GPU price history: how graphics cards got so expensive

In 2014 you could buy NVIDIA's second-fastest gaming GPU, the GTX 970, for $329. In September 2026, $329 doesn't reliably buy the cheapest card in NVIDIA's current lineup, and the flagship RTX 5090, launched at $1,999, sold for 2.2 to 2.6 times that at US retail in September, roughly $4,400 to $5,200. On our own tracker, a typical new RTX 5090 listing cost about $7,700 on Oct 27, 2026. This article traces how graphics cards got here: when prices started climbing, why, whether every manufacturer moved the same way, what it's doing to the people who just want to play games, and what could bring prices back down, if anything.

For live prices while you read, our GPU best-price hub tracks the cheapest current cards, and the NVIDIA, AMD and Intel price lists rank what retailers are charging today.

Check today's GPU prices

What our tracker shows: RTX 50 and RX 9000 prices today

Here is what new graphics cards actually sell for at the stores we track, compared with the week of September 12-18, when the street prices quoted below were reported. Every number comes from our own listings.

Every model with enough listings got more expensive over those six weeks. The RTX 5080 moved most, from about $1,840 to about $2,510, and the RTX 5060 moved least, from about $470 to about $490. These cards are sold by at least two of the stores we track, with today's lowest price and the lowest we've seen in 90 days:

Yes, prices have risen, in two different ways

It helps to separate two things that both get called "GPU prices going up."

The first is list price (MSRP) creep: what NVIDIA and AMD officially charge for each tier. That story is visible in one chart:

The flagship tripled in eleven years, from $649 to $1,999. The x80 class nearly doubled. The midrange x70 class went from $329 to $549. Only the budget x60 tier held roughly still in nominal dollars: $199 then, $299 now. Even that masks a real-terms squeeze, because those dollars buy a smaller slice of each generation's silicon than they used to.

The second is the street price: what you actually pay at checkout. For most of GPU history, street price tracked MSRP closely outside of brief launch windows. Since 2020 the two have repeatedly detached, and in 2026 the detachment is the worst it has ever been. In September, VideoCardz put US median prices for RTX 50-series cards at 35% to 135% above MSRP. Our tracker, which counts new single cards at eBay and Walmart, reads higher: in the week of September 12-18 the typical $549 RTX 5070 listing was about $895 and the $999 RTX 5080 about $1,840. By October 27 they were about $975 and $2,510. We wrote a whole guide on why MSRP is fiction; GPUs are its starkest example.

When did it start, and why

Each step up has a specific cause. The short timeline:

2017-2018: the first crypto boom. Ethereum mining made GPUs money-printing machines, and street prices roughly doubled while MSRPs stayed put. It ended in a mid-2018 crash and a flood of used cards, the first proof that GPU demand was no longer just gamers.

September 2018: Turing raises the list price. With no competition at the high end and new ray-tracing silicon to pay for, NVIDIA moved the official tiers themselves: the RTX 2080 Ti launched at $999 against the 1080 Ti's $699, and every class beneath it shifted up. This is the first structural break in the chart above, the moment higher prices became policy rather than market fever.

2020-2022: pandemic plus crypto boom two. Lockdown demand, a global chip shortage, and a second Ethereum boom stacked on top of each other. Cards sold at two to three times MSRP for about eighteen months; the RTX 3080's $699 list price was almost theoretical. Scalping became an industry.

Late 2022: the crash that didn't reset prices. Ethereum's move to proof-of-stake killed GPU mining, used cards flooded back, and shortages ended. But the RTX 40 series launched with the crypto era priced in: $1,199 for the RTX 4080 that had been $699 as a 3080. List prices never went back down; 2023-2024 was the closest thing to normal since 2016, and "normal" now meant permanently higher tiers.

2025: tariffs and a thin launch. The RTX 50 series and AMD's RX 9000 series launched into tight supply, US import tariffs, and immediate markups. Cards spent most of 2025 well above MSRP.

2026: the memory supercycle. This is the current, and sharpest, phase, and it deserves its own section.

2026: AI bought the memory, and memory is most of the card

The force driving prices right now has little to do with gaming demand. AI datacenter buildouts are consuming memory, HBM for accelerators and DDR5 for servers, faster than the industry can make it. Graphics memory is collateral damage: GDDR7 and HBM compete for the same wafer starts and packaging capacity, and a fab can earn dramatically more selling HBM to a datacenter customer than GDDR7 for a gaming card. We covered the same squeeze from the RAM side in our RAM price forecast, where DRAM contract prices roughly doubled in a single quarter. A modern GPU is, in bill-of-materials terms, substantially a memory product wearing a graphics chip.

The result moved from street markups to official prices. TrendForce reported in January that NVIDIA and AMD planned phased price hikes starting Q1 2026. AMD notified partners of roughly 10% increases on GPU-and-memory kits in July; NVIDIA followed weeks later with increases across the entire GDDR7-based RTX 50 line. By September, retail prices had climbed for the third time this year, and rising VRAM costs keep every current-gen card above launch price. The stunt that got headlines was NVIDIA selling cards at MSRP as a convention perk: the list price is now the deal you queue for.

No, it hasn't been uniform across tiers or brands

The pain is real everywhere, but it is not evenly distributed.

By tier, the top is worst. The RTX 5090 is partly an AI story of its own: with 32GB of VRAM it doubles as a workstation and local-AI card, so gamers bid against professionals. On Oct 27 the typical new 5090 on our tracker cost about $7,700, nearly four times its $1,999 list price. The RTX 5080 ran about 2.5 times list (about $2,510 against $999) and the 5070 Ti about twice (about $1,500 against $749). Further down, the markups narrow but don't disappear: the RTX 5070 at about 1.8 times list and the RTX 5060 at about 1.6 times.

By brand, NVIDIA has stretched furthest. It raised list prices earlier and harder (Turing in 2018, Ada in 2022, phased hikes through 2026), and its cards carry the largest street premiums. AMD mostly held its list prices below NVIDIA's tier for tier (the RX 9070 XT launched at $599 against the 5070 Ti's $749), but it has not escaped the memory squeeze: after this year's rounds of increases, the typical new RX 9070 XT on our tracker cost about $950 on Oct 27, against a $599 launch price. Intel is the outlier. The Arc B580, a $249 GDDR6 budget card, has stayed closest to its list price of any current GPU, the best value-per-dollar story in the market, though it competes only in the entry tier and supply is thin.

For what the market charges right now, the NVIDIA, AMD and Intel price lists track live retail listings on the cards we monitor.

What this is doing to gamers

The clearest evidence is what gamers are not doing: upgrading. The most popular GPU on Steam in 2026 is still the RTX 3060, a card from early 2021, and the 2019 GTX 1650 still holds a top-five spot. Over 40% of Steam users are on RTX 30 and 40 series cards, riding out the market on hardware approaching six years old.

That changes behavior in ways that would have seemed strange in 2016:

• The upgrade cycle stretched from 2-3 years to 5-6. A GPU is now treated like a major appliance, not a periodic refresh.
• Upscaling replaced upgrading. DLSS and FSR turn software into the performance bump a new card used to provide, which is a big reason older cards remain viable and their owners feel no urgency.
• The used market became the default midrange. A used RTX 3080 or 6800 XT is many builders' answer to a $900 new midrange card, and used prices have firmed up accordingly.
• Settings came down. 1080p remains the dominant Steam resolution in an era when 1440p was supposed to be mainstream; our 1080p GPU guide exists for exactly this reason, with the 1440p guide for those stretching further.
• The whole build got harder, not just the GPU. RAM and SSD prices are riding the same memory supercycle, so a mid-tier build that cost roughly $1,000-1,200 in 2020 now runs closer to double that. Some buyers are moving to prebuilts (which get component allocations at scale), consoles, or cloud gaming instead.

A decision nobody agonized over in 2016, "just get the x70 card," is now a four-way tradeoff between paying the markup, buying used, buying down a tier, or not buying at all. Mostly, people are choosing "not yet."

Is there hope for prices to level off or fall?

Not soon, and anyone telling you otherwise is guessing optimistically.

The near-term signals point the wrong way. Vendor price increases arrived in phases all year and analysts expect further rounds into 2027. Memory contract prices are still rising quarter over quarter, more slowly than early 2026's doubling, but moderation means smaller increases, not decreases. No major analyst forecasts meaningful memory-cost relief before 2028. And the usual pressure valve, a next generation with better price-performance, has moved away: the RTX 60 series has reportedly slipped to 2028 because the memory to build it isn't there.

Our own six weeks of data don't show a plateau yet. Between mid-September and late October, every card in our tracker table with enough listings got more expensive, from about 5% for the RTX 5060 to about 37% for the RTX 5080. That's an observation, not a forecast, but it's the opposite of what you'd see if the pressure were easing.

The realistic near-term hope is a plateau: hikes decelerating through 2027 as memory supply and demand find a balance, with current-gen cards stabilizing at today's elevated prices rather than climbing further. A genuine decline waits on the same thing RAM prices wait on: new memory fabs (SK Hynix, Samsung, Micron all have capacity arriving 2027) ramping enough output to matter on shelves, which analysts put at 2028.

GPU shortages have always ended eventually; 2018's did, 2022's did. But this one is underwritten by AI capital spending rather than a speculative bubble, so it unwinds on fab timelines, not sentiment.

What has to change

Four things, roughly in order of impact:

1. Memory supply has to catch up. This is the structural fix. GDDR7 and HBM stop competing so brutally for wafers only when total DRAM capacity grows. The 2027 fab wave reaching volume in 2028 is the honest timeline.
2. AI demand has to stop absorbing every marginal wafer. Whether through the buildout maturing, or simply slowing from its current pace, consumer parts get relief only when datacenters stop pre-buying years of output.
3. Competition has to reach the tiers people actually buy. Intel's B-series shows discipline is possible at $250; an Intel or AMD that seriously contests the $400-700 range would cap what NVIDIA can charge there. NVIDIA's 90%-plus share of the add-in-board market is itself a price.
4. List prices have to mean something again. Restocked Founders/reference cards at MSRP, retailer anti-scalping measures, and honest regional pricing don't fix supply, but they anchor the market. Today's "MSRP as convention perk" is the opposite.

None of these are things a buyer controls. What you can control is knowing the real market price, not the sticker. If you need a card now, track the model you want, set an alert, and buy the moment you see anything near list price; in this market, MSRP is the deal, not the baseline. If your current card still runs your games, waiting is the rational move, and the timelines above say relief means 2028, not next quarter. Our GPU best-price hub and the per-brand price lists below show what "a good price" means this week, because that definition moves monthly.