EAFEasyStack AI Infrastructure Suite
Put heterogeneous compute, model serving, and agents into production — on the open cloud infrastructure you already run.
EAF is EasyStack's enterprise AI infrastructure platform designed to build AI-ready infrastructure that is more integrated and delivers higher performance for AI applications. Built on ECF and ECNF, it adds unified management of heterogeneous GPU, NPU and DCU accelerators via the EOS digital native engine and ESCL Cloud Operating System, compute pooling with Volcano scheduling, and a model gateway with token metering. An existing ECF or ECNF deployment evolves smoothly into AI-ready infrastructure instead of growing a second, separate stack.
In the EasyStack family, EAF is the intelligent stage between foundation and agents. The four products — ECF for cloud infrastructure, ECNF for cloud-native workloads, EAF for AI infrastructure, and Cortex for the agent platform — deploy independently and combine on demand. EAF carries GPU pooling, training and inference scheduling, and the model runtime; Cortex builds industry agents and applications for energy, government, telecom, finance, education, and internet scenarios on top of it.
Everything below EAF stays shared: EOS as the cloud operating system, the ECF and ECNF base, and the physical resource pool of x86 and Arm CPUs, GPU, NPU, and DCU accelerators, NVMe and SSD storage, and 25/100/200 GbE with RoCE V2 networking. AI capacity is added without changing the operating model the platform already runs on, with 10 years of production experience across 100,000 nodes.
One Open Foundation
ECF, ECNF, EAF, and Cortex are four independent products: deploy them standalone, combine them on demand, and evolve an existing ECF or ECNF base into AI-ready infrastructure.
Inference Optimization Decides Cost
Infrastructure sets the ceiling; inference optimization decides the unit cost and experience of every token — from quantization and KV-cache management to continuous batching and prefill-decode disaggregation.
Open Source DNA
Envoy AI Gateway and Gateway API provide Token as a Service; vLLM, SGLang, PyTorch, KServe, and llm-d carry model serving; Kubernetes, Harbor, Kata Containers, Cilium, and Prometheus run underneath — enterprise grade.
Core components
Heterogeneous AI Compute Base
GPU, NPU, and DCU accelerators from NVIDIA, Ascend, Hygon DCU, and other vendors are managed as one pool on a one-cloud-many-chips basis, alongside x86, Arm, Kunpeng, Hygon, and Phytium CPUs. RoCE V2 fabric, high-performance storage, and NVMe-based model caching prepare the base for scale-out inference.
- GPU, NPU, and DCU managed as one pool
- One cloud, many chip ecosystems
- RoCE V2 fabric and NVMe model caching
MaaS Model Services
A model marketplace, repository, deployment pipeline, evaluation, and elastic scaling serve the model lifecycle, with an ecosystem spanning DeepSeek, Qwen, GLM, Kimi, MiniMax, and Gemma. Locally optimized open-source models are quantized to FP8, INT8, or INT4, tuned with LoRA, and served by vLLM and SGLang with KV-cache management and continuous batching.
- Model marketplace, repository, evaluation, elastic scaling
- FP8/INT8/INT4 quantization, LoRA, MoE, distillation
- vLLM and SGLang engines with continuous batching
Intelligent Inference Scheduling
Load-aware and cache-aware routing places every request on the best replica, elastic scaling tracks demand, and prefill-decode disaggregation separates prompt processing from token generation. The outcome is higher utilization and stability instead of isolated GPU islands.
- Load-aware and cache-aware routing
- Elastic scaling of inference services
- Prefill-decode disaggregation
Model Gateway and Token Operations
A unified gateway handles access endpoints, model routing, traffic and access control, and API keys — built on Envoy AI Gateway and Gateway API. Token usage is metered and billed per tenant, user, model, and API key, with audit logs and SLA reporting; local models carry high-frequency, privacy-sensitive traffic while cloud models absorb high-value, complex workloads.
- Unified access, rate limiting, quotas, and audit
- Token metering and billing per tenant, model, and key
- Local and cloud model scheduling by task value
Ready to join our partner network?
Would you like to talk about opportunities? Reach out and let's explore how we can grow together.
Become a Partner