GPU Token Platform for Data Centers

A data center with GPUs has compute. What it usually doesn’t have is a product: a way for a developer to sign up, get an API key, call a model, and pay for exactly the tokens they used. Without that layer, GPU capacity gets sold the old way, by the server and by the month, to the few customers big enough to run their own inference stack. We build every layer between your users and your GPUs so the capacity can be sold by the token.

What sits between your users and your GPUs?

Three layers, built on top of the cluster you already run.

• Access layer: an OpenAI-compatible API gateway with TLS and streaming (SSE), API keys issued and revoked per organisation and project, and RPM/TPM rate limits, spend caps, and alerts per key
• Token core: a model router that picks the right cluster for each model with load balancing and fallback, metering of input and output tokens, and a wallet and ledger that every request settles against
• Commercial layer: a pricing engine per 1M tokens with tiers and enterprise rates, billing by PromptPay, card, or transfer with invoices and e-Tax Invoice, and a developer portal plus an admin console with reporting

Underneath all three sit observability, an audit log, notifications, and a daily reconciliation job that checks usage, ledger, and payments agree.

Why metering and the ledger are the core

Everything a customer is billed for traces back to two components. Metering records every request’s input and output tokens as a usage event that can’t be edited after the fact. The wallet reserves an estimated cost before a request runs and settles the actual cost when it finishes, so a prepaid customer can’t run past their balance. Both write to a double-entry ledger, which means every baht can be traced from the invoice back to the requests that produced it. Get these two right and pricing, billing, and reporting are straightforward. Get them wrong and no dashboard will fix the disputes.

If you can’t trace an invoice line back to the requests that produced it, you aren’t selling tokens. You’re selling estimates.

Who runs the inference layer?

The inference layer serves the models with vLLM or SGLang, keeps a registry of models and versions, and reports token counts back to metering through an internal API. We agree who operates it at the start of the engagement.

• Option A: your team operates inference on your cluster, and we integrate the platform with it through the internal API
• Option B: HarmonyX deploys and operates the inference layer on your GPU cluster as part of the engagement

What stays with you

The data center itself: GPU servers procured by you or your vendor, the Kubernetes cluster with its compute, storage, and network, and the facility’s power, cooling, and physical security. The platform runs on top of that infrastructure. It doesn’t replace it.

If you have GPU capacity and no way to sell it by usage, start with an architecture workshop. We map your cluster, the models you want to serve, and who your customers are, then agree the scope, including who runs inference.