Skip to content
Ardelva
中文

Private AI · Open enterprise AI platform

Ardelva

Solution as a Generated Service.

An open enterprise AI platform. Privately deployed. Under your control.

Keep core knowledge inside the enterprise — and put tokens to work on the business with confidence.

Talk to usSee the test data
INFERENCE CLUSTERprofile v1.4switch · NIC · NCCL · vLLMPrefillDecodeP18 GPUNICP28 GPUNICD18 GPUNICD28 GPUNICLeaf · 400GLeaf · 400GSpine · 800GSpine · 800G

Product

AI already works. Core business needs three certainties.

Customers aren’t buying GPU clusters or an API. They are buying certainty that AI can be wired into core workflows — each certainty backed by evidence you can review.

  • 01

    Data & execution boundary

    Processing, retrieval, logs and tool calls all happen inside the control domain your enterprise chooses.

    Evidence · Data-flow map · permission matrix · egress audit

  • 02

    Domain knowledge assets

    Source knowledge and everything derived from it — answers, expert corrections, eval sets, adapters — stay in a closed loop, used only as you authorize.

    Evidence · Knowledge loop · retention rules · usage terms

  • 03

    Stable quality of service

    Critical workloads respond on time at peak; upgrades and failures are handled to plan, not improvised.

    Evidence · SLO report · capacity table · recovery drills

Platform

One platform. Every control point inside your boundary.

Open interfaces connect existing systems; knowledge, models and service quality are governed by a single layer. External APIs are an optional, approved egress — never a silent fallback.

Customer-selected deployment & data control domain

  • Business entry

    • Code & R&D assistants
    • Knowledge Q&A / internal workflows
    • Enterprise apps / agents

  • Governance & access

    • Identity / permissions / quotas
    • Routing / approval / audit
    • OpenAI- & Anthropic-compatible API

  • Knowledge & execution

    • Permission-aware local RAG
    • Controlled tools / isolated secrets
    • Eval sets / version management

  • Inference & resources

    • Open-weight model serving
    • Scale-out GPU pools
    • RoCE / caching / observability

End to end: data classification · egress policy · security logs · model regression · service targets · cost metering

Drop-in API

Change one base URL. Move from cloud to private.

One gateway serves OpenAI- and Anthropic-format APIs. Model names, prompts, tool definitions and session logic stay the same; identity and quotas are issued per person or department.

.env
OPENAI_BASE_URL    = https://ai.corp.internal/v1
OPENAI_API_KEY     = <key issued by your gateway>
ANTHROPIC_BASE_URL = same gateway
  • Coding agents: Claude Code · Codex CLI · Cursor · Cline · OpenCode
  • Agent / workflow frameworks: Dify · n8n · LangChain · LlamaIndex · OpenAI Agents SDK
  • Enterprise apps: IM bots · knowledge Q&A · internal integrations · RAG services

Data boundaries

Three levels of control. Private execution is the widest.

  1. Level 1

    Not used for training

    Constrains how data is used. Feedback, tools and contract scope still need checking.

  2. Level 2

    Zero data retention

    Constrains how long data is kept. “Not retained” is not the same as “not processed externally”.

  3. Level 3

    Private execution

    Inference, knowledge retrieval and logs sit in a control domain you choose — and you take on the matching operational responsibility.

Commodity GPUs · Recompiled Ethernet

Three product forms, from desk-side pilot to scalable pool.

Standard PCIe GPUs and Ethernet you can actually buy and cable. Same platform, same governance layer, same acceptance method — each form ships with its switch and a versioned RoCE network profile.

  • Form A · Compact

    4 / 8 × DGX Spark

    Memory

    512 / 1,024 GB unified memory

    Network

    CX204Q-2D16Y

    • 4 × 400G
    • 2 × 200G
    • 16 × 25G

    Fit

    1–8 interactive sessions

    R&D teams, departmental knowledge bases, air-gapped pilots.

    Scale-out

    • 4 units
    • 8 units

  • Form B · Enterprise cluster

    4 / 8 / 16 × RTX PRO 5000 72GB

    Memory

    288 / 576 / 1,152 GB GPU memory

    Network

    CX732Q-N

    • 32 × 400G

    Fit

    16–128 concurrent sessions

    department to business-unit RAG, agents and code assistants.

    Scale-out

    • 4 GPU · 1 node
    • 8 GPU · 2 nodes
    • 16 GPU · 4 nodes

  • Form C · Scalable pool

    8 / 16 / 32 × MetaX C550 64GB

    Memory

    512 / 1,024 / 2,048 GB GPU memory

    Network

    CX864E-N

    • 64 × 800G RoCEv2

    Fit

    64–256 concurrent sessions

    domestic-compute requirements, long-document workloads, shared multi-department pools.

    Scale-out

    • 8 GPU · 1 node
    • 16 GPU · 2 nodes
    • 32 GPU · 4 nodes

Recompiled Ethernet · model-defined networking

A versioned workload and fabric profile.

Starting from the workload and its QoS / SLO targets, we generate a versioned configuration for the whole fabric — pushed down to switches, NICs, NCCL and vLLM — then deploy it. When the model changes, the fabric is recompiled.

  1. 1

    Describe the workload

    • Model architecture: layers, experts, KV size
    • Parallelism: TP / EP / DP
    • P:D ratio, context length, concurrency
    • GPU and NIC inventory
  2. 2

    Compile the fabric

    • Traffic matrix per inference phase
    • Buffer sizing and PFC headroom
    • ECN / DCQCN thresholds
    • Priority, queue and ECMP plan
  3. 3

    Ship a versioned profile

    • Switch + NIC configuration artifacts, plus NCCL and vLLM parameters
    • Automated validation test suite
    • SLO report for customer sign-off
    • Re-compile when the model changes

AsterFlow is the profile workflow used in this solution: configuration, validation and change control.

Test data

Measured on Ethernet.

Headline results from our own test runs. Each figure comes with its key condition.

Form C · MetaX C550

10,046 tok/s

2 nodes · 1P1D · DeepSeek V4.0 Flash

Single node
5,300 tok/s

Form B · RTX PRO 5000

1,142 tok/s

4 GPUs · 16 concurrent · Qwen3.8-Flash-Next

GLM 5.3 Flash
808 tok/s

Signed-off tok/s

Performance you sign off on your workload — not a datasheet number.

Model evaluation

3 models avg / 100

  • Claude Opus 5.595.5
  • Qwen3.8-Flash-Next (xhigh)94.7
  • GLM 5.3 Flash (high)93.7

Six execution-graded systems-programming tasks; best tier per model.

Artificial Analysis Intelligence Index

  • Claude Opus 5.5 (max)58
  • GLM-5.3-Flash42
  • Qwen3.8-Flash-Next40
  • DeepSeek V4.1 Flash (max)39

Source: Artificial Analysis Intelligence Index v4.3.2, snapshot 2026-09-25; max = highest reasoning setting.

Unit economics

≤1.2×*

cost per million output tokens vs. DeepSeek API, at ~75% utilization

Synthetic workloads; results vary with model, workload and configuration. Full test reports are available on request under NDA. * Scenario estimate; details on request.

Deploy

Start with one high-value workflow. Deliver an operable service.

Solution as a Generated Service is not a one-off demo — it is a delivery model of continuous evaluation and iteration. You own business, data and security authorization; Ardelva and delivery partners own integration, adaptation and acceptance evidence.

  1. 01

    Define the business

    Fix the data boundary, task set, business quality and service targets.

    → Business & acceptance baseline

  2. 02

    Two-node validation

    Model adaptation, real-workload testing, security and cost review.

    → PoC evidence pack

  3. 03

    Go live

    Connect identity, data and tools; deploy monitoring, rollback and operations runbooks.

    → An operable service

  4. 04

    Run and expand

    Review quality, usage and cost; expand to 4 / 8 nodes once conditions are met.

    → Capacity & optimization plan