Costed scenario: DeepSeek V4 on DGX Station for an 80-person SME

What DeepSeek V4 Flash 0731 on a DGX Station GB300 would cost and deliver for an 80-person SME: a sizing scenario built on explicit assumptions, not a customer story, with estimated throughput, 3-year total cost of ownership and break-even against APIs.

By QDNA · Published on 2 August 2026 · Updated on 8 October 2026 · 8 min read

Updated on 8 October 2026: reference prices raised to €150,000 excl. VAT for the DGX Station and €7,000 excl. VAT for the DGX Spark; the total cost of ownership, the daily cost, the cost per million tokens, the break-even thresholds and the cost of the 2× DGX Spark cluster are recomputed. Same-day correction: the gap with the cloud bill subtracted €23,300, an amount the article never derived; it now rests on the recalculated bill, 150 M tokens × €10.52/M × 12 = €18,936 per year, i.e. €56,340 more per year on-premises (€51,606 with 25% growth), and that growth gives ~€23,700/year instead of the €29,100 previously shown.

Updated on 26 September 2026: API prices realigned on the 26 September 2026 price lists (Claude Opus 5.5 at $4 and $20, GPT-6 Astra and Sol, DeepSeek V4.1 Flash, which replaced V4 Flash 0731 in the API on 10 September), ECB rate of 25 September, DGX Spark price from LDLC.pro, pi3g page of 14 September; the reference cloud bill and the break-even thresholds are recomputed.

Updated on September 2, 2026: this article, published as a "case study", is requalified as a costed scenario. There is no customer behind it: the SME, its volumes and its previous provider are working assumptions. Throughput figures that do not come from our measurements section are marked "estimated, to be measured"; the DeepSeek API price and the break-even formula are updated. Same-day re-read: the assumed cloud bill (€8,000/month) is recalculated at the Anthropic price (€1,950/month), which reverses the economic conclusion; the GPT-5.5 price is taken from the OpenAI price list; "estimated" throughputs, latencies and user counts without a formula are replaced by "to be measured"; the cost per million is given by its formula. Same-day harmonisation: the DGX Station price (€95,000 excl. VAT at pi3g, page dated 14 September 2026) and the electricity tariff (€0.1624/kWh excl. VAT, Tarif Bleu for professionals) are aligned with the site's shared assumptions, replacing the €120,000 and the 30 cents per kWh set as conservative assumptions; the TCO and the thresholds are recomputed.

Scenario assumptions

We assume an SME publishing logistics software, 80 employees in France on two sites, whose 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).

We assume prior use of Claude Opus 4.5 through the Anthropic API for ~150 M tokens/month (chat + RAG), billed ~€1,580/month (~€18,900/year) at the Anthropic price of Claude Opus 5.5 on 26 September 2026, $4 input and $20 output per million mixed 50/50, i.e. €10.52/M: a working assumption, not an invoice, and the 2 August version assumed €8,000/month, an amount no price list supports. A 25% growth would bring that cost to ~€23,700/year (€18,936 × 1.25 = €23,670), not counting rate limits at peak times.

The scenario retains 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

The deployed configuration in this scenario pairs a DGX Station GB300 with a quantised DeepSeek V4 Flash 0731, served by vLLM behind the LiteLLM gateway, each choice being justified in the table.

Estimated throughput and latency, to be measured

No DeepSeek V4 Flash 0731 / DGX Station GB300 pair appears in our measurements section: the values below are estimates derived from memory bandwidth and KV-cache sizing, to be confirmed by a reproducible measurement, and concurrent user counts are memory calculations, not observations.

Total cost of ownership (TCO) over 3 years

The total cost of ownership over three years adds the station's acquisition, integration, electricity at the Tarif Bleu rate, maintenance and training, which comes to €225,828 detailed year by year.

Calculated ROI

The scenario assumes ~150 M tokens/month (chat + RAG), or about 5 M tokens/day, which gives a recalculated cloud bill of about €1,578/month or €18,936/year, rounded to ~€18,900 (150 M × €10.52/M × 12).

The local cost amortised over 3 years comes to €75,276/year (TCO €225,828 / 3), that is €206 per day (€225,828 / 1,095 days), a figure that every break-even below divides by an API price.

Break-even, visible formula: tokens per day = daily local cost (€206) / API price per million tokens, at €1 = $1.1403 (ECB, 25 September 2026), which gives, against Claude Opus 5.5 at $12/M (€10.52/M), 19.6 M tokens/day, and against GPT-6 Astra at $30/M ($10 input, $50 output, OpenAI price list of 26 September 2026; €26.31/M): 7.8 M tokens/day; against GPT-6 Sol at $6/M ($2 and $10; €5.26/M): 39.2 M tokens/day, and against the DeepSeek API (deepseek-flash model served by V4.1 Flash since 10 September 2026, price list read on 26 September 2026, $0.15 input and $0.60 output per million off-peak, mixed 50/50, i.e. $0.375/M or €0.33/M; double at peak hours): 625 M tokens/day off-peak, 312 M at peak hours. At the assumed 5 M tokens/day, none of these thresholds is reached.

Against the recalculated cloud bill, on-premises costs €75,276 − €18,936 = €56,340 more per year, and still €75,276 − €23,670 = €51,606 with 25% growth; the 2 August version, with a €96,000 cloud bill no price list supported, concluded on a saving. The scenario is therefore not justified by cost: its arguments are data sovereignty (no transfer off-site), the absence of a vendor-imposed rate limit and the ability to fine-tune the model on internal data.

Points of attention

Four points of attention decide whether the deployment succeeds: the runtime memory margin, CPU offload for large models, FP8 calibration of the KV cache and the station's real memory.

• 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 at the pace of NVIDIA security updates; plan a maintenance window.
• Train the internal team: the scenario budgets €10,000 and two weeks to train 10 people (5 dev, 3 ops, 2 data), a condition for autonomous operation.

Conclusion

In this scenario, DeepSeek V4 Flash 0731 on a DGX Station GB300 holds technically (167 GB of published weights for 748 GB of memory, one-million-token context) with throughput still to be measured.

Economically, it costs €75,276 per year amortised against a cloud bill recalculated at €18,936, i.e. €56,340 more per year, at the Anthropic price of Claude Opus 5.5 on 26 September 2026, far from the break-even thresholds against APIs at the retained volume. What justifies it is data sovereignty (no transfer off-site), the absence of a vendor rate limit and the stability of a pinned version.

The next step of such a scenario would be a 2× DGX Spark cluster (€14,000 excl. VAT of machines, 2 × €7,000 excl. VAT at the price used from October 2026) for redundancy, sized in our DeepSeek V4 Flash guide for DGX Station and 2× Spark cluster.

References

• DeepSeek V4 Flash 0731, official Hugging Face repository
• DeepSeek V4 technical report (arXiv 2606.19348)
• Unsloth, DeepSeek V4 documentation
• NVIDIA, DGX Spark performance blog
• ServeTheHome, DGX Station GTC 2026
• QDNA sheet: DGX Station
• QDNA sheet: vLLM
• QDNA sheet: DeepSeek V4