AI on-premises: What the much-quoted 82 percent really mean

A number has been running through the trade press for months: 82 percent of companies want to operate AI at least partially on-premises or at the edge. It sounds like a turnaround – away from the cloud, back into your own data center. We looked at where the number came from. The result is instructive, in both directions.

What the Number Really Says

The source is the IDC CIO Playbook 2026, created on behalf of Lenovo – i.e. by a manufacturer that sells exactly the hardware in question. This does not make the study worthless, but belongs to the classification.

The sample is more important: 800 companies in Europe and the Middle East were surveyed, all with more than 1,000 employees. Not a single medium-sized company in today’s sense. And the figures for Germany are clearly different than the often cited European ones:

	Hybrid	on-prem/Edge total	only public cloud
Europe and Middle East	58 %	82 %	18 %
Germany	50 %	75 %	25 %

Germany is therefore in the hybrid approach underground the European average and has the highest proportion of pure public cloud users. Who leads the 82 percent as evidence of a German on-premises trend, cites the wrong line.

The Counter-Evidence You Should Know

It gets even more uncomfortable. The cloud monitor from KPMG and Bitkom surveys German companies with more than 50 employees – in other words, SMEs. There are 96 Percent of Cloud Users Purchase AI from the Cloud. The classic own data center plays only a role for AI workloads at 13 percent, and the trend is falling.

Both findings side by side give a realistic picture: companies build hybrid AI architectures. Medium-sized companies predominantly use AI from the cloud. Anyone who claims otherwise will sell you something.

Why the topic still belongs on the agenda

Because something else moves in parallel, and that is well documented. The Bitkom Cloud Report 2026 shows that 85 percent of German companies surveyed feel too dependent on US providers and 64 percent rethink their cloud strategy because of the political situation. In Bitkom’s AI study, 93 percent say they would prefer German AI providers – while 70 percent actually use ChatGPT.

This gap between desire and practice is the real point. And the options to close them have become significantly better in 2026:

• The AWS European Sovereign Cloud has been live since January 2026 with a region in Brandenburg.
• Microsoft and Google are expanding sovereign offerings in Germany.
• Above all, however, the open models have grown up: Mistral Large 3, Qwen, DeepSeek, gpt-oss and Gemma are under real open source licenses (Apache 2.0 or MIT) – no longer under craft licenses. The quality gap to the closed top models is about four to six months.

For the tasks that actually arise in SMEs – Make documents searchable, classify content, extract information, answer questions on your own knowledge base – this distance is practically meaningless. It counts for highly complex agent tasks, not for RAG.

The only question that determines the cost

Here it becomes unpleasantly concrete, because in the net circulate rules of thumb ("from two million tokens per day worth on-premises"), which are methodologically not resilient. What can be said seriously:

The price gap between self-hosted open models and the major API providers is real and several times per token. But the break-even depends almost completely on the utilization. A five percent utilized GPU costs you several times more than what the same workload would have cost via an API. Hardware only pays for itself when it runs. Deloitte cites as an orientation: self-operation becomes interesting when the cloud costs exceed 60 to 70 percent of the acquisition costs of an equivalent own system.

The practical consequence: Do not expect token prices, count on utilization. Those who use AI sporadically drive cheaper with the cloud – significantly.

When on-premises is still the right answer

Not because of the cost, but because of the data. In the IDC survey, data protection and compliance in Germany is the number one driver for on-premises operations – not the price. And this is in line with our project experience: Personal data, construction data, patient data or contracts that are not allowed to leave the house are a sufficient reason, regardless of any economic calculation.

The pragmatic middle ground in most projects is not “all or nothing”, but: processing sensitive data in-house, uncritical workloads in the cloud – exactly what the studies call “hybrid”.

Conclusion

The 82 percent is not a request to bring your AI into your own data center. You are a group finding, commissioned by a hardware manufacturer. The right question for a medium-sized company is not “cloud or on-premises?”, but: Which data are not allowed to leave the house – and how high is the load? The answers give the architecture itself. It is usually hybrid.

How this can be implemented with verifiable answers on your own data, we describe in the AI Consulting; to the question Public, Private or Hybrid we go on the page Cloud Migration in detail.

Sources: IDC CIO Playbook 2026 (“The Race for Enterprise AI”, Europe & Middle East, on behalf of Lenovo, n=800, companies with more than 1,000 employees, survey September/October 2025); KPMG/Bitkom Cloud Monitor; Bitkom Cloud Report 2026; Deloitte Tech Trends 2026. As of 14 July 2026.