n

NVIDIA's flagship gaming graphics card, the RTX 5090, has virtually disappeared from the US online retail market in recent days, with third-party sellers asking as much as US$9,500 — close to 4.75 times its official starting price of US$1,999.

According to a recent investigation by technology publication Tom's Hardware, official stock at major e-commerce platforms including Newegg and Amazon has completely sold out. Consumers can now only buy the graphics card from third-party sellers, with asking prices ranging from US$6,500 to US$9,500, and some even higher. In June, the median market price of the RTX 5090 was still US$4,299. By early September, the lowest tracked online price had risen to US$5,199. In less than two weeks, regular online stock had vanished altogether and third-party asking prices surged again.

One point needs to be clarified first: US$4,299 was not the RTX 5090's official recommended retail price, or MSRP, but its median market transaction price in June. The RTX 5090 officially went on sale in January 2025, with a starting price of US$1,999. In other words, this is not a new product that was simply snapped up immediately after launch — it has been on the market for nearly 20 months. The real news is that a consumer graphics card launched at US$1,999 is now being bid up to US$9,500, representing an increase of nearly 3.75 times over its starting price (or 4.75 times when calculated as 9,500 divided by 1,999). That gap is considerably larger than the roughly 1.2-fold difference suggested by comparing US$4,299 with US$9,500.

Memory, not chip production, is at the heart of the shortage

Some reports have blamed the shortage on insufficient advanced packaging capacity, or CoWoS, at Taiwan Semiconductor Manufacturing Co (TSMC). That explanation, however, conflates two separate issues. CoWoS packaging is mainly used for AI accelerator chips that are paired with high-bandwidth memory, such as the B200 and GB200 used in data centres. As a consumer gaming card, the RTX 5090 uses 32GB of GDDR7 video memory and does not use CoWoS packaging. Its manufacturing bottleneck therefore does not lie in chip packaging.

The real issue is the memory itself. The three major memory manufacturers — Samsung, SK Hynix and Micron — have in recent years allocated an increasing share of their wafer capacity to HBM for AI data centres. This has squeezed capacity for GDDR7 used in consumer graphics cards. An industry source said previously that NVIDIA prioritises its GPU products based on the revenue generated per gigabyte of memory. High-end data-centre products, such as the RTX Pro 6000, generate the highest return per gigabyte and are allocated memory first. Although the RTX 5090 also generates a substantial return, it still has to give way to higher-priority enterprise products. Multiple industry reports indicate that NVIDIA cut overall RTX 50-series production by 30% to 40% in the first half of this year because of GDDR7 shortages.

On top of this structural supply gap is another source of demand: some technology enthusiasts are using the RTX 5090 as a component in low-cost AI inference servers. The enterprise RTX Pro 5000, which uses the same GB202 core, typically sells for between US$7,000 and US$9,000. By comparison, the RTX 5090, priced at several thousand US dollars but equipped with 32GB of video memory and capable of running large language models locally, has naturally become an alternative for budget-conscious AI developers. Scalping is a third force layered on top of the memory shortage and cross-sector AI demand. Together, the three factors have created the extreme shortage seen today.

NVIDIA has so far made no official comment on the RTX 5090 shortage at any of its earnings calls.

An offline exception

US computer retailer Micro Center, which insists on selling only through physical shops, has so far escaped the worst of the buying frenzy. Its stores are still selling the RTX 5090 for US$4,299 — the level of the June median market price — and some branches still have stock. Because Micro Center has operated without an online sales channel for years, it has avoided the mass purchasing of stock by online shopping bots and professional scalpers.

Tom's Hardware also warned that buying an RTX 5090 at an inflated price from a third-party seller carries risks. In January, more than 40 complaints were reportedly filed by buyers who had “ordered” an RTX 5090 at a low price on Amazon but ultimately received counterfeit products. More recently, some buyers have reported receiving graphics cards from which the GPU core and memory chips had been removed.

Indirect impact on the mainland market

The supply crunch is also affecting the mainland Chinese market. Many mainland research institutions, internet companies and independent developers with substantial computing needs have long been seeking high-performance hardware. But against the backdrop of US export controls on advanced computing chips, specially supplied products such as the H20 are compliant but do not fully meet performance requirements in practice. If the RTX 5090 shortage continues worldwide, mainland consumers are likely to face greater difficulty and higher costs when obtaining the card through unofficial channels, commonly known as parallel imports. This could further widen the gap between the mainland and overseas markets in personal end-user computing experiences. Motherboard manufacturers and cooling-module suppliers will also have to deliver under even tighter supply-chain conditions.