Sell AI inference: requirements and costs
You can prepare to sell AI inference by checking your model rights, measuring the equipment you already have and calculating the full cost of serving requests. Start with these checks before committing hardware or relying on future demand.
By Gyblos · Reviewed
What you need before offering inference
Start with equipment you already have. Choose a model you can lawfully serve, measure it on your exact hardware and decide how much time and capacity you can offer. A working local model is a starting point, not proof of a working marketplace integration.
- Check the model licence and, for an upstream API, its resale terms.
- Record model version, quantization, context length and engine version.
- Measure memory use, response speed and simultaneous requests on representative, non-sensitive tasks.
- Estimate operating costs and choose hours you can reliably support.
- Document data handling, failure recovery and any upstream dependencies.
Selling inference is different from renting a GPU
When you sell inference, you operate the model engine and deliver responses to requests. The buyer selects a model offer rather than taking over your machine. You remain responsible for availability, licensing and your service’s data policy.
Gyblos is designed to connect an engine URL, including compatible local engines or an upstream API under the seller’s responsibility. Running Ollama does not establish that a particular model, streaming mode or tool-calling workflow is supported by Gyblos. Compatibility needs a tested release.
Measure your model, not just your GPU
Memory requirements depend on the model, quantization, context and concurrent work. Test the combination you intend to offer instead of treating a model’s download size or a single speed result as its full operating requirement.
Record the hardware and software versions, task, input/output token counts, context, concurrency and measurement date. Use synthetic or public test inputs. Keep prompts, responses, keys and customer data out of public reports.
Check connectivity and availability
Measure upload capacity, latency and uptime under the conditions in which you will serve requests. A fast local response does not measure the complete buyer experience.
Home routers and carrier-grade NAT can complicate direct connections. The Gyblos design includes direct connections where possible and relay fallback where needed; it does not guarantee that every connection will work. Account for any relay and bandwidth costs. Do not publish an unprotected engine URL or API key to test demand.
Count all costs, including idle time
Compare revenue and costs over the same period. If electricity or equipment costs are in another currency, use an explicit, dated conversion before comparing them with USDC revenue. Demand can be zero while costs continue.
| Cost | What to record |
|---|---|
| Electricity and cooling | Measured power in kW × operating hours × tariff per kWh; include idle hours. |
| Hardware | Depreciation or rental cost for the period, plus repairs and maintenance. |
| Connectivity | Bandwidth, internet plan and relay charges attributable to the service. |
| Upstream services | API usage, licences and charges for failed or retried work where applicable. |
| Fees and operations | Protocol and network costs, support time and applicable taxes. |
A margin example, not an earnings forecast
Assume the seller’s net inference amount is 10.00 USDC over a period. The planned 4% P2P fee is 0.40 USDC, so the buyer pays 10.40 USDC including that fee. The seller receives 10.00 USDC before operating costs. This illustration precedes atomic-unit rounding and network costs; the reserved buyer budget is not revenue.
For illustration only, subtract 2.40 USDC in electricity and 1.30 USDC in other operating costs for that same period: 10.00 − 2.40 − 1.30 = 6.30 USDC before depreciation, maintenance, support and taxes. Those remaining costs can eliminate the margin. These invented amounts are not current offers, measured earnings or a currency-conversion assumption.
Operating result = seller’s net inference amount − all operating costs. Run scenarios with no demand, occasional requests and sustained use. A listing, an idle GPU or a low token price does not guarantee requests or profit. Do not buy hardware based on this example.
Describe the offer and its limits honestly
State the exact model, input/output price units, availability and data policy. Gyblos uses prices per million input tokens and per million output tokens. Check what the displayed price includes and how failures, retries and rounding are handled.
The planned offer levels distinguish verified model, measurable usage and declared usage. They describe evidence about the model or usage, never a verified person. Statistical audits do not prove every response or factual correctness. In the direct route the seller processes and can read the prompt; P2P does not itself provide confidentiality from the seller.
Questions before you commit
Can I sell inference with Ollama? It is a candidate engine connection; a local model working in Ollama is not proof of a released Gyblos integration. Check the tested release and model licence first.
Do I need to buy a GPU? First measure what your existing equipment can do. This guide does not prescribe a hardware purchase or promise that it will pay for itself.
How much will I earn? That depends on paid demand, price, successful delivery and all your costs. There is no earnings figure established by the website examples.
Can I offer free inference? Giving access is part of the Gyblos design, with limits and availability to check. Public free offers cannot be created on this website today.
What you can do now
First seller beta invitations are planned for November 16, 2026. Contact @gyblos on X with your operating system, model and engine to request an invitation. Requests and invitations are handled manually. Public seller onboarding is not open.
Already invited? Follow the beta quickstart for installation options on your platform, then continue to registration and publication. Use the access details you received privately with your invitation.
Registration and payments use devnet and test assets, with no real payouts. Bring your own compatible setup on Windows, macOS, Linux or another compatible environment and cover your own operating costs.
This is a preparation guide, not an installation tutorial. Use the linked quickstart for setup; the public paper distinguishes built, tested and planned capabilities.
Sources and further reading
Engine documentation supports the preparation checks; it does not certify Gyblos compatibility. Read the exact licence for the model or upstream service you choose.