Amazon US · weekly listed prices
GPU Price Tracker
Compare RTX 5090, 5080 and 5070 Ti card prices. Choose a model and date range below.
OBSERVED RETAIL SERIES
RTX 50 high-VRAM prices since 2025
18 monthly archive points plus the live weekly series.
MEDIAN LISTED PRICE · USD
BASELINE INITIALIZINGFirst validated collection pending
The previous snapshot stays in place if a required model lacks two eligible listings.
Historical median Live median Live mean
Monthly archive → marked handoff → Monday-dated weekly points
Some weeks use different cards. Gaps mark baskets with no shared boards; their price levels are not a like-for-like change.
Data correction: a complete desktop PC was excluded from the RTX 5090 basket. September 14 was withdrawn because only one selected card remained; September 21 was recalculated from two cards. Original source records are retained.
History: eight exact Amazon US ASIN archives · 45-day maximum carry · at least two cards per model-month · four implausible source values preserved but excluded
Scope: new standalone GDDR7 cards · RTX 5090 32GB · RTX 5080 16GB · RTX 5070 Ti 16GB · established board partners
Visible basket
The boards behind the medians
Up to three distinct board families per GPU model. Selection favors identity, availability, ratings, and demand evidence—not the lowest price.
3 model tiersNever blended into one dollar price16GB / 32GBExplicit GDDR7 capacityNew cards onlyNo laptops, PCs, used cards, or accessoriesWeeklyFits the provider budget
| Board listing | GPU / VRAM | Listed price | USD / VRAM GB | State |
|---|---|---|---|---|
| ASUS TUF Gaming RTX 5090ASUSASIN B0DS2X13PH | RTX 509032GB GDDR7 | $7,439.00 | $232.47 | thin stock |
| ASUS ROG Astral RTX 5090ASUSASIN B0F4ZCN7FV | RTX 509032GB GDDR7 | $7,499.00 | $234.34 | thin stock |
| ASUS Reference family RTX 5080ASUSASIN B0FSB9GY9Q | RTX 508016GB GDDR7 | $1,869.46 | $116.84 | in stock |
| Gigabyte Gaming OC RTX 5080GigabyteASIN B0DS2R6948 | RTX 508016GB GDDR7 | $1,699.99 | $106.25 | in stock |
| MSI Suprim RTX 5080MSIASIN B0DSWR8WMB | RTX 508016GB GDDR7 | $1,999.99 | $125.00 | in stock |
| ASUS TUF Gaming RTX 5070 TiASUSASIN B0DS6WTXGP | RTX 5070 Ti16GB GDDR7 | $1,269.00 | $79.31 | in stock |
| Gigabyte Windforce RTX 5070 TiGigabyteASIN B0DTGNB9Q5 | RTX 5070 Ti16GB GDDR7 | $1,495.00 | $93.44 | in stock |
| MSI Ventus RTX 5070 TiMSIASIN B0DV9GC9SR | RTX 5070 Ti16GB GDDR7 | $1,429.99 | $89.37 | in stock |
Why this basket
Close to AI memory—without pretending GDDR7 is HBM.
These cards use high-capacity GDDR7 and suit local AI, creator, and graphics workloads. AI accelerators generally use stacked HBM instead. The defensible connection is upstream allocation: DRAM makers divide capital, wafer capacity, engineering, and packaging resources across products.
Official context: NVIDIA specifications, SK hynix outlook, and capacity explainer.
Retail signalGDDR7 cardsShared DRAM resourcesHBM demandConnected constraints ≠ proven causality
Methods · gpu-backcast-v1-eight-card + retail-v3-mainstream-gpu1
A longer series with a visible source handoff.
The fixed historical archive and live weekly basket stay separate, meet at a marked August 2026 handoff, and preserve three distinct GPU models.
Read the GPU methodology →- 01
Archive
Eight exact Amazon US ASIN histories, reviewed once and stored immutably.
- 02
Validate
Same model bounds as live collection; 45-day carry and monthly coverage gates.
- 03
Chain-link
At least two exact cards per model-month, rebased to 100 at handoff.
- 04
Continue live
Weekly screened Amazon collection extends the chart without rewriting history.
Questions, answered
GPU price tracker FAQ
Current prices, AI-memory context, buying limits, and the retail method behind the lines.
Reading the chart
What does the GPU price tracker measure?
It measures weekly Amazon US listed prices for eligible new standalone RTX 5090, RTX 5080, and RTX 5070 Ti graphics cards. Each model stays separate; this is not a wholesale memory quote, transaction-price dataset, or cheapest-card ranking.
What are current RTX 5090, 5080, and 5070 Ti prices?
Use the dated model cards and visible board basket above. A universal GPU price would hide large differences between model tiers, cooling designs, and availability.
Are high-end GPU prices going up or down?
Use the monthly exact-ASIN history from March 2025 and the weekly live segment after the marked August 2026 handoff. Compare each model separately because one GPU tier can move while another is flat.
Which graphics cards are tracked?
The scope is GeForce RTX 5090 32GB, RTX 5080 16GB, and RTX 5070 Ti 16GB desktop add-in cards using GDDR7. NVIDIA RTX 50 specifications.
Why show both mean and median GPU prices?
The mean reacts to every eligible board listing. The median is less sensitive to one unusually expensive factory-overclocked or liquid-cooled design. Their spread shows whether movement is broad or concentrated.
What does an index level of 100 mean?
The historical and live model segments are both rebased to 100 at the August 2026 handoff. A level of 110 means the chain-linked trend is 10% above that handoff—not that the card costs $110.
AI & memory
Is GDDR7 the same memory as HBM used by AI accelerators?
No. These GeForce cards use discrete GDDR7 packages around the GPU. Data-center AI accelerators commonly use stacked HBM with advanced packaging. Both are DRAM, but their interfaces, packaging, and production economics differ. NVIDIA RTX 50 specifications.
Can AI memory demand cannibalize graphics-memory supply?
It can create an upstream allocation tradeoff, but not a one-for-one substitution. Memory makers divide capital, cleanroom space, process capacity, engineering, and packaging resources across HBM and conventional DRAM products. SK hynix capacity explainer.
Does high AI demand automatically make graphics cards more expensive?
No. AI demand can tighten shared upstream resources and high-VRAM cards can be bought for local AI, but retail prices also depend on GPU die supply, boards, tariffs, inventory, seller behavior, launches, and promotions. SK hynix 2026 memory outlook.
Why track RTX 5090, 5080, and 5070 Ti instead of every GPU?
They are current Blackwell desktop cards with 16GB or 32GB of GDDR7 and meaningful creator or local-AI usefulness. That is a narrower AI-memory-adjacent task than mixing low-VRAM cards, old GDDR6 generations, and broad semiconductor data. NVIDIA RTX 50 specifications.
Why is the RTX 5090 especially relevant to local AI?
Its 32GB of GDDR7 can fit larger local models and workloads than the 16GB cards. It remains a GeForce retail card, not a substitute for every HBM accelerator or enterprise workload. NVIDIA RTX 50 specifications.
Does this chart prove that HBM demand caused a GPU price change?
No. It can show the timing, magnitude, breadth, and persistence of retail movement. It cannot isolate HBM allocation from GPU chip supply, tariffs, seller mix, inventory, gaming demand, or creator demand.
Cost & buying
Can I use the tracker to check a GPU deal?
Use the dated model median and basket as market context, then verify the exact board, seller, condition, warranty, shipping, return terms, power requirements, and live checkout price.
What does price per VRAM GB mean?
It is listed card price divided by 16GB or 32GB of onboard GDDR7. It normalizes capacity cost but does not measure compute speed, bandwidth, cooling, power, noise, warranty, or performance per dollar.
Is the tracked price the NVIDIA launch price or MSRP?
No. It is an observed Amazon listing for a complete board-partner card. Factory overclocks, large coolers, limited availability, seller markups, and channel conditions can move listings far from launch pricing.
Why not show only the cheapest GPU listing?
The lowest offer can be thin stock, a weak seller, an accessory, a misleading variant, or a temporary anomaly. A screened multi-brand median is more stable for tracking a tier.
Do the listed prices include sales tax and shipping?
Sales tax is excluded. Shipping is excluded unless a provider exposes it consistently, so the basket price is a listed card amount rather than an all-in checkout total.
How fresh is the GPU price data?
The historical archive is a fixed, reviewed reconstruction through the live handoff. Current Amazon US collection is weekly to stay within the provider budget. A failed run does not insert a zero price or pretend old data is fresh.
Data & method
Where does the GPU price data come from?
The historical segment uses fixed public PriceHistory.app archives for eight exact Amazon US ASINs. The live segment uses server-side Amazon US search collection through the site’s ordered provider adapters. Public browsers receive only the validated same-origin snapshot.
How are GPU listings filtered?
A listing needs an exact supported GPU, the correct 16GB or 32GB GDDR7 specification, new condition, plausible price, no displayed rating below 4.0, and credible market or fulfilment evidence. “Only one or two left” remains visible as thin stock because scarcity matters here, but is ranked below deeper stock.
How are board-partner variants handled?
Listings are grouped by GPU model, board brand, and recognizable product family. Closely related duplicates do not get multiple basket slots, and selection favors market evidence rather than the lowest price.
How are historical and weekly GPU prices calculated?
Historical daily archive values carry forward for at most 45 days and collapse to exact-ASIN monthly medians; each model-month requires at least two cards. Live weekly mean and median values use the screened active board basket. Both segments chain-link overlapping products and meet at a visibly marked handoff.
What happens when collection or validation fails?
The run is recorded as failed, the prior public snapshot remains intact, and no zero-price point is inserted. A required model needs at least two active eligible listings before publication.
Why exclude workstation GPUs, older cards, AMD cards, and lower tiers?
They broaden the page into different product markets and memory generations. Workstation Amazon samples are often thin, and older, AMD, or lower-tier cards need their own reviewed basket rather than being silently mixed into this one.