Your data wants to stay home

For three years, the architecture of corporate AI had one shape: your data travels to someone else's computer, and you trust the contract more than you can verify the practice. This arrangement had one decisive defense. The hosted models were simply better, and nothing you could run yourself came close.

That defense expired this year. Open-weight models now match or beat the hosted APIs on the tasks most businesses actually run: extraction, classification, summarization of domain documents, retrieval over internal knowledge. Not on every benchmark, and not on frontier reasoning, but on the ordinary work that fills ordinary backlogs. The quality gap that justified the data exodus has closed for the median task.

What remains is a different calculation, and many organizations are surprised by how it lands. Running locally costs more in hardware and expertise than the API pricing pages suggest, and less in risk than the security team's stomach suggests. Sensitive documents stop appearing in third-party logs. Compliance stops depending on a vendor's roadmap. The per-seat rent stops scaling with your headcount.

The honest trade-off is operational. Someone has to own the machine. Model updates, GPU drivers, monitoring, fallback behavior: the API made these someone else's problem, and local deployment makes them yours. This is real work, and organizations with no one to do it should not romanticize the perimeter. The correct answer for a ten-person firm is usually still an API with a good contract.

But the default has inverted. The question used to be why would you run your own models. For regulated industries, sensitive workloads, and anyone whose inference bill has become a board topic, the question is now why would you not. The perimeter is back, not as nostalgia, but as arithmetic.

The lab's expectation: hybrid becomes the normal architecture. Local models handle the bulk of routine tasks inside the perimeter, frontier APIs handle the hard cases under explicit policy, and a routing layer decides which is which. Data residency stops being a legal footnote and becomes an engineering property. Your data wanted to stay home all along. The models finally let it.