A company can store and process its data in Europe and still use the strongest AI available. The question is what that costs in practice, who it’s for, and where the trade-offs sit. Here’s an overly comprehensive review of the prices, the contracts and the fine print.
In May 2025, the chief prosecutor of the International Criminal Court lost access to his work email. Karim Khan had been sanctioned by the United States three months earlier.1 According to the Associated Press, his Microsoft email was cancelled;2 Microsoft said the disconnection concerned a sanctioned official, not the court, and that it had at no point ceased services to the ICC.3 Microsoft’s own senior director later gave the UK Parliament a different account, which Microsoft then asked to be corrected.4 The court reportedly began moving about 1,800 workstations to openDesk, a German open-source suite.5 In October 2026, the US sanctioned the court itself, with a licence authorising cloud and email services to continue.6 (So the court can still use the email. Whether it wants to is another question.)
This was email, not AI. But the question it raises applies to any hosted service: can your provider’s government switch you off, and can you leave? The ICC’s answer was to start moving. It is a story about dependency and exit, not about someone reading the data.
On the data question, the record is calmer than the architecture suggests. On 10 June 2025, a French Senate commission asked Microsoft France’s director of public and legal affairs whether he could guarantee that French citizens’ data would never be handed to US authorities without French agreement. Anton Carniaux said no. He added that it had never happened, that Microsoft contests such requests, and that its transparency reports show no European business or public-sector customer has been affected.7,8 That is Microsoft’s statement about its own customers, and it is the best base rate we have. GDPR Article 48 pushes the other way too: a foreign court order is not, by itself, a lawful basis for a transfer.9 The risk is real and structural. On the evidence, it has not materialised for a European company.
We have a view on what to do. For most companies, the first and cheapest step is to get everyone onto business terms. A business seat costs about €18–26 per user per month;10,11,12 for staff already on a paid personal plan, the switch is almost free. It closes more than the training question. All four major providers say business-plan data is not used to train their models by default, and Google, Microsoft and Anthropic put it in their terms; OpenAI states it for the API.13,14,15,16,17 But a consumer account also typically lacks admin controls: you cannot set retention, audit usage or remove a departing employee’s history. (The history leaves with the employee, which is not what you want.) A business plan gives you the controls and the contractual commitments together. It does not change what people paste into the chat, which is the next thing to think about.
After that, sort the data. Not into two neat piles (that is harder than it sounds, since most office text is someone’s personal data under GDPR), but into categories with a short policy for each and admin controls that enforce it. Spend attention on the data that needs it, and test before paying for more control.
Figure 1 compares seven ways a company can use AI, from a personal account to a server of its own. For each route it answers seven questions, from where the data is stored to who built the model, and it shows the strongest AI that route can reach.
Figure 1
What you gain and what you give up
Seven questions about where your data goes, who has jurisdiction, and who built the model, answered for seven routes from a personal account to your own server, with the strongest AI each route can reach. Four routes reach the Top band, all using US providers’ models. Your own server answers yes to six of the seven, more than any other route. The seventh, whether the model is built in Europe, is no for the models we costed on both the self-hosted and EU-cloud routes. A table view follows.
No route answers yes to all seven questions. Your own server comes closest, at six. The seventh, whether the model itself is built in Europe, is no for the models we costed, though Mistral’s smaller open models can run on a single card. Cost is in Figure 2.
The last row places each route’s strongest model in a band by its score on the Artificial Analysis Intelligence Index (v4.3.2), a public index of hard reasoning, coding and professional tasks. The scores in that row’s notes are on this index, at each model’s default or medium setting. We draw the bands at 50 and above for Top, 40–49 for Strong, 30–39 for Mid and below 30 for Basic, as in Figure 2. On routine office work the differences between models are probably smaller. *DeepSeek V4.1 Flash has no published default score, so it is placed by its best-setting score.
The answers come from providers’ published terms and price pages, read in October 2026. Whether a foreign government could switch a route off is our assessment. This is a general picture, not advice on any provider or contract.
| Question | Consumer account | Business plan | US model, EU region | EU router or workspace | European-made model (Mistral) | Open model on an EU cloud | Your own server |
|---|---|---|---|---|---|---|---|
| Stored in the EU | No. Usually stored outside the EU. | Partly, or it depends. Varies by plan. Copilot and Vibe store in the EU. ChatGPT Business and Claude Team do not. | Yes. Stored in the chosen EU region. | Yes. The router and the regions behind it are in the EU. | Yes. Hosted in the EU by default. | Yes. Stored in EU data centres. | Yes. On hardware you control. |
| Processed in the EU | No. Usually processed outside the EU. | Partly, or it depends. Varies by plan. None is fully processed in the EU at the priced tier. | Yes. Inference runs in the EU region you choose, where offered. | Yes. Enforces EU regions for you, which a plain cloud account does not do by default. Routers also send work to EU clouds for open-weight models. | Partly, or it depends. Processed in the EU by default, but Mistral’s own documentation says some features send data outside the EU. | Yes. Scaleway says everything is hosted in Europe. OVHcloud’s technical documentation says the infrastructure is in Gravelines, France. | Yes. On hardware you control. |
| Support staff outside the EU blocked | No. No such limit. | No. No such limit on most plans. | Partly, or it depends. Some providers allow remote access by non-EU staff. | Partly, or it depends. Depends on the router and on the US cloud underneath. | Partly, or it depends. Mistral monitors API traffic for abuse unless zero data retention is on, and zero retention is available only on pay-as-you-go API calls. A limited number of staff see reported content. Mistral does not say where they sit. | Partly, or it depends. Scaleway keeps no data by default, but may access request content for up to two weeks on suspected misuse. We found nothing published by OVHcloud on staff access. Neither says where staff sit. | Yes. Only your own people. |
| Not used to train models | No. Providers train on your inputs unless you opt out. | Partly, or it depends. Off by default on the US plans. Mistral’s Vibe Team leaves it on unless an administrator turns it off. | Yes. Excluded by the business terms. | Yes. No-training and zero-retention terms are published. | Partly, or it depends. Off by default on Enterprise. Switchable by an administrator on Team and pay-as-you-go API. On by default for free and individual use. | Yes. Scaleway’s documentation and OVHcloud’s product page both say customer data is not used for training. | Yes. Nothing leaves your servers. |
| Provider under EU law only | No. Most providers are US companies. | Partly, or it depends. US companies, except Mistral’s Vibe. | No. The US CLOUD Act follows the company, not the server. | Partly, or it depends. Moves the contract and the invoice to Europe, not the US cloud’s jurisdiction. For EU-cloud models, the provider is already under EU law. | Partly, or it depends. Mistral is French, under French law. Which data centres serve its standard API is not publicly confirmed. | Yes. French companies such as Scaleway and OVHcloud. | Yes. Your own company. |
| A foreign government can’t switch it off (our assessment) | No. Any US provider is bound by US sanctions and court orders. | Partly, or it depends. A paid contract, but the provider can change terms, retire models, or comply with its government’s orders. Mistral’s Vibe is a French business plan. | No. The data is in the EU but the provider is a US company that can be directed by its government. | Partly, or it depends. The router is European, but Claude, GPT and Gemini still run on a US cloud. For EU-cloud models, no US provider is in the chain. | Partly, or it depends. Mistral is French. Which data centres serve its standard API is not publicly confirmed, so the full chain cannot be verified. | Yes. French companies running the model on their own servers, with no US company in the chain. | Yes. A rented server has a hosting provider (Hetzner, a German company); your own hardware has no provider at all. |
| The model itself is built in Europe | No. Most consumer AI tools run models from US or Chinese labs. Mistral’s free Vibe is the European exception. | Partly, or it depends. Mistral’s Vibe is the one European-built option among the business plans. | No. These are models built by US labs. | Partly, or it depends. Routers also offer Mistral’s models alongside the US ones. | Yes. Mistral is a French AI lab building its own models. | No. DeepSeek, Qwen and GLM come from Chinese labs. gpt-oss comes from OpenAI. Scaleway also hosts Mistral’s models. | No. The costed models, gpt-oss and Qwen, are not European. Mistral’s Ministral 3 (3B to 14B) fits on a single card but scores 5–6 on the index, against 10–12 for gpt-oss. |
| Strongest AI you can get | Top. The providers’ own apps carry their strongest models. | Top. Frontier-class models; the plan pages don’t say which. | Top. Claude Opus 5.5 and GPT-6 Astra are offered in EU regions. They score 51 and 50 on the index. | Top. Routers list Claude Opus 5.5 on EU endpoints. It scores 51 on the index. | Mid (European-made). Mistral’s best model, Large 4 (a preview), scores 38 on the index. | Strong (Chinese-lab models*). Scaleway runs DeepSeek V4.1 Flash, which scores 45 on the index at its best setting. No default score is published. | Basic. Qwen3.8-27B fits on one server and scores 28 on the index at its medium setting. Strong (~€100k a year*). DeepSeek V4.1 Flash needs two large servers, about €100,000 to €110,000 a year. It scores 45 on the index at its best setting. |
When providers say “EU data residency” they mean different things: where the data is stored, where the model runs, who can see it, whether it trains models, and whose law governs the provider.18 Each is a row in Figure 1.
The question underneath is jurisdiction. A US company can be compelled to hand over data it controls, wherever the server sits.19 EU processing closes the storage and processing question. It does not close the jurisdiction question, and remote support access from outside the EU is itself a transfer under GDPR.20
What EU jurisdiction actually buys is not immunity. European governments can reach data held in their territory or by their companies, through court orders and, under the e-Evidence Regulation, through cross-border production orders.21 What a provider under EU law gives you is a known procedure and a court you can reach. Your own server means no provider at all, which is a different thing (and a different set of duties). And the strongest European-hosted model in the strong band is a Chinese-lab model, whose home government’s reach over its own companies is a question in its own right.
GDPR does not require data to stay in the EU.22 It requires a legal basis for transfers. The Data Privacy Framework, a Commission adequacy decision of 2023,23 provides one; standard contractual clauses provide another.24 The framework is under pressure: an appeal is pending at the Court of Justice (Case C-703/25 P),25 the US oversight board for the arrangement has lacked a quorum since January 2025,26 and the whole thing rests on a US executive order.
Before asking where AI should process your data, it is worth asking where the data already sits. If your company keeps its email, documents and CRM in Microsoft 365 or Google Workspace, the contract you would paste into an AI assistant is already in a US cloud. The customer list is in a US CRM. All of it sits with a US provider under exactly the jurisdiction exposure this article describes, accepted when the company signed.
Putting that same contract through Copilot inside the tenant it already lives in, or through Claude in an EU region on AWS, adds no new jurisdiction.30 For data already in a US cloud, the sovereignty question was settled when you signed the cloud contract. The routes further down this page are for data you have kept out, or for a company deciding it is time to bring the whole estate back.
For a company already on Microsoft 365, Copilot is the path of least resistance: a paid add-on at about €18–26 per user per month,12 EU Data Boundary for storage and partly for processing, no new processor to onboard.32 (The exception is Anthropic models selected inside Copilot, which are processed in the US.)20 For a Workspace tenant, Gemini is included in Business Standard at no extra cost,33 though in-region processing is limited to the Frontline Plus and Enterprise Plus editions.33 Neither gives you EU jurisdiction over the provider, but neither adds jurisdiction you haven’t already accepted.
The most capable AI models are American, and all now offer EU processing, typically at about 10% over the global price.34,35 That premium is confirmed for OpenAI and for Claude on Bedrock and Vertex; reported for Azure; assumed for Gemini. Anthropic’s own platform has no EU option.36 EU versions of the newest models can arrive later than the global release.37 The models are the same when they get there, under business terms that prevent training.
European companies like Langdock (Berlin) and EUrouter (Amsterdam) sit between you and these US models, offering one EU contract, admin enforcement of EU-only regions, and multiple models in one place.39,40 Routers also send work to EU-headquartered clouds (Scaleway, OVHcloud, IONOS and others) for open-weight models; only Claude, GPT and Gemini run on US clouds underneath.40 In Finland, Solita resells Claude on AWS with consulting and managed governance.41 A field note by Cloud2, a Finnish consultancy, walks through the EU setup on Bedrock and the traps that silently send data outside Europe.42 These services cost roughly 0–15% over the cloud’s own price,40,43,44 or about €25 a month for a workspace seat.39 What they don’t change, for the US-model routes, is the layer underneath: Claude still runs on AWS, controlled by a US parent. A European contract moves the invoice, not the jurisdiction.
Mistral, based in Paris, makes Large 4, the highest-scoring European-built model on the Artificial Analysis Intelligence Index (v4.3.2).45 It came out in preview in October 202646 and scores 38. Claude Opus 5.5 scores 51 at its default setting and 58 at maximum effort, on the same hard tasks.45 A contract with Mistral is a contract with a French company under EU law, though which data centres serve its standard API is not publicly confirmed. How hard a model thinks is a separate dial now, and turning it up raises both the score and the bill.
EU clouds like Scaleway47 sell open-weight models (models whose makers publish the model itself for others to download and run) by the token. We found no independent study of how these models handle routine office work like summarising, extracting and drafting; model cards and the Finnish leaderboard suggest mid-sized models do it well.48 For Finnish text in particular, some open models match older closed models and sit ahead of some older frontier ones on EuroEval.48
The strongest open-weight models come from Chinese labs.45 For many Finnish buyers this is not just a preference but a procurement bar: public-sector, defence and industrial supply-chain contracts often exclude them, and NIS2 supply-chain concerns push the same way.49 For companies that accept them, DeepSeek V4.1 Flash on Scaleway is the cheapest strong-band option.38 For those that don’t, the best non-Chinese open model (Meta’s Muse Spark Open, scoring 3945) and Mistral’s own models are the alternatives, scoring lower on the hardest tasks.
Running a model on your own hardware gives the most control over your data. Self-hosting the same model costs roughly three to 14 times as much a year as running it through an EU cloud,38 depending on company size and who does the work. Any company using AI has deployer duties under the AI Act.54,55 Most companies will never need this, which is reassuring, because most companies do not have a spare GPU engineer either.
Figure 2 sets out the yearly cost of the routes we priced, model by model. The models are grouped by how well they score on hard tasks, and the per-user plans form a group of their own.
Figure 2
What each route costs, by capability
Year-two cost for a company with 50 daily users, one line per route, grouped into four capability bands and per-user plans. Only US providers’ models reach the Top band, on the hardest tasks. In the Strong band the cheapest route is a Chinese lab’s model on a French cloud. The same open model on your own rented server costs about five times as much as by the token. Seats and metered API routes cost about the same at base use. A table view follows.
Yearly cost from year two, after setup, for a company of 200 staff with 50 daily users, before VAT, at list prices read 22–23 October 2026. Bands are scores on one public index of hard tasks (Artificial Analysis, v4.3.2) at each model’s default or medium setting (Top 50+, Strong 40–49, Mid 30–39, Basic below 30). On routine office work the differences between models are probably smaller. Seats and metered API routes cost about the same at base use, and the seat includes an application.
Each line runs from an in-house rate to a consultant rate for the people who keep the route running. How much people use it moves the totals more. At a fifth of the base usage, GPT-6 Astra falls from €26,900–29,300 to €8,300–10,700. *DeepSeek V4.1 Flash has no published default score, so it is placed by its best-setting score. Discounts and evaluation are not costed. Bought servers have no line of their own, and their costs are in the table.
Each row gives a year’s cost from the second year on, for a company of 200 staff with 50 daily users, in euros before VAT, ranging from an in-house labour rate to a consultant rate. On every route except the per-user plans, “Of which” is the token bill, the server’s rent or, for a bought server, a third of its purchase price plus power, and the rest is people’s time. A per-user plan’s cost is what the plan charges, since we count no upkeep for it after the first year’s setup. The assumptions and sources are in “Assumptions and method” at the end of the note.
| Option | Route | Score [best] | Year two | Of which | Stored in the EU | Processed in the EU | Provider under EU law only | Note |
|---|---|---|---|---|---|---|---|---|
| Top, 50 and above | ||||||||
| Claude Opus 5.5 | US provider’s model in an EU region | 51 [58] | €12,930–15,330 | €9,330 | yes | yes | no | |
| GPT-6 Astra | US provider’s model in an EU region | 50 [53] | €26,930–29,330 | €23,330 | yes | yes | no | |
| Strong, 40 to 49 | ||||||||
| DeepSeek V4.1 Flash, Scaleway | Open model on an EU cloud | [45]* | €5,040–7,440 | €1,440 | yes | yes | yes | The model is from a Chinese lab. |
| Gemini 3.8 Flash | US provider’s model in an EU region | 40 [41] | €7,100–9,500 | €3,500 | yes | yes | no | Tokens are costed at Google’s 2027 price, which is twice its price until the end of 2026. |
| GPT-6.1 Sol | US provider’s model in an EU region | 48 [52] | €8,270–10,670 | €4,670 | yes | yes | no | |
| Claude Sonnet 5.5 | US provider’s model in an EU region | 41 [56] | €8,270–10,670 | €4,670 | yes | yes | no | |
| DeepSeek V4.1 Flash, your own servers | Your own server, rented in Helsinki | [45]* | €100,000–110,000 | n/a | yes | yes | yes | This is indicative, for two rented four-GPU servers, and peak use would need more. The model is from a Chinese lab. |
| Mid, 30 to 39 | ||||||||
| Claude Haiku 5.5 | US provider’s model in an EU region | 34 [43] | €3,830–6,230 | €230 | yes | yes | no | |
| Mistral Large 4 (preview) | European-made model (Mistral) | 38 | €5,750–8,150 | €2,150 | yes | partly | partly | |
| GLM-5.3, Scaleway | Open model on an EU cloud | 34 [45] | €7,960–10,360 | €4,360 | yes | yes | yes | The model is from a Chinese lab. |
| Basic, below 30 | ||||||||
| gpt-oss-120b, OVHcloud | Open model on an EU cloud | 12 | €3,840–6,240 | €240 | yes | yes | yes | |
| Mistral Small 4 | European-made model (Mistral) | 11 | €3,880–6,280 | €280 | yes | partly | partly | |
| gpt-oss-120b, Scaleway | Open model on an EU cloud | 12 | €4,010–6,410 | €410 | yes | yes | yes | |
| Qwen3.8-27B, Scaleway | Open model on an EU cloud | 28 [34] | €5,520–7,920 | €1,920 | yes | yes | yes | The model is from a Chinese lab. |
| Qwen3.5 397B, Scaleway | Open model on an EU cloud | 18 | €5,610–8,010 | €2,010 | yes | yes | yes | The model is from a Chinese lab. |
| Qwen3.5 397B, OVHcloud | Open model on an EU cloud | 18 | €5,610–8,010 | €2,010 | yes | yes | yes | The model is from a Chinese lab. |
| Mistral Medium 3.5 | European-made model (Mistral) | 14 | €6,780–9,180 | €3,180 | yes | partly | partly | |
| gpt-oss-120b, your own server, bought | Your own server, bought | 12 | €17,530–24,730 | €6,730 | yes | yes | yes | Year two includes a third of the purchase, which we put at about €18,400, and year one all of it. |
| Qwen3.8-27B, your own server, bought | Your own server, bought | 28 [34] | €17,530–24,730 | €6,730 | yes | yes | yes | This is one server, and year two includes a third of its purchase, which we put at about €18,400. The model is from a Chinese lab. |
| gpt-oss-120b, your own server | Your own server, rented in Helsinki | 12 | €22,790–29,990 | €11,990 | yes | yes | yes | |
| Qwen3.8-27B, your own servers (one server) | Your own server, rented in Helsinki | 28 [34] | €22,790–29,990 | €11,990 | yes | yes | yes | One server may be enough for this load. The model is from a Chinese lab. |
| Qwen3.8-27B, your own servers (two servers) | Your own server, rented in Helsinki | 28 [34] | €36,580–44,980 | €23,980 | yes | yes | yes | The throughput benchmark suggests two servers. The model is from a Chinese lab. |
| Per-user plans, model varies | ||||||||
| Claude Team | Per-user plan | varies | €10,800 | n/a | no | no | no | |
| Copilot Business | Per-user plan | varies | €10,920 | n/a | yes | partly | no | It is an add-on, and the Microsoft 365 licence it needs is not counted. |
| ChatGPT Business | Per-user plan | varies | €12,600 | n/a | no | no | no | |
| Claude Enterprise | Per-user plan | varies | €14,920 | n/a | no | no | no | The cost is a seat plus usage at Sonnet 5.5’s API rates. |
| Mistral Vibe Team | Per-user plan | varies | €14,990 | n/a | yes | partly | partly | |
| Microsoft 365 Copilot | Per-user plan | varies | €15,600 | n/a | yes | partly | no | It is an add-on, and the Microsoft 365 licence it needs is not counted. |
| ChatGPT Enterprise | Per-user plan | varies | no public price | n/a | yes | partly | no | |
The sensitive work (the contract review, the customer-data clean-up) is also the hard work, where a weaker model’s errors cost most and are hardest to see. There are two honest routes.
One is the strongest model under the tightest available terms: an EU region, zero data retention, no training,13,14 with the residual US-jurisdiction risk accepted. For data already sitting in SharePoint that risk is the same risk the company has already accepted. The other is a European-law provider (Mistral, or an EU cloud), accepting lower capability on the hardest tasks and infrastructure whose details are not all public. For data subject to sector rules that require processing in approved EU environments, the European route is the only one.
In practice a company can split the two, routing sensitive requests to a European model and the rest to a frontier model. That pattern has an obvious failure mode: the routing is only as good as the detector that decides what is sensitive, and a missed name is a leak.56 (Names in Finnish do not always cooperate.)
We estimated what each route costs a year for a mid-sized company (200 staff, 50 daily AI users), from published list prices in October 2026.38 Figure 2, above, shows the estimates, and the table under it lists every figure. Our assumptions are at the end of the note.
On the hardest tasks, no model run by a European provider or hostable by a company reaches the top band as we draw it (index 50 and above) at any reasoning setting.45 On routine office work, we found no independent measurement, and the evidence from model cards and the Finnish leaderboard suggests the gap is much smaller.48
Below the top band, cost is not what separates the options. The model choice moves the total less than you’d expect, because the people who build and maintain the interface cost about the same on every per-token route.38 In Figure 2, a per-user seat plan and a metered API route cost roughly the same at base use once the API route’s upkeep is counted, and the seat comes as a finished product with search, apps and admin built in. The API route gives you a chat window you own. In our view, a company of 50 office users usually buys a product (Copilot, ChatGPT Business, Langdock); the per-use routes are for companies with a developer and a reason to build. For many the right answer is both: seats for heavy users, a cheap metered route behind a shared tool for the occasional ones.
None of the priced seat plans is fully processed in the EU.10,11,12,57 Two store and partly process there (Copilot12 and Mistral’s Vibe57); two offer no EU residency (ChatGPT Business10 and Claude Team11). The EU-resident ChatGPT Enterprise seat has no public price.
These are back-of-the-envelope estimates from list prices. Prices fall fast: one research group found them dropping by a factor of 9–900 a year.58 Treat every figure as a snapshot dated October 2026, and challenge our assumptions.
Start with where your data lives. If your documents and mail are already in Microsoft 365 or Google Workspace, the jurisdiction question for that data was decided when you signed. Adding AI through the same tenant or an EU region adds no new jurisdiction. Get everyone onto business terms and read what you signed.
Sort by category and write a short policy: what may be pasted, what goes to which route, what stays out of hosted services entirely. Enforce it with the admin controls the business plan gives you.
For the work that matters most, test. Run the five or ten tasks you would actually use the system for on three models: a cheap EU-hosted one, a mid-range US one and the most expensive US one. Compare the results against what a person would write, and see what the money buys. The strongest reason to test is cost: on our numbers the token bill per person ranges from about €0.40 to €39 a month,38 and a model at €0.40 that does the job saves a year’s worth of the difference over one at €16.
Leaving is cheap. Leaving a seat plan is trivial. For per-token routes, Mistral, Scaleway and (by its own statement) OVHcloud accept the widely used OpenAI client library with a changed address, key and model name;59,60,61 moving between them is mostly a configuration change. Moving to or from Claude is a code change, because Anthropic uses its own format in production.62 Either way, the cost of trying a European model and deciding it isn’t good enough is close to nil. Try the cheap one first.
Below the top band, the model choice moves the total less than you’d expect. Buy a product unless you have a reason to build, and spend the decision on terms and capability, not price. A first pass on your own tasks, with your own documents, is in our estimate an afternoon’s work.63
The costs in Figure 2 and its table are estimates, not quotes, built from list prices read on 22 October 2026. Seat plans are priced as shown to a buyer in Finland on 23 October, and dollar prices are converted to euros at the ECB rate of 21 October. Gemini 3.8 Flash is costed at the price Google lists from 1 January 2027, which is twice what it charges until the end of 2026.64 Every cost is a range from an in-house labour rate (€60 an hour) to a consultant rate (€100 an hour); both are assumptions, and the labour rate moves the figures more than any model choice below the top band.
We assumed 200 staff with 50 daily AI users, each generating about 150,000 tokens a day before reasoning (an estimate from published conversation sizes,65 not a measurement), split 88% input and 12% output. Models now produce hidden reasoning tokens billed as output; we applied a uniform 1.6× multiplier at default effort, based on about 35–40% reasoning in one gateway’s developer traffic.66 We assumed 50% of input is cached at about a tenth of the input price, for providers that publish cached-input pricing (Anthropic and Mistral do;34,67 OpenAI and Google are assumed to price it similarly;35 Scaleway and OVHcloud publish no cached-input price, so none is applied there47,68). For heavy or coding use, a metered route can cost far more than a flat seat. Anthropic’s published average for Claude Code, its coding agent, is about $13 per developer per active day.69 The EU-region premium is 10%, confirmed for OpenAI35 and Claude on Bedrock and Vertex;34 assumed for Gemini; reported for Azure.70 Capability scores are from Artificial Analysis Intelligence Index v4.3.2,45 each model at its default or medium setting (Mistral Large 4 has a single published setting). The bands (Top 50+, Strong 40–49, Mid 30–39, Basic below 30) are ours. The index is built from hard reasoning, coding and professional tasks. On routine office work the differences between models are probably smaller.48
Setting up takes eight hours for a seat plan and 40 for an API route, which also needs five hours a month of upkeep. Self-hosting adds 80 setup hours and 10 upkeep hours a month, growing with each extra server. A rented server is Hetzner’s GEX63 in Helsinki, with one 96 GB graphics card, at €999 a month plus a one-off setup fee of €999.71 We assume the rent covers power and add nothing for it. One is enough for gpt-oss-120b. Qwen3.8-27B needs one or two. The throughput benchmark we used measures a model writing text, while 88% of this load is text the model reads, which goes faster, so one may be enough. A bought server is an RTX PRO 6000 card in a host machine. The card costs about €14,700, the midpoint of two retail listings without VAT,72 and we put the host at about €3,700, about €18,400 in all. In year two the hardware counts as a third of that, the purchase spread over three years, plus about €600 of power, also our estimate. That third is an accounting figure, not cash. The strongest open models, such as DeepSeek V4.1 Flash, need two or more four-GPU servers, and running one yourself costs €100,000 or more a year.
Year one costs more than year two because it includes the setup, which comes to €480–800 for a seat plan, €2,400–4,000 for an API route and €8,200–13,000 for one rented server, Hetzner’s fee included. For a bought server, year one includes the setup hours and the whole purchase, where year two includes a third. Claude Opus 5.5 in an EU region, for example, costs €15,330–19,330 in year one and €12,930–15,330 in year two. One rented server for gpt-oss-120b costs €30,990–42,990 in year one and €22,790–29,990 in year two.
Not included: VAT, enterprise discounts, failover, base Microsoft 365 licences for Copilot,12 evaluation costs, exit costs, or the cost of a model being less capable for a given task. Gemini has no row among the per-user plans because it comes with Google Workspace (Business Standard is €13.60 per user a month),33 so its price is that of a whole office suite, not of an AI add-on. The Azure premium, and a premium of 17–18% reported for some models on Bedrock, are from a secondary source.70
Sources