Rafay Token Factory

Turn every GPU-hour into token revenue

A growing share of AI is bought as tokens through an API. Rafay Token Factory serves models on your GPUs, meters every token, and runs alongside GPU rental on one control plane.

Open laptop displaying an operations console dashboard with token usage statistics, model deployment settings, and charts.
Why it matters

Get more value from every GPU-hour

The same GPU-hour can earn very different amounts. The difference comes from what the operator sells. Rented capacity earns a set rate for time, and contract discounts and idle hours can reduce that rate further. When the operator sells model output as tokens, the hour earns based on how much the GPU produces.

GPU rental
Rafay Token Factory
Billing unit
The hour, whatever the GPU produces
The output, metered as tokens
Revenue ceiling
Fixed at hourly rate × hours rented
Rises with throughput, model mix, and demand
Faster GPUs
Current
Faster
Earn the same hourly rate
Current
Faster
Serve more tokens and earn more each hour
Customers per GPU
One tenant at a time
Many tenants billed on the same GPU through shared endpoints
Pricing power
One rate
One hourly rate per GPU type, set by a crowded market
LightStandardPremiumReasoning
A price for each model, with higher prices for premium and reasoning models
Both run on the same Rafay control plane
Token Factory capability

Route every request by policy with Rafay Token Router

Token Router checks each request against capacity, site health, budget, and provider rules, and then sends it to the right model.

Burst capacity+
Users
→
Token Router
→
Self-hosted model
+
Public cloud

Your self-hosted model gets traffic first. Above its capacity, requests burst to AWS Bedrock, Anthropic, OpenAI, or any OpenAI-compatible provider. Bursting happens only when your policy allows it.

Multi-datacenter high availability+
Users
→
Token Router
→
Data center 1
⇄
Data center 2

One endpoint load-balances across data centers or edge sites. If one site goes down, traffic moves to a healthy site.

Token budgets by team+
Teams
→
Token Router
→
Premium model
or
Economy model

Teams use premium models within a daily or monthly budget. At the limit, requests fall back to an economy model and keep being served.

One endpoint for all providers+
Users
→
Token Router
→
Token Factory
Anthropic
OpenAI-compatible

Users reach models from many providers, including Rafay Token Factory, through one endpoint. The routing policy decides which provider serves each request.

Interactive demo: token budgets with Rafay Token Router
Global demand

Token demand is going global. So are Rafay Token Factories.

North AmericaTier 1 telcoAfricaMultinational technology groupMiddle EastTelco group
~70% lower cost
After moving AI workloads away from hyperscaler hosting
~10x token growth
Within weeks of deployment, with multiple business units sharing one Rafay Token Factory platform

Four levers that decide what a GPU earns

01Token price

A catalog with premium and reasoning models raises the average price per token.

02Tokens per GPU-hour

A better serving setup gets more output from each GPU. Rafay scales replicas automatically as demand grows.

See it in practiceNVIDIA Dynamo on Rafay Token FactoryRead the post →
03Utilization

Busy GPUs earn more. One shared pool lets one tenant's peak fill another tenant's quiet time.

04Model choice

More models means more workloads you can serve. Add, meter, and bill new models from one control plane.

Go deeperEconomics of the Rafay Token FactoryRead the white paper →
Open-weight models

Put new open-weight models into service quickly

Capable open-weight models arrive every few weeks, and many workloads no longer need costly frontier APIs.

Rafay Token Factory lets operators add each new model and serve it to tenants right away.

Add a model, start billing+

Publish a new open-weight model to the catalog, then meter and bill it from the same control plane.

Price each model on its own terms+

Set pricing per model and track usage per tenant, from lightweight models to reasoning models.

Approve the models you trust+

Curate the catalog, control which tenants use which models, and keep data in-country with auditable usage.

How it works

From endpoint to billing data in four steps

Choose a model

Pick a model from the catalog and publish it as an endpoint on your GPUs.

→
Open access

Give each tenant access through an OpenAI-compatible API, so apps that already use that format can connect.

→
Meter usage

Rafay counts input and output tokens for each tenant and model, in real time.

→
Bill tenants

Export usage data that is ready for your billing process.

Interactive demo: the end-user experience

Everything you need to run inference as a business

Rafay provides the operational and economic control plane to deliver inference as governed, scalable AI services.

One catalog for every model+

Publish, meter, and bill open models from one place. Each model you add becomes a new product.

Token-level metering+

Count input and output tokens by tenant, model, and endpoint for usage visibility, cost allocation, and billing.

Shared and reserved endpoints+

Serve on-demand traffic from shared endpoints, and give tenants reserved endpoints with guaranteed capacity.

Elastic, policy-driven scaling+

Inference autoscales with demand. Tenant quotas and rate limits keep consumption within policy.

Flexible pricing models+

Price per million tokens or per request, packaged into tiers and on-demand, reserved, or committed-spend plans.

Runs alongside GPU rental+

One control plane moves capacity between rental and token serving, so hours rental leaves empty can still earn.

Who it is for

Built for how you sell AI

Rafay Token Factory runs on the same control plane for every operator. What changes is who buys the tokens and how you sell them.

GPU hours, one dayGPU rentalToken serving fills the gaps

Neoclouds

More revenue from each GPU, with rental contracts left in place.

What you can do+
  • Sell tokens on the GPU hours that rental leaves empty.
  • Add a per-token product next to GPU rental on one control plane.
  • Offer new open-weight models to developers as they arrive.
IN-COUNTRYYour data centerApproved modelApproved modelNot approved

Sovereign AI clouds and telcos

Model choice under your own rules, with all inference kept in-country.

What you can do+
  • Approve which models each tenant can use.
  • Keep inference in your own data centers, under your own policy.
  • Keep auditable usage records by tenant and model.
One internalendpointBusiness unit ABusiness unit BBusiness unit CBars show token budget used

Enterprises

Control token spend and allocate cost fairly across business units.

What you can do+
  • Give every team one internal, OpenAI-compatible endpoint.
  • Set token budgets by team with Rafay Token Router.
  • Allocate cost by team and model from token-level metering.

Token Factory FAQs

An AI Token Factory transforms inference infrastructure into token-based AI services, governed and monetized at the unit of consumption. It shifts organizations from managing compute capacity to delivering measurable AI outcomes as scalable, revenue-generating services.

What is Rafay Token Factory?+

Rafay Token Factory turns the GPUs you already operate into metered model services. You publish models from a catalog as endpoints, give tenants access through an OpenAI-compatible API, and meter every token for billing. It is part of the Rafay Platform and runs on the same control plane as your GPU rental.

How is selling tokens different from renting GPUs?+

GPU rental sells time on a GPU at a set hourly rate. Token Factory sells model output, metered as tokens. With tokens, a faster or busier GPU earns more each hour, because it serves more tokens.

Can I run Token Factory and GPU rental on the same infrastructure?+

Yes. Both run on the same Rafay control plane. You can move capacity between rental and token serving, so GPU hours that rental leaves empty can still earn.

Which models can I serve?+

You can publish open-weight models from the catalog and add new models as they arrive. You set the price for each model and approve which tenants can use which models.

What is Rafay Token Router?+

Token Router sends each request to the right model, based on your policy. It checks capacity, site health, team budgets, and provider rules. For example, it can send requests to a public provider when your own capacity is full, or move a team to an economy model when it reaches its budget.

How do metering and billing work?+

Rafay counts input and output tokens by tenant, model, and endpoint, in real time. You can price per million tokens or per request, in tiers and as on-demand, reserved, or committed-spend plans. You can then export usage data that is ready for your billing process.

Can I keep inference in my own country or data center?+

Yes. You serve models on your own GPUs, approve the models each tenant can use, and keep auditable usage records. Token Router sends requests to an outside provider only when your policy allows it.

Do my tenants need to change their applications?+

Tenants call your endpoints through an OpenAI-compatible API. Applications that already use that API format can connect to your endpoints.

“Token Factories are the new cellphone companies. Similar to how cellphone companies used to sell pre- and post-paid minute plans, AI Factories are beginning to sell pre- and post-paid token plans. Team Rafay is looking forward to supporting the success of a thousand AI factories across the world with our Token Factory offering.”

Haseeb Budhani
Haseeb Budhani
CEO and Co-founder
,
Rafay