Xavier Lyu: Are AI Models Done Getting Smarter? Inference Optimization, Pareton SN10, dTAO | Ep. 103
About SN10 Pareton
In this episode, Xavier Lyu of Pareton (Subnet 10) details the development of a permissionless competition that optimizes the inference engines serving AI models, making the same hardware run meaningfully faster and cheaper. He discusses his thesis that model quality is converging and competition is shifting to cost and latency, the strategy of targeting underutilized non-CUDA hardware such as AMD, Intel and Chinese GPUs, and the architectural decision to shrink each campaign's surface area so miners search a smaller optimization space. Xavier also details his time at the Opentensor Foundation designing dTAO and what it feels like to now build under a mechanism he helped write, the mechanics of campaign-based competitions pinned to a single model and GPU class, the move toward vertical integration and on-premise deployments, and the unsolved problem of pricing a speedup for enterprise clients. *In this episode, you'll learn* - Why Xavier believes models are commoditizing and the real competition has shifted to cost and latency, which in business terms is margin - How a Pareton campaign works, pinning one model and one GPU class so miners optimize against a narrow stack rather than generalizing across hardware - The 54% throughput improvement one campaign produced in a single week, and why the harder problem is converting that result into paying clients - Why Pareton is chasing underutilized hardware, including AMD chips, Intel B70s, Chinese GPUs and repurposed Bitcoin mining rigs rather than the well-researched NVIDIA stack - The pricing problem of charging for an optimization, taking a cut of client savings versus a long-term B2B contract, and why neither answer is settled - Xavier's view that most subnets produce shiny benchmark results and fail to convert them into stable revenue, and what distribution actually requires - What designing dTAO at the Opentensor Foundation taught him about the limits of simulation, and how it feels to run a subnet subject to his own
Watch the video on Ventura Labs (YouTube) ↗
SN10 Pareton at a glance
SN10, Pareton (α token Ⱂ), is one of the independent markets on the Bittensor network, each with its own α token and its own price. Its α currently trades at 0.006481 τ (≈ $1.87), against 16,591 τ of pool liquidity. On-chain, the best-delegated validator on it yields about 35.84% APR (simple) and 43.10% APY (compounded), emissions add 0.000000 τ to the pool each block, epochs run every 360 blocks. It was registered at block #2,869,647.
More from SN10 Pareton
Taosis indexes announcements, articles, posts and releases about Bittensor subnets from their own channels and keeps a record of each. The text above is the source’s; the figures are read from the chain by Taosis.