SayGM

SN28 SayGM is a resellers' market for the big AI models, where brokers holding their own Claude, GPT, and Gemini accounts underbid the list price for the same answer, and the deal is done inside a sealed booth.

The commodity, explained from zero

The commodity is inference from the biggest closed models: sending a prompt to Claude, GPT, or Gemini and getting the reply, metered per million tokens, the word-pieces those models count. SayGM sells the same replies for less than the makers' list prices. Its front page says "27% Cheaper than list, on average" and "Never more than list," with a live price table.

It can be cheaper because SayGM does not run the models. Miners are the brokers: "you bring your own provider capacity, bid a discount off gm's retail price, and earn on the traffic you serve."

Who buys: developers. The front page tells them to "Point Cursor, Cline, Claude Code, or your AI SDK at the gm gateway. One base URL. No rewrites." No customer is named, but the miners page publishes every finalized epoch: the one ending 2026-09-10 02:17 UTC carried $798.51 of traffic across 5,219 requests and 42 miners.

Why it is on Bittensor at all

First, price competition: instead of one middleman's margin, many miners bid against each other, and the miners page shows each one's discount. Second, privacy, which the subnet states as its difference: "Verified by hardware, not by policy." Every prompt is decrypted only inside a trusted execution environment, a locked compartment in an Intel processor that can prove what code it ran, so neither SayGM nor the miner's machine can read it. Here the comparison breaks: a ticket broker knows exactly who is travelling, while at SayGM neither the gateway nor the host machine is meant to see the buyer's prompt or the miner's own provider keys.

How the work gets done

Miners rent a confidential virtual machine on Phala Cloud, load their provider API keys into it, and declare which models they serve at what discount; "a 10% discount means you keep 90% of each per-Mtok dollar." The gateway routes buyer requests to miners, checks that each reply echoes the model it claims to be, and records reliability. What the materials do not describe is how validators turn that into on-chain weights: the miners page says every earning is "recomputable from the public epoch artifacts," and the README covers attestation and key handling, but no document states a formula for emission, the subnet's share of newly minted TAO (inferred: the share and reliability columns on the miners page are the inputs).

How you would know it works

The miners page at https://saygm.com/miners shows a dated, finalized epoch with traffic value, request count, and per-miner share, reliability, and discount. Each model page, such as https://saygm.com/models/claude-sonnet-4-6, carries benchmarks dated 2026-08-20 and measured throughput over recent requests.

What is missing

There is no incentive document: nothing in prose says how weights or emissions are set. No hardware or Phala cost is stated for miners, and the registration cost is not mentioned. There is no status page or SLA; the terms promise no uninterrupted access. No Discord is set on chain.

Go deeper

• Front door with live prices: https://saygm.com/
• Miner overview, updated August 2026: https://docs.saygm.com/mining/overview/
• Per-epoch miner economics page: https://saygm.com/miners

Metaphor: a resellers' market where brokers underbid the list price for the same seat. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.