Sell AI inference from your own GPU
Selling inference means serving model requests from hardware you control and charging for the tokens processed. Gyblos is designed to let independent providers, including home GPU owners, publish their own offers and receive USDC on Solana. Public seller onboarding is not available yet.
Selling tokens versus renting out a GPU
With GPU rental, a customer leases compute capacity and manages a workload. With inference, you operate the model engine and expose a request interface. You are responsible for the model, availability, licensing and data policy. Buyers choose a model offer rather than administer your machine.
What to prepare
Choose a model you have the right to serve. Check memory needs for its exact weights, quantization and context length, then measure throughput and concurrency on your hardware. An engine such as vLLM or Ollama can serve local models, but a working engine alone does not prove Gyblos integration. The exact supported combinations must be published with tested releases.
Connectivity from home
Home routers and carrier-grade NAT can prevent inbound access. The Gyblos design includes direct connectivity where possible and relays where needed. This is planned networking behavior, not a guarantee that every connection works. Measure upload bandwidth, latency and availability, and account for relay costs before offering service.
Price input and output separately
Define prices in USDC per million input tokens and per million output tokens. For illustration, 1,000 input tokens at 0.10 USDC per million and 500 output tokens at 0.30 USDC per million cost 0.00025 USDC together. These are arithmetic examples, not current Gyblos offers. Published prices must make the protocol fee clear.
Check the economics before committing hardware
Revenue depends on actual paid demand, uptime and successful delivery. Subtract protocol and relay fees, electricity, bandwidth, cooling, hardware depreciation and maintenance. A listed model or an idle GPU does not guarantee requests or profit. Do not buy hardware based on illustrative marketplace activity.
Reputation and model audits
The design separates measured service reputation from the verified-model seal. Any seller-declared model may be offered under the seller’s licensing responsibility, but only eligible open-weight offers can be audited against public weights. Stake is optional; it does not itself establish service quality or model authenticity.