NVIDIA has quietly done something it rarely does: it raised costs on essentially every graphics card it sells. Board partners like ASUS, MSI, and Gigabyte received a fresh notification of GPU kit price increases covering both the GDDR7-based RTX 50 series and older GeForce products still using GDDR6. That second group matters because it includes the revived RTX 3060 12GB, the one budget card that looked like a genuine value deal in 2026. That window is closing.
This is the second kit-price notification of the year. In May, NVIDIA raised kit prices for the RTX 5090 and the China-market RTX 5090D V2. This round is broader: it reaches the mid-range and entry tiers that were sheltered during the first wave. If you have been waiting for mid-range prices to stabilize before upgrading, the math has just changed.
Why the whole lineup is repricing
The root cause is not a GPU shortage. It is a memory reallocation. High-Bandwidth Memory, the memory that feeds AI data-center accelerators, and GDDR7 and GDDR6, the memory that feeds gaming GPUs, compete for the same fabrication capacity at Samsung, SK Hynix, and Micron. Those three companies effectively control the world's DRAM supply, and hyperscalers such as Google, Microsoft, Meta, and Amazon are absorbing essentially all available HBM with open-ended orders.
HBM dies are physically huge. Micron CEO Sanjay Mehrotra confirms an HBM3E die is roughly twice the size of an equivalent DDR5 module. Every wafer shifted to HBM produces far fewer memory bits than the same wafer would in GDDR7, and memory makers optimize for margin. Consumer graphics memory loses that allocation battle every time.
TSMC's CoWoS advanced packaging is the second bottleneck. It is expanding to an estimated 120,000 to 140,000 wafers per month by late 2026, but the capacity is fully booked, with NVIDIA alone consuming roughly 60 percent of the allocation for its H100, B200, and GB200 data-center accelerators.
The concrete numbers behind the hold
The most visible casualty is the RTX 50 SUPER refresh. The hardware is reportedly ready, at least one board partner has received production-ready GPUs, and NVIDIA has told partners to hold the launch. The reason is simple arithmetic. A 3 GB GDDR7 module currently costs $60 to $70, versus roughly $20 for the standard 2 GB chip. For the RTX 5080 SUPER, which would use eight 3 GB modules for 24 GB of VRAM, the memory bill alone runs $480 to $560, an increase of $320 to $400 over the 2 GB configuration.
The entry tier is not spared. The RTX 5050 9GB would need three 3 GB modules, costing $180 to $210 just for memory, approaching the retail price of last-generation mid-range cards. NVIDIA is also reportedly steering available 3 GB GDDR7 toward higher-margin products such as the RTX PRO 6000 and Rubin CPX AI cards, where the cost is easier to pass through to enterprise buyers.
The pain is industry-wide. AMD raised its GDDR kit prices to partners by about 10 percent effective July 2026, its second increase in six months, with 10 to 15 percent more expected in the second half. Intel's Arc lineup is equally exposed. No consumer GPU vendor is insulated.
How to decide: buy, hold, or wait
Here is a practical framework for your decision, based on the supply-chain data rather than the marketing:
- If your current GPU handles your workload, hold. IDC has classified this shortage as a potentially permanent strategic reallocation, not a cyclical dip. Intel CEO Lip-Bu Tan says relief will not arrive before 2028.
- If you must buy today, favor last-generation cards. RTX 4070, RTX 4070 Ti, RX 7800 XT, and RX 7900 GRE cards use memory bought before the acute shortage and are experiencing far less pressure.
- If you were waiting for the RTX 50 SUPER, plan as if it does not exist. The credible window is late 2026 or CES 2027, and only if memory pricing cooperates. No official timeline has been given.
- If you were waiting for the RTX 6000 series, expect a long wait. Rubin gaming cards are now widely expected no earlier than the second half of 2027, and some reports push mass production into 2028.
Run this checklist before you click checkout: confirm the card's memory type, check whether the price you see is a post-notice price, and ask whether your current card still meets your workload targets. If the answer is yes, the most financially defensible move is to sit tight.
One number is worth remembering: $60 to $70 for a 3 GB GDDR7 module is more than double the price of the 2 GB chip, and the premium is non-linear. Higher-density dies have lower yields and more defect sensitivity, so costs compound during a shortage.
TrendForce data shows conventional DRAM contract prices rose 90 to 95 percent quarter over quarter in Q1 2026, with further increases for graphics DRAM in Q2. SK Hynix has warned that 2027 will be the worst year of the shortage, with pressure potentially persisting into 2030. New fabs from Micron, Samsung, and SK Hynix will not add meaningful output before late 2027 at the earliest.
NVIDIA's official statement says the company continues to ship all GeForce SKUs and is working closely with suppliers to maximize memory availability. It does not address pricing. Board partners will work through existing inventory first, and some will absorb part of the increase, so retail prices may not move instantly. But the direction is unambiguous: every GPU assembled from kits ordered after this notice will cost more than the equivalent card built before it.
The era of stable, predictable GPU pricing is over, and AI demand is the reason. The takeaway is simple: if you need a new card now, buy smart and buy soon, and prefer last-generation silicon. If you do not need one, hold what you have. Waiting for prices to fall is now a bet against the structural direction of the semiconductor industry.
Comments