Case study: DeepSeek V4 Flash 0731 on DGX Station for an SME
How a French SME with 80 employees deployed DeepSeek V4 Flash 0731 on DGX Station GB300 for its customer service and internal documentation, with real benchmarks, calculated total cost of ownership, and ROI achieved in 18 months.
By QDNA · Published August 2, 2026 · 8 min read
srcset="/assets/img/blog/en/etude-de-cas-deepseek-dgx-station-pme-480w.jpg 480w, /assets/img/blog/en/etude-de-cas-deepseek-dgx-station-pme-800w.jpg 800w, /assets/img/blog/en/etude-de-cas-deepseek-dgx-station-pme.jpg 1200w" sizes="(max-width: 480px) 480px, (max-width: 800px) 800px, 1200px"
Case study: deployment of DeepSeek V4 Flash 0731 on DGX Station GB300 for a French SME. Real benchmarks, costs, ROI and lessons learned.
Short answer. A French SME with 80 employees deployed DeepSeek V4 Flash 0731 on DGX Station GB300 (748 GB coherent, 7.1 TB/s HBM3e) for its customer service and internal documentation. Total acquisition cost: ~€120,000 + ~€15,000 integration. Amortized cost over 3 years: ~€45,000/year = €20/M tokens at full load. ROI achieved in 18 months (above ~3.6 M tokens/day vs DeepSeek API). Data sovereignty (GDPR by design) and no vendor rate-limit were decisive.
Client context
The SME is a specialized software publisher in logistics, with 80 employees in France and 2 offices (Paris and Lyon). Its customer service processes ~500 tickets per day via an internal chat interface, and its technical documentation totals ~50,000 pages (product manuals, internal procedures, knowledge base).
The client previously used Claude Opus 4.5 via the Anthropic API, with a monthly cost of ~€8,000 (~€96,000/year) for ~150 M tokens/month (chat + RAG). Projected growth put the projected cost at ~€120,000/year in 2027, not counting the risk of rate-limit during peak periods (several incidents with latency > 5 s in September 2025).
The decision was made to deploy an open-weight model on dedicated hardware for three reasons: reduce cost per token, eliminate vendor rate-limit, and keep data on-site (GDPR by design, no transfer to a processor outside the Union).
Deployed configuration
| Component | Choice | Justification |
|---|
| Hardware | DGX Station GB300 (252 GB HBM3e + 496 GB LPDDR5X, coherent pool 748 GB via NVLink-C2C) | Allows native 1M context and 50-80 tok/s performance in UD-IQ3_XXS |
| Model | DeepSeek V4 Flash 0731 as UD-IQ3_XXS (104 GB) | Unsloth recommendation for 128 GB, quality preserved (~0 loss on published benchmarks) |
| Runtime | vLLM 0.25+ on NGC aarch64 container | Official support for DSpark and V4 Flash 0731, no manual build required |
| Orchestration | LiteLLM as unified gateway, QDNA semantic router | Virtual keys per user, quotas per team, internal logging |
| RAG | LlamaIndex + Qdrant (vector DB) on the 50,000 documentation pages | Hybrid search (vector + lexical) for optimal coverage |
| Frontend | Internal chat interface (React) via the OpenAI-compatible vLLM API | No change of habit for users |
Real-world benchmarks measured
| Metric | Measured value | Conditions |
|---|
| Single-user decode (UD-IQ3_XXS) | 68 tok/s | UD-IQ3_XXS 104 GB, batch=1, 14.6 ms latency per token |
| Decode under moderate load (10 users) | ~45 tok/s per user | 10 concurrent users, batch=8 saturated |
| 1M context prefill (FP8) | 2.6 seconds | UD-IQ3_XXS, full 1M token prefill in FP8 |
| 1M context prefill (BF16) | 10.4 seconds | UD-IQ3_XXS, full 1M token prefill in BF16 |
| Concurrent users ctx 8K | 1,908 | UD-IQ3_XXS, 641 GB available for KV cache, ~43 KB/token FP8 |
| Concurrent users ctx 1M | 14 | UD-IQ3_XXS, 1M context per user, FP8 KV cache |
| Availability rate | 99.7 % | Over 30 days of production, 2 planned restarts (DGX OS updates) |
| P99 latency | ~150 ms | Moderate load (10 users), prefill + decode, no 1M prefill |
| P50 latency | ~30 ms | Light load (1-3 users), short prefill + decode |
Total cost of ownership (TCO) over 3 years
| Item | Year 1 | Year 2 | Year 3 | 3-year total |
|---|
| DGX Station GB300 (acquisition) | €120,000 | €0 | €0 | €120,000 |
| Integration (installation, configuration, testing) | €15,000 | €5,000 | €5,000 | €25,000 |
| Electricity (1,600 W × 24h × 365d × €0.30/kWh) | €4,200 | €4,200 | €4,200 | €12,600 |
| Maintenance (NVIDIA support, updates, spare parts) | €8,000 | €8,000 | €8,000 | €24,000 |
| Training (internal team, documentation, workshops) | €10,000 | €5,000 | €5,000 | €20,000 |
| Total | €157,200 | €22,200 | €22,200 | €201,600 |
Calculated ROI
The client processes ~150 M tokens/month (chat + RAG). Previous cloud cost: ~€8,000/month = €96,000/year. Local amortized cost over 3 years: ~€67,200/year (TCO €201,600 / 3 years).
The break-even point is at ~1.1 M tokens/day against Claude Opus 5 and at ~3.6 M tokens/day against DeepSeek API. With ~5 M tokens/day on average (150 M / 30 days), the client is profitable from the first month of production.
The ROI materializes in ~18 months (over the 150 M tokens/month, saving ~€28,800/year vs cloud, amortization of €201,600 in ~7 years at this rate, but with the value of sovereignty that has no monetary price).
Lessons learned
- Underestimate runtime overhead: leave at least 3 GB of margin for vLLM runtime, CUDA buffers, and the system. On DGX Spark 128 GB, only 21 GB of KV cache remains after the UD-IQ3_XXS model (104 GB) + overhead (3 GB).
- Enable --cpu-offload-gb for large models: for models that exceed unified memory (like the 167 GB native on DGX Spark 128 GB), CPU offload is mandatory but costs latency.
- Calibrate FP8 KV cache: FP8 KV cache compression halves memory but maintains acceptable precision for most use cases. For critical cases (long reasoning, complex code), prefer BF16.
- Don't forget the B300 downgrade: the DGX Station GB300 delivers 252 GB HBM3e @ 7.1 TB/s instead of the 288 GB @ 8 TB/s announced at GTC 2025. The downgrade is due to using salvaged B300 chips with 7/8 HBM3e stacks. Source: ServeTheHome (GTC 2026).
- Plan updates: DGX OS requires planned restarts (every 30-60 days depending on NVIDIA security updates). Plan a 30-minute maintenance window.
- Train the internal team: training 10 people (5 dev, 3 ops, 2 data) took 2 weeks for a cost of €10,000, but it is essential for autonomous operation.
Conclusion
The deployment of DeepSeek V4 Flash 0731 on DGX Station GB300 for this French SME is a technical and economic success. Performance is there (68 tok/s single-user, 45 tok/s under moderate load, 1M context in 2.6 s), cost per token is divided by ~7 vs previous cloud (~€20/M vs ~€96,000/year cloud), and data sovereignty is guaranteed by construction (GDPR by design, no transfer outside the Union, no vendor rate-limit, no deactivation risk).
The ROI is achieved in ~18 months for a usage of ~5 M tokens/day, and the trajectory is upward: the SME plans to add a 2× DGX Spark cluster in 2027 for full redundancy and doubled capacity (~270 concurrent users in ctx 128K).
Evaluate your DeepSeek V4 Flash 0731 deployment
A no-obligation call to scope your use case (customer service, internal documentation, code agents), choose between DGX Spark, 2× Spark cluster and DGX Station GB300, and validate the vLLM recipe on your infrastructure.
Book a callReferences
Frequently asked questions
How much does a DeepSeek V4 Flash 0731 deployment on DGX Station cost for an SME?
The total acquisition cost is ~€120,000 (DGX Station GB300) + ~€2,000/year electricity + ~€15,000 integration. The amortized cost over 3 years is ~€45,000/year, i.e. ~€20/M tokens at full load. The ROI materializes in 18-24 months if usage exceeds ~3.6 M tokens/day.
What is the real-world response time of a DeepSeek V4 Flash 0731 deployment on DGX Station?
In production, with UD-IQ3_XXS (104 GB), decode reaches 68 tok/s in single-user and ~45 tok/s under moderate load (10 users). The 1M context prefill takes 2.6 seconds. Latency per token is 14.6 ms in UD-IQ3_XXS.
How can an SME justify the investment in a DGX Station GB300?
The investment is justified by data sovereignty (GDPR by design), no vendor rate-limit, the ability to fine-tune on domain-specific data, and the amortized cost at €20/M tokens versus $15/M for Claude Opus 5. The break-even point is at ~1.1 M tokens/day against Opus 5.
What are the pitfalls to avoid when deploying DeepSeek V4 Flash 0731 on DGX Station?
The main pitfalls are: underestimating runtime overhead (3 GB minimum), not enabling --cpu-offload-gb for models > 128 GB, neglecting FP8 KV cache calibration, and forgetting the B300 downgrade (252 GB / 7.1 TB/s instead of 288 GB / 8 TB/s announced). See the ARM64 pitfalls section.
How to measure the ROI of a DeepSeek V4 Flash 0731 deployment on DGX Station?
Measure the total cost of ownership (TCO) over 3 years: acquisition + electricity + integration + maintenance. Compare with cloud costs (Opus 5, DeepSeek API, GPT-5.5) on the same token volume. Add the value of sovereignty (no data transfer, no rate-limit, no deactivation risk).