The portfolio approach to enterprise AI
When companies adopt AI, they often think in all-or-nothing terms: either you rent everything from a cloud provider, or you host everything yourself. But reality is more nuanced. The right architecture is a portfolio of models matched to tasks, and not every model in that portfolio needs to be rented.
High-volume, well-defined workloads such as classification, extraction, code completion, and document processing make up the bulk of enterprise token volume. These workloads are ideal candidates for hosting on your own infrastructure with open-weight models, writes Zentera.
Why your own infrastructure makes sense for high volume
- Cost control: High-volume workloads generate a lot of tokens. Hosting them on your own infrastructure with open-weight models eliminates recurring API fees.
- Data protection: Local hosting gives you full control over data, which is critical under GDPR and the EU AI Act.
- Latency: Locally hosted models can offer lower latency for high-volume workloads.
- Customization: Open-weight models can be fine-tuned for specific tasks without depending on an external provider.
The European perspective
Nvidia's paper on open weights and US AI leadership has put open weights on the infrastructure policy agenda. For Europe, that means open weights aren't just a technological option, but a strategic choice about digital sovereignty, writes noze.it.
For Danish companies, this means the portfolio approach isn't just an economic decision, but also a strategic position within a European AI landscape where data sovereignty matters more and more.
What does this mean for your hosting strategy?
- Identify high-volume workloads: Classification, extraction, code completion, and document processing are typical candidates.
- Choose open-weight models: Find models that are good enough for the task, even if they aren't frontier models.
- Host on your own infrastructure or with a local provider: Use EU-based hosting to ensure data sovereignty.
- Only rent what makes sense: Low-volume, specialized workloads can still be rented from cloud providers.
The portfolio approach is about optimizing both cost and control. For most companies, it's the high-volume workloads that drive the bill, and those are precisely the workloads that benefit most from running on your own infrastructure.
At Liviate, we help Danish companies build the right portfolio architecture with open-weight models hosted on European infrastructure.