id
gotchas/sent-regret-buying-without-validating-the-workload
Concepts
capexrent-vs-buydepreciationhobbylisttdp
Symptom
The strongest cross-cutting consensus in the whole slice is not about any particular GPU. It is that the buy decision should be preceded by renting, and the most-voted piece of practical advice in a 'talk me out of buying an RTX Pro 6000' thread (141 upvotes, 333 comments) is 97 upvotes of advice not to buy: 'Treat yourself to an RTX Pro 6000... ...by renting one. Then if you use it too much, you've completed validated a reason to buy it.' The follow-on cost argument at 45 upvotes: 'You can often rent a full server with an H200 (which blows the RTX 6000 out of the water) for a little over $2/hour, including electricity, bandwidth etc. Unless you're planning to run the GPU pretty much 24/7, the math doesn't work out.'
Root cause
The regret being described is capital-absorption regret rather than performance regret: buying a dedicated inference box for a workload that is bursty, experimental, or model-dependent means owning a depreciating asset before knowing what it will run. A separate thread makes the same point from the other end, at 85 upvotes: 'buying hardware to run specific models is not sustainable unless you are 110% satisfied with the model at hand and will not find the urge to upgrade to something even better. Its a race against big datacenters and you might as well just rent the hardware.'
Workaround
Rent the class of hardware you are considering for the workload you actually have, for long enough to establish the usage level. Buy when the usage is established and continuous, not when the specs are attractive.
notes
Recorded as split rather than consensus, because the same thread's counter-argument is real and is also highly upvoted (181 upvotes, from someone who has now bought three): 'If you want a real reason not to do it, it's that you'll never be satisfied. No matter how much hardware you have, you'll always want more.' The threads here are Hobbyist-flavoured, so rent-first is advice for individual buyers, not for a team with a steady serving load. A second cost dimension that is purely non-technical and shows up repeatedly: resale value and depreciation - 'Being older and less popular, both alternative options are more likely to depreciate over time' - and the 2026 memory-price effect, where a commenter notes 'RAM Prices are skyrocketing' and still lands on used 3090s because VRAM per dollar is what binds.