Ai TokenAi Token

Deployment Options

Pick the deployment model that matches where you are

One AiToken gateway, six delivery shapes: PoC, standard production, cluster scale, Xinchuang full stack, hybrid cloud and turnkey appliance. Start where you are and scale smoothly as usage grows.

Ranges below are indicative; final configuration and pricing depend on on-site assessment and POC results.

Deployment Options

Six delivery shapes

Different compute, model size, concurrency and compliance requirements lead to different rollouts. Pick the closest fit — or combine them.

1–2 weeks

PoC Validation

Single-node Docker Compose — the lowest-cost way to validate a use case or a bid POC.

  • Single-node vLLM inference + AiToken gateway (OpenAI-compatible)
  • PostgreSQL / Redis / MinIO orchestrated by one compose file
  • Built-in model management, key distribution, usage and audit logs

Indicative budget

CNY 150k–400k (hardware excluded)

Best for

Internal RAG Q&A, single-department pilot, pre-sales POC

3–4 weeks

Standard Production

Three-node high availability for day-to-day production traffic.

  • Active-active gateways with load balancing and automatic failover
  • Full-stack monitoring and alerting with Prometheus / Grafana / Loki
  • Scheduled backups, canary upgrades and rollback runbooks

Indicative budget

CNY 600k–1.5M (hardware excluded)

Best for

Support assistant, coding copilot, enterprise document Q&A

6–8 weeks

Cluster Scale

Kubernetes with multi-node, multi-GPU serving for group-level concurrency and MoE models.

  • GPU Operator + LeaderWorkerSet for multi-node inference
  • vLLM / SGLang dual engines selected per model workload
  • HPA autoscaling, tenant isolation, quotas and rate limiting
  • Dedicated SRE on-call and capacity planning

Indicative budget

CNY 2M–5M (hardware excluded)

Best for

Group multi-tenancy, model-as-a-service, external AI platform

4–6 weeks

Xinchuang Full Stack

Domestic chips, OS and databases for strict compliance and air-gapped delivery.

  • Kylin / UOS with Dameng / KingbaseES / OceanBase adaptation
  • SM2/SM3/SM4 cryptography and MLPS Level-3 hardening
  • Offline installer, source-level audit and SBOM

Indicative budget

CNY 1.5M–4M (domestic compute adaptation included)

Best for

Government, finance, energy and defense with strict compliance

~1 week

Hybrid Gateway

No GPU build-out: deploy only the gateway and audit locally, with models over private links or dedicated instances.

  • AiToken Gateway + audit + semantic cache + DLP masking
  • VPC private link / PrivateLink, data stays in your domain
  • Local small model fallback keeps sensitive requests on-prem

Indicative budget

CNY 100k–300k/year + usage-based billing

Best for

No GPU room yet, compliance-first rollout, or spiky traffic

3–5 days

Turnkey Appliance

Pre-integrated hardware and software, ready out of the box for branches and edge sites.

  • Pre-loaded AiToken + vLLM + mainstream open models
  • 3-year warranty, remote operations and spare parts
  • Power-on ready, fully air-gapped capable

Indicative budget

CNY 500k–2M (appliance incl. software)

Best for

Branch offices, edge inference, fast delivery

Matrix

Comparison matrix

Delivery modelComputeModel sizeTypical concurrencyLead timeIndicative budget
PoC Validation1 GPU server (1×L40S 48GB / 1×A100 80GB)7B–32B (Qwen3-32B, GLM-4-9B, DeepSeek-R1-Distill-32B)20–50 concurrent1–2 weeksCNY 150k–400k (hardware excluded)
Standard Production3 nodes, 2–4×A100 / H20 / L40S32B–70B (Qwen3-32B/72B, Llama 3.3-70B)100–200 concurrent3–4 weeksCNY 600k–1.5M (hardware excluded)
Cluster ScaleK8s cluster, 8+ GPUs, multi-node TP / EP70B–671B MoE (DeepSeek V3/R1, Qwen3-235B)500+ concurrent6–8 weeksCNY 2M–5M (hardware excluded)
Xinchuang Full StackKunpeng / Hygon / Phytium + Ascend 910B or Hygon DCU7B–70B domestically adapted (Qwen / GLM / DeepSeek distilled)50–150 concurrent4–6 weeksCNY 1.5M–4M (domestic compute adaptation included)
Hybrid Gateway2–3 general-purpose servers on-prem (no GPU)Cloud or dedicated-instance models with hybrid local/cloud routingElastic with cloud quotas~1 weekCNY 100k–300k/year + usage-based billing
Turnkey ApplianceSingle box with 4 / 8 GPUs (L40S / L20 / Ascend)7B–70B pre-installed, expandable50–150 concurrent3–5 daysCNY 500k–2M (appliance incl. software)

Delivery

Delivery process

1

Assessment

Review use cases, data sensitivity, concurrency and model size; produce a recommendation.

2

Design

Define compute BOM, network topology, HA and compliance baseline, with quote and schedule.

3

Implementation

On-site installation, model import, gateway integration, load testing and hardening.

4

Acceptance & Ops

Accept on concurrency, TTFT, throughput and availability; hand over docs and training.

FAQ

FAQ

Yes. The gateway routes local and cloud models together: sensitive requests stay on-prem, general requests go to the cloud, split by policy.

Not sure which tier fits?

Tell us your scenario and budget, and we will come back within one business day with 2–3 options and price ranges.

⚡ Ai Token AI Gateway

Ready to get started?

Try Ai Token AI Gateway for free — simplify multi-model integration and unlock low-latency global access.