I'm planning to build a PC mainly for machine-learning experiments, but current component prices are making it difficult. A basic 32GB DDR5 kit costs around $400 unless I choose an unfamiliar brand, and a 16GB VRAM GPU such as the RTX 5060 Ti is roughly $750. Do prices seem likely to improve within the next two or three years, or should I buy what I can now and start working?
4 Answers
The most realistic expectation is that prices may stabilize before they meaningfully decline. An AI-market slowdown could release some capacity and lower memory prices, but that outcome and its timing are impossible to know, and a market crash would bring its own economic problems. If you already have a working computer, waiting is reasonable; if you have a concrete need and can afford the parts, buy a balanced system now rather than paying extra for hardware you won’t use.
Used systems, prebuilts, and complete upgrade platforms may offer better value than buying every component separately. A previous-generation machine with DDR4 and a used GPU could be perfectly adequate for learning and experimentation, especially if you can add RAM or storage later. You don’t have to build the ideal long-term machine on day one.
There’s no reliable way to predict the timing. Demand from AI and data centers is keeping memory and GPU supply tight, while new manufacturing capacity takes years to come online. Prices could fall if demand suddenly weakens, but they could also rise first. If you can afford the system and need it now, buying a sensible configuration is safer than waiting for a specific future price.
For local machine-learning experiments, you may not need the most expensive setup. Traditional ML can work fine with 16–32GB of system memory, and many smaller models can run with 8GB or more of VRAM. Consider starting with a midrange AM5 system, buying only the memory and GPU you genuinely need, and upgrading later. For large models or serious training, renting cloud GPU time may be cheaper than buying high-end hardware outright.
The right amount depends heavily on the workload. Large datasets, virtual machines, and containers may justify 64GB of RAM, while lighter experiments can work with 32GB or even less.
Exactly. Even if prices eventually drop, nobody knows whether that will happen in six months or several years, so trying to time a complete build is mostly guesswork.