PSA: NVIDIA DGX Spark has terrible CUDA & software compatibility; and seems like a handheld gaming chip.

> Personally I'm interested in setting one of these devices / clusters up for making a Claude Code 5x plan go further. Using Haiku to route simple queries against my 'free' local GPT 120 OSS / Qwen3. And use APIs like Sonnet / Opus for the heavy thinking. in that case there really is not much case right now for spark IMO > Even then, the business case of a $100 delta between 5x and 20x plans equates to 1 to 2 years before such a setup breaks even based on my napkin math - depreciation, power cost, APIs becoming cheaper etc. Without the local emphasis, we'd be better off just using APIs. there are factors that go beyond cost per token or per task, like having stuff you don't want to send to the cloud like passwords, reliability of your workflow as each model has its quirks you get used to but the online ones they can just come and go or change overnight, and there's also models that are bad at certain things or just refuse to respond at times, for some stuff that you may incorporate in your workflow like translation or reference checks latency also makes a huge difference and as you say the mental drag of not knowing how much are you spending on some task (although there are subscription plans that partially take care of that) this mental drag extends to having to adapt your workflow and your knowledge of models not just to the changes you make but the changes 3rd party providers make, which in my experience happen extremely often - if you start using these systems sufficiently, you start having a level of dependency that should account for the risk of 3rd parties having outages, or making changes (including temporary nerfing) I'm very conservative with my LLM usage, for instance in the case of coding i review everything - there's stuff that GPT and Claude models do better or get more consistently correct faster than say Qwen3 models, but I check and review everything the models do anyway, it's stuff I would have been able to do on my own with enough time otherwise I wouldn't just use it blindly for critical stuff in my work - that just sounds crazy to me - and then there's stuff like generative audio, images, translation, etc which although they're unlikely to become critical in my personal workflow, it's nice to be able to do offline too