NVIDIA shelves RTX 50 Super indefinitely; GDDR7 shortage forces gaming GPU cuts for AI demand

NVIDIA has indefinitely delayed the RTX 50 Super lineup—including a completed 24GB RTX 5080 Super design—due to a structural GDDR7 memory shortage driven by AI data-center demand. According to The Information and corroborated by PCWorld, TrendForce, and PC Gamer, NVIDIA cut RTX 50-series consumer GPU production by 30–40% in the first half of 2026 and shelved the Super refresh in December 2025, redirecting memory capacity toward higher-margin AI accelerators (H200, Blackwell). This marks the first time in nearly three decades that NVIDIA is not shipping a new gaming GPU architecture in a calendar year. The RTX 5090, once $1,999 MSRP, now sells for $4,329 on Amazon, with premium board partners listing above $5,000.

The root cause is not fab capacity but memory: SK Hynix, Samsung, and Micron have allocated the vast majority of 2026 HBM (High Bandwidth Memory) production to data-center customers, and GDDR7 modules have become scarce. A 3GB GDDR7 die (used in high-end gaming cards) now costs $60–70 each, compared to ~$20 for 2GB chips; for an RTX 5080 Super with eight 3GB modules, memory alone runs $480–560 in bill-of-materials. IDC forecasts AI data centers will consume 70% of global memory output in 2026, up from 20–30% in 2022. Intel CEO Lip-Bu Tan and SK Hynix both signaled relief won't arrive until 2027–2028.

AMD is also hit: the company raised GDDR memory prices to AIB partners by ~10% effective July 2026 (second increase in six months) and announced 10–15% price hikes on the Radeon RX 9000 lineup in H2 2026. Console makers (Nintendo, Sony) and PC component vendors are all facing the same reallocation. IDC classified the shortage as a "potentially permanent strategic reallocation," not a cyclical correction.

For practitioners: the shortage is reshaping the entire compute market. If you need a GPU today, last-gen cards (RTX 4070-class, RX 7800 XT) using pre-shortage memory face less pricing pressure. If you're designing for 2027, assume higher memory costs are structural—they constrain single-GPU configurations and favor architectures using FP8 quantization (~50% memory savings) or inference-optimized SKUs. The gaming GPU market has become secondary to AI; NVIDIA will continue directing supply upstream until data-center utilization matures.