If you’ve tried to buy an RTX 5090 at MSRP lately, you know the situation. Scalper prices are 40-60% above retail. The RTX 5080 is sporadically available but rarely in stock for more than minutes. GDDR7 memory shortages, surging AI demand, and tariff uncertainty have created the tightest GPU market since the crypto mining craze of 2021.
So if you’re building a local LLM rig in 2026, the question is straightforward: buy now, or wait it out?
NVIDIA GeForce RTX 3090
24GB GDDR6XUsed RTX 3090 at ~$820 — 24GB VRAM is future-proof and supply is unaffected by the GDDR7 shortage, and it won't depreciate further. The shortage-proof pick.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
What’s causing the 2026 GPU shortage
Three factors are converging:
1. GDDR7 supply constraints. Samsung and SK Hynix are the two main GDDR7 producers. Yields on the new memory standard are lower than expected, and every RTX 50 series card needs it. NVIDIA can’t build enough cards to meet demand until GDDR7 production stabilizes — likely Q4 2026 at the earliest.
2. AI demand from datacenters. Enterprise buyers are purchasing consumer GPUs (especially RTX 4090s and 5090s) in bulk for AI inference clusters. This was supposed to stop when NVIDIA cracked down on consumer-GPU-in-datacenter usage, but enforcement has been inconsistent.
3. Tariff and trade uncertainty. New US tariffs on Chinese-manufactured electronics components have disrupted supply chains for AIBs (add-in board partners) like ASUS, MSI, and Gigabyte. Even when NVIDIA ships enough chips, board partners struggle with component costs.
The result: new RTX 50 series cards are hard to find at MSRP, and when available, they sell instantly. The RTX 40 series is no longer manufactured. The used market is the most stable segment right now.
Current GPU pricing landscape (September 2026)
| GPU | Launch MSRP | Street price now | Availability |
|---|---|---|---|
| RTX 5090 | $1,999 | ~$4,900 | Very scarce |
| RTX 5080 | $999 | ~$1,400 | Sporadic |
| RTX 5070 Ti | $749 | ~$1,050 | Limited |
| RTX 5070 | $549 | ~$875 | Limited |
| RTX 4090 (used) | — | close to new | Available |
| RTX 3090 (used) | — | ~$820 | Available |
| RTX 3060 12GB (used) | — | ~$250 | Available |
| RTX 3060 12GB (relaunch) | $329 | $329-480 | Available |
Every current-generation card sells well above the price it launched at, and the gap has widened through 2026 rather than closing. The RTX 5090 is the extreme case: it launched at $1,999 and now goes for roughly two and a half times that.
Notice the pattern: used cards have the most stable and predictable pricing. The RTX 3090 has drifted up to around $820 over 2026, but slowly and without the availability swings of increasing. The RTX 4090 used market hovers around $1,400-1,600. These prices are set by actual supply and demand from individual sellers, not by GDDR7 availability or scalper bots.
VRAM capacity memory bandwidth Specs are manufacturer figures. Bar lengths are scaled independently per metric.
My recommendations for each budget
Around $250: Buy a used RTX 3060 12GB now
At around $250, the used RTX 3060 12GB is the cheapest path to running 7B models locally. The RTX 3060 relaunch arrived in June 2026 at $329 and did not push used prices down — the memory shortage lifted the whole 12GB segment instead, so used cards cost more now than before new stock landed. If you find one under $250 in good condition, take it.
Check NVIDIA GeForce RTX 3060 12GB on Amazon→Buy on Shopee SG→Around $820: Buy a used RTX 3090
This is the single best GPU purchase you can make during a shortage. The RTX 3090 at ~$820 used gives you 24GB VRAM — enough for 34B models — and its price has completely stabilized. It won’t go up because newer cards exist. It won’t go down much because 24GB at this price is irreplaceable.
The 3090 is shortage-proof. It doesn’t need GDDR7. It’s not manufactured anymore, so NVIDIA’s production issues don’t affect supply. The used market has steady, predictable inventory from miners and gamers who upgraded. For a buying checklist, see the used RTX 3090 guide.
Check NVIDIA GeForce RTX 3090 on Amazon→Buy on Shopee SG→Around $1,050: RTX 5070 Ti, at whatever it actually costs
The RTX 5070 Ti is a strong card — 16GB GDDR7, 5th-gen tensor cores, and excellent efficiency. It launched at $749 and now sells for around $1,050, and waiting for it to return to MSRP has been a losing bet for most of 2026. If your local retailer has stock or you can get on a notification list, it’s worth buying.
Compare it against 16GB alternatives before buying. The RTX 4070 Ti Super is the same 16GB at around $800, and a used RTX 3090 is 24GB at around $820 — both worth weighing against a $1,050 5070 Ti, and you get 24GB VRAM and faster performance. The premium over MSRP doesn’t justify what you’re getting.
Check NVIDIA GeForce RTX 5070 Ti on Amazon→Buy on Shopee SG→Around $2,200: RTX 4090
The RTX 4090 is the performance king for single-GPU LLM inference. It is no longer manufactured, and used prices have converged toward what a new one costs — roughly $2,200 — so the used discount that made it attractive in early 2026 has largely gone. It is still the fastest consumer GPU for token generation at ~65 tok/s on 7B models. Its 24GB VRAM handles 34B models comfortably.
Around $4,900: the RTX 5090 question
The RTX 5090 at its $1,999 MSRP would be a reasonable buy — 32GB GDDR7, fastest consumer GPU available. That price no longer exists. It sells for roughly $4,900, and at two and a half times MSRP you’re paying a 50%+ premium for 32GB when a used 4090 gives you 24GB at half the price. The 8GB VRAM difference does not justify a $1,200+ premium.
Wait for retail availability to normalize. If you need 32GB+ now, consider a used RTX A6000 (48GB) from the workstation market at ~$2,500.
What about NVIDIA Rubin?
NVIDIA’s next-gen Rubin architecture is expected in H2 2026, but it’s datacenter-only (B300/GB300 successors). There will be no consumer Rubin GPUs in 2026. The RTX 50 series is the consumer lineup for at least another 12-18 months.
Don’t wait for Rubin if you’re building a local LLM setup. It won’t help you.
The bottom line
| Action | When it makes sense |
|---|---|
| Buy used RTX 3090 now | Best overall value play. Stable price, 24GB VRAM, zero shortage risk. |
| Buy used RTX 3060 now | Budget entry point at ~$250. Prices rose after the relaunch, not before it. |
| Buy RTX 5070 Ti at MSRP | Great card if available. Don’t overpay. |
| Buy the relaunched RTX 3060 | Only near its $329 launch price, and only if you want new with warranty. |
| Wait for 5090 prices to drop | Reasonable if you can wait 6+ months for GDDR7 supply to stabilize. |
| Buy scalped 5090 | Never. Used 4090 is better value at half the price. |
My strongest recommendation: if you’re building for local LLM right now, the used RTX 3090 at ~$820 is still the move. It’s the only GPU where price, VRAM, and availability all align right now. Everything else is either overpriced, out of stock, or compromises on VRAM.
For budget buyers, the best budget GPU for local LLM roundup covers options under $300.
NVIDIA GeForce RTX 3090
24GB GDDR6XUsed RTX 3090 at ~$820. 24GB VRAM fits 34B models, and used supply is steady. The safest buy in a volatile market.
Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.
Frequently asked questions
Will GPU prices drop in 2026?
New RTX 50 series prices should normalize once GDDR7 supply improves, likely Q4 2026. Used GPU prices (3090, 4090) are already stable and unlikely to drop significantly.
Should I wait for NVIDIA Rubin GPUs?
No. Rubin is datacenter-only in 2026. Consumer GPUs will remain RTX 50 series for at least 12-18 more months.
Is a used RTX 3090 still worth buying in 2026?
Yes. At ~$820 for 24GB VRAM, it’s still the best value GPU for local LLM inference. It has drifted up from about $600 earlier in 2026 — the memory shortage lifted used prices rather than lowering them — but nothing else offers 24GB anywhere near this money.