Buy nice or buy twice
For a couple months, the most disciplined sentence in my homelab’s documentation was about a purchase I kept refusing to make. The RTX PRO 5000, a 48 GB workstation card, went on hold in my hardware log in mid-June with the release condition written into the hold itself: buy when a real workload is concretely blocked by the 24 GB on my RTX 3090. Over the following weeks I reaffirmed that hold three separate times.
The hold encoded two beliefs from earlier that month. The first was that I would never need more than 24 GB of VRAM, which history suggests is never going to be true; you’re always going to need more memory, more compute. Lesson learned again. The second was quieter: I hadn’t yet proven to myself that my local AI projects had legs, so a bigger card would have been a bet on a hobby. My hourly mail triage and the voice satellites in the house now run on models served from the garage machine, and both earn their keep daily.
Then two model releases in two days made the wall concrete. Muse Glimmer 30B fit only when I cleared the card for it, which made it useless beside the workloads already there; Nemotron 3.5 Lightning had no artifact that fit the 3090 cleanly at all. Both ship their native quantizations in a format built for newer silicon than the 3090’s. The buy trigger the hold had named all along, a consumer concretely blocked on 24 GB, fired twice inside 72 hours. And while I waited to save money, the market moved the other way: the card climbed from just under $4,750 when the hold was written to $5,999 the week it broke, with no comparable newer card coming to market for less, while the 3090’s resale value sat at its peak. I had also just retired a bunch of enterprise hard drives now worth more than I’d paid for them four years earlier, so waiting was no longer buying me anything except a worse trade, and there was a narrow window to trade one set of hardware for the next without spending the MSRP equivalent of a 1981 Ford Escort.
I made myself write the case out before ordering. If I’m honest, some slice of the decision was driven by impatience, and a larger slice by genuine excitement about what’s happening in the local-model space and the value the agents in this house keep producing. What pushed it over the edge was my bi-weekly model sweep, which kept surfacing releases in a class my 3090 could never serve. It felt like the train leaving the station whether I was on it or not, and I wanted my ticket punched. The rest of the case mattered more: this is not an economical purchase so much as a learning one. Understanding how these systems work, when to apply them, and how people actually benefit from them is my professional mission, whether or not I ever formally carry an AI title.
I mean for the purchase to fund itself the unglamorous way, by selling the hardware it replaces. The retired NAS’s drives were wiped, documented, and listed; the 3090 goes next; together the sales should net the cost down to less than the card was going for when the hold was written. I have a personal policy of buy nice or buy twice, and I genuinely thought the 3090 was me buying nice. It was me buying with the knowledge I had at the time: back then I could not have told you what makes one GPU better suited to AI work than another. The second purchase was an educated one.
The card has been in the machine since mid-August, running a 27-billion-parameter model (Qwen3.8) with headroom the 3090 never had under real load. The next day it took over the mail triage and beat the 14B I had spent July fine-tuning, without any fine-tuning of its own; the 14B and its adapter retired together. The retired router got its exam back too: the new model passed it, and the seat is re-armed, in shadow. The 3090 itself hasn’t sold yet; work commitments and travel keep pushing the listing out. I’ll get to it, and I’ll be honest: I won’t miss its rainbow LEDs lighting up my garage at night.