Explainer · GPU prices · with video · Updated October 2026
Why GPUs Will Never Get Cheaper
For decades, every new chip generation made graphics cards cheaper for the performance. That era is over. Here's why prices aren't coming down, what our own rental-price tracking shows, and what to buy instead.
Watch: Why GPUs Will Never Get Cheaper
The old deal is dead. Shrinking the chip used to make each transistor cheaper, so every generation gave you more for the same money. Now every step down in nanometers makes chips harder to make. A wafer costs more than it used to, and the factories that make them, the fabs, cost tens of billions of dollars. Samsung's new cutting-edge fab project in Texas alone is more than $37 billion. Only a handful of companies on earth can build them, and TSMC makes nearly everything at the smallest sizes.
No price war. Think about how a T-shirt gets cheaper: you get a quote from one supplier, then another undercuts it, then another undercuts that, until the price hits the bottom. That only works when there are lots of suppliers. Desktop GPUs have three: NVIDIA, AMD and Intel. Nobody has much reason to race to the bottom, and nobody does.
AMD is honestly great for gaming right now. For AI you're mostly locked into NVIDIA's software. Intel is the late straggler, getting better, but it's not clear they'll have the volume to support it long term, which is worth thinking about before you buy an Intel card.
The real bottleneck is memory. The chip itself isn't the hardest part anymore. It's the HBM, the stacked high-bandwidth memory that sits next to every data-center GPU, and how many memory chips each GPU needs. As chips get smaller and denser, more of them come out with flaws, and they get binned and sold as different parts. Every step depends on several supply chains at once, which is also why RAM prices are up across the board.
Cheapest H100 and A100 rental we track, weekly median
Median of the daily cheapest hourly rate on RunPod and Vast.ai for each week; the last point covers October 4 and 5 only.
Cheapest consumer-card rental we track, weekly median
Cheapest listing each day, so single days are noisy; the weekly median shows the direction.
Even renting isn't getting cheaper right now. We log the cheapest hourly rate for every card every day. Since August the cheapest H100 has gone from about $1.34 to $2.14 an hour, and the RTX 3090 and 4090 have roughly doubled from their lows. What does get cheaper is old silicon: the A100 just dropped under a dollar an hour as newer cards take its place. That's the pattern to expect. Prices fall when a generation stops being the one everybody wants, not before.
What it means for you. For a gaming-only card, buy something AI doesn't want: an AMD card, or an 8GB-class card sized for your resolution. Once you get into 12 and 16GB NVIDIA cards, AI demand is setting the price. I'm on an RTX 3060 myself, and honestly AMD would do the job just as well for gaming; FSR has come a long way, and not every game even supports DLSS.
For AI, rent the big cards when you need them. The longer-term good news is software: models keep getting smaller and better optimized, and more of them run usefully on 12 and 16GB cards every month. Eventually you won't need to rent data-center GPUs for most things, because the card you already own will be good enough.
Our verdict
GPUs aren't getting cheaper because fabs cost tens of billions, only three companies make desktop GPUs, HBM memory is the bottleneck, and NVIDIA earns far more selling to data centers. Even rental rates for current cards are rising: the cheapest H100 we track went from $1.34 to $2.14 an hour since August. Buy an AMD or lower-VRAM card for gaming, and rent for AI.
FAQ
- Will GPU prices go down in 2026?
- Not for current-generation cards. Demand from AI sets the price of anything with lots of VRAM. Older architectures do get cheaper once newer ones replace them, the way the A100 just fell under a dollar an hour to rent.
- Why are GPUs so expensive?
- Cutting-edge chips cost more to make every generation, the fabs that make them cost tens of billions of dollars, high-bandwidth memory is in short supply, and with only three desktop GPU makers there is no price war.
- Is AMD good enough for gaming?
- Yes. AMD cards are excellent for gaming and don't carry the AI premium NVIDIA cards do. NVIDIA is still the safer pick if you also want to run AI locally.
- Are GPU rental prices going down?
- Not right now. The cheapest H100 rental we track rose from about $1.34 to $2.14 an hour between August and October 2026, and consumer cards like the RTX 3090 and 4090 roughly doubled from their lows.
How we test
Rental trends are weekly medians of the cheapest hourly rate we logged each day on RunPod and Vast.ai from August 14 to October 5, 2026 (the rentals page has live numbers). A single cheapest listing can jump day to day, which is why the charts use weekly medians. Industry figures (fab investment, margins, memory) are the ones discussed in the GPU Battle video 'Why GPUs Will Never Get Cheaper'.