Search pages, courses, and articles
Enter your monthly token volume, pick a cloud model class and a local hardware shape, get a real 3 year total cost of ownership for both. Plus the breakeven point where local starts to pay for itself, and the EU AI Act sovereignty trade-offs cloud pricing alone never captures.
Defaults are May 2026 list pricing from Anthropic, OpenAI, and Google, converted from USD at roughly 0.93. Override every field for your real numbers. Not legal or procurement advice.
One million tokens is roughly 750,000 English words, or about 350 pages of dense text.
Pick a model class. EUR rates are May 2026 list pricing; override either field for your negotiated rate.
Source: anthropic.com/pricing
Pick a hardware shape, then tune the numbers. Runs 70B class models comfortably. Needs a server room or colo.
At your current cloud rate of 2.79 in / 13.95 out EUR per million tokens, local breaks even at roughly 422 million tokens per month. Below that, cloud wins on cost; above that, your local stack pays for itself.
Volume is only one axis. Latency, data residency, vendor lock-in, and Article 10 data governance all push the decision around.
If your AI system falls into Annex III high-risk, Article 10 of Regulation (EU) 2024/1689 requires that training, validation, and testing datasets are subject to data governance with explicit attention to data origin, processing operations, relevance, and bias correction. Sending personal data through a US-based cloud LLM under Standard Contractual Clauses requires a documented Transfer Impact Assessment; under the EU-US Data Privacy Framework, careful self-certification checks. Both add cost the calculator above does not show.
The deepest cost of cloud is not the per-token bill, it is the ongoing legal load. Local + EU-hosted is sometimes the cheapest answer once you add a fair DPIA and TIA cost back in.
We build private AI infrastructure on your data, your servers, your rules. EU-hosted, GDPR-compliant, full audit trail. 12 month embedded partnership.
The crossover depends on volume and your hardware shape. For a dual A100 server (roughly EUR 35,000 capex, 15% ops time, 60% utilisation), local typically beats Anthropic and OpenAI list pricing somewhere between 60 and 120 million tokens per month. Below that, cloud is cheaper and the operational simplicity wins. Above that, the amortised hardware cost falls below per-token cloud billing and local pays for itself within the 36 month amortisation window.
Yes, materially. The EU AI Act does not require local hosting by default, but Article 10 of Regulation (EU) 2024/1689 imposes strict data governance obligations on high-risk AI systems, and GDPR transfer rules (Chapter V) restrict where personal data can be processed. Several EU regulators have flagged that sending personal data to US-based cloud LLMs without a valid transfer mechanism is a GDPR violation. If you are deploying a high-risk Annex III system or processing sensitive personal data, the real comparison is not cloud vs local at the same cost, it is cloud + a robust DPIA + transfer impact assessment vs local with full data control.
A 50-person team running an internal AI assistant for 20-30 daily active users typically burns 30 to 80 million tokens per month, weighted heavily toward input (RAG context, prompt scaffolding). A 500-person team running a customer-facing chatbot at moderate volume sits around 200 to 500 million tokens per month. Heavy-use cases (transcription, embeddings, autonomous agents, multi-step pipelines) can hit 1 billion tokens per month or more.
Patching the OS, updating CUDA / drivers, monitoring GPU temperature and memory pressure, model version upgrades, prompt cache invalidation, observability and incident response, periodic re-quantisation when newer model weights drop, and capacity planning. A well-run dual-GPU setup needs roughly 10 to 20% of an SRE FTE in steady state. A multi-node cluster needs more. The calculator defaults to 15%.
It depends on amortisation. At 36 months on a EUR 35,000 dual A100, hardware adds roughly EUR 970 per month to your TCO. Electricity at NL business rates (EUR 0.20 per kWh, 1.2 kW draw, 60% utilisation) adds about EUR 100. Ops at 15% of a EUR 95,000 FTE adds EUR 1,200. So ops time, not hardware, is usually the biggest single line. The further you can automate the stack, the more local pays off.
The defaults are May 2026 list pricing on the Anthropic, OpenAI, and Google websites, converted to EUR at roughly USD * 0.93. Negotiated enterprise rates and committed-spend discounts can be 30 to 60% lower; batch and prompt-caching can knock another 30 to 90% off input cost. Override the cloud price fields with your actual negotiated rate for a real comparison.
No. The calculator outputs a financial TCO comparison. The EU AI Act, GDPR transfer rules, sector-specific obligations (NIS2, DORA, EHDS), and your company's own data classification policy all affect the legally defensible answer. For a binding view on whether cloud or local is the correct architecture for a specific use case, consult a qualified EU technology lawyer.
When to host locally, when to stay in the cloud, how EU-based providers compare, and how Article 10 data governance changes procurement.
