The Training Bill Has Come Due
A recent discussion on Hacker News, If I were 17, I'd learn how to build LLMs from scratch, puts numbers on something most companies already know: training an LLM from scratch is economically unrealistic for anyone but the largest labs.
Even renting hardware for a 1-2 month training run, the costs are out of reach for the vast majority of companies. As one comment in the thread puts it: it's an interesting learning exercise, but some of what you learn can be applied elsewhere in the process.
What This Means for Your Hosting Strategy
For Danish companies, the conclusion is simple: training isn't your competitive advantage. It's open weights combined with self-hosting that gives you control, data security, and cost predictability.
- Training from scratch requires billions in investment and months of compute time. Only the giants can afford it.
- Fine-tuning open weights is possible with moderate hardware, but requires specialized expertise.
- Self-hosting open weights is the practical path for most companies.
The EU Dimension
Infercom puts the distinction precisely: when you run an open-weight model on EU infrastructure, the model's developer has published the weights with no insight into how they're run. Your data stays on European soil. That's the decisive difference between an API call to an American cloud and self-hosting on EU infrastructure.
The Practical Takeaway
Open weights have completely changed the rules. Training economics make it impossible to compete on model building, but they give everyone access to top-tier models for self-hosting. The question is no longer whether you can build AI, but whether you can host it securely and in compliance with the law.
For most companies, this comes down to choosing the right open-weight models, finding EU-based hosting that can deliver GPU capacity, and ensuring EU AI Act compliance without losing control over your data.
Liviate helps Danish companies do exactly that: get open-weight models into production on EU-hosted infrastructure, without having to build anything from scratch.