Local AI or cloud AI? The 2026 cost comparison for SMEs

Two paths to AI in the company

Cloud AI such as Microsoft 365 Copilot, ChatGPT Business or Claude Team is billed per user per month: ready immediately, always the latest models, but the costs grow with every employee and keep running. Local AI flips the model: a machine with plenty of fast memory in your own building runs open language models, costs a one-off purchase plus electricity and maintenance, and the data never leaves the premises. The interesting question is not which sounds better, but from what point which path pays off, and for which tasks the local models are good enough.

What cloud AI costs in 2026

The list prices of the three big business offerings, as of September 2026, each billed annually:

• Microsoft 365 Copilot Business: 18.20 EUR per user per month plus VAT on an annual subscription, for up to 300 users (promotional price 15.60 EUR for the first year when purchased between 1 July and 31 December 2026; 21.84 EUR with monthly payment). Requires a Microsoft 365 Business Basic, Standard or Premium plan, whose costs come on top. The Business Standard plus Copilot and Business Premium plus Copilot bundles sit at 20.36 and 27.73 EUR respectively.
• ChatGPT Business: since 2 April 2026, 20 USD per user per month billed annually (25 USD monthly), for 2 to 200 seats. Deep research and image generation are included within the usage limits; if you need more, you buy usage-based workspace credits. Since August 2026 there has also been a premium seat with five times the usage for 100 USD.
• Claude Team (Anthropic): standard seat 20 USD per user per month billed annually (25 USD monthly), premium seat with five times the usage 100 USD (125 USD monthly), for teams of 2 to 150 people, Claude Code included.

What a local AI server costs

Local AI needs one thing above all: plenty of fast graphics or system memory. In October 2025, the IT magazine heise concluded that local models had become genuinely usable and named as entry points a used RTX 3090 with 24 GB of VRAM from around 700 EUR or complete systems with AMD Ryzen AI Max+ 395 from about 1,800 EUR. Those prices are out of date because memory has become much more expensive since late 2025: according to heise, a used RTX 3090 cost under 1,000 EUR in August 2026, and a Ryzen AI Max+ 395 system with 128 GB costs around 3,200 to 3,900 EUR including VAT in September 2026 depending on the manufacturer, i.e. around 2,700 to 3,300 EUR net; with 64 GB it is about 1,950 to 2,200 EUR including VAT or 1,650 to 1,850 EUR net. Memory bandwidth remains the most important performance factor, not raw compute.

That yields a simple rule of thumb: a local server lands at roughly 2,800 to 4,500 EUR net over three years plus maintenance. For 5 users, including maintenance, that is usually more expensive than a cloud subscription. From around 10 users it plays in the same league including maintenance, from 25 users clearly below it, because the cloud bill keeps growing per head and the local one does not. One caveat: when many people use the server at the same time, answers slow down, and at some point you need a second machine or a stronger GPU.

Which models are genuinely usable locally in 2026

The cost comparison is useless if the quality is not there. The sober assessment: the best cloud models remain ahead of the open ones, especially for complex reasoning. For typical SME tasks such as summarising, rewriting, email drafts, minutes and searching your own documents (RAG), open models are now genuinely usable, though; in August 2026 heise recommended Qwen 3.8 27B, for example, for a used RTX 3090. Examples as of September 2026, with rough memory needs at 4-bit quantisation:

• Qwen 3.8 27B from Alibaba (August 2026, Apache 2.0): also understands images and video and needs around 16–19 GB.
• Gemma 4 from Google (April 2026, Apache 2.0) in several sizes up to 31B: the largest variant needs around 17–20 GB, the small ones run with just 4–8 GB.
• gpt-oss from OpenAI (August 2025, Apache 2.0): the 20B variant gets by with around 16 GB, the 120B variant needs more than 60 GB and therefore an 80 GB GPU or a system with 128 GB of unified memory.
• Mistral (Apache 2.0): the compact Ministral 3 models with 3 to 14 billion parameters (December 2025) run on small hardware; Mistral Small 4 (March 2026), despite its name, is a large model with 119 billion parameters meant for data-centre hardware.
• Watch the licences: for Meta’s multimodal Llama 4 models, the acceptable use policy grants no licence rights to companies with their principal place of business in the EU. Check the licence of every model before you use it.

Data protection: when local is the better choice

With local AI, prompts and documents never leave the building: no data processing agreement with an AI provider, no third-country transfer, no dependence on adequacy decisions. If an external provider maintains the system with access to data, however, a data processing agreement may be needed for that. For medical practices, law firms and anyone with sensitive customer data, that is the strongest argument. Cloud AI can also be used in line with data protection law, but demands homework: the right tier and region, a data processing agreement and clear team rules. Which provider options exist for that, from the EU data zone on Azure via OpenAI’s EU data residency for enterprise customers to open models hosted in Germany such as the IONOS AI Model Hub, is covered in the guide Using ChatGPT and co. in line with the GDPR.

Hybrid in practice: the best of both worlds

In practice it is rarely either-or. The pattern that works: sensitive and recurring tasks run locally (document search, internal summaries, receipt processing), peak tasks with high quality demands run in the cloud, on a proper tier with a data processing agreement. That keeps the cloud bill small, because not every employee needs a full-time subscription, and the delicate data stays in-house.

These are exactly the setups I build as part of my AI automation services: from sizing the AI server via model selection to connecting the models to your documents and workflows. Which tasks are concretely worth it is shown in the article on AI use cases in SMEs.