Quick question: whats the cheapest way to run DSv4 flash locally at 10tps?

This is the bare minimum needed for daily driving 100% on a production codebase reasonably well for me. Been wondering whats the cheapest anyone has gotten this range of tps. My current bet would probably be some specialise inferrence software with a 12-16gb vram gpu and rest on ram with one or 2 workstation cpus like xeon or something with lots of memory channels (cant remember but i think an intel one has 16 channels, cant confirm) Still wont be under 4k isd with current rates, what do you think? Is this the cheapest or are my estimates off?