Also, I would anticipate at least a 5 year lifespan for a current generation card. The 3090 is still respectable simply because it has 24GB of RAM which, for years, has been the limiting factor for ML at home. If you got a 6000, sure it’s going to cost 7-8k, but the resale value is likely to be very good. Even the 3090 is 50%+ of RRP still. And if you’re not doing LLMs, it’s an interesting value proposition for “classic” CNN vision model training. You can fit enormous batch sizes on 96 GB. The biggest reason to upgrade is perf/watt has about doubled (eg 4000 pro Blackwell is half the 3090 for similar).

People tend to assume the capex is thrown away but as we’ve seen with RAM, don’t be so sure you won’t be able flip it if you need to.