Three months in with the Spark: what I actually use it for
The novelty of running big models on your own hardware wears off after a few weeks, and what is left is the honest question: what do you actually reach for it to do? Three months into living with the DGX Spark, here is where it earned a permanent place in the workflow, and where the cloud is still the right call.
Where local won
Anything with data I would hesitate to paste elsewhere. This is the biggest one and it is not close. The moment a task involves internal logs, config, incident history, or anything with even a whiff of sensitivity, the local box is not a compromise, it is the only option that lets me use a model at all. The hesitation that normally makes me not use AI on sensitive work just disappears when the model physically cannot phone home.
High-volume, repetitive grunt work. Summarizing, classifying, drafting a hundred variations, tagging a pile of records. The kind of thing where a metered per-token cloud bill makes you ration, and rationing makes you not bother. On local hardware the marginal cost of another thousand tokens is basically zero, so I just run it, and that changes what I am willing to automate.
Interactive, fast-loop work. When I am iterating quickly, the absence of a network round trip and a rate limit makes the whole thing feel like a local tool instead of a remote service. Small latency differences change behavior more than people expect.
Where the cloud still wins
I am not pretending local is strictly better. For the genuinely hardest reasoning, the frontier hosted models are still ahead, and when a task needs the absolute best available intelligence and the data is not sensitive, that is the right tool. The honest framing is not local versus cloud. It is a portfolio: the local box for private, high-volume, interactive work, the frontier models for the hardest reasoning on non-sensitive tasks. Most weeks I use both, for different jobs.
The thing I did not expect three months ago is how much the mere existence of a private, always-available model changed what I am willing to try. When using AI on sensitive work carries no exposure and no marginal cost, you experiment more, and the experiments are where the useful patterns come from.