SayGM

LLM providers charge different rates for identical models, and buyers pay retail without access to cheaper sources. Infrastructure operators running inference workloads have no way to monetize spare capacity or arbitrage pricing differences across providers. Centralized gateways also create vendor lock-in and expose request traffic to intermediaries.

A decentralized network where independent operators supply inference capacity from multiple providers simultaneously solves this through cryptographic attestation and per-request settlement. Each operator deploys a secure container that forwards buyer requests to their chosen upstream provider, keeps their API keys encrypted inside a TEE, and earns the difference between buyer retail and their upstream cost. The network's registry tracks which models each operator serves, validates their infrastructure through ARM/Azure configuration checks and image measurement attestation, and settles earnings on-chain.

gm implements this on Bittensor (subnet 28) by having miners deploy container workloads to Phala Cloud's confidential VMs, configure API keys for Anthropic, OpenAI, Google, Chutes, Z.ai, Moonshot, DeepInfra, KubeTEE, Engy, Moonmath, NEAR, and cloud backends (Azure OpenAI, Foundry, Bedrock), and declare which models they serve at what discount. The gateway (also TEE-secured) routes each buyer request to an eligible miner's declared model, verifies the response echoes the correct model identity, and the registry settles what the buyer paid minus the miner's declared discount.

gm is a Bittensor subnet (netuid 28 mainnet, 482 testnet) where miners supply inference capacity from upstream LLM providers (Anthropic, OpenAI, Google, and 8+ others) through confidential VMs, while buyers access identical APIs through a gm gateway. Miners configure provider API keys, deploy containers to Phala Cloud, declare supported models, and earn the price spread between buyer and upstream costs.