nvidia's flagship GPU just crossed $5,100 — for local AI, that math doesn't work anymore

The RTX 5090 just crossed 7.3 million won in South Korea — about $5,112 at today's exchange rate — for a card Nvidia still officially prices at $1,999. That's not a scalper listing or a launch-week fluke; it's the going retail rate as of this week, per ZDNet Korea's tracking and Tom's Hardware's reporting. If you've been eyeing a 5090 for a local-AI build because 32GB of VRAM sounded like the last upgrade you'd ever need, the number that actually matters just changed: at 2.5x sticker price, does the extra memory still pay for itself?

What actually happened this week

Tom's Hardware reports the RTX 5090 climbed roughly 1.5 million won — about $1,050 — in a single month, pushing it past 7.3 million won at South Korean retail. TechPowerUp's independent tracking confirms the same up-to-30% swing across the whole RTX 50 lineup, not just the flagship. The mechanism is dull but important: Nvidia doesn't sell the GPU die and GDDR7 memory as separate line items to its board partners — it ships a bundled kit, die plus memory, at one price. When GDDR7 modules climb to around $20 apiece and TSMC raises what it charges for the wafers those dies are cut from, both costs land on the kit at once, with no lever for a partner like Gigabyte or Asus to absorb just one of them. We've been tracking this thread since the RTX 50 Super line got shelved over the same memory-chip math — this is that story continuing, not a new one.

32GB sounds great — until you price it per gigabyte

Here's the part nobody selling you a 5090 wants to walk through. At $1,999, 32GB of GDDR7 worked out to about $62 per gigabyte — expensive, but defensible if you were serious about 70B-class models at Q4 with real context headroom. At $5,112, that same 32GB now costs $160 per gigabyte. A used RTX 3090 — same 2020-era architecture, same 24GB, no warranty — is going for $1,050 to $1,296 on eBay right now, itself pulled higher by the same memory shortage. That's roughly $50 per gigabyte, and two of them gets you 48GB of usable VRAM for less than the price of one 5090, if your case and power supply can take it.

Before this hike, the pitch for a 5090 was simple: if you're burning $150–$200 a month on cloud API calls, a $1,999 card pays for itself in under a year, then it's free. At $5,112, that payback stretches past two years for most people's actual usage — and a used RTX 3090 or the RTX 5070 Ti gets there in months, because you're not paying for memory you won't touch until context windows grow further. If your workload genuinely needs everything 32GB unlocks — fine-tuning, long-context agents, multiple models loaded at once — the math still works. For everyone else asking what's the biggest model they can comfortably run, the cheapest path to 70B-class inference doesn't touch a 5090 at these prices.

Who this actually makes sense for

I'd skip the 5090 at $5,100 unless you're already certain you need everything it offers — and honestly, that's a small slice of the people asking whether to buy one. Most local-AI buyers are chasing a specific model size, not maximum headroom. If South Korea's pattern repeats in the US — and every 2026 hike so far has landed here within weeks — waiting won't make the card cheaper either. The realistic move is buying a tier down and living with 16GB or 24GB, or splitting the budget across two cheaper GPUs instead of chasing one card that does everything.

None of this means local AI got more expensive across the board — it means the top of the stack did. The $220 cards that actually run local AI well haven't moved nearly as much, and that's still where most people should look first. I'll keep tracking whether this hike crosses the Pacific; if the 5090 settles back under $3,000 in the US, the math in this piece changes, and I'll say so.