From "usable" to "useful"
Large models entered production for text generation, code assistance and Q&A.
AI Token Factory Solution
Turn AI Infrastructure into Governable Token Services
EasyStack AI Token Factory is a production-grade solution built on ECF (Cloud Infrastructure) and EAF (AI Infrastructure). It transforms GPU, NPU and DCU capacity into OpenAI-compatible token APIs with unified compute governance, model lifecycle management, intelligent routing, token metering, chargeback and enterprise-grade security.
Executive Summary
Industrialize token production from private AI infrastructure
AI is moving from experimentation to industrialization. As inference workloads are projected to exceed 70% of AI compute demand by 2028, enterprises and service providers need a way to turn heterogeneous AI infrastructure into standardized, governable and monetizable token services.
Large models are no longer the end product — they are the engine. The real output is tokens: the unit of AI consumption, cost, governance and business value. The AI Token Factory combines ECF for cloud foundation, EAF for heterogeneous AI compute and MaaS, and a dedicated Token Service Layer for API gateway, metering, billing and governance.
An open, hardware-agnostic, sovereign-ready AI token factory that lets every enterprise or cloud provider produce, control and monetize AI tokens at scale.
Token industrialization
Produce standardized, OpenAI-compatible token services from private AI infrastructure.
Higher utilization
Improve accelerator utilization from ~25% to 70–80% with pooling and intelligent scheduling.
Model-to-API in 10 minutes
Deliver model-to-API services quickly with pre-configured engines and one-click deployment.
Sovereign & auditable
Maintain data sovereignty, compliance and full auditability with multi-tenant isolation.
The Token Economy
From large models to token services
AI application evolution has moved through three phases, each raising the bar for infrastructure.
From "general" to "specialized"
RAG, fine-tuning and prompt engineering adapted models to vertical scenarios.
From "conversation" to "execution"
AI Agents perform complex tasks, call tools and invoke models repeatedly — creating bursty, non-deterministic inference demand.
Challenges
Five core challenges for token factories
Low GPU utilization
Coarse allocation and isolated workloads leave accelerators underused.
Fragmented model serving
Separate systems for compute, inference, gateway and applications.
Heterogeneous complexity
NVIDIA GPU, Hygon DCU, Ascend NPU and others require unified management.
No token-level cost visibility
Token consumption is not tracked by user, project, department or model.
Security & sovereignty risks
Public AI services create data leakage, compliance and lock-in concerns.
Solution Architecture
Three-in-one plus a token service layer
Compute governance, model service, a dedicated token layer and an AI application platform — open, hardware-agnostic and sovereign-ready.
Compute Governance
Unified management, scheduling and observability of heterogeneous AI accelerators.
- Unified pooling of NVIDIA GPU, Hygon DCU, Ascend NPU and emerging accelerators
- Fine-grained partitioning with NVIDIA MIG and HAMI software sharing
- Binpack, Spread and topology-aware scheduling in under 2 seconds
GPU · NPU · DCUMIG + HAMITopology-aware
Model Service (MaaS)
Model lifecycle, inference engines, deployment and unified gateway.
- One-click import from HuggingFace and ModelScope with version management
- Production-grade engines: vLLM, SGLang and Llama.cpp
- Fine-tuning with PEFT (LoRA) and a seamless path to production inference
vLLMSGLangLlama.cppLoRA
Token Service Layer
OpenAI-compatible API, intelligent routing, token metering, billing and quota.
- OpenAI-compatible /v1/chat/completions — existing apps connect unchanged
- Intelligent routing by weight, cost, latency or complexity with automatic fallback
- Token metering, chargeback, showback and external monetization
OpenAI APIRoutingMeteringBilling
AI Application Platform
RAG, Open WebUI, Agent templates and one-click deployment.
- Enterprise knowledge-base RAG with private document retrieval
- Open WebUI conversational interface with multi-model switching
- AI Agent secure container sandbox for multi-step business tasks
RAGOpen WebUIAgent sandbox
Core Capabilities
Eight capabilities that make tokens producible and governable
Capability
What the AI Token Factory delivers
One Cloud, Multi-Accelerator
Unified infrastructure for NVIDIA GPU, Hygon DCU and emerging accelerators with PXE auto-discovery.
Fine-grained Partitioning
NVIDIA MIG hardware profiles plus HAMI software compute/VRAM sharing; DCU and NPU passthrough.
Intelligent Scheduling
Binpack, Spread and topology-aware strategies avoid cross-NUMA/PCIe performance loss.
Production Inference Engines
vLLM, SGLang and Llama.cpp pre-configured for online, reasoning and edge workloads.
Unified Token Gateway
OpenAI-compatible API with intelligent routing, fallback and fine-grained rate limiting.
Token Metering & Billing
Input/output usage by API key, user, project, department and model with report export.
Multi-tenant Governance
RBAC, namespace isolation, quota management and audit logs retained ≥180 days.
AI App Marketplace
Versioned application templates with one-click deployment in under 10 minutes.
Customer benefits at a glance
Measured outcomes across compute, delivery speed, cost visibility and revenue.
75%+accelerator utilization
10 minmodel deployment, down from weeks
Fulltoken cost visibility by key, project, dept, model
Newrevenue stream for hosting & sovereign cloud operators
Enterprisegrade security and compliance
Compute & Token Platform
From pooled accelerators to metered token APIs
Four building blocks cover the full path from raw AI compute to a billable token service.
Compute Resource Pool
One cloud, multi-accelerator infrastructure with automatic discovery and unified governance.
- NVIDIA A100/H100/H200/L40/L20, Hygon BW1100/BW150 and emerging accelerators
- PXE auto-discovery reporting card type, VRAM, topology and system info
- Three-layer observability: node, inference instance and token business level
<60 snode-level fault alerts
3-layerobservability model
Virtualization & Scheduling
Fine-grained accelerator sharing plus intelligent placement for every workload profile.
- NVIDIA MIG profiles (1g.10gb, 2g.20gb, 3g.40gb) and HAMI software partitioning
- Binpack for cost-sensitive batch, Spread for latency-sensitive online services
- AI Workspaces with Jupyter, VS Code remote and SSH in under 3 minutes
<2 sscheduling decision time
75%+single-card utilization
Model Service & Token Gateway
Full model lifecycle behind a single OpenAI-compatible, intelligently routed gateway.
- Pre-configured DeepSeek, Qwen, Llama and other mainstream models
- Bind up to 10 backend models; route by weight, cost, latency or complexity
- Automatic fallback and QPS/TPS limits with whitelist/blacklist control
<10 msgateway latency increase (P99)
>5,000QPS per gateway instance
Metering, Billing & API Keys
Turn AI infrastructure from a cost center into a measurable, monetizable service.
- Input/output token usage by API key with multi-dimensional analysis
- Report export for chargeback, showback and external billing
- Keys visible only at creation, one-way hash storage, rotation in 5 seconds
<5 minstatistics latency
5 skey disablement / account revocation
Enterprise Governance
Sovereign, isolated and fully auditable
Governance and security controls keep every token inside your jurisdiction.
Data sovereignty
Models and data stay on-premises; input/output tokens never leave the private domain.
Multi-tenant isolation
Namespace, MIG hardware, Kata sandbox and network/storage/compute isolation.
Access control & audit
RBAC fine-grained control; audit logs ≥180 days with masked API keys.
Vulnerability management
Regular image scanning; high-risk repairs completed within 7 days.
High availability
99.9% control-plane availability; multi-replica inference with RTO <60 s.
Differentiated Competitiveness
Where the AI Token Factory stands apart
A neutral, on-premises alternative to cloud AI platforms and inference services.
Vendor typeRepresentativeCore characteristics
Enterprise Cloud AI PlatformRed Hat OpenShift AIKubernetes ecosystem, MLOps lifecycle
Public Cloud AI PlatformQingCloud AI CloudPublic cloud focus, compute scheduling and training
AI Inference ProviderSiliconFlowStrong inference cost competitiveness, rich API ecosystem
Hyperconverged VendorArcfraHyperconvergence driving AI
EasyStack AI Token FactoryECF + EAFOn-premises, open, neutral, token-level governance and monetization
Industry Practices
Four proven paths to a token economy
Explore how organizations across industries produce, govern and monetize AI tokens.
Cloud Hosting Provider — Token-as-a-Service
A regional hosting provider deploys ECF + EAF to offer OpenAI-compatible token APIs to SMEs. Token metering enables usage-based billing while multi-tenant isolation protects customer workloads.
Government — Sovereign AI Token Factory
A provincial government builds a private token factory on Hygon DCU and NVIDIA GPU. Citizen Q&A and policy interpretation run with response latency under 300 ms, and all data remains within national borders.
Financial Services — Internal AI Token Platform
A national bank deploys DeepSeek as an internal API service for investment advisory, compliance review and customer service. Token usage is tracked by department for cost allocation, improving service efficiency by 60%.
Manufacturing — Private Knowledge Token Service
A large manufacturer uses RAG and AI Agents for equipment fault diagnosis and maintenance planning, reducing fault handling time by 40%.
Ready to build your AI token factory?
Talk to our solution architects about turning heterogeneous AI infrastructure into standardized, governable and monetizable token services — or download the full solution whitepaper.