Many companies in regulated industries may only run AI on their own infrastructure. The question then is not whether, but how far behind the frontier.
The gap in numbers
We compared when the best open model that fits a single eight-GPU server reached the performance a cloud-only model already had (as of October 2026):
- Coding: about 3 to 5 months behind.
- Without models of Chinese origin: about 8 to 10 months.
- The harder the task, the wider the gap.
Notably, open models have caught up strongly - but the catch-up came mainly from China. For companies that may only use Western models for compliance reasons, the gap is considerably larger.
The options
- Open models on own hardware - full control, but a lag and own operations.
- Closed frontier models in air-gapped environments - some vendors now license their models for operation without internet access. Close to the frontier, but tied to one vendor and its hardware.
- Cloud with contractual guarantees - EU processing, no data retention, no training on customer data. Sufficient for many use cases if governance approves it.
What we recommend
Benchmarks measure model capability, not project results. What matters is which tasks a model handles reliably - and for many well-defined tasks, a model a few months behind the frontier is enough. Design the architecture so the model stays swappable. Then the lag becomes an upgrade path.
Sources of the analysis: public benchmarks and model cards, as of October 2026. Benchmarks do not reflect delivery results.
