GPU Token Platform to sell your data-center GPUs by the token

GPUs in a data center are usually sold the way servers always were: by the machine, by the month, to the few customers big enough to run their own inference stack. The GPU Token Platform turns that capacity into an AI service developers can actually buy. They sign up, get an API key, call a model through an OpenAI-compatible endpoint, and pay for exactly the tokens they used. The operator sees every request, every token, and every baht.

What does the platform include?

• Access layer: an OpenAI-Compatible API Gateway with TLS and SSE streaming, API keys per organisation and project, and RPM/TPM rate limits with spend caps and alerts
• Token core: a Model Router with Fallback across clusters, Per-Token Metering of input and output tokens, and a Prepaid Wallet & Double-Entry Ledger
• Commercial layer: pricing per 1M tokens with tiers and enterprise rates, Thai Billing & e-Tax Invoice by PromptPay, card, or transfer, and a developer portal plus an admin console with reporting
• Operations: observability, an audit log, notifications, and a daily reconciliation job across usage, ledger, and payments

Why is an OpenAI-compatible API the default?

Because that’s what developers already write against. Existing SDKs, agent frameworks, and tools work by changing the base URL and the key. Developers don’t need to learn a new API before they can buy capacity from you.

Metering and the ledger: the part customers dispute

Every request’s input and output tokens are recorded as a usage event that can’t be edited afterwards. Before a request runs, the wallet reserves an estimated cost, and when it finishes it settles the actual one, so a prepaid customer can’t run past their balance. Both sides write to a double-entry ledger, which means any invoice line traces back to the requests that produced it.

If you can’t trace an invoice line back to a request, you aren’t selling tokens. You’re selling estimates.

Where does inference run?

Models are served with vLLM or SGLang on your Kubernetes cluster, with a model registry for versions and a usage report sent back to metering through an internal API. Your team can operate that inference layer, or HarmonyX can operate it for you. GPU servers, the cluster, and the facility stay yours either way.

HarmonyX builds and integrates the platform on top of your existing GPU capacity. To scope a build, see our GPU Token Platform service.