Computer component prices

Start with a price we can define

“RAM prices” can mean a memory chip sold to a manufacturer, a finished kit on a retailer’s shelf, or an average across several kit sizes. Those are different measurements. Here the historical series follows four specifications: a 32GB DDR5-6000 kit, a 32GB DDR4-3200 kit, a 1TB NVMe solid-state drive and a 2TB NVMe drive. DDR identifies the memory generation; 6000 and 3200 identify the transfer-rate class. GB and TB describe capacity. NVMe is the storage interface.

These figures are dated observations, not today’s cheapest offer or a promise that a particular model costs that amount. They summarize the tracked products within each specification. Read the MemRadar source notes and download the original monthly basket CSV to inspect the product counts and definitions.

Observed prices and budgeting scenarios

Loading the latest source- and schema-checked data release.

Solid lines show monthly sample price levels. Dashed lines show separately named conditional scenarios. These are not calibrated forecasts or probability intervals. Historical errors were high; actual prices can fall outside every displayed path.

History shown Last 24 observed monthsAll observed history

Show optional lower-price and higher-price illustrative budgeting paths

Source: MemRadar aggregate medians. Scenarios are not forecasts or probability bounds; historical errors are high and actual outcomes can lie outside every path.

Read the historical line with its limits attached

A fixed specification is useful for budgeting, but it does not hold every product constant. MemRadar assembled its catalog in July 2026. Earlier rows therefore contain products that remained on sale in 2026. That survivor sample can miss discontinued, unsuccessful or replaced products. It cannot reconstruct the full choice a shopper had years earlier.

Each product contributes its monthly median of recorded daily prices; the specification then takes the median across those products. Months begin only when the basket’s components meet the source’s minimum sample rule. Regular Amazon listings are included, while member-only offers, flash sales and Warehouse prices are excluded. Newegg’s current feed is not part of this historical file.

Product mix and sample size can change. Comparing two monthly price levels does not measure the same products’ inflation. For that question, use the separate matched-product price index. The source methodology explains the distinction.

Why memory and storage need separate outlooks

DRAM is the volatile memory used in RAM; NAND flash is the non-volatile storage technology inside SSDs. Each depends on its own manufacturing capacity and its own sources of demand. In its September 30, 2026 outlook, TrendForce projected fourth-quarter contract-price increases of 10–15% for conventional DRAM and 15–20% for NAND flash. Its explanation centers on server demand and suppliers’ allocation of capacity, alongside weaker consumer buying. That gives a direction for upstream pressure. It does not mean retail memory kits will rise by the same percentage. Finished products also reflect inventories, channel margins, specifications and promotions. TrendForce, September 30 outlook.

The longer view is less uniform. TrendForce’s July 30 analysis expected DRAM supply to remain tight in 2027, with substantial new output contributing in 2028. It anticipated looser NAND conditions in the second half of 2027 as output grew against weaker consumer demand, while acknowledging that stronger AI storage demand could absorb that supply. RAM and SSDs can rise together and still come off that peak at different times. TrendForce, July 30 outlook.

IDC’s June 2 forecast was more cautious about broad relief, anticipating memory shortages through the end of 2027 and a 17% increase in whole-PC average selling prices during 2026. The forecast has a different date and scope: a whole-PC average also moves when customers buy different types of machines. These sources support a discussion of risk; their dates and measures do not support combining them into one retail growth rate. IDC, June 2 forecast.

A new factory is not an instant cheaper upgrade

Micron’s fiscal 2026 filing puts first DRAM wafer output from its first Idaho fab in mid-calendar 2027. It expects meaningful shipments from the acquired Taiwan site from mid-2027, while additional cleanroom space at its Singapore NAND facility is expected to become operational in the second half of 2028. These are company projections, not completed milestones. They also describe different stages of capacity becoming useful. Micron fiscal 2026 Form 10-K.

A capacity announcement can improve the supply outlook for later years without adding retail stock in the next few months. New capacity must ramp up and qualify, and its output may go to products other than consumer PC parts. The next year’s retail price still depends on demand as well as supply.

The GPU evidence is real, but too short for this model

TechSpot’s August 31, 2026 price survey listed the cheapest in-stock US GeForce RTX 5070 at $550 in February, $600 in July and $800 in August. That is a dated view of available offers for a GPU class, rather than a matched individual card or a market-wide average. It shows why a builder should monitor graphics separately. It cannot support a defensible 12-month statistical forecast from three observations. TechSpot’s original GPU survey.

No adequate reusable CPU history has been verified for this article. Motherboards, power supplies and other parts are also outside the modeled sample. Leaving those charts absent is a limit on the analysis, not evidence that their prices will stay flat. A complete PC budget needs dated offers for those items alongside the memory and storage evidence.

A useful horizon is not the same as a reliable forecast

A 12-month planning horizon is reasonable for a purchase or upgrade budget. None of the tested formulas performed well enough to justify naming a specific future price. Holding the latest median unchanged produced 12-month retrospective mean absolute percentage errors of 76.9% for DDR5, 72.8% for DDR4, 52.9% for 1TB NVMe and 53.3% for 2TB NVMe. Each score uses only ten overlapping origins. Those errors are too large to present the flat line as a reliable expectation.

These ten targets run from December 2025 through September 2026, and all flat predictions undershot the realized sample levels. The consecutive origins overlap, mostly testing one rising-price episode rather than independent market regimes. The same current survivor catalog underlies earlier data, so these scores cannot establish performance across future regimes. In the October 12 audit, the minimum and maximum of the candidate paths contained none of the ten realized targets for any specification. That failed envelope is not shown as a likely range. That result rules out describing those paths as a likely range. Trend formulas did not consistently improve on the simple benchmark.

The chart therefore presents assumptions openly. The flat benchmark asks what happens if the latest sample median persists. Optional illustrative budgeting paths reach 25% below or above that level after twelve months. They are arithmetic budgeting exercises, not estimates of what is probable. Their endpoints are arbitrary and narrower than past errors; neither path is a likely limit. A buyer should still compare dated offers for the exact part they need.

Methods, downloads and the monthly plan

The flat benchmark holds the latest complete-month median unchanged. Optional illustrative budgeting paths use geometric interpolation to reach illustrative minus 25% and plus 25% endpoints at month twelve. Neither is a forecast or confidence limit. Trend formulas were evaluated for the audit but are not presented as expected prices.

For the audit only, six- and twelve-month trends fit log prices, cap the monthly slope at plus or minus ln(1.1), and damp each successive monthly increment by 0.8. These trend paths are not displayed as expected prices. Rolling tests begin after 24 observations. Predictions at horizons of 1, 3, 6 and 12 months are scored only where a target observation exists. For each origin, calculate 100 × |prediction − observed target price| / observed target price, then average those per-origin percentages. This is mean absolute percentage error (MAPE); it is not the average dollar error divided by a single price. Changing-sample levels and survivor bias apply to those tests too.

The update scaffold uses free aggregate data, advances only completed months and preserves source and method versions. Scheduled monthly checks for two years. Updates appear only when validated new or revised data are available. Source availability determines whether new completed observations are available; the published release date and observation month remain unchanged when a check finds no new or revised data. A failed or invalid refresh leaves the last successful release available, with its actual date and an overdue warning. A new download does not turn an unfinished month into an observation.

Source data are free with attribution. Credit: MemRadar aggregate medians; underlying Amazon history licensed from Keepa. No private product history is reproduced. Download the original source CSV; the chart also links its scenario assumptions and checked source data and source release. The dataset includes full rolling backtest records. Download the October 12 publication audit for the failed candidate-envelope check.

Sources and reproducibility

• MemRadar public aggregate data and reuse terms
• MemRadar methodology
• Matched-product price index
• Keepa underlying licensed price history

The distinction matters throughout: observed prices document this sample; analyst outlooks describe possible market conditions; scenario paths expose assumptions. Before buying, compare a dated offer for the exact product with these sample levels; use the scenarios to test your budget, not to time the market.