Question 1

What upstream providers can a gm miner source inference from?

Accepted Answer

Miners can supply capacity from Anthropic (direct API or AWS Bedrock or Microsoft Foundry), OpenAI (direct API or Azure OpenAI), Google (Gemini), and ten sourcing upstreams: Chutes, Z.ai, Moonshot, DeepInfra, KubeTEE, Engy, Moonmath, NEAR, Bedrock, and Foundry. Each provider's API key is set once via gmcli, encrypted into the TEE at deploy time, and never exposed to gm operators or the gateway.

Question 2

How do miners get paid and how is the payout calculated?

Accepted Answer

Miners declare a discount percentage (0–99.90%) on each model they serve. Settlement is: miner rate = buyer retail price × (100% − discount) / 100%. For example, at 5% discount on a model retailing $1.40/input Mtok, the miner receives $1.33/Mtok. The registry resolves the percentage against its own current retail price, so miners earn the declared spread even if retail prices change. A miner's profit is that figure minus whatever the upstream provider actually charges them for the same tokens.

Question 3

Can one miner supply the same model from multiple upstream providers?

Accepted Answer

Yes. A single worker can declare multiple routes to the same buyer model by configuring different upstream providers (e.g., DeepSeek V4 Flash from both DeepInfra and Engy). The gateway picks the cheapest eligible route per request. To serve two different upstreams for the same model concurrently, run two separate workers under the same hotkey using gmcli worker add.

Question 4

How does gm verify that miners are honest about what they can serve?

Accepted Answer

The registry runs periodic capability probes against each declared product. For direct API routes, it sends a 1-token request and checks that the response echoes the correct model. For Azure OpenAI and Foundry routes, gmcli doctor and the registry's attestation server verify through ARM that a deployment named X actually serves model X with the correct format (OpenAI or Anthropic), using exact ASCII matching. A catalog-named mismatch blocks boot or takes a worker offline when detected.

Question 5

What attestation guarantees does gm provide to buyers?

Accepted Answer

Each miner runs in an Intel TDX confidential VM. At startup, attestd verifies that the miner's cloud account (Azure, AWS) has no telemetry, diagnostic settings, or Application Insights connections that could expose request content. The TEE measures the miner's container composition and OS image; the gateway verifies this measurement in attestation quotes. The gateway also checks that each response from a miner echoes the correct model ID, proving the upstream was actually called. For NEAR routes, the in-image verifier validates the NEAR endpoint's Intel TDX quote before forwarding any request.

Question 6

Do I need to modify my code to use gm as a buyer?

Accepted Answer

No. Buyers use standard OpenAI or Anthropic Python/TypeScript SDKs pointed at the gm gateway URL instead of the provider's URL. The gateway translates requests and responses transparently, so existing code works unchanged. API keys and model names in buyer code remain the same; the gateway routes them to eligible miners behind the scenes.

Question 7

How are API keys managed and kept secure?

Accepted Answer

Miners set their provider API keys once using gmcli set-api-keys. Keys are encrypted and baked into the TEE container at deploy time; they never appear in logs, config files readable outside the TEE, or in network traffic. gm operators and validators never see the keys. Multi-key slots allow one provider to have up to 8 semicolon-separated keys, and the gateway selects a slot per request using the x-gm-upstream-slot header. If no slot is specified, the first key is used.

Question 8

What happens if a miner's upstream provider account is suspended or rate-limited?

Accepted Answer

The registry runs periodic capability probes on each declared product. If a probe fails (e.g., invalid API key, rate limit, or account suspension), the registry detects it and removes that product from the miner's eligible offerings. Buyers will stop seeing that model from that miner. The miner remains registered but cannot earn on that product until the probe passes again. Persistent failures may require the miner to update their API key or choose a different upstream.

Question 9

What is a 'sourcing route' and how do I pick one?

Accepted Answer

A sourcing route is one of multiple upstream providers serving the same buyer product. For example, GLM-5.2 is available from Z.ai, DeepInfra, Engy, KubeTEE, and Moonmath. Each route has different latency, pricing, and regional availability. gmcli sources lists all routes you can serve, their upstream hosts, and your current worker eligibility. You pick a route by setting the upstream key (gmcli set-api-keys --engy sk-...), deploying, and declaring that route (gmcli declare-product --provider engy --model glm-5.2 --discount-pct 5). The gateway routes each request to your cheapest eligible option.

SayGM

SayGM

Hardware Requirements

Hardware Requirements

Miner Setup

Participate

Participate

Hardware Requirements

Hardware Requirements

Miner Setup