9 min read
One 16GB laptop memory module now costs $265 on contract, as of late July 2026[1]. Not the laptop. The stick of RAM. That is up 16.7% from the previous quarter, which was itself up from the quarter before.
Server memory has gone somewhere else entirely. A 32GB server module (RDIMM) was averaging $1,900 on the spot market on September 7[1].
The reason converges on one thing: AI data centers are taking the world’s memory first. This piece follows that money as a cost structure. What the money is spent on, where it turns into profit, and who ends up actually paying it.
Three terms first
DRAM is the memory a computer uses for whatever it is working on right now. Think of it as desk surface. The wider it is, the more things you can have spread out at once. Cut the power and it all disappears. Long-term storage is a different kind of memory called NAND, which is what sits inside an SSD.
HBM (high bandwidth memory) is a specialized DRAM built for AI chips. Ordinary DRAM dies are stacked vertically and placed right next to the GPU. A single NVIDIA B300 accelerator carries eight HBM stacks, and each stack is 12 DRAM dies[2]. That is 96 DRAM dies in one accelerator.
Contract price vs. spot price. The contract price is what a manufacturer like Samsung or SK hynix negotiates quarterly with a large buyer such as Dell or Apple. The spot price is what the open market charges on a given day. When you buy a stick of RAM on Newegg, you are feeling the spot price.
1. Where the money goes: wafers moved to AI
A memory plant etches chips out of silicon wafers. Its capacity is fixed, and the manufacturer decides what that capacity makes.
Right now they are pointing it at AI. TrendForce estimates that HBM and server modules (RDIMM) together account for 51% of global DRAM bit supply in 2026[3]. More than half of the DRAM being made goes straight into AI servers and data centers. Every PC, phone, console and car on earth splits what is left.
The reason is not complicated. HBM sells for far more. The semiconductor analysis firm SemiAnalysis estimates HBM costs roughly three times as much as other memory and makes up half or more of the cost of a packaged GPU[2]. Given the same wafer, there is no case for making ordinary DRAM out of it.
That is also where the price inversion came from. On September 18 spot pricing, an older DDR4 16Gb chip averaged $85.75 while the newer DDR5 16Gb chip averaged $56.00[1]. The obsolete part costs 1.5 times the current one. Manufacturers wound down legacy lines to free up capacity for server parts, so legacy supply dried up first.
2. Unit prices: how far they moved
| Item | Change | Period |
|---|---|---|
| Server DRAM contract price | roughly +270% (forecast) | cumulative across 2026[3] |
| Enterprise SSD price | roughly +235% (forecast) | cumulative across 2026[3] |
| Conventional DRAM contract price | +13–18% | Q3 2026, QoQ[4] |
| NAND contract price | +10–15% | Q3 2026, QoQ[4] |
| HBM contract price | +70–140% (forecast) | 2027[3] |
The 2026 and 2027 numbers above are TrendForce forecasts, and the Q3 figures were published in early July. They are not settled results.
The settled side shows the scale more clearly. Global DRAM industry revenue in Q2 2026 came to $154.7B, up 59.5% quarter over quarter[5]. An entire industry grew 1.6x in three months.
By vendor: Samsung at $61.0B for a 39.4% share, SK hynix at $38.6B for 24.9%, and Micron at $36.0B for 23.3%[5].
That revenue growth did not come from shipping more. TrendForce says Q2 DRAM bit shipments grew only “modestly,” and attributes the revenue jump to contract price increases[5]. Roughly the same volume, sold for much more.
3. Profit: 76 cents on the dollar
What price increases do to margins shows up plainly in filings.
SK hynix, Q2 2026 (company-reported)[6]
| Q2 2026 | Year earlier | |
|---|---|---|
| Revenue | ₩79.32 trillion | ₩22.23 trillion (+257%) |
| Operating profit | ₩60.54 trillion | ₩9.21 trillion (+557%) |
| Operating margin | 76% | 41% |
Seventy-six cents of operating profit on every dollar of revenue. The prior quarter was 72%[6]. First-half revenue crossed ₩100 trillion for the first time.
For comparison, NVIDIA’s operating margin last quarter was 66.2%. The memory makers are out-earning the company with a near-monopoly on AI chips.
Micron, from its SEC quarterly filings (Form 10-Q)[7]:
| Quarter | Revenue |
|---|---|
| FY2025 Q3 (Feb–May 2025) | $9.30B |
| FY2026 Q1 (Aug–Nov 2025) | $13.6B |
| FY2026 Q2 (Nov 2025–Feb 2026) | $23.9B |
| FY2026 Q3 (Feb–May 2026) | $41.5B |
Quarterly revenue is 4.5 times what it was a year earlier. Nine months into fiscal 2026, cumulative revenue of $79.0B is already more than double the $37.4B Micron booked in all of fiscal 2025[7]. Fiscal Q4 results land on September 30.
4. Capex: why nobody built more fabs
With demand like this, building more plants seems obvious. Why that does not happen is the heart of the problem.
First, nobody built during the preceding years. Memory analyst Thomas Coughlin notes that after the 2023 downturn, “all the memory and storage companies were very wary of increasing their production capacity again,” with the result that “there was little or no investment in new production capacity in 2024 and through most of 2025”[2]. A crash three years ago is constraining supply today.
Second, fabs are slow. A new memory plant runs $15 billion or more and takes 18 months or longer to build and ramp[2]. Breaking ground now means output in 2028. The published schedules agree: Micron’s Singapore HBM fab in 2027, its Taiwan fab in the second half of 2027, and its New York complex not at full production until 2030. Samsung’s new Pyeongtaek line starts in 2028; SK hynix’s Cheongju HBM fab completes in 2027 and its West Lafayette, Indiana site by the end of 2028[2].
Intel CEO Lip-Bu Tan put it bluntly at an event in February 2026: “There’s no relief until 2028.”[2]
Third, the shape of expansion changed. TrendForce expects the big three to add bit supply through 2026 and 2027 mainly by migrating to finer processes, squeezing more chips out of the same wafer, rather than by raising wafer starts much[5]. That method has a low ceiling.
5. Who profits and who pays
This is where the structure resolves.
Cloud providers absorb it first. TrendForce projects that DRAM and NAND will rise from 47% of major cloud service providers’ capital expenditure (capex, the money spent on assets like buildings and servers) in 2026 to 68% in 2027[3]. Two-thirds of data center spending becomes a memory bill. Over the same span it expects total CSP capex to grow 98% in 2026 and another 50% in 2027.
TrendForce frames two responses available to those buyers: raise capex further to protect the AI chip volumes they planned on, or “more aggressively optimize AI system memory architectures by reducing memory capacity per system”[3]. The second is already underway. Server module capacities are stepping down from 96GB and 128GB to 32GB and 64GB[8]. On August 4, TrendForce published an analysis titled around exactly this: DRAM supply staying tight through 2027 is prompting NVIDIA to lower the HBM configuration on Rubin Ultra.
Contract terms split the pain. Several large US cloud providers locked in multi-year long-term agreements (LTAs) that bar suppliers from raising their prices. So from Q3 2026 onward, TrendForce expects the primary source of server DRAM price increases to shift toward “customers without LTAs”[8]. The big buyers fixed their price. Everyone else pays the increase.
Consumers pay last. It arrives three ways.
First, price increases. Look at HP’s fiscal Q3 2026. Personal Systems revenue was $11.8B, up 18%, while unit shipments fell 16%[9]. Fewer machines, more money. That is a price increase. CFO Karen Parkhill, noting gross margin of 18.8% was down year over year, told analysts: “We continue to expect memory and storage costs to increase further as a percentage of the bill of materials.”[9]
Apple is not exempt. TrendForce estimates that memory costs for the 256GB iPhone 18 Pro in Q3 2026 will run nearly 400% higher than a year earlier, and projects iPhone 18 series prices rising 10–20%[10]. The first foldable model is estimated to start at $2,099–2,299. These are analyst projections, not Apple announcements.
Second, quiet downgrades. Instead of raising the price, shrink what is inside it. TrendForce research VP Avril Wu says “Both approaches are likely to occur,” with premium models taking outright increases while mid- and low-tier models cut specs to hold a price point[11]. Counterpoint Research senior analyst Shenghao Bai adds that “In some models, we are seeing downgrades of components like camera modules, displays, and audio components”[11]. The same model name at the same price may now ship with 8GB of RAM where it had 16GB.
Third, buying parts directly. Consumer Reports tracked one 32GB RAM kit in late 2025: $90 on September 11, $269.99 by November 21, and $349.99 the week the article ran[11]. Just under 4x in three months.
What to actually do about it
Nobody buying a laptop can change this structure, but buying informed beats buying blind.
1. Check the RAM and storage spec, every time. Last year’s model name at this year’s familiar price may have less of both. A downgrade is harder to notice than a price tag.
2. Assume you cannot upgrade RAM later. Thin laptops now solder memory to the board. Paying one tier up at purchase beats replacing the machine early.
3. Keeping your current device longer got more valuable. Few analysts see supply loosening before 2028.
4. Cloud and subscription pricing sits under the same pressure. If two-thirds of data center capex becomes memory, that has to surface somewhere.
One more thought
The profit moved.
People picture GPU makers when they talk about the AI buildout. But on Q2 2026 numbers, the company earning 76 cents on the dollar was not the GPU company, it was a memory company. If more than half the cost of an AI chip is memory, then a large share of the AI boom was a memory boom to begin with.
And the money comes from somewhere. It comes out of cloud capex, and upstream of that out of AI service fees, and downstream of that out of whoever buys the $600 laptop that quietly has 8GB now. Data centers and the laptop aisle look unrelated, but they are queuing for the same wafers from the same plants.
Asked when prices come down, economist Mina Kim answered that “prices come down much more slowly and reluctantly than they go up.”[2]
Should AI service prices rise by as much as memory did?
—
This article is for information only and is not investment advice. Investment decisions and their outcomes are your own responsibility.
Related: NVIDIA’s $96.2B Quarter, 2.06x in a Year · Arm’s CEO on Why the CPU Still Anchors the AI Data Center
Notes
- TrendForce memory price tracking. DRAM spot prices as updated 2026-09-18 18:10 (GMT+8); module spot prices as of September 7; DRAM contract prices as of the second half of July. DRAM prices · NAND prices ↩
- Samuel K. Moore, “How and When the Memory Chip Shortage Will End,” IEEE Spectrum, February 10, 2026 (corrected March 9). HBM stack structure and B300 die counts, the SemiAnalysis cost estimate, the Coughlin and Kim remarks, the Lip-Bu Tan quote, and fab costs and schedules are all from this article. Source ↩
- “Memory Prices Soar; DRAM and NAND Flash to Account for 68% of Major CSP CapEx in 2027,” TrendForce, August 25, 2026. Source ↩
- “AI Server Demand Continues to Support Memory Prices in 3Q26, but Gains Moderate as Consumer Demand Weakens and High Base Effects Take Hold,” TrendForce, July 3, 2026. Source ↩
- “DRAM Industry Revenue Rises 59.5% QoQ in 2Q26 as Supply Expansion Continues to Lag Demand Growth,” TrendForce, September 7, 2026. Source ↩
- SK hynix Q2 2026 business results (K-IFRS consolidated, preliminary), July 29, 2026. Q1 figures from the April 22, 2026 release. Source ↩
- Micron SEC quarterly reports (Form 10-Q), XBRL data. The FY2026 Q3 report was filed June 25, 2026 covering 2026-02-27 to 2026-05-28; the FY2025 annual report (10-K) was filed October 3, 2025. SEC XBRL API ↩
- “Long-Term Agreements Cap Price Increases; Server DRAM Contract Prices Expected to Rise 13-18% QoQ in 3Q26,” TrendForce, July 9, 2026. The Rubin Ultra point is from “DRAM Supply to Remain Tight in 2027, Prompting NVIDIA to Lower HBM Configurations for Rubin Ultra,” August 4, 2026. Source ↩
- HP Inc. fiscal Q3 2026 earnings call, August 26, 2026. Transcript ↩
- “Memory Costs Drive Up iPhone 18 Series BOM; AI and Pricing to Shape Upgrade Demand,” TrendForce, September 3, 2026. TrendForce states these are analyst assessments subject to Apple’s official announcements. Source ↩
- Nicholas De Leon, “With AI Data Centers Scooping up RAM, Laptop Prices Could Spike in 2026,” Consumer Reports, December 19, 2025. Source ↩
Source: https://www.trendforce.com/presscenter/news/20260825-13198.html