MCG Live · June 9, 2026 · 49 min
Dolphin is an AI lab that makes uncensored large language models and is building a distributed inference network using idle consumer GPUs. Founder Luke, formerly a GPU miner, explains how Dolphin became Venice's default uncensored model provider with 5 million monthly Hugging Face downloads, and how the peer-to-pool network model could undercut centralized inference providers like OpenAI by 25x on pricing.
Luke's background: ran a GPU mining facility in Australia with 400+ 1070 TIs mining privacy coins Beam and Grin in 2017, then took a break from crypto after FTX.
Dolphin sent a cold email to Venice suggesting their uncensored model, got a reply in 2-3 weeks, and became their default model after Venice users tested and preferred it.
Dolphin's refusal benchmark: ChatGPT and Gemini answer only 5-15% of 5,000 commonly refused prompts, while Dolphin answers 100% of them using reinforcement learning over 1-2 weeks instead of the faster obliteration technique.
Dolphin's most popular model is a de-censored Chinese open-source model with over 4 million downloads per month for at least 6 months, combining removal of both CCP censorship and refusal layers.
Distributed inference beta ran for 6 weeks with peak 600 RTX 6000 Pros, processing 3.5 billion tokens per day using synthetic data; version 2 launching in 2 weeks with remote updates capability.
Qwen 35B runs on single consumer GPUs and performs at 85% of Claude Opus quality while costing 25 times less ($1 vs $25 per million tokens on OpenRouter).
Peer-to-pool model lets GPUs join and leave without uptime requirements, unlike previous rental markets, enabling capture of idle gaming GPUs that are only used 4 hours per day.
Dolphin raised $900K seed funding 2.5 years ago pre-product and earned 40K $VVB across four Venice model partnerships; currently worth approximately $680K at $17 per token.
100% of inference revenue goes to buybacks of the $DOLPHIN token, which also receives emissions for node contributions; chose token-only model over split equity after observing Morpho's higher valuation than Aave.
Nodes paid 50 cents per million tokens for Qwen 35B, half the $1 OpenRouter price, targeting GPU owners' idle capacity rather than enterprise data centers who face Nvidia restrictions on consumer GPU use.
This summary was written from the recording's transcript by a language model. For the exact words, ask the MCG Live archive: every answer there is quoted word for word and linked to the second it was said. Market Bubble Search is independent and not affiliated with the show. Nothing here is advice.