AI news and RAM prices: what moves them and what only moves shares

AI news moves memory makers’ shares within the day and reaches the price of RAM only when it changes orders, production or supply contracts. Across 36 dated events since late 2022 (AI news, makers’ own announcements and wider market routs), memory makers’ shares moved 5% or more in a day in about three cases in four and by as much as 30%, and where a weekly spot price (what a DRAM chip sells for in the open market that week) could be matched to the event, which was about 15 cases, it moved about 2% or less in every case outside the price run of autumn 2025. What has moved RAM prices is the makers’ own supply decisions, buyers’ inventories and the terms of supply contracts. In the summer of 2026 SK hynix shares fell about 55% between June 22 and July 30, and TrendForce’s spot reading for a mainstream DDR4 chip was about 30% higher on October 21 than on June 30. Nvidia announced on November 12, 2026 that its board had accepted the resignation of its chief executive; the last spot table TrendForce printed before the announcement (stamped 18:10 Taipei time on November 12, about four hours earlier) showed the DDR5 chip unchanged on the day, and the first test of orders is Nvidia’s results on November 17. To sort a headline, ask three things: does it change what a buyer has ordered, what a maker will produce, or what buyers hold in stock and what their supply contracts say? News that does none of the three is share news. This guide is about buying memory and gives no view on shares.

Evidence as of November 12, 2026: TrendForce’s spot prices of November 12 and its PC DRAM contract forecast of November 11, the third-quarter results of SK hynix (October 28) and Samsung (October 29), Micron’s fiscal fourth-quarter results (September 30), the third-quarter filings of the four largest US cloud companies, and Nvidia’s Form 8-K and press release of November 12. Trends shows how this site’s like-for-like price per gigabyte has moved since October 3, 2026.

What moved RAM prices and what moved only shares, in one table

These terms recur. DRAM is the kind of memory chip inside every RAM module; a 16Gb (gigabit) chip holds 2GB, so a 16GB module carries eight of them. The contract price is what a large buyer (a PC maker, a cloud company, a module brand) pays a memory maker under a negotiated deal that is reset each month or quarter; most memory is sold this way. The spot price is what a chip or module sells for today, in small lots, between traders and module makers; it is a thin market and it moves first. An LTA (long-term agreement) is a supply contract that runs for years and fixes a volume and a price band. HBM (high-bandwidth memory) is DRAM stacked for AI accelerators. An RDIMM (registered module) is the standard server memory stick. TrendForce is a research firm in Taiwan whose price tables and quarterly forecasts are cited throughout. 4Q26 means the fourth quarter of 2026. An OEM is a company that builds PCs or servers. The KOSPI is Korea’s main share index, where Samsung and SK hynix trade.

Each row links to the section that argues it.

The rows that moved prices are all decisions about supply, stock or contracts; the rows that moved only shares are all news about the companies that buy AI hardware. AI demand as a level is why RAM costs what it costs, and the shortage guide covers that. This guide is about the news.

In a hurry: sort a headline, what to do by buyer type, the dates ahead, where five possible developments would show first, how to use this site’s tables.

Sort any AI headline with three questions

A headline can reach the price of a memory kit by three routes only, so ask about each in turn.

1. Does it change an order? An order is a number in a filing: a cloud company’s capital spending guidance, the supply commitments in Nvidia’s quarterly report, or a memory maker’s long-term agreements.
2. Does it change production? That means a maker cutting output, ending a product, moving wafers from one kind of memory to another, or opening or delaying a plant.
3. Does it change buyers’ stock of memory or the terms of a memory supply contract? That means buyers holding more or fewer weeks of memory than they need, or a contract’s floor, ceiling or deposit being rewritten.

A headline that answers no to all three has not touched the price of memory, whatever it did to share prices that day.

Price readings are not headlines: a spot price under contract is a price signal in its own right, read in the watchlist.

A fall in memory shares says little about memory prices. Micron’s closing price moved 5% or more on 70 of the 217 trading days of 2026 through November 11, which is about one day in three, and the median daily move in 2026 was 3.4% at Micron, 3.4% at SK hynix and 3.0% at Samsung (arithmetic on Nasdaq and Naver closing prices).

If a headline changes no order, no production plan, no buyer’s stock and no contract term, change nothing about when you buy; if it changes one of them, the first place to see it is TrendForce’s Wednesday spot note and the server tables, weeks before a desktop kit. The live prices are on RAM prices and their movement on Trends.

Each company on the chain decides one thing

The price of a kit is the sum of a few decisions made by named companies. Knowing who decides what tells you whose news counts.

The companies in the news most often (the labs and the accelerator designers) sit two or three steps from the price of a kit. The companies that set it (the three large makers and the module brands) are in the news least, and every large price move on this record began with one of their decisions.

When news breaks, find the company in this table: if its decision is two steps from the wafer, the price of memory has not been decided yet. The shortage guide explains why three makers and one kind of wafer leave so little slack.

On November 12, 2026 contracts were rising, spot was flat and the Goldman Sachs memory basket was about half its June level

Five layers of the same market disagreed on November 12, 2026, and a buyer pays one of the lower four, never the share price.

Each layer answers a different question. Shares price the makers’ profits in 2027 and after; JPMorgan’s reading, in investingLive’s summary of November 9, 2026, was that the market “appears to be anticipating a memory downcycle driven by AI funding and power constraints as early as the first half of 2027”. Contracts record memory already allocated. Spot records what the last small buyer paid this week, and TrendForce’s note of November 4 said that with spot above the makers’ official prices and end demand slow, buyers “generally believe prices have room for corrections”.

Three readings point down: DDR4 chip spot, server module spot and Shenzhen street prices, and the last two are local. Street prices sit about 20% under the reported contract price (arithmetic on the two figures above), and server module spot fell in its latest update. Against them, several leading makers are reported to have “verbally informed distributors of plans to raise prices for enterprise-grade memory chips”, with no formal notice issued (Lanjing News, November 11, relayed by TrendForce’s news desk, November 12, 2026).

A second source gives a higher figure for the fourth-quarter contract. Taiwan’s Commercial Times, relayed by TrendForce’s news desk on October 20, 2026, put it at about 20%, after about 80% in the second quarter and 40% in the third. Set beside the makers’ own reports, those figures sit closer to the smaller DDR4 makers than to the three large ones (this guide’s comparison), whose reported DRAM price rises ran from the high teens in percent at Micron for its quarter to September 3 (Micron, September 30, 2026) to the high twenties at SK hynix for the third quarter (Yonhap Infomax, relayed by TrendForce’s news desk, October 28, 2026).

Research firms and banks read the same evidence two ways. Counterpoint wrote on November 5, 2026 that “Memory prices should stay strong next year.” Morgan Stanley’s view, in the South China Morning Post’s summary of August 10, 2026, was that “the memory cycle was set to enter its late stage in the fourth quarter, as price increases moderate and inventories build”. The first is a statement about prices and the second about the cycle that share investors price; neither is a dated price reading.

As of November 12, 2026 no contract series has turned down; what has softened is DDR4 chip spot, server module spot and street prices in China, and those are the places to watch. Trends carries this site’s own like-for-like record, and the wait guide lists the dated forecasts.

The chain from AI demand to a kit has six links, and each one has a lag

AI demand reaches a memory kit through orders, and every link in that chain is a contract, a factory schedule or a warehouse. That is why news travels along it slowly and why some news never arrives.

The lags are ranges, and one executive’s estimate for the last link is longer: the wait guide cites a Lexar manager who put the retail lag at eight to nine months. Used prices follow the same chain from its last link, and the used market is where a buyer can read the result directly, on this site’s tables.

Stock in buyers’ warehouses works on the price as well. By early 2019 buyers held over seven weeks of stock and contract prices fell more than 20% in a quarter (TrendForce, March 25, 2019).

Long-term agreements damp the price in both directions

The link between the makers and their large buyers is the one that changed in 2026. TrendForce wrote on October 7, 2026 that suppliers and cloud buyers “have concluded multi-year server DRAM LTAs”, with much of that supply under price bands, “while the remainder stays on quarterly pricing”. The makers’ own figures show the scale: more than half of SK hynix’s capacity for 2027 and 2028, with prices moving “within a price range agreed upon with customers” (SK hynix, October 28, 2026), and over 35% of Micron’s revenue through 2030, three-quarters of it under a defined pricing framework, a majority of which has “pricing bands with floor and ceiling prices”, and backed by $32 billion of customer commitments, the “vast majority of which are cash deposits” (Micron, September 30, 2026).

A ceiling holds the price down while the market rises. Nanya’s selling price jumped more than 60% in the third quarter, a rise TrendForce’s news desk described on October 12, 2026 as “largely driven by the expiration of older LTAs in June”. A floor and a cash deposit are written to hold the price up when the market falls, since a buyer who has paid a deposit has reason to take the memory; no downturn has yet tested them. The shortage guide covers how these agreements came about.

For a kit buyer: prices tied to contracts (new modules from the large brands, prebuilt PCs, laptops) will move late and by less, in both directions, and prices tied to spot (used modules, server pulls, small-brand kits) will move first. That is this guide’s reading of the contract terms above.

A change in AI demand would show in server module spot within weeks and in a new retail kit some months later, while JPMorgan expects makers’ earnings to hold for about six months; the spot lag is documented, and the retail lag is an inference from 2018 and 2025. The used server tables, DDR5 registered and DDR4 registered, sit at the fast end of the chain.

Shares moved 5% to 30% in a day while the weekly spot reading moved about 2% or less in most cases

Take each dated event, record what memory makers’ shares did that day, and set beside it what TrendForce’s weekly spot reading did that week and what the next quarter’s contract did. The chart shows five of the largest one-day falls in Seoul for which both numbers come from a dated publication.

• SK hynix closing price, that day
• DDR4 8Gb spot chip, that week's note

Sources: Share closes: Korea Times (Jan 31, 2025; Oct 15, 2026), Korea Herald via Asia News Network (Mar 5, 2026), Financial News (Jun 23, 2026), Kyunghyang Shinmun (Jul 28, 2026); TrendForce spot note, Feb 5, 2025; TrendForce spot note, Mar 4, 2026; TrendForce spot note, Jun 24, 2026; TrendForce spot note, Jul 29, 2026; TrendForce spot note, Oct 21, 2026. Figures as of November 12, 2026. January 31, 2025 was Seoul's first session after the DeepSeek sell-off of January 27; the Korean market was closed from January 27 to 30 for the Lunar New Year. Spot is the week-on-week change printed in TrendForce's Wednesday note for the mainstream DDR4 8Gb chip. The March 4, 2026 spot figure was read on March 3, the day before the fall; the next reading was about 1.2% higher. The October 15, 2026 bar is the close reported by the Korea Times; a series computed from Naver closing prices gives a larger fall, about 13%.

Show the figures

On five rout days in Seoul SK hynix fell 9.6% to 14.7%, and the DDR4 spot chip moved 2.1% or less that week (percent change)

Period	SK hynix closing price, that day	DDR4 8Gb spot chip, that week's note
Jan 31, 2025	-9.86%	0.00%
Mar 4, 2026	-9.58%	+2.10%
Jun 23, 2026	-12.47%	+0.28%
Jul 28, 2026	-14.65%	+1.39%
Oct 15, 2026	-11.06%	+0.84%

The same holds across the longer record. The spot figure is from the weekly note nearest the event; where that note was printed on or before the event day the row says so.

The exceptions are the weeks inside the price run of autumn 2025, when the spot chip rose 5.40%, 7.21% and 9.86% in the notes of October 1, 15 and 22 and 4.75% in the note of November 19, 2025. That run was under way before OpenAI’s letters were announced: the 5.40% reading was taken the day before the announcement, the notes give the cause as sellers holding back stock for higher prices, and TrendForce’s news desk reported on October 8, 2025, citing Commercial Times and Economic Daily News, that makers had “suspended new enterprise quotes for a week”.

Summer 2026 is the cleanest case. From June 22 to July 30 SK hynix’s closing price fell from 2,919,000 won to 1,322,000 won, about 55%; Samsung fell about 43% from its June 18 close to July 30 and Micron about 39% from its June 25 close to July 29 (arithmetic on Naver and Nasdaq closing prices). Over the same weeks the weekly DDR4 spot reading went from $36.00 on June 30 to $42.08 on July 28, and it stood at $46.79 on October 21. On July 29 SK hynix shares fell another 9.61% “despite announcing record quarterly earnings” (Korea Times), and on July 31 they rose 29.95% (Yonhap).

TrendForce’s weekly spot notes name contract timing, sellers holding stock, buyers refusing quotes and holidays as causes; none of those cited here names an AI-company event.

Retail kits did not follow the shares down either. Of six share shocks between January 2025 and November 2026, none was followed by a consistent fall in US retail prices: after the rout of July 28, 2026 the cheapest 32GB DDR5-6000 kit went from $384 on July 27 to $402 on August 24 (Tom’s Hardware price index).

No source on this record ties a fall in memory prices to a one-day fall in memory shares. Where prices did soften afterwards (the 2Q25 contract forecast of flat to down 5% after the DeepSeek sell-off), TrendForce gave supply and inventory as the cause, so a share fall alone is no reason to delay a memory purchase. What a sustained turn looks like, and the order its signals arrive in, is in the wait guide.

A change of chief executive did not move memory prices in three earlier cases, and Nvidia’s came on November 12, 2026

Nvidia’s quarterly report put its supply commitments at $279 billion on July 26, 2026 and says they are for “data center infrastructure systems, primarily memory and manufacturing facilities” (Form 10-Q, August 26, 2026), so its news bears on memory orders more directly than most companies’ outside the three makers. On November 12, 2026 it filed a Form 8-K (Item 5.02) and issued a press release. The filing says that on November 11 the board “accepted the resignation of Jen-Hsun Huang, as NVIDIA’s President, Chief Executive Officer and member of the Board, effective immediately”. The press release says the board “announced today that Jensen Huang has resigned as President, Chief Executive Officer, and a member of the Board, effective immediately”. The company gave no reason.

The rest of the announcement, in the company’s words and figures:

• Colette Kress, chief financial officer since September 2013, is interim chief executive and remains chief financial officer.
• Harvey Jones, a director since 1993, became the board’s first Chair. He said “we expect the company’s operations, customer commitments and product roadmap to continue without interruption”.
• The board “is initiating a search for a permanent Chief Executive Officer and will consider internal and external candidates”.
• Third-quarter fiscal 2027 results stay “on Tuesday, November 17, at approximately 1:20 p.m. PT, as previously scheduled”, with the call at 2 p.m. PT.

The share readings below are closing prices for November 12, 2026, the day of the announcement. The other rows carry their own time stamps.

The memory readings are all from before the news. The first spot table that can show a reaction is the one for November 13, and a spot table can only show what small traders did. The first test of orders is the results call on November 17, and three numbers in it bear on memory: revenue against the outlook of $108.0 billion, plus or minus 2%, given on August 26; the outlook for the following quarter; and the supply commitments in the quarterly report.

Earlier cases give a thin base rate. There are three, and no source ties a memory price move to any of them.

Two other headlines about OpenAI’s leadership in 2026 show the scale. On July 9 its chief of applications left her full-time role (The Next Web, July 10, 2026); Micron closed up 4.5% and SK hynix up 5.3% that day, and the spot note printed the day before showed the DDR4 chip up 2.88% on the week (TrendForce, July 8, 2026). On Saturday, September 12 its chief executive ruled out a share listing in 2026 (Fortune, relayed by TrendForce’s news desk on September 14, 2026); on the Monday Micron closed down 5.3% and SK hynix down 7.1%, and the spot note of September 16 showed a rise of 0.47%. The share figures are arithmetic on Nasdaq and Naver closing prices, on days with other news as well, and they sit close to the makers’ median daily move of about 3% in 2026.

The memory in an Nvidia system is fixed in its design (a Vera Rubin rack’s 20.7 TB of HBM4 is a specification), and it is bought under commitments that a filing reports each quarter. A change of leadership becomes memory news on the day one of those numbers changes.

Until Nvidia changes its guidance or its supply commitments, the resignation is share news, and the date to check is November 17. The same rule covers any future leadership headline at a lab, a cloud company or a memory maker.

AI service prices and memory prices have no documented link in either direction

AI service prices are what users pay: the price of a million tokens from a model, a monthly subscription, or an hour of a rented GPU. As of November 12, 2026 no published source ties a change in any of them to a change in a DRAM price, up or down.

Cheaper tokens could mean less memory, if the same work were done on fewer machines. So far they have come with more use, and no source ties the one to the other: Google’s token count rose more than sixfold in a year while unit prices fell, and the servers that produce tokens are the ones that hold the memory. Where sources explain a price rise in 2026, they explain it by scarce GPU capacity.

What is documented is memory as a cost to the companies that sell computing, which is not the same as a link between an AI service price and a memory price. OVHcloud raised most of its catalog, one server range by 40% to 59% from September 2026, citing memory and storage costs (InfoQ, August 23, 2026). Amazon’s chief executive said on October 29, 2026 that AWS’s margins on AI were improving “even with heightened component costs like memory”. Memory is a documented cost of running these services; no published source ties an AI service price to a memory price in either direction.

The one line worth keeping in view is TrendForce’s on monetization, written about server buyers in China; of US buyers the same report says subscription revenue “supports continued deployment expansion”. If paid AI services stayed unprofitable for long enough, the route to memory would be a cut in capital spending, and that is order news with its own dates.

A token price cut or a GPU rental move is no reason to buy or to wait; the thing to watch on the AI-service side is volume, and the latest count here, Google’s for May 2026, was more than six times the year before. The memory behind those tokens is server memory, listed on DDR5 server memory.

Capital spending is the link that counts, and part of the 2026 raises was the memory bill

Capital spending is what a cloud company pays for data centers and the servers inside them. It is the one AI-market number that is also a memory order, so it gets four checks.

The level. Microsoft, Alphabet, Amazon and Meta together spent $33.4 billion in cash on property and equipment in the second quarter of 2023 and $201.9 billion in the third quarter of 2026 (their cash-flow statements filed with the SEC; the sums are arithmetic). TrendForce’s count of nine cloud companies puts 2026 at $957 billion, up 106%, and 2027 at over $1.5 trillion (October 14, 2026).

201.9 / 33.4 = 6.04    about six times

What a raise is made of. In 2026 three of the four said on their own calls that component prices, memory above all, were part of the figure; Alphabet’s July raise, from $180-$190 billion to $195-$205 billion, was “primarily due to an acceleration in the delivery of capacity” (July 22, 2026).

A raised figure is therefore partly the memory price arriving in the buyers’ accounts, and a higher number can buy the same number of servers. Neither of Microsoft’s two trims was a cut in orders: July’s was accounting (more of its data-center leases became operating leases, which are not counted as capital spending) and October’s was late deliveries of capacity built by others; it guided the next quarter above $50 billion and expects spending “to accelerate primarily from CPUs, inclusive of higher memory and storage costs, and GPUs”, and each server CPU brings DDR5 with it (TrendForce, October 29, 2026).

The cash. Alphabet’s free cash flow was negative $9.1 billion in the third quarter, Meta’s negative $2.1 billion, and Amazon’s was an outflow of $42.4 billion over twelve months (company results, October 2026). The companies name their own lever. Meta, on April 29, 2026: “we can choose to bring it online more slowly or reduce our spending in future years”. Amazon’s chief executive, on July 30, 2026: “If the demand isn’t there, we won’t spend the capital.”

The precedent. Cloud cuts have reached memory prices once on this record. On January 31, 2023 TrendForce cut the four North American cloud companies’ server growth for the year from 6.9% to 4.4% and forecast server DRAM contracts down 20% to 25% in the first quarter, writing that the “downward corrections exacerbate the oversupply situation”. Meta lowered its capital spending guidance the next day, to $30-$33 billion from $34-$37 billion. The buyers deepened a fall that oversupply had started.

Spending news has also moved shares with no effect on memory. Oracle’s shares fell as much as 17% when it raised its spending plan (Reuters, December 15, 2025), and DRAM contracts for the following quarter rose 93% to 98% (TrendForce, June 1, 2026).

A trim the company explains by late deliveries, with the next quarter guided higher, is timing. A cut explained by demand, or a next-year figure below this year’s, is a change in orders.

A cut in server spending by a large cloud buyer would be the first AI headline in this guide that is an order; until one is filed, read a raise as partly the memory bill, and check whether a trim is an accounting change, a delivery date or a cut in orders. The next round of guidance, with the first 2027 figures, comes in late January and early February 2027, and the memory those budgets buy is listed on DDR5 registered modules.

Financing news moved shares, and contracts kept rising through every credit scare

More of the AI build-out was paid for with borrowed money in 2026, and the terms got worse through the year. Through November 12, 2026 every contract forecast published after a credit scare still pointed up.

Two things explain the gap. A lab’s promise is not an order: Nvidia’s backing of OpenAI’s Ohio sites is a guarantee of leases “cumulatively capped at $105 billion” (Nvidia Form 8-K, August 17, 2026), and of OpenAI’s seven US Stargate sites one was running, at about 0.3 gigawatts, on April 23, 2026 (Epoch AI). The memory makers’ own agreements, by contrast, carry cash deposits.

Credit trouble reaches memory by one route: a lender tightens, a cloud company trims server spending, and server orders fall. The Bank of England’s committee record of September 30, 2026 cited a Morgan Stanley estimate that AI-related debt issuance worldwide had reached about $450 billion by early September, more than double all of 2025, and the IMF’s October 2026 stability report lists “circular financing arrangements” among the risks. The one place the squeeze has shown in a memory price is server module spot.

A financing headline is a reason to watch cloud spending guidance and server module spot, and no reason yet to expect a cheaper kit. Server module asking prices are on DDR5 registered and DDR4 registered.

Efficiency scares cut share prices and left contracts where they were

An efficiency scare is news that AI can do the same work with less hardware. Two are on record. In the first the memory price was already falling for its own reasons; in the second the cause of a spot dip is disputed and the contract forecast still rose.

The March 2026 dip is often credited to TurboQuant, and the evidence is mixed. DigiTimes named two causes for the fall in spot, distributors selling stock they had built up and TurboQuant prompting some to clear out (relayed by TechSpot, April 10, 2026). Against that, the retail low of about $370 arrived about two weeks before the paper, TrendForce’s notes for those weeks name flat March contracts and spot sitting above contract, and the contract forecast issued a week after the paper was a rise. The wait guide follows that dip and its reversal. A line the Korea Times printed in the DeepSeek week has held up: “HBM memory is absolutely necessary with the current technology.” (Korea Times, January 31, 2025).

Design changes that use less memory are real and slow. TrendForce wrote on September 29, 2026 that chip designers “are now considering reducing HBM capacity per device”, and AMD says its flash-based MEXT software could cut the DRAM a server needs by 20% to 75% (Wccftech, relayed by TrendForce’s news desk, November 11, 2026). The HBM change is under evaluation for 2027 products and AMD’s software began rolling out in November 2026; Micron’s view is that such steps “have diminishing return for any further optimizations” (September 30, 2026).

An efficiency headline has not lowered a contract forecast on this record; if a design change does lower demand, it will arrive as a specification in a product launch before it shows in supply. For a kit buyer the check is the price itself, on 32GB DDR5 kits.

Policy changed who may buy, and it has reached contract prices only through what buyers did ahead of a rule

Governments reach memory in three ways: who may buy it (export controls), what it costs to move it across a border (tariffs), and how fast data centers can be built (permits and power). As of November 12, 2026 a rule has moved a contract price only through what buyers did ahead of it.

The election is the newest item and the one still open. TrendForce’s assessment of November 11, 2026 was that “Texas and Georgia have been raised to severe risk, and Wisconsin to high risk” for data-center regulation. A pause on building would delay server orders, which is order news. The same note records one statewide pause already in place, on data-center permits in Oregon, and says “the delay risk falls mainly on projects planned for 2028 and beyond”; no cloud company has changed its guidance for one.

Two later US steps, yearly licenses for Samsung’s and SK hynix’s plants in China (eeNews Europe, December 30, 2025) and a 25% tariff on a narrow set of advanced computing chips from January 14, 2026 (Morgan Lewis, February 18, 2026), have no price effect attributed to them. Power already limits how fast servers can be installed: transformer lead times were near 128 weeks (The Next Web, July 9, 2026), which moves memory demand later in time.

A new rule is worth acting on when it changes what buyers do before it takes effect, as tariffs did in April 2025; a rule that only moves shares has not changed what a kit costs. RAM prices by country compares the same mix of modules by where the item is located.

Every large price move on this record began with a maker’s own decision

Every large move in memory prices on this record starts with a decision by the companies that make it.

Not every maker headline is a decision. On November 6, 2026 the Taipei Times carried a Bloomberg report that Micron had “all but halted construction of a plant in Taoyuan” that was due in 2028. Micron’s statement, which did not address the Taoyuan site, was that “our operations today continue as normal”. The report came during the week’s share rout, no spot reading moved on it, and a plant due in 2028 changes nothing about supply in 2026 or 2027.

New supply is the decision that ends a shortage, and it is dated. SEMI counts 23 new 300mm fabs of all kinds starting operation in 2027 (October 13, 2026), CXMT is reported to be building 10, and Counterpoint’s view is that “major new fabs will not start production until 2028” (November 5, 2026). The shortage guide follows those dates.

Watch what Samsung, SK hynix and Micron do with wafers, quotes and product lines: a maker’s notice has moved the spot chip within a week, and no AI headline on this record has.

Weak PC and phone sales have not lowered contract prices, and makers tentatively plan less PC memory for 2027

High memory prices have cut sales of the devices that use memory. Phone shipments fell 22% from a year earlier in the third quarter of 2026 (IDC, October 15, 2026) and PC shipments about 20.1% (IDC figures relayed by guru3d, October 16, 2026). In an ordinary cycle that would bring prices down. In this one the contract is still rising, and TrendForce reports that makers plan to send less memory to PCs.

TrendForce’s release of November 11, 2026 sets it out. Notebook shipments fell 15.8% on the quarter in 3Q26. Earlier price rises had “prompted some PC brands to place orders beyond their actual requirements”, so that “finished-goods inventories have been gradually rising among OEMs and in North American distribution channels”. On the supply side, TrendForce says demand from AI applications “has led memory manufacturers to tentatively plan cuts to PC DRAM supply in 2027”, and the PC DRAM contract is still forecast up 13% to 18% this quarter, “with increases slowing from the previous quarter”.

The slack shows where stock changes hands outside contracts. A 16GB DDR5 module is “generally quoted at around RMB 1,600 in Shenzhen’s Huaqiangbei electronics market, below the reported average contract factory price of RMB 2,000” (Lanjing News, relayed by TrendForce’s news desk, November 12, 2026). Mobile DRAM contracts for the fourth quarter are “flat to slightly higher” (TrendForce, October 30, 2026), and Samsung is reported to be cutting Galaxy A shipments by 20% this quarter (TrendForce news desk, November 10, 2026).

Buyers holding more stock than they needed is how the 2018 fall began (TrendForce named double ordering on September 26, 2018), so this is the indicator to take most seriously. It has not turned a contract yet, and it works on PC and phone memory first.

Falling PC sales alone are no reason to expect a cheaper desktop kit in 2027: the contract is still forecast up and makers tentatively plan to send less memory to PCs. That is this guide’s reading of TrendForce’s release; short-lived discounts on overstocked kits are the realistic gain, and the desktop DDR5 table is where they show.

Server memory is where a cash squeeze among Chinese cloud buyers showed first

This is the one place in 2026 where a research firm tied money trouble at cloud buyers to a memory market. TrendForce reported cash constraints at Chinese cloud providers and modules redirected into spot; the spot price of server modules then fell, and the contract price rose through it. ChinaFlashMarket (CFM), the other source for these prices, reads the fall as thin trading.

The record runs in order. By July 9, 2026 some cloud providers and server makers had moved their RDIMM orders from 96GB and 128GB modules to 32GB and 64GB, to hold down cost and to match processor supply (UDN, relaying TrendForce). On September 30 TrendForce’s server module report said “Chinese cloud service providers face liquidity bottlenecks that prevent them from placing add-on orders”. Its bulletin of October 28 said “Some CSPs face cash flow constraints. High-capacity modules are being redirected to and absorbed by other CSPs and OEMs.” (A CSP is a cloud service provider; an OEM here is a server maker.) Two days later it wrote that “some Chinese CSPs have cut high-capacity pickups, prompting suppliers to redirect volume to OEMs and other regions while spot-market circulation rises”. TrendForce describes them as “Chinese cloud service providers” and names none.

The spot price fell within a week of the October bulletins; that the two are cause and effect is this guide’s inference. TrendForce’s spot quote for a DDR5 32GB RDIMM fell 6.47% to $2,025 in its table stamped November 2, 2026. CFM, which publishes a monthly range for each module, shows the two layers moving in opposite directions.

• Month price, Oct 13
• Month price, Nov 10
• Trade-spot high, Oct 13
• Trade-spot high, Nov 10

Sources: ChinaFlashMarket monthly quotes, DDR5 RDIMM 32GB, rows of October 13 and November 10, 2026; ChinaFlashMarket monthly quotes, DDR5 RDIMM 64GB, rows of October 13 and November 10, 2026; ChinaFlashMarket monthly quotes, DDR5 RDIMM 96GB, rows of October 13 and November 10, 2026. Figures as of November 10, 2026. CFM states that its monthly low mainly references the contract prices of large customers such as cloud providers, that its month price reflects the trend of contract prices, and that its monthly high is the trade-spot market for small lots delivered at once.

Show the figures

DDR5 server modules, October 13 to November 10, 2026: month price up, trade-spot high down (US dollars per module)

Period	Month price, Oct 13	Month price, Nov 10	Trade-spot high, Oct 13	Trade-spot high, Nov 10
32GB	$930	$1,030	$2,100	$2,000
64GB	$1,630	$1,800	$3,300	$2,900
96GB	$2,650	$3,025	$3,500	$3,300

Between the two dates the trade-spot high fell 4.8%, 12.1% and 5.7% for the 32GB, 64GB and 96GB modules, and the month price rose 10.8%, 10.4% and 14.2% (arithmetic on CFM’s rows). For the 64GB module the trade-spot high went from about twice the month price to about 1.6 times:

Oct 13:  3,300 / 1,630 = 2.02
Nov 10:  2,900 / 1,800 = 1.61

Spot is coming down onto a rising contract price. The 96GB module’s trade-spot high was $4,000 in April 2026, so its slide began before October. CFM gives a different reason from TrendForce: thinner trading and uncertainty around the US midterm elections, with no mention of buyers’ cash. Both explanations describe the same layer, the small-lot market, and neither claims a contract fell. TrendForce’s bulletin of November 4 reads “Server DRAM contract prices are rising, with gains varying by RDIMM capacity”. In Shenzhen’s Huaqiangbei market, a trader at stalls selling enterprise memory put prices 10% to 20% under their early-year high (Shanghai Securities News, November 11, 2026).

A server buyer gains from this in the small-lot spot market, in 64GB and 96GB DDR5 modules first; a contract quote will not fall because spot did. Compare registered DDR5 and DDR5 server memory listings against any quote in hand, and see how fleets bargain over the quote window.

A DDR4 chip costs more than a DDR5 chip in spot, and a DDR4 kit is still the cheaper purchase

DDR4 is the older generation the three large makers are leaving, and its price is set by that exit far more than by AI news. The chip and the kit give opposite readings, so name the layer before drawing a conclusion.

In spot, a DDR4 16Gb chip was $79.75 on November 12, 2026 and a DDR5 16Gb chip $60.233 (TrendForce spot table, read November 12, 2026). At retail, the cheapest new 32GB kit in the US was $197 for DDR4-3200 and $438 for DDR5-6000 (Tom’s Hardware price index, November 11, 2026).

The DDR4 chip costs more because few are still made. TrendForce’s bulletin of October 28, 2026 says “As the three major suppliers phase out mature process nodes, DDR4 supply is set to tighten notably”, and the makers that remain, Nanya and Winbond, are expected to raise prices again this quarter (Nanya’s president and investors cited by Commercial Times, relayed by TrendForce’s news desk on October 12 and October 6, 2026). The DDR4 16Gb chip has slipped 4.7% in spot in four weeks (arithmetic on TrendForce’s readings). TrendForce’s explanation is that spot had run ahead of the makers’ official prices while end demand was slow, so that buyers “generally believe prices have room for corrections” (TrendForce, November 4, 2026). For DDR5 the same notes say “supply remains tight and suppliers are clearly reluctant to sell” (October 21, 2026). One Amazon listing for a named 32GB DDR4-3200 kit still reached its highest recorded price, $269.99, on November 9 (camelcamelcamel).

DDR4 is the cheaper kit to buy today and its supply is shrinking, so buy it for a machine you already own and do not count on it getting cheaper. Compare DDR4 and DDR5 listings directly, see DDR4 or DDR5 for which one a given board takes, and DDR4, used memory and upgradeability for the used route.

Spot prices and makers’ decisions turned first at each of the last three turns

At each of the last three turns in memory prices, the first dated signal inside the memory market was a spot price or a maker’s supply decision; no AI headline or demand tracker came first.

• 2018 peak. Spot turned first: it fell below contract at the end of June 2018, and contracts fell in October.
• 2023 low. Production cuts came first, then a higher contract in the fourth quarter. Korea’s chip exports turned positive only in November 2023, in figures published on December 1.
• 2025 rise. End-of-life notices for DDR4 came first, then a 22.2% week for the DDR4 spot chip in June while the DDR5 chip stood still.

All of these indicators are free to read.

Sources for the readings: TrendForce’s price page and releases on the dates shown; Korea Times, October 1, November 1 and November 11, 2026; KED Global, December 1, 2023; Nvidia’s release of August 26, 2026; SEC filings for the capital spending sum.

Read the Wednesday spot note each week, the server module spot table when it updates, and the contract forecast each quarter; everything else on the list confirms a turn after it has happened or measures something one step removed. This site’s Trends page is the same idea applied to listings: a like-for-like record that began on October 3, 2026, so it shows every event from that date forward.

Five developments would each show first in a different place

The table applies the lags above to five things that could happen next. The timings are this guide’s inference from those lags and from the precedents named; none is a published forecast.

Check the first place in the row before acting on any of these: a cut in cloud spending would reach server module spot first, and the other four have no documented route to what a kit costs.

The dated tests run from November 13, 2026 to March 31, 2027

Each date below is a point where one of the three sorting questions gets an answer. “Confirmed” means the organizer has published the date; “usual” means the publisher’s pattern points to it.

Nothing on this calendar before late January is likely to lower a contract price: TrendForce’s fourth-quarter forecasts are all rises, most PC DRAM deals were expected to close in November “at slightly higher levels” (TrendForce, October 30, 2026), and its one statement on the first quarter points up. That is this guide’s reading of those releases. No tariff or export-rule decision with a published date falls inside the window. One is pending without a date: US approvals for chipmaking equipment at Samsung’s and SK hynix’s plants in China are reassessed each year (eeNews Europe, December 30, 2025), and for SK hynix the approval for 2027 “has yet to be issued” (TrendForce news desk, October 15, 2026).

The two dates that count most are November 17, for whether orders changed, and the cloud results in late January, for whether budgets did. The wider order in which a turn would show is in the wait guide, and Trends shows what listings did across each date once it has passed.

This site’s tables show asking prices, which change over days and follow the news late

The tables on this site list eBay asking prices, the last link in the chain. Asking prices on eBay change over days and hardly at all within hours, so the day of a headline is the wrong day to look and a week later is the right one. Trends keeps a like-for-like record from October 3, 2026: it prices the same kinds of listing every time (generation, kind of module, condition, size and pack) and labels each change a “real rise”, a “real fall” or “no clear change”. The record covers the share rout and the elections of early November 2026 and Nvidia’s announcement of November 12, so the chart shows what listings did across each of them.

Used kits are where eBay listings most often undercut new retail. For a listing marked new, check a retailer’s price first: none of the six share shocks between January 2025 and November 2026 was followed by a consistent fall in US retail prices, so a rout is no reason to expect the retail price to have dropped.

Use these tables a week after a headline to see what sellers ask now; a listing price is the end of the chain, and on this record it has followed contracts and supply and no share price.

What to do depends on what you are buying

The evidence above gives a different answer for each kind of buyer. None of it rests on a forecast of shares.

Two readings would change this table: a TrendForce conventional DRAM contract forecast with a flat or negative range, which would be the first since the one for 2Q25 (flat to down 5%, March 25, 2025), and a cloud company guiding 2027 spending below 2026. On this record, either would be expected to show in server module spot within weeks and in a retail kit later.

The share market and the makers disagree on timing. JPMorgan’s reading, relayed by investingLive on November 9, 2026, is that shares are pricing a memory downcycle “as early as the first half of 2027”. Micron said on September 30, 2026 “we do not have line of sight to when supply and demand will return to balance”, and SK hynix said on October 28 that it expects demand to exceed supply through 2030. The wait guide sets out the published forecasts.

If you need memory in the next three months, the AI news of November 2026 is no reason to wait and no reason to rush; the price you will pay rests on contracts and supply plans made before it.

Short answers

Five common beliefs, each with the check that tests it:

• A fall in memory shares means cheaper RAM next week. Check the Wednesday spot note: SK hynix shares fell about 55% in summer 2026 while the weekly spot reading rose.
• TurboQuant caused the March 2026 dip. The retail low arrived about two weeks before the paper, and DigiTimes put distributors selling stock first.
• Cheaper AI tokens mean less memory demand. Check use: Google’s monthly tokens rose more than sixfold in the year to May 2026 while unit prices fell.
• A letter of intent is an order. Look for a deposit, a filing or a delivery date: one of OpenAI’s seven US Stargate sites was running on April 23, 2026 (Epoch AI).
• DDR4 is a refuge from AI demand. The three large makers are ending it, and its chip costs more in spot than a DDR5 chip.

Does AI news move RAM prices? It moves memory makers’ shares the same day. It has reached RAM prices only by changing orders, production or contracts, and then over weeks to quarters.

Did Nvidia’s chief executive resigning change RAM prices? No memory price reading taken after the announcement existed on November 12, 2026. The DDR5 spot chip was unchanged in the last table before it (stamped 18:10 Taipei time), the first table that can show a reaction is November 13, and the first test of orders is November 17.

Samsung and SK hynix shares fell sharply. Is RAM about to get cheaper? No source on this record ties a fall in memory prices to a one-day fall in memory shares. SK hynix shares fell about 55% in summer 2026 while the weekly DDR4 spot reading rose, and the contract forecasts published since (September 30 and November 11, 2026) were rises.

If AI services get cheaper, does memory get cheaper? No published source ties the one to the other in either direction. Cheaper AI has so far come with more use, and the servers that do the work are the ones that hold the memory.

Is server memory getting cheaper? In the trade-spot market, yes: ChinaFlashMarket’s monthly highs for 64GB and 96GB DDR5 modules fell 12.1% and 5.7% between October 13 and November 10, 2026. Contract prices rose over the same weeks.

Is DDR4 a way around the shortage? It is the cheaper kit today. The three large makers are ending it, so the supply behind it is shrinking.

When is the next date that could change the answer? November 17, 2026 for Nvidia’s orders, the turn of the year for the 1Q27 contract forecast, and late January 2027 for cloud spending plans.

Where can I see what memory costs today? RAM prices lists current offers by type, Trends shows how they have moved since October 3, 2026, and prices by country compares the same mix of modules by item location.