| PoC Validation | 1 GPU server (1×L40S 48GB / 1×A100 80GB) | 7B–32B (Qwen3-32B, GLM-4-9B, DeepSeek-R1-Distill-32B) | 20–50 concurrent | 1–2 weeks | CNY 150k–400k (hardware excluded) |
| Standard Production | 3 nodes, 2–4×A100 / H20 / L40S | 32B–70B (Qwen3-32B/72B, Llama 3.3-70B) | 100–200 concurrent | 3–4 weeks | CNY 600k–1.5M (hardware excluded) |
| Cluster Scale | K8s cluster, 8+ GPUs, multi-node TP / EP | 70B–671B MoE (DeepSeek V3/R1, Qwen3-235B) | 500+ concurrent | 6–8 weeks | CNY 2M–5M (hardware excluded) |
| Xinchuang Full Stack | Kunpeng / Hygon / Phytium + Ascend 910B or Hygon DCU | 7B–70B domestically adapted (Qwen / GLM / DeepSeek distilled) | 50–150 concurrent | 4–6 weeks | CNY 1.5M–4M (domestic compute adaptation included) |
| Hybrid Gateway | 2–3 general-purpose servers on-prem (no GPU) | Cloud or dedicated-instance models with hybrid local/cloud routing | Elastic with cloud quotas | ~1 week | CNY 100k–300k/year + usage-based billing |
| Turnkey Appliance | Single box with 4 / 8 GPUs (L40S / L20 / Ascend) | 7B–70B pre-installed, expandable | 50–150 concurrent | 3–5 days | CNY 500k–2M (appliance incl. software) |