This article was originally published on runaihome.com
TL;DR: Multiple supply-chain reports say NVIDIA will not launch any new consumer RTX cards in 2026 — the RTX 50 SUPER refresh is pushed to Q3 2026 at the earliest, the RTX 5060 Ti 16GB and 5070 Ti are reportedly being wound down, and the next-gen RTX 60 series may not arrive until 2028. The cause is a brutal GDDR7/DRAM shortage driven by AI datacenter demand. For home labs, that means used-card prices are climbing, not falling, and the buy-now-or-wait math just flipped.
| Buy now (used 24GB) | Buy now (new 16GB) | Wait for SUPER / RTX 60 | |
|---|---|---|---|
| Best for | Most local-AI builders | Tight budgets, warranty | Gamers who can wait |
| Price / Cost | Used RTX 3090 ~$966–$1,189 (Jun 2026) | RTX 5060 Ti 16GB ~$400–$500, stock shrinking | RTX 50 SUPER not before Q3 2026; RTX 60 ~2028 |
| The catch | Prices up ~72% over 4 months | 16GB ceiling; may be discontinued | No firm date, no VRAM bump guaranteed at MSRP |
Honest take: If your models fit in 24GB, buy a used RTX 3090 now — waiting for "more VRAM at the same price" is betting against a memory shortage that's getting worse, not better.
What's actually happening
The short version: NVIDIA's annual cadence of new consumer graphics cards has stalled. According to a report covered by Tom's Hardware, NVIDIA is not expected to release any new RTX gaming GPUs during 2026, and the next architecture (the Rubin-based RTX 60 series) has slipped on the internal roadmap toward a 2028 debut rather than the late-2027 mass production originally planned.
That doesn't mean the lineup is frozen in place — it means the moves that are happening are subtractions, not additions:
- The RTX 50 SUPER refresh (rumored RTX 5070 SUPER, 5070 Ti SUPER, 5080 SUPER) has been postponed from its expected Q1 2026 window to Q3 2026 at the earliest, with some outlets describing it as "delayed indefinitely." NVIDIA reportedly told its board partners the plan is not canceled, just pushed.
- NVIDIA has reportedly ended RTX 5070 Ti production, and the RTX 5060 Ti 16GB — the single most recommended budget local-AI card on this site — is next in line for a production pause or discontinuation.
- Industry estimates point to NVIDIA cutting gaming GPU production by 30–40% in 2026.
For a gamer, this is an annoyance. For someone building a local-AI box, it reshapes the entire decision tree, because the thing you actually buy a GPU for — VRAM — is exactly what's getting rationed.
Why: the memory shortage, not the silicon
The GPU dies are not the bottleneck. Memory is.
The RTX 50 SUPER cards were designed around 3GB GDDR7 modules — denser chips that pack 50% more capacity per module than the 2GB chips on the current RTX 50 series. That's how a 5070 SUPER was supposed to jump from 12GB to 18GB and a 5080 SUPER from 16GB to 24GB. Those high-density modules are precisely the parts that suppliers can't make enough of, because the same fabs are being redirected toward HBM and high-margin server DRAM for AI accelerators.
The macro numbers explain the pressure. Per TrendForce, conventional DRAM contract prices rose roughly 93–98% quarter-over-quarter in Q1 2026, and are projected to climb another 58–63% in Q2 2026. NAND flash jumped around 95% in Q1 with another 70–75% forecast for Q2. TrendForce pegged total DRAM industry revenue up 81% QoQ in the first quarter of 2026. When memory of every type is selling for nearly double quarter-on-quarter, a memory maker has no incentive to spin up extra GDDR7 lines for $400 gaming cards.
So NVIDIA is doing the rational thing for its margins: reallocating the scarce GDDR7 it can get toward fewer, higher-end SKUs. The reported plan to wind down the RTX 5060 Ti 16GB while keeping the 8GB version alive is the clearest signal — they'd rather put those memory chips on an RTX 5070 than "waste" them doubling up VRAM on a midrange card.
That logic is bad news for us specifically. Local AI is the one consumer workload where the 16GB-over-8GB difference isn't a luxury — it's the line between running a 13B model and not.
What this does to the used market
Here's the part that catches people off guard: a supply freeze on new cards doesn't leave used prices alone. It pushes them up.
The used RTX 3090 — 24GB of GDDR6X at 936 GB/s of bandwidth, still the value king for local inference — is the clearest example. As of June 2026, price trackers put the lowest average around $1,070, with a range of roughly $966 to $1,189. The notable part is the trajectory: the lowest average is up about 71.9% over the past four months. Back in early 2026 you could find clean 3090s in the $700–$850 range. That window is closing.
The RTX 4090 tells the same story from higher up: used units run around $2,000–$2,250, and new stock — where it exists — sits near $2,755, well above its original MSRP. When no new high-VRAM cards are coming, every used 24GB card on eBay becomes more valuable, not less.
This inverts the usual advice. Normally "wait for the next generation" is sound — prices fall, VRAM goes up, you get more for your money. In a memory shortage with no new launches, waiting means watching used prices climb while the new cards you're waiting for either don't ship or ship without the VRAM bump you wanted.
Buy now or wait? A budget-by-budget framework
The decision comes down to whether your target models fit in 24GB and how much risk you want to carry on a moving market.
Under ~$550 — buy a RTX 5060 Ti 16GB while it's in stock. At 16GB and 448 GB/s, it runs every 7B–13B Q4 model comfortably and many 30B MoE models. With production reportedly being paused, the ~$400–$500 price is more likely to rise or vanish than to drop. If this is your card, don't sit on the decision. We compared it head to head with the 3090 in our 5060 Ti 16GB vs used 3090 total-cost breakdown.
~$1,000–$1,200 — buy a used RTX 3090. This is the sweet spot for local AI and the recommendation that survives the shortage best. 24GB lets you run 32B-class models at Q4 and gives breathing room on context length; the 936 GB/s bandwidth delivers roughly 95 tok/s on a 7B model at Q4_K_M in llama.cpp. Yes, it costs more than it did in January. It will likely cost more still in September. Full analysis in our used RTX 3090 value piece.
~$2,000+ — used RTX 4090, or two 3090s. The 4090 buys you speed (around 127 tok/s on an 8B model) and 24GB. But two used 3090s give you 48GB of pooled VRAM for similar money, which matters more than raw speed once you're loading 70B-class models. The catch is power and complexity — two 285W cards versus one 450W card, plus a PSU sized accordingly.
Whatever your budget — don't wait for the RTX 50 SUPER as a VRAM play. Even if it lands in Q3 2026, two things are likely: it'll launch into the same memory shortage (so expect inflated street prices, not MSRP), and the extra VRAM that justified it is exactly the component in shortest supply. Waiting six-plus months for a card that may ship late and above sticker is a weak bet when used 24GB hardware is on the shelf today.
The escape hatch: rent before you overpay
If the used market has priced you out, or you only need big-VRAM hardware for occasional jobs — a weekend of fine-tuning, testing whether a 70B model is even worth running locally — renting is the rational hedge in a supply crunch. A few hours on a cloud GPU through RunPod costs less than the shipping on a used 3090, and it lets you mea
Top comments (0)