The frontier is moving toward smaller models running closer to your data. In Alaska that is not a cost story, it is a permission story.

In June 2025, the Deep Learning Efficiency Research Group at NVIDIA Research published Small Language Models are the Future of Agentic AI. It is a position piece rather than a benchmark, and its argument rests on three claims: small models are already capable enough for many agent tasks, they are inherently better suited to agentic systems, and they are more economical at the volumes agents actually run at.

The definition they propose is useful because it is operational rather than a parameter count: a small model is one that runs on common consumer devices with responses fast enough for a single user. Anything else is a large model. That framing puts the emphasis where it belongs, on where the model can physically live.

The paper does not argue that large models go away. It argues for heterogeneous systems, where general conversation and genuinely hard reasoning route to a large model while the repetitive specialized steps - the bulk of the invocations - run small. NVIDIA also maintains a public page collecting critiques of their own position, which is a posture worth noticing in a field where most claims arrive without a rebuttal channel.

This paper reads as a cost optimization if your business is in a city with fibre and no data governance obligations. In Alaska it reads as a change in what is permitted at all.

A system built entirely from small models will disappoint on tasks that genuinely require breadth: open-ended reasoning, unusual language, synthesis across domains that were never in the training data. The design that works is small by default with a route to a larger model when a step needs it, and the engineering discipline is in the routing rather than in the models.

The other honest point is that a locally deployed model is infrastructure you now own. Somebody updates it, monitors it, and notices when it degrades. For an organization with an IT function that is a small addition. For a six-person office it is a real consideration, and we would rather raise it during scoping than after the hardware arrives.

GENOME, the memory server we wrote and benchmarked in public, has an offline mode because we needed one before there was a fashionable argument for it. Kryos, our agent language, was built so an agent’s actions are auditable line by line, which matters far more when the agent is running inside a regulated organization’s own boundary than when it is a hosted service somebody else is responsible for.

That is why a tenant-resident or fully air-gapped deployment is a configuration choice for us rather than a custom engineering project with a custom price. A firm assembling a product from third-party services cannot make that offer, because the answer to where the data goes is set by a vendor they do not control.