The short answer
- An owned eight-GPU B200 server costs about €14,790 a month over three years. It costs the same as a premium closed model at under one billion tokens a month, and as a cheap hosted open model at over twenty billion.
- One server cannot always handle that volume. Extra servers, staff, power and idle time raise the cost; the calculator adds servers and staff.
- For most organisations, control over data decides the case for self-hosting.
How the calculation works
API cost is monthly tokens times a price per million tokens. Input and output prices blend three to one, typical for document and search work.
Self-hosting cost is the server plus its staff. A vendor puts an owned eight-GPU B200 system at $22.70 (€20.22) per server-hour, at 70 percent utilisation over three years.1 That is about €14,790 a month. Dollar prices use the ECB reference rate (€1 = $1.1225 on 2 October 2026). Rental figures are list prices at Verda, a Finnish provider,2 tracked in European cloud GPU prices.
Break-even is the monthly volume where both bills are equal. A server handles only so many tokens. Above that, the calculator adds a server, and the API can be cheaper again until the new server is busy. Compare both bills at your own volume.
Break-even by API
The cheaper the API, the more volume a server needs
Each bar shows the monthly volume at which one server costs the same as the API. Above a server's capacity, more servers are needed; the calculator includes them.
Break-even ranges from under 1 to over 22 billion tokens a month.
Monthly volume at which one B200 server costs the same as the API, billion tokens
1 Owned 8-GPU B200 server at about €14,560 per month (3 years, 70% utilisation), divided by blended API price per million tokens (3 input : 1 output). Excludes staff costs. Dollar prices converted at the ECB reference rate of 25 September 2026.
Source: Lindstead calculation; server cost from Mercatus (vendor estimate, July 2026); API list prices, 27 September 2026
View data table
| Category | Series 1 | Series 2 |
|---|---|---|
| GPT-6 Astra | 0.8 | Not reported |
| Claude Sonnet 5 | 4.2 | Not reported |
| DeepSeek V4-Pro API | Not reported | 8.4 |
| Mistral Large API | Not reported | 22.1 |
Series 1: Closed model API. Series 2: Hosted open-weight model API. Unit: billion tokens
Compare like with like. Measure model quality and server throughput on your own workload. If each extra server costs more than the API use it replaces, the API stays cheaper at every volume. Independent research finds payback in about three months for small models and often more than two years for the largest.3 The Model Index lists open-weight models that can replace premium closed models, with their hardware.
Calculator
Amounts are in euro, excluding VAT. Dollar list prices use the latest ECB reference rate (€1 = $1.1225 on 2 October 2026).
Result
Capacity is your estimate. Each extra server can wipe out the savings. The result says nothing about model quality or legal fit.
Discuss this cost modelWhat the numbers leave out
- Quality. A cheaper model that needs more human review is not cheaper.
- Utilisation. An idle server still costs money at night. An API does not.
- Price trends. API prices for the same capability have fallen steeply every year.4 GPU rental prices rose in 2026 (see the European GPU price index).
- Control. Keeping data out of third-party hands has no line in a cost model. It often decides the case.
Control weighs most in financial services and healthcare, with their strict confidentiality and exit duties. A full self-hosted AI total cost of ownership adds staff, energy and utilisation per workload. If the numbers favour own infrastructure, see AI deployment on infrastructure the organisation controls. The wider trade-offs are in Where Should Your AI Run? The 2026 European Briefing.
Sources
- Mercatus, B200 server price and cost per hour, 16 July 2026 (vendor analysis). mercatus-ai.com/blog/b200-server-price
- Verda GPU pricing, accessed 27 September 2026. verda.com/pricing
- Pan et al. (Carnegie Mellon University and independent researchers), cost-benefit analysis of on-premise LLM deployment, arXiv 2509.18101. arxiv.org/abs/2509.18101
- Epoch AI, LLM inference price trends, 12 March 2025. epoch.ai/data-insights/llm-inference-price-trends
- API list prices: Anthropic, OpenAI, DeepSeek and Mistral pricing pages, accessed 27 September 2026.