The Resurgence of Decentralized AI | Roundup

Decentralized AI has shifted from a position all three hosts were fading a year ago to a theme they now take seriously, driven by visible frontier model limitations and growing platform risk concerns. Miles argues the era of token-maximizing state-of-the-art models is effectively over because harnesses have become much better at smart routing, spreading usage across more models than even two or three months ago. Frontier labs spend roughly one hundred billion dollars training a state-of-the-art model expecting one hundred fifty to two hundred billion in return, but net new revenue from each successive frontier model is predicted to decline as users route cheaper models to lower-complexity tasks. Hardcore agent engineers are already largely using open source models for execution tasks and reserving closed source frontier models only for evaluation or planning.

Anthropic's Fable launch illustrated the tensions around frontier model restrictions. A team of reportedly Amazon researchers jailbroke it quickly after launch. Anthropic heavily restricted Fable use cases related to biology, basic science, and cybersecurity, prohibited using it to train other models, and was nerfing responses without indicating to users that responses were being downgraded. Anthropic and Thropa also announced data collection policies that could put healthcare providers using Claude in violation of HIPAA, and Fable is unavailable in Canada due to export controls. Whether these restrictions reflect genuine safety concerns or incumbency strategy remains disputed, with some viewing them as regulatory capture and others as legitimate cybersecurity concern.

Miles remains bearish on decentralized inference because it will always perform worse than centralized inference, but says decentralized training is more interesting than it was a year or two ago. GPU clusters are extremely expensive, and spreading that capital expenditure across many participants could break the hundred-billion-dollar training cost bottleneck. Projects including Prime Intellect, Pluralis, and Hermes have been training open source models for approximately two years and are expected to produce viable results soon. Jensen is noted to have identified the need to crack decentralized training approximately a year and a half ago, and Jake Bergman of Coin Finance is cited as an early advocate when it was a widely disbelieved concept.

The speakers draw a distinction between decentralized AI, decentralized training, and open source AI. Open source models such as those from Alibaba and Qwen currently offer better capability than decentralized models but can still censor users or change licensing terms. Censorship resistance at the business level, framed as platform risk elimination, is described as a stronger commercial driver than individual-level censorship arguments. PayPal is cited as an example of a large corporation taking open-sourced model weights and training its own finance and payments specific foundational model on proprietary transaction data. A key reason developers may want to leave a dominant AI platform is fear that the platform will turn their product into a feature, reducing their company to irrelevance.

Nous Research is described as arguably better than Claude for hard research tasks, built by an open source and decentralized team operating with crypto ideals. Its strategy is to use a harness as a user-facing wedge that owns the customer relationship while routing to multiple open and closed source models, with plans to vertically integrate by routing to its own model once it reaches competitive quality, at which point revenue is expected to increase significantly. The harness is compared to a car while the model is the engine, and significant value in decentralized AI is expected to accrue at the harness layer. Nous Research is valued at approximately one billion dollars and has crypto investors including Polychain.

Venice is described as probably the most talked about crypto project at the moment and possibly the best performing consumer crypto application currently. Founded by Eric Hughes, an early crypto ethos figure, Venice operates with client-side processing so user information does not go server side, gives users access to open source models, and removes user identity from queries to provide a privacy and anti-censorship layer. Venice is characterized as the first of many apps that will look like regular apps with crypto mechanics improving the back-end user experience, and it is described as the only current way to get token exposure to something with real usage at the crypto and AI intersection.

The speakers argue the bottleneck for AI adoption is not compute but product creativity, and that the industry has not moved past a naive chatbot user experience. SaaS companies routing users through the same underlying models are creating redundant integrations rather than differentiated value. Privacy and security in agentic AI workflows is identified as a wedge opportunity for startups against hyperscalers, with cryptography seen as particularly well suited for organizational AI systems that touch sensitive data. Decentralized AI is expected to be a major theme in the next crypto cycle, with wearables cited as a likely coming catalyst that will increase demand for local AI models. The speakers predict that if Nous Research and similar companies launch tokens, they will be significant launches following demonstrated product-market fit, with the caveat that it remains unclear whether crypto AI teams will launch tokens if they pivot hard into serving Fortune 500 customers exclusively.