About the data

870 parts across 11 categories, each with a street price recorded on 2026-09-03. Price history is kept for the three categories that move: GPU, CPU, memory; cases, coolers, fans and monitors barely change.

Prices are street estimates, not quotes. A number without a date is worse than none, so the date travels with every figure, exports included. Parts added after the snapshot carry their own date and source, and the test suite rejects any price without both.

• Tom's Hardware GPU price indexGraphics card street prices and MSRPs
• Tom's Hardware RAM price indexDDR4 and DDR5 kit prices by capacity
• NAND / SSD pricing coverageNVMe and SATA drive prices
• TrackaLacker processor listingsDesktop CPU street prices
• PassMark CPU value chartRetail prices of CPUs on sale new, as the chart listed them on the day read
• Newegg listing, read by handThe price as newegg.com listed it, sold and shipped by Newegg, on the date given
• Newegg marketplace listing, read by handThe price a marketplace seller listed on Newegg's product page, new, on the date given, where Newegg sells none itself
• amd.com store listing, read by handA Ryzen processor's price in AMD's own store, in stock, read by hand on the date given; used where it is below the latest Newegg read, for 14 days
• Newegg listing, read automaticallyThe cheapest in-stock listing sold and shipped by Newegg whose title carries the exact model, read from the search page on the date given
• Modelled from the chipset indexA card no reader has priced, at its chipset's latest read-day index; ×1.05 for a halo air-cooled card, ×1.10 for a liquid-cooled one
• Product listingPrebuilt prices, read from the maker's or retailer's page on the date given
• Launch MSRPThe maker's launch price, dated to the launch (AMD's suggested price; the low end of Intel's recommended customer price range), for a processor added from TechPowerUp's review charts and not yet read at a retailer; the Newegg reader replaces it where the part is still sold new, and a part no longer in production keeps it

GPU, CPU and memory prices are read from Newegg's search pages: a run every 4 hours, the parts least recently read first, one page every 1.2 s, as PCBuilderTool/0.3 (+https://pcbuildertool.com; personal project). A listing counts when it is new, in stock and names the part. Newegg's own listing comes first; a marketplace seller only when Newegg has none, and the figure says so. A listing over 2× or under half the last figure is held for a hand check. Every read is dated under its figure.

A part no longer made (10 processors, 74 cards on 15 chipsets) is priced at its launch MSRP and not read: a leftover listing says nothing about what it costs. A part added since the snapshot shows its launch MSRP until its first read. The seed history holds only defensible points: launch MSRPs dated to launch, street prices dated to the snapshot, and one pre-shortage memory reference labelled a category-level estimate wherever it appears.

Prebuilt systems are priced from their own listing, read by hand on the date the page prints. A listing with no price has none here. A part the maker names resolves to the catalog and its price. A part the maker does not name ("brand varies", or the maker's own board) is recorded in the listing's words and priced at nothing, never as the nearest part we carry. A card named only by its silicon is priced at that chipset's index, and says so.

Specifications are the makers' pages as recorded; a figure remembered rather than read is marked “recalled” on its page. The silicon figures of 49 of 49 chipsets and the records of 15 of 94 processors are TechPowerUp's hardware database, free for a student project with attribution, read 2026-09-26; 16 of those chipsets the free API does not carry were read by hand from the database's pages, pending TechPowerUp's reply; each such page says so under its table. Every processor's figures, power limits (AMD's PPT, Intel's PL1 and PL2) and die were read by hand from its page in the same database on 2026-09-27, pending the same reply; the power estimate and the cooler check use those limits where the page gives them.

• DDR5 has roughly quadrupled since late 2025: a 32 GB kit was about $90 and is now about $400, because of the DRAM shortage.
• NVMe SSDs have roughly doubled: 2 TB drives moved from $130-150 to $290-400.
• Most graphics cards sell well above MSRP. An RTX 5090 lists at $1,999, sold near $5,200 at the snapshot, and by late September 2026 was out of stock at Newegg itself and $7,000 to $9,300 from marketplace sellers.
• CPUs are the exception. They have stayed near launch pricing, so they are unusually good value right now.

Unusual, and real. They are not typos.

A chipset ships as several board-partner cards, 49 chipsets become 324 here, priced hundreds of dollars apart for the same silicon. One card misrepresents the chipset; all of them are unreadable. So each chipset gets an index: per capture date, the mean of the cards near the market that day, where the market is the median of the cards a retailer was read for. A card read that day counts within 25% of the median; a card not read that day counts on its last figure within 15%, since a stale figure earns its place only by still agreeing with the market. A date with no reads treats every card priced that day as read.

The rule anchors on the cheapest card rather than the average on purpose. An average is dragged upward by exactly the halo models the rule exists to exclude, so an average-based threshold would quietly let them back in. Every GPU page lists its siblings and says which ones the index counts.

Card-level prices are modelled from the chipset's street price by each card's cooler tier, not verified one by one. Read the index as a trend and a single card's figure as an estimate.

Measured: the Compare tab's Benchmarks card prints a processor's or a card's Cycles render score from Blender Open Data (CC0): the median over every public run on Blender 4.5.0, read from the source once a day; a median on fewer than 10 runs is shown with a note that it may be inaccurate, and only medians on 10 or more count toward auto build. A processor under that with a twin on the same die (a KF and its K) that clears it is read as the twin, and Compare says so. It measures render throughput, a productivity figure, not gaming. Auto build's productivity weighting uses it.

Card performance: the Benchmarks card also prints TechPowerUp's relative performance, the GPU database's own index from the site's reviews with the GeForce RTX 3080 at 100, read 2026-09-26 for 49 of the 49 chipsets here, with permission given by email on 2026-09-23; a figure TechPowerUp marks as its estimate is marked so on the card. It is the card index the frame-rate estimate and auto build's gaming weighting use, halved onto the catalog's scale; the 0 chipsets the free set does not score keep a hand-set index that is never printed.

Processor performance: for two processors the card prints TechPowerUp's CPU reviews' relative performance in games at 1920×1080, read by hand from three reviews' charts on 2026-09-24 and re-based from each review's own processor to the Ryzen 7 9800X3D at 100, with the review, its date and its test system named beside the figure; two figures from different test systems are shown but not compared. It is the processor index the frame-rate estimate and auto build's gaming weighting use for the 59 of 94 processors those reviews tested; the rest keep a hand-set index that is never printed. A KF is read as its K, the same silicon with the graphics off, and the figure says so.

Makers' peaks: the Compare sheet's throughput rows (FP32 and ray-tracing TFLOPS, AI TOPS) are the makers' own theoretical peak figures, read from their spec pages on 2026-09-23, not measurements. Each maker counts AI on a different precision (NVIDIA FP4 with sparsity on Blackwell and INT8 with sparsity on Ada, AMD INT8 with sparsity, Intel INT8), so the sheet names the basis and compares AI within a brand only; a maker that publishes no such figure gets a dash.

Modelled: the frame-rate estimate is a model: an index for the card and one for the processor (TechPowerUp's where it has one, hand-set where not), and a resolution cost. It is labelled indicative wherever it appears and answers "which build is faster", not "how many frames". It remains a model even with measured inputs.

Sockets, memory type and slot count, card length against the case, cooler height and radiator size, power draw against the supply, form factors, M.2 and SATA counts, and more. Every rule runs on every change, and over the whole catalog, so a part that cannot fit is marked before you pick it.

Four rules exist that never fire today: memory height against cooler clearance, radiator thickness against the case, a front USB-C port against the board's header, and M.2 slots disabling SATA ports. Their fields are unset on every part: those are per-model facts that need a spec sheet, and this catalog does not guess. They start firing as fields are filled from cited sheets.

Product pages link to each shop's own search (Newegg, Amazon, Best Buy, B&H, Micro Center) and list no offers. A retailer is named beside a price only where the price was read from that retailer's own page, with the date of the read; stock and today's price are the shop's to answer. No link carries an affiliate tag; nothing here earns.

The same rule rules out product photos: maker and retailer images are copyrighted, so pages draw a placeholder that says so.