Will graphics card prices go down?

We will not tell you, over one to six months or over any other period. We built that answer, tested it against 1,252 past decisions, and it was wrong more often than a coin toss.

what we publishNo forecast

the horizon asked aboutone to six months

shop prices we hold0 of 10

evidence current to26 July 2026

Retail pricenot publishedWhy

A shop's price is the shop's data, and showing it needs that shop's permission. No shop has approved a price feed for this site: Rakuten Product Catalog technical approval, and the separate Newegg advertiser-feed approval, neither of which has been supplied for this property. Rather than copy prices we have no right to publish, we left them out and built the parts of the job that do not need them.

Every source we use, limit or refuse, each with the reason.

The working

The five things that answer rests on

Each one is a question you would reasonably ask before trusting anybody on this subject.

1. What is actually measured today?

Category price index, not the price of a part

Semiconductor and related device manufacturing (PPI)

29.695index level, June 2026

Match to this categorygood

Months held114

Months missing, left missing0

Against its own 12-month average-0.4%

January 2017June 2026

Source: US Bureau of Labor Statistics, public domain, PCU334413334413, read 2 August 2026.

Month by month

That is a US government index for a whole category of goods, which we are free to keep and calculate from. It is not the price of a board model, and it must not be read as one: a category can move while the exact part you want does not, and the other way round. The published match rating for this series against graphics card is good, and it stays on the panel rather than the index being quietly treated as exact.

2. What is not measured?

0 of 10shop prices held for one exact part

A real timing answer needs the same exact part priced over and over, from a source we are allowed to keep. We hold none, because no shop has approved a feed for this site yet. Nothing on this page is loading or broken; the record simply does not exist, and it cannot be filled in backwards once it does. That is the honest difference between this page and a chart built from prices someone copied without permission.

3. Why not just call it anyway?

60.4%of our own wait calls were wrong at three months

We wrote the buy-or-wait rule and tested it against 1,252 past decisions across 11 index series, with no hindsight: every decision used only what had been published before that date. At three months, 29 of 48 wait signals were followed by a higher index. At six months, 74.5% were. A rule that is wrong more often than a coin toss is worse than useless, because people would act on it, so it is switched off in code and a test keeps it off.

The full method, every horizon, and the caveat we hold against our own result.

4. What would change this page?

Three specific thresholds, all of which have to be met: at least 10 usable prices for one exact board model, gathered across at least 14 days, covering at least 60% of the offers we would expect to find. Any one of them failing keeps this page exactly as it is.

Note what all three passing would produce: a statement about where today's price sits against that part's own recent past. Still not a forecast. We would need a different and much higher bar before publishing one of those, and we may never clear it.

5. How complete would the data be?

Coverage is the share of the offers we would expect to find that a collection run actually matched. We report it as unavailable rather than as zero, because zero would suggest we ran a collection and found nothing. No collection has run: no source has given permission for one.

This matters more than it sounds. If a large shop drops out of a feed, the average price jumps and a naive tracker reports a rise that never happened. When collection does start, a run covering less than 60% of the comparison period suspends the answer rather than recalculating a direction from the smaller sample.

Index series kept for this category: 5. A series carrying a weak match rating is shown with that rating rather than quietly promoted. See every series mapped to graphics cards.

In detail

What the evidence does say about graphics cards

The retained evidence cannot date the next change in retail graphics-card prices. It can show whether two upstream semiconductor indexes are moving together, but it has no current, retention-cleared history for a named GPU, board partner, or cooler variant.

The June 2026 read

The BLS commodity index for integrated-circuit packages, including microprocessors, was 31.201 in June 2026. It was 0.1% above May 2026, 3.2% above its trailing 12-month mean of 30.238, and 4.4% above June 2025. The broader semiconductor-manufacturing index was 29.695 in June 2026. It was 1.5% below May 2026, 0.4% below its trailing 12-month mean of 29.829, and 0.6% below June 2025. March through June 2026 are provisional in both files. BLS WPU117839 and BLS PCU334413334413, retained observations fetched 25 July 2026.

That is a split upstream picture. The integrated-circuit-package index is above its recent mean and nearly unchanged in the latest month. The wider semiconductor index is slightly below its recent mean and recorded a second lower month after April. Averaging the two would hide the disagreement.

Why neither line is a GPU price

WPU117839 includes integrated-circuit packages beyond graphics processors. PCU334413334413 covers a broader manufacturing industry. Neither source identifies the GPU chip, board partner, memory capacity, board revision, cooler, power limit, condition, merchant, available stock, or promotion. BLS definitions and source registry, reviewed 2026-07-25.

Those distinctions matter in this category. A reference card, a compact partner card, and a factory-overclocked card using the same chip are not one retail product. A source disappearing from stock is not a price observation. Any future GPU price record must preserve the exact manufacturer part number and seller cohort before its trend can be trusted. Product-identity methodology, published 2026-07-26.

What coordinated movement looked like before

Both retained upstream series reached local lows in August 2025 and then recorded higher values across later releases. WPU117839 moved from 29.538 in August 2025 to 31.225 in April 2026, a 5.7% increase. PCU334413334413 moved from 29.392 in August 2025 to 30.602 in April 2026, a 4.1% increase. The paths were not identical, but the direction over that interval agreed. BLS WPU117839 and BLS PCU334413334413, monthly history read 25 July 2026.

That historical agreement preceded the April local highs in these indexes. It did not measure graphics-card inventory or prove a matching retail move. It is useful as a standard for what coordinated upstream evidence looks like.

The condition to watch

Check both BLS releases. The current split would change to coordinated lower upstream movement only if both series record two successive month-over-month decreases after June and later revisions preserve them. The June 2026 comparison points are 31.201 for WPU117839 and 29.695 for PCU334413334413. BLS WPU117839 and BLS PCU334413334413, retained values read 25 July 2026.

For an actual GPU price state, watch the source registry for a source approved for current display, historical retention, and derived analysis. A same-SKU price record must then reach at least 14 days, 10 usable price readings, and 60% coverage. Until those published thresholds are met, a claim based on launch price, a different board variant, or a stock alert is not a like-for-like trend. Price-state methodology, published 2026-07-26.

Why there is no timing verdict

The recommendation rule was backtested without lookahead over 1,252 observed series-months across 11 BLS series. At three months, 29 of 48 testable wait signals were contradicted. The 60.4% failure rate exceeded the pre-declared 40% limit, so the rule is switched off in code. Backtest and disabled rule, run 2026-07-26.

The useful result is therefore a check, not a guess: the upstream indexes currently disagree, and two aligned monthly moves would change that public context. Retail GPU timing remains unmeasured until like-for-like offer history exists.

Where next