Key findings
- The three largest memory makers expect the memory shortage to last into 2028 or beyond. Micron sees supply improving “gradually in 2028,” Samsung says the shortage lasts through 2028, and SK hynix expects demand to exceed supply beyond 2030.
- New fabs take three to five years to reach volume production, so most new capacity arrives between 2027 and the end of the decade.
- Price forecasts point earlier. Gartner said in March that prices may stabilize by early 2027, and Citi expects them to peak in the second quarter of 2027.
- Prices are easing mainly in the spot market for older DDR4 chips and in China’s used-module market. Contract prices are still rising: TrendForce has raised its fourth-quarter DRAM outlook, with PC and server DRAM leading the gains, and expects NAND contract prices to rise 15–20%.
- The biggest risk to the forecasts is AI spending. Hyperscaler budgets are still rising, but more of that spending is now funded by borrowing.
Contents
- Memory shortage forecasts, side by side
- Why new supply arrives slowly
- The first signs of easing
- What could make the forecasts wrong
- Three ways this could play out
- What it means for buyers and sellers of hardware
When will the memory shortage end? It is probably one of the questions the industry asks most often, and over the past year the memory makers’ answer has moved further out. In October 2025, SK hynix said supply of high-bandwidth memory (HBM), the stacked DRAM packaged next to AI accelerators, would stay tight into 2027. In July 2026, its CEO said 2027 will be “the worst year in the industry’s history from the supply perspective” and that demand would exceed supply “even beyond 2030.” The same month, Samsung told investors that the shortage “will persist through 2028.”
But not everyone agrees. In mid-September, Acer’s CEO said memory makers talk up distant end dates to protect their margins, and that memory and SSD supplies are already ample. Acer buys memory for its PCs. Behind both views sits a larger question: whether AI data-center spending, the main driver of the shortage, will keep growing.
Part of the reason the forecasts differ is that they measure different things. Most of them forecast how long memory supply will stay tight. A smaller group forecasts when prices will peak or fall, and those point to earlier dates. Prices can ease while supply is still tight, because prices also depend on inventory, how much buyers are willing to pay, contract terms and the type of memory.
Memory shortage forecasts, side by side
The two tables below collect 20 public forecasts made between October 2025 and September 2026, split into supply forecasts and price forecasts. The “Covers” column shows which memory each forecast refers to, because HBM, server DRAM, consumer DRAM and NAND flash are moving at different speeds.
Supply forecasts: how long supply stays tight
| Who (what they sell), when | Covers | Forecast |
|---|---|---|
| Phison CEO (SSD controllers), Oct 2025 | NAND | “Tight for the next ten years” |
| SK hynix (memory maker), Oct 2025 | HBM | Tight into 2027 |
| Micron (memory maker), Dec 2025 | DRAM and NAND | Constraints “persist beyond calendar 2026” |
| Samsung (memory maker), Jan 2026 | All memory | Supply growth limited in 2026 and 2027 |
| SK Group chairman (memory maker’s parent), Mar 2026 | All memory | Shortages for “another four to five years” |
| Former head of Samsung’s chip business (retired), May 2026 | All memory | Market could turn as supply surges in late 2027 or early 2028 |
| Goldman Sachs (bank), Jun 2026 | DRAM, NAND and HBM | Shortages “extend into 2028” |
| IDC (research firm), Jun 2026 | Memory for PCs and phones | No relief “before the end of 2027” |
| Micron (memory maker), Jun 2026 | DRAM and NAND | Tight “beyond calendar 2027,” improving “gradually in 2028” |
| SK hynix CEO (memory maker), Jul 2026 | All memory | 2027 the worst year; demand above supply beyond 2030 |
| ADATA chairman (memory modules), Jul 2026 | DRAM | Shortage for “another 10 years” |
| TrendForce (research firm), Jul 2026 | NAND vs DRAM | NAND supply catches up with demand in 2027; DRAM gap widens |
| Samsung (memory maker), Jul 2026 | All memory | Worse in 2027 than 2026, lasting through 2028 |
| Nvidia CFO (memory buyer), Aug 2026 | All supply, not memory only | A bottleneck “at least through the end of fiscal year 2028”, which ends January 2028 |
| KB Securities (Korean brokerage), Sep 2026 | DRAM and NAND | “The tightest supply conditions in history” in 2027 |
| Acer CEO (PC maker, memory buyer), Sep 2026 | Memory and SSDs | Shortage already over |
Price forecasts: when prices peak or fall
| Who (what they sell), when | Covers | Forecast |
|---|---|---|
| Gartner (research firm), Mar 2026 | Memory prices | May stabilize by 1Q27, with a small decline in 2H27 |
| AMD (CPUs and GPUs), Jun 2026 | DDR5 prices | Back to normal in “about another two years” |
| Citi (bank), Jul 2026 | DRAM and NAND prices | Peak in 2Q27, then edge down |
| TrendForce (research firm), Jul 2026 | Server DRAM contract prices | Rising each quarter through 2H27, by smaller amounts |
The three largest memory makers all point to 2028 or later. Micron expects supply to improve “gradually in 2028,” Samsung says the shortage lasts through 2028, and SK hynix expects demand to exceed supply beyond 2030. Only three voices in the supply table expect an earlier end. A former head of Samsung’s chip business expects the market could turn in late 2027 or early 2028. TrendForce expects NAND flash supply, but not DRAM, to catch up with demand in 2027. Acer’s CEO says the shortage is already over.
The price forecasts point to earlier dates. Gartner said in March that prices may stabilize by early 2027, Citi expects them to peak in the second quarter of 2027, and an AMD executive expects DDR5 prices to return to normal around 2028. Server DRAM is the exception. TrendForce expects its contract prices to keep rising every quarter through the second half of 2027, although the increases are likely to get smaller.
The makers have also pushed their own end dates later over the past year. SK hynix moved from HBM staying tight into 2027 to overall demand exceeding supply beyond 2030. Samsung moved from 2027 to 2028, and Micron from “beyond calendar 2026” to “beyond calendar 2027.” Two explanations fit: demand keeps running ahead of the makers’ plans, or a longer shortage gives them a stronger hand in contract negotiations. The dates alone don’t show which.
Each speaker’s business also shapes how higher prices affect them. Memory makers gain pricing power in contract talks when buyers expect a long shortage. PC makers sell more when buyers expect prices to ease. The two ten-year forecasts came from Phison and ADATA, which make controllers and modules and buy their memory chips from others. Higher prices raise the value of the inventory they hold and also raise what they pay to replace it. That context is worth keeping in mind when weighing each forecast.
Why new supply arrives slowly
If memory is this scarce, why don’t the makers simply build more? They do. Micron alone expects about $27 billion in capital spending in its 2026 fiscal year, and each of the three largest makers has announced new fabs. Two things slow the arrival of new supply: AI memory uses up much of the capacity that already exists, and a new fab takes years to build.
The first problem is HBM. According to Micron, HBM3E needs roughly three times the wafer area of DDR5 to deliver the same memory capacity, and the gap widens with HBM4, the next generation. Micron presented the figure in August at Hot Chips, an annual chip-design conference. TrendForce’s figures show the same effect: HBM will take about 22% of the three big makers’ DRAM wafer input by the end of 2026 but supply only about 9% of DRAM bit output. Working from those two figures, each gigabyte of HBM uses roughly three times the wafers of other DRAM, close to Micron’s own estimate. An earlier article explains why HBM squeezes PC and server memory in more detail.
The second problem is time. A modern memory fab typically takes three to five years from groundbreaking to high-volume production. TechInsights, a semiconductor research firm, puts construction and cleanroom installation at 18–30 months, and equipment installation and qualification at another 9–15 months. A finished cleanroom is therefore still about a year away from shipping chips. SK hynix’s M15X fab in Cheongju shows the timeline even on an accelerated schedule: construction began in April 2024, and full operation is expected around mid-2027, about three years later.
The fabs already announced will come online in a wave from 2027 to the end of the decade. The table lists the next milestone for each. A “cleanroom ready” date comes about a year before first output.
| Fab | Location | Next milestone |
|---|---|---|
| SK hynix Yongin, first fab | Yongin, Korea | Cleanroom ready February 2027 |
| Micron ID1 | Boise, Idaho | First wafer output mid-2027 |
| Samsung P5, phase 1 | Pyeongtaek, Korea | Output as early as 2027; other reports say late 2028 |
| Micron ID2 | Boise, Idaho | Output late 2028 |
| SK hynix M17 (NAND) | Cheongju, Korea | Cleanroom ready December 2028 |
| SK hynix Yongin, second fab | Yongin, Korea | Cleanroom ready June 2029 |
| Micron New York | Clay, New York | Output from 2030 |
China is adding capacity outside this list. CXMT, China’s largest DRAM maker, has capacity for about 320,000 wafers a month, which Taiwan’s Commercial Times expects to reach 420,000 by 2027, almost a third more. Most of its output is mainstream DDR5 and LPDDR5. In early September it was reported to have begun small-scale HBM3E production, and Chinese AI chip designers are testing the parts, while the Korean makers already mass-produce the next generation, HBM4. The same report says China’s own AI build-out could absorb much of CXMT’s new output through 2027, which would limit how much reaches buyers elsewhere.
These fab timelines support the makers’ 2028 date. They also set up a possible glut later, because every one of these fabs was planned during a shortage.
The first signs of easing
TrendForce’s figures show contract price increases shrinking through most of 2026. Conventional DRAM contract prices, the negotiated prices large buyers pay, rose roughly 93–98% in the first quarter, and a further 58–63% rise for the second. For the third quarter, TrendForce forecast 13–18% and said PC and phone makers were “reaching their affordability limit.”
The fourth quarter breaks that pattern for NAND and possibly for DRAM. On September 24, TrendForce said cloud-provider demand “keeps 4Q26 DRAM and NAND prices rising, led by AI.” It forecasts NAND flash contract prices up 15–20%, faster than its 10–15% forecast for the third quarter, with enterprise SSDs up 23–28%. The day before, it raised its fourth-quarter DRAM contract price outlook, saying tight supply and strong cloud-provider demand keep DRAM undersupplied. It expects PC and server DRAM to lead the gains, while increases for mobile and consumer DRAM cool from high levels. The DRAM percentage is only in TrendForce’s paid report. The public figures show that NAND price increases are accelerating and that TrendForce now expects larger DRAM increases than it did before.
Where prices are easing, it is in the spot market for older DDR4 chips and in used modules. Spot prices, the day-to-day market for chips sold outside contracts, have softened for DDR4. DDR4 is being phased out, so its spot market is thin and its prices are a noisy signal. In mid-September, DDR5 quote inquiries slowed and some DDR4 chip prices retreated, with “no clear buying interest” in DDR4. BuySellRam’s September 1 market update put the spot price of a 16-gigabit DDR4 chip at $91.78. On September 24, TrendForce listed the same chip at $83.78, about 9% lower. In China, used 16GB DDR4 modules had fallen more than 30% from their early-2026 peak by July, though they still cost five to six times what they did a year earlier.
These declines start from a very high level. The same DDR4 chip had a contract price of $48 in late July, so its September spot price is still roughly 75% higher, based on TrendForce’s published figures. A real turn would show two things: spot prices falling to contract levels, and contract increases dropping to zero. Neither has happened yet. At the high end of the market, a 36GB HBM3E stack reportedly trades on the spot market at four to five times its contract price.
Consumer demand shows the strain most clearly. IDC expects smartphone shipments to fall 16.7% in 2026, the steepest drop it has recorded. TrendForce forecasts a smaller 14% decline in production, because consumers bought early to avoid expected 2027 price increases, which pulls demand forward from next year. In notebooks, the CPU, DRAM and SSD now make up 68% of the parts cost, up from about 45% in early 2025. Apple CEO described the situation as a “100-year flood on memory pricing.”
Long-term contracts will also slow price declines when they come. Micron has signed 16 multi-year agreements covering roughly 20% of its DRAM and a third of its NAND through 2030, backed by about $22 billion in customer deposits and financial commitments. Samsung plans to sell 60–70% of its capacity under long-term contracts, reportedly with price floors for commodity memory. When prices fall, spot buyers and customers without contracts will see the drop first. That is the two-tier pricing pattern long-term contracts have created in server DRAM.
What could make the forecasts wrong
The biggest risk to these forecasts is AI spending. The long-shortage forecasts assume AI data-center spending keeps growing. If it slows, for example because the AI boom turns out to be a bubble, demand would weaken just as the new fabs come online, and the shortage could end sooner, possibly in a glut like the one in 2019.
So far the spending is still growing. In July, each of the four largest cloud companies, often called hyperscalers, raised or held its 2026 budget. Alphabet raised its plan to $195–205 billion, up from $91.4 billion in 2025, and Amazon raised its plan to about $220 billion, up from $131.8 billion. In May, Moody’s projected that hyperscaler capital spending would reach $785 billion this year and close in on $1 trillion in 2027.
Part of that growth is the cost of memory itself. Microsoft attributed $25 billion of its budget to higher memory and component prices, and TrendForce estimates DRAM and NAND at 47% of major cloud providers’ capital spending in 2026 and 68% in 2027. Record spending therefore partly reflects record memory prices. If memory prices ease, total spending could flatten while the number of servers keeps growing. A flat budget in 2027 could simply reflect cheaper memory.
The concern is how that spending is paid for. Free cash flow, the cash left after operating costs and capital spending, was negative $5.9 billion at Alphabet in the second quarter and negative $7.6 billion at Amazon over the past twelve months. Borrowing is covering the gap. By late July, Goldman had counted $489 billion of AI-related bond and loan issuance this year, well above its $322 billion estimate for all of 2025.
AI labs carry the same risk. OpenAI, one of the largest renters of AI capacity, had an operating loss of about $20.9 billion on $13.1 billion of revenue in 2025, according to audited figures reported in June. Its widely quoted $38.5 billion net loss includes a one-time non-cash charge from its conversion to a for-profit company. A Financial Times report says OpenAI expects to burn about $278 billion in cash from 2026 to 2030.
Part of OpenAI’s funding is linked to its suppliers. In a 2025 letter of intent, Nvidia said it intended to invest up to $100 billion in OpenAI progressively as Nvidia systems are deployed. Jensen Huang later said the figure “was never a commitment” and that investments would be decided in stages. When a supplier invests in a customer that buys its hardware, demand signals become harder to read. If OpenAI or a similar lab slowed its build-out, memory makers would feel it through fewer server orders.
The same filings also show strong demand. Cloud revenue is still growing fast: Google Cloud grew 82% in the second quarter, and AWS grew 37%, its fastest rate in 18 quarters. Microsoft said demand “continues to exceed available supply,” and Meta expects to remain supply constrained for the foreseeable future. That is one key difference from the late-1990s telecom overbuild, when network capacity expanded far ahead of the traffic that used it.
Older AI hardware is also still in demand. Oracle reported 97.9% utilization of its GPU fleet and said GPUs coming up for renewal, most of them four years old or more, were re-leased at a 20% premium. H100 rental prices rose about 22% in a month to $3.28 an hour in September, although that is still far below roughly $8 an hour in early 2024. Older accelerators rent for much less than two years ago, yet Oracle’s are almost fully used and the rental market has regained some pricing power. That is evidence against an overbuild in Oracle’s fleet, although one operator cannot settle the question for the whole industry.
The last memory boom offers a warning. In 2018, Samsung said continued demand and supply constraints were likely to sustain its profit margins for the foreseeable future. By October that year, TrendForce was forecasting a 15–20% DRAM price decline for 2019, and SK hynix’s operating profit fell 87% in 2019. HBM and multi-year contracts make this cycle different. Even so, consensus estimates already show Micron’s earnings per share peaking in fiscal 2028 and falling in 2029.
On balance, the evidence today points to continued spending, with financing as the main risk to watch.
Three ways this could play out
Most forecasts fit one of three paths, similar to the outcomes the Center for Strategic and International Studies (CSIS) outlined in a September report.
AI spending keeps climbing. Hyperscalers reach the roughly $1 trillion forecast for 2027, and demand absorbs new fabs as they open. The shortage lasts through 2028, as Samsung and Micron expect.
AI spending slows as new fabs open. If returns disappoint or credit tightens, budgets could flatten in 2027 or 2028, just as new capacity arrives. That would repeat the 2019 pattern. CSIS warns that “the investment needed to relieve today’s shortage may eventually recreate the conditions for the industry’s next glut.”
Older and consumer memory eases first. Weak phone and PC demand and new Chinese capacity push down DDR4 and consumer DRAM prices, while server DRAM and HBM stay tight.
The current data fits the third path only for DDR4 in the spot and used markets. Contract prices fit the first path: TrendForce’s raised fourth-quarter outlook has PC and server DRAM leading the gains, and HBM is still tight. A cut to 2027 budgets at the big four’s late-October earnings would point to the second path.
These public signals will show which path the market is taking:
- Quarterly contract price changes, and whether they reach zero or turn negative
- The gap between spot and contract prices for DDR4 and DDR5
- 2027 capital spending guidance from the big four in late October
- Rental and resale prices for older GPUs such as the H100
- Memory makers’ finished inventory, which KB Securities puts below 10 days at Samsung and SK hynix as of the third quarter
What it means for buyers and sellers of hardware
Most server buyers see OEM or reseller quotes rather than contract or spot prices, and those quotes have risen with memory costs. In December 2025, channel sources reported that Dell, Lenovo, HP and HPE planned server price increases of around 15%. TrendForce expects server DRAM contract prices to keep rising through the second half of 2027, with smaller increases along the way. Delaying a server memory purchase is therefore a bet against that forecast. Waiting can make sense for a discretionary upgrade that can move into 2028, if the buyer trusts the forecasts of a 2027 price peak. For a refresh needed within the next few quarters, there is little to gain by waiting while server DRAM remains tight.
For anyone holding decommissioned hardware or planning to sell RAM, the softening in DDR4 spot prices and Chinese used-module prices is a signal to watch secondary-market quotes closely. Server memory as a whole has not turned yet. Two forces pull DDR4 prices in opposite directions: Samsung and SK hynix are winding down DDR4 production, which supports prices, while falling spot prices and new Chinese capacity push them down. Anyone planning to sell server memory from a refresh should compare quotes over time.
The next data point arrives on September 30, when Micron reports its fiscal fourth-quarter results and gives its first outlook for fiscal 2027. If Micron repeats that supply will “remain tight beyond calendar 2027” and raises capital spending again, the long-shortage view holds for another quarter. If it trims capacity plans or signals flat prices, the forecasters expecting a 2027 turn will have their first hard evidence.