot about the Robinhood agentic experience not being as good. Maybe it’s changed after yesterday or whatever.
No, I’m using that more.
My belief is that every SaaS company either has to have a harness or be a router, because they need to get into the token order flow. Otherwise, they’re kind of screwed if they’re not in the token order flow game. That’s the easiest thing to monetize. In that world, Hermes is the white-label harness for anyone. No one wants to start from zero. There’s a security and brand risk: if your harness has been hacked, no one’s going to use it, much like a wallet. I think having this really strong brand means other people will want to use you as their backend beyond the prosumer market.
Mm.
That’s where I see the growth. One area is enterprise usage; another might be consumer usage. Another might be apps built on top of Hermes, because they don’t want to start from scratch.
So true. If every company becomes a harness, and Hermes is the best harness to build on, why does every company want to become a harness? Or why do they all need it?
They bring users and data, and they get a revenue share from someone who’s managing the token order flow for them. It’s really like a fintech and a bank. Most fintechs don’t have banking licenses, so they do revenue shares with the bank underneath.
Love it. Tarun, you took us down such a unique and differentiated view of Hermes today.
I could go on. I think this is where it looks like crypto.
No, it’s cool. No, it’s interesting, because a lot of people just talk about what they’re building, but—
The last thing I’ll end on, which is why I think this crypto thing is true, is that Muse is going to monetize in a very crypto way. It’s charging transaction fees on stuff you buy in the app.
And it’s basically doing that by doing revenue shares with Stripe and Shopify and whatever. How is that any different from this world I just presented? Not really, right? So I think a lot of the ways order flow management is done in crypto—in any open system where there are competing counterparties—will end up looking very similar to what crypto has. By taking the lessons you’ve learned from there, you can make some pretty, hopefully useful hypotheses.
I love it. Tarun, thank you for coming on. Huge congrats on leading this new round. We’re co-investors with you, and we’re so excited. I’m also excited for humanity, because this is open source and where I want the world to go and to be.
Your comments were just so unique. People come on and talk about what they’re building or what they want to do. You brought us down the eras of crypto—why this is so interesting, the differentiated build on the dev side and the app side, and the relationship to crypto and how it could be related. It’s really unique.
Definitely. Thanks so much.
So, I appreciate your time.