Robot Ventures just led a $90 million round in Nous Research, alongside Microsoft and Nvidia — and Nvidia has already made Nous's software the default on its DGX Spark machines.
Tarun Chitra, founder of the risk-management firm Gauntlet and managing partner at Robot Ventures, told The Delphi Podcast that two earlier waves of crypto people trying to build AI both got it wrong. Nous Research got it right, he said, by doing the opposite of what crypto usually does: no token, no hype, years of narrow research most investors never bother to check.
"OpenClaw is kind of the worst developer experience of anything I've used personally... it makes you have to make too many decisions to use it." — Tarun Chitra
Tarun also diligenced this bet the way he runs Gauntlet, which advises some of decentralized finance's biggest protocols on risk: rather than trust revenue figures he calls unreliable across AI right now, he counted GitHub forks of Nous's agent software to gauge real demand.
The full interview is covered here so you can skip it. 1 hour 16 minutes of audio, 9 minutes of reading.
Here are the 7 takeaways that matter.
Delphi · Oct 26, 2026 · 1h 16m · 67m saved
With Tarun Chitra, Founder of Gauntlet and Managing Partner at Robot Ventures
Key Takeaways
Robot Ventures diligenced a $90 million bet by counting GitHub forks, not revenue
Chinese robotics manufacturers had forked Nous's Hermes software to run embedded hardware without Nous even knowing
OpenClaw lost to Hermes on product design, not on underlying technology
Tarun calls OpenClaw "the worst developer experience" he's used; Hermes auto-generates skills and learns a user's preferences instead
Crypto's first two AI bets both failed, for different reasons
2023's GPU-rental projects assumed language models would behave like image generators; 2024-25's token-launched AI agents shipped a year before usable agent software existed
Nous Research never launched a token, which Tarun says is exactly why it worked
He says Nous and Prime Intellect were "the only people doing real research" among decentralized-AI projects
Tarun's AI market-structure thesis maps directly onto crypto: the money concentrates at the top and the bottom
Harnesses sit on top like a crypto wallet; inference providers sit on the bottom like liquidity providers; routers in the middle get squeezed
Some inference providers are gaming their cached-token pricing on OpenRouter, quoting a low headline price while not passing on the savings of caching repeated calls
Every SaaS company will eventually need to run its own AI harness or become a router, Tarun argues, because owning "token order flow" is the easiest layer left to monetize
1. The $90M Nous Bet
Robot Ventures led Nous Research's $90 million funding round, with Microsoft and Nvidia both participating alongside it.
Nvidia already ships Nous's software by default on its DGX Spark machines
"NVIDIA and Microsoft both participated in this round... NVIDIA has installed it on their hardware by default for DGX Spark machines." — Tarun Chitra
Tarun says Nous's open-source inference router has served a cumulative 2 trillion tokens, and that tens of millions of unique GitHub accounts access its products every day — a figure he treats as a reliable usage signal because, unlike most social platforms, running multiple GitHub accounts is costly rather than free.
