AI hardware boom makes local model experimentation harder for PC users - Big Tech and the digital economy

Graphics card prices have risen sharply as demand for AI-related components puts pressure on the wider hardware market, making local AI increasingly expensive for enthusiasts.

The rising cost of graphics cards is making it harder for PC users to build systems capable of running AI models locally. Enthusiasts looking for a powerful card for both gaming and experimentation with local AI are facing higher prices across new and used hardware.

According to figures from Price Squirrel cited via Reddit, graphics card prices have increased by more than 30% over the past three months. The RTX 5090 is reported to have doubled in price, while older models have also become more expensive. Used RTX 3090 cards, for example, are being listed above €1,000 on platforms such as Wallapop, although listing prices do not necessarily reflect completed sales.

The pressure is not limited to consumer graphics cards. The article attributes part of the problem to manufacturers prioritising high-bandwidth memory, or HBM, for AI accelerators over GDDR memory used in graphics cards. Companies including Samsung, SK Hynix and Micron can find it more profitable to produce components for the AI market, reducing supply available for consumer hardware and contributing to higher prices.

Data centres also face increased component costs, but their large-volume and long-term purchasing agreements can help reduce part of that impact. Individual buyers do not have the same negotiating power, leaving gamers and local AI users more exposed to price increases.

This situation also affects access to local AI. Graphics cards, particularly Nvidia RTX models because of their CUDA support, can run a range of open models directly on a personal computer. However, the higher entry cost means that some users are abandoning plans to build systems for local inference and are instead more likely to rely on cloud services.

OpenAI, xAI and Anthropic could benefit indirectly from this trend because their services run on company servers. The article contrasts that approach with providers such as DeepSeek and Z.ai, which make models available openly while also offering cloud access.

OpenAI chief executive Sam Altman recently described local inference as “a terrible allocation of scarce resources.” The statement reflects the disagreement over where AI should run: on centralised infrastructure operated by large companies or on users’ own hardware. With graphics cards becoming less affordable, the cost of choosing the local option is increasing, leaving consumers with fewer practical alternatives.

Sources: Xataka.