I’ve been looking for a powerful graphics card for quite some time not only to improve the quality of my games on PC, but also as a test bench for native AI. And it’s impossible. I wanted an RTX 5090, then I lowered my expectations until I settled for a much older model (I have a PC with a “modest” RTX 4060). But nothing, I don’t have the budget to get decent equipment. All due to material shortages caused by the AI itself.
the numbers. Hardware growth is something that all of us who monitor the PC market already know. And you can put numbers to that increase: the RTX 5090, in particular, has doubled in price. Generally speaking, graphics cards have risen more than 30% in the last three months (data from Price Squirrel via Reddit).
What’s interesting is that not only have the most current and powerful graphics become more expensive, but that second-hand has also increased for previous generations. Finding an RTX 3090 at a good price is an impossible mission, for example: it is enough to search on sites like Wallapop that they are over a thousand euros (whether they will be paid is another story). Paying over a thousand euros for six-year-old models, without knowing what they’ve been used for, seems crazy to me.
Graphics cards are the new gold rush. If renewal is only within reach of those with a whim-proof bank account, those who already have a card are in for a treat. It’s almost literal, because reselling can make you a huge profit. And this affects players, who find it impossible to update their equipment due to the assumptions generated by the AI market.
Data centers don’t use graphics cards for personal use, but they cannibalize component manufacturing efforts. Brands like Samsung, SK Hynix or Micron find it more profitable to make HBM for AI accelerators than GDDR for graphics cards. This decision makes all technology more expensive, especially graphics cards.
Rebound, benefits for those who create crises. Components for data centers are also more expensive, but they have one advantage: high volume and future purchases. With this they manage to reduce the share of cost increase. Furthermore, the idea is that they recoup this investment with benefits, at least in theory. Practice does not make it so clear.
While companies seek a return on profit from hardware purchases, users only want improved experience. The price increase loses the latter twice
Graphics cards not only allow you to play at high quality, but also run native AI models to avoid cloud. Not all models perform equally well, Nvidia’s RTX is especially good at it (blessed with CUDA). And, due to the price hike, those of us looking for a card for the local model find it impossible. It is suitable for companies like OpenAI, xAI and Anthropic. Like a token sent from heaven.
Sam Altman doesn’t want native AI. He made it clear last week: “Local speculation is a terrible allocation of scarce resources.” The big AI magnates want us to use their tools, which don’t offer open source models that work on our local computers. Your fear that we are self-sufficient is justified: most of what we use ChatGPT or Claude for is a good card and can be fully covered with models like GLM, DeepSeek or Qwen. That’s why Sam doesn’t like it.
Perfect for those with hardware that’s impossible. I highly doubt that the entire spectrum of uses can be covered with local models. Nevertheless, having this option adds greater freedom to users, contrary to what big companies are looking for, who have invested a lot of money in their LLM training. Return on investment should be equal. Unlike DeepSeek or Z.ai, who publish their models publicly and charge for their use in the cloud, OpenAI or Anthropic only provides access through their servers.
Companies that raised the price of graphics cards are lucky because these cards now have a huge barrier to entry. It’s not something to ask for, but this way they ensure that users can’t run local models with enough power and have to hire ChatGPT, Claude and company. It is an unpredictable cause/effect relationship that is only suitable for one party. As is often the case, we the users suffer the most.
Cover image Edited by Laura Lopez
In Xataka | SK Hynix’s operating profit rises thanks to massive memory and AI bottlenecks
