A mid-market company with 500 employees rolls out Microsoft Copilot. Year-one bill, licenses only: EUR 168,000. Add training, implementation, support. By year five the line item passes EUR 650,000. Run the same exercise honestly across cloud and on-premise, and the result is the same shape every time.
This is the math IT directors are running right now — and it rarely favors cloud. The deeper truth: most companies don’t actually know their AI costs. The invoice is one part of the story. Implementation, training, compliance, and creeping vendor lock-in add the rest.
What you’re actually paying
Vendor list prices look manageable. The all-in numbers don’t.
Microsoft Copilot
- License: $30 USD per user per month (~EUR 28). (Microsoft 365 Copilot.)
- Prerequisite: Microsoft 365 E3 or E5 license (additional EUR 36–57 per user per month).
- 500 users, 5 years: EUR 840,000 (Copilot + E5 base).
- Copilot share alone: subtract existing M365 spend → realistically EUR 504,000–672,000 pure Copilot cost.
What gets missed: Copilot licenses are billed per user, not per use. Gartner (2024) finds only 30–40% of licensed users use Copilot regularly. You pay for 500 seats, ~300 of them barely touch the tool.
ChatGPT Enterprise
- License: ~USD 60 per user per month (with minimum commit).
- 500 users, 5 years: ~USD 1,800,000 (≈ EUR 1,650,000).
- Limitation: No direct access to internal systems. Data fed in manually or via API.
Google Gemini for Workspace
- License: USD 30 per user per month (Business) or USD 36 (Enterprise).
- 500 users, 5 years: USD 900,000–1,080,000 (≈ EUR 825,000–990,000).
- Prerequisite: Google Workspace. Microsoft shops add migration cost.
API-based usage (OpenAI, Anthropic, Google)
- Cost: per token / per request. Variable, hard to plan.
- Example: GPT-4o at $2.50 / 1M input tokens, $10 / 1M output (as of 2025).
- Problem: heavy workloads explode quickly. 500 documents per day hits EUR 5,000–15,000 per month before anyone notices.
The hidden costs
Licensing is the tip:
| Cost factor | Typical share | Often forgotten? |
|---|---|---|
| Licenses | 60–70% | No |
| Implementation & integration | 15–25% | Yes |
| Training | 5–10% | Yes |
| Ongoing support | 5–10% / year | Yes |
| Compliance (DPA, DPIA, audits) | Variable | Yes |
| Vendor lock-in (migration cost) | Unknown until you try | Yes |
On top: a cost no vendor will quantify. Data sovereignty. When contracts, customer communications, and internal documents are processed in a US vendor’s cloud, you don’t have full control anymore. Not a classic line item — but a risk that can show up as a GDPR fine.
Another risk that rarely makes the TCO sheet: cloud spend regularly overshoots budget. Per the Flexera State of the Cloud Report 2024, actual enterprise cloud spend runs on average 32% over budget (Flexera, 2024).
What on-premise AI costs
Higher upfront, no per-user fees. That changes the curve.
Turnkey on-premise appliance
- Hardware + software + setup: from ~EUR 52,000
- Ongoing cost: maintenance and power, ~EUR 3,000–5,000 / year
- 5-year TCO: ~EUR 67,000–77,000
- Users: unlimited. No per-user license. No token caps.
Self-hosted open source (Llama, Mistral)
- Hardware: GPU server from ~EUR 15,000–40,000 (NVIDIA A100/H100)
- Setup: internal IT or integrator — EUR 20,000–80,000
- Maintenance: in-house team required — ongoing personnel cost
- 5-year TCO: EUR 100,000–300,000, heavily dependent on complexity and what you do yourself
The full TCO: 500 users, 5 years
| Solution | 5-year TCO | Per user / month | GDPR risk | Maintenance |
|---|---|---|---|---|
| Microsoft Copilot | EUR 504K–672K | EUR 17–22 | Medium (third-country transfer) | Low (SaaS) |
| ChatGPT Enterprise | ~EUR 1,650K | ~EUR 55 | High (US cloud) | Low |
| Google Gemini Enterprise | EUR 825K–990K | EUR 28–33 | High (US cloud) | Low |
| On-premise (turnkey appliance) | EUR 67K–77K | EUR 2–3 | None | Medium |
| Self-hosted open source | EUR 100K–300K | EUR 3–10 | None | High |
Same gap, every time: on-premise is 7–20× cheaper than cloud AI at this scale — and removes the data-protection problem at the same time.
Why the gap is this big
Cloud’s per-user model scales linearly: 2× users, 2× cost. On-premise scales the hardware requirement, not the bill — and a single powerful GPU server for 500 users costs a fraction of the cumulative license fees.
There’s also a structural issue with cloud pricing: it’s calibrated for the US market, where data-protection requirements are lower. European companies pay the same list price plus the cost of GDPR compliance — DPAs, DPIAs, third-country safeguards, works council agreements. This “hidden compliance tax” doesn’t appear in any cloud TCO calculator. It just shows up in your team’s calendar.
The catch with on-premise: upfront investment and local infrastructure. What surprises people: no ML team, no data center, a server rack is enough. Turnkey solutions like contboxx Vault bundle the hardware.
Run the TCO for your specific case What does your AI spend look like today — and what could you save? Fifteen minutes, we’ll show you the math. Book a free demo
The factor nobody quantifies: opportunity cost
What does it cost when your AI can only search emails and Teams chats — but not the contracts in SAP, the documentation in Confluence, or the files on the network share? Copilot only knows Microsoft 365. For everything else, employees fall back on manual search.
Turnkey on-premise platforms with broad system reach (~40 sources) unlock the full body of enterprise knowledge. That productivity gain isn’t on the TCO sheet. It’s the actual reason companies switch.
When cloud AI still makes sense
To be fair:
Cloud makes sense when:
- You have fewer than 50 users (per-user cost stays small)
- Your data isn’t sensitive (no personal data, no IP)
- You need to start in days, not weeks, and you don’t have a data center
- You’re already deep in the vendor’s ecosystem
On-premise makes sense when:
- You have more than 200 users (fixed-cost amortization kicks in)
- You handle sensitive data (contracts, HR, customer comms)
- Compliance is non-negotiable
- You don’t want to be locked to one vendor
- You need to search across more than one ecosystem (SharePoint + SAP + Confluence + file shares)
FAQ
What does AI cost per employee per month?
Cloud AI: EUR 17–55 per user per month (Copilot ~EUR 28, ChatGPT Enterprise ~EUR 55). On-premise: ~EUR 2–3 per user per month at 500 users over 5 years. Cloud scales linearly with user count. On-premise stays mostly fixed — that’s where the divergence comes from.
Is on-premise AI really cheaper than cloud AI?
From around 200 users on, yes — significantly. The fixed cost (from EUR 52,000) spreads across all users without per-seat licensing. At 500 users over 5 years, on-premise is 7–20× cheaper. Below 50 users, cloud can win because there’s no upfront investment to amortize.
What hidden costs does cloud AI have?
Implementation (15–25% of total), training (5–10%), ongoing support (5–10% per year), compliance work, vendor lock-in. Plus shelfware: per Gartner, only 30–40% of Copilot license holders use it regularly. Actual cloud spend runs 32% over budget per Flexera. None of that is on the vendor’s price page.
Bottom line
AI cost isn’t a technical question. It’s strategic. Cloud wins on day one and loses on year three. On-premise asks for the upfront investment and pays back 80–95% over the lifecycle — with full data control thrown in.
The right question isn’t “can we afford on-premise?” It’s “can we afford cloud, long-term?”
If you’re replacing shadow AI with a controlled solution, start with the cost math, not the feature list.
TCO from EUR 2 per user per month contboxx Vault: on-premise, no per-user licenses, no token caps.