Recommended Free Tools
AI infrastructure is increasing demand for several kinds of memory, not just high-bandwidth memory (HBM). HBM sits close to AI accelerators, while conventional server DRAM supplies host systems and NAND flash stores datasets, model checkpoints and inference data. Memory manufacturers say this demand is prompting new wafer capacity, cleanrooms, process changes and advanced packaging. Their announcements show the direction and scale of investment, but they do not establish how much of total memory-market growth is caused by AI or guarantee that planned capacity will arrive on schedule.
How AI systems create memory demand
Training and serving AI models requires moving large volumes of data between processors, memory and storage. That workload changes both the type and location of memory used in a data center. Manufacturers’ disclosures identify three main product categories affected by the expansion of AI infrastructure.
As an Amazon Associate I earn from qualifying purchases.
HBM supplies data to AI accelerators
HBM is a stacked DRAM product packaged alongside a processor or accelerator. Its wide interface is designed to provide very high bandwidth in a relatively compact connection, making it a premium component for many AI computing systems. Producing it requires memory wafers, stacking and testing, and specialized back-end packaging. A constraint in any of those steps can limit finished HBM output.
Crashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minutePC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Server DRAM remains part of the AI build-out
AI servers still need conventional DRAM for host CPUs, operating systems, data preprocessing and other tasks around the accelerator. SK hynix reported that conventional server-memory demand increased sharply in 2025 alongside more than a doubling of its HBM revenue year over year. Its outlook for AI expansion includes server DRAM as well as HBM, particularly as workloads move toward inference and distributed architectures.
NAND and enterprise SSDs store AI data
Training corpora, checkpoints, logs and inference data reside on storage systems built with NAND flash, usually in enterprise solid-state drives. NAND does not perform the same job as HBM or DRAM, but AI data centers can require substantial high-performance storage. Samsung’s FY2025 update listed AI-related NAND among its 2026 priorities, alongside HBM4, DDR5, SOCAMM2 and GDDR7.
What the evidence does not quantify
Company releases establish that executives are citing AI as a demand driver and that product mix is shifting toward data-center markets. They do not provide an independently measured percentage of industry growth attributable to AI. It is therefore accurate to describe AI as an important cited driver, not to say that all memory growth or every shortage is caused by AI.
What memory companies have reported about demand
The strongest public evidence comes from company earnings releases and regulatory filings. These sources are authoritative for what each company reported, but a company’s description is not an independent measurement of the entire market.
| Company disclosure | Reported figure or observation | How to interpret it |
|---|---|---|
| SK hynix FY2025 results (reported in 2026) | Revenue of KRW 97.1467 trillion and operating profit of KRW 47.2063 trillion; HBM revenue more than doubled year over year in 2025 | Shows the company’s financial performance and HBM growth, not an industry-wide total |
| SK hynix preliminary Q2 2026 results, speaking as of July 29, 2026 | Revenue of KRW 79,318.7 billion and operating profit of KRW 60,542.6 billion | Preliminary quarterly figures; they should not be treated as a full-year result |
| Samsung FY2025 release | High quarterly memory revenue and operating profit in Q4 2025, limited supply availability, strong conventional DRAM demand and expanding HBM sales | Samsung’s account of its quarter and priorities for 2026 |
| Micron 2026 filing | Some DRAM supply was shifted toward data-center and hyperscale-cloud markets during 2025; strong DRAM demand growth was linked in part to AI advancement | Evidence of a product-mix response, without a quantified displacement of consumer products |
SK hynix said in its Q2 2026 results that customer demand exceeded its supply capability. Samsung described limited supply availability in its FY2025 release. Those statements indicate tight conditions as experienced by the companies; they are not an independently verified industry-wide shortage statistic.
Rank #2
How manufacturers add memory capacity
“Capacity expansion” covers more than a new wafer-fabrication building. AI-oriented memory output can require coordinated investment across front-end manufacturing and back-end assembly.
New wafer fabs
A wafer fab processes silicon through many lithography, deposition, etching and cleaning steps. New fabs add long-term volume but require years of site work, equipment installation and qualification. A building that is announced or under construction is not equivalent to saleable chips.
Cleanrooms and existing-site expansion
Manufacturers can add cleanroom space at an existing campus, modernize an older line or transition equipment to a newer process. Cleanroom projects may be reported separately from a fab because the room, utilities and manufacturing tools come online in stages.
Advanced packaging
HBM requires stacking memory dies and connecting them to an accelerator package. As a result, packaging capacity can be a limiting step even when wafer output is available. Investment plans that mention HBM packaging, testing or assembly are addressing this part of the supply chain rather than adding another conventional fab.
Process and product transitions
Manufacturers can redirect tools or wafers from one product to another. Micron’s filing describes shifting some DRAM supply toward data-center and hyperscale-cloud customers during 2025. Such a mix change can support AI infrastructure without creating an equal amount of net new wafer capacity.
| Manufacturing lever | What it adds | Why timing matters |
|---|---|---|
| New wafer fab | Long-term front-end wafer capacity | Construction, tool installation and qualification precede production |
| Cleanroom expansion | Prepared floor space, utilities and controlled environments | Space may open before all tools are installed and qualified |
| Advanced packaging | HBM stacking, assembly and test capability | Packaging output must be synchronized with wafer supply and customer qualification |
| Process transition or product-mix change | More output of a targeted product from existing assets | May improve availability of one category while reducing another |
Company projects and expected dates
The following projects illustrate how announcements differ by product, geography and stage. Dates are the companies’ expectations, not guarantees of commercial supply.
| Company and project | Location and focus | Stated timing or scope | Status qualification |
|---|---|---|---|
| SK hynix M15X | South Korea; memory production | Accelerated mass-production preparation in the Q2 2026 update | Company plan; the release did not state that the added output was already available |
| SK hynix Yongin Phase 1 | South Korea; cleanroom and related capacity | Preparation for the cleanroom for early 2027 | Planned milestone, subject to phased execution |
| Micron HBM advanced-packaging facility | Singapore; HBM packaging | Packaging-capacity expansion beginning in early 2027 | Expected expansion described in Micron’s September 2026 filing |
| Micron acquired wafer fab | Taiwan; DRAM wafer production | Existing-fab shipments expected to contribute meaningfully from mid-2027 | Company expectation after the March 2026 acquisition |
| Micron Hiroshima cleanroom | Japan; memory manufacturing | Initial output expected in late 2028 | Forward-looking date in the September 2026 filing |
| Micron additional NAND cleanroom | Singapore; NAND | Expected to be operational in the second half of 2028 | Cleanroom milestone, not a statement that all tools and qualified product will be shipping then |
SK hynix said it was executing other mid- to long-term projects in phases according to customer demand and investment efficiency. Its June 2026 explainer put the construction challenge plainly: “Since building semiconductor fabrication facilities requires many years, preparation for new production sites has become increasingly important.”
Do these 3 things before closing this tab:
1Fix the driver behind crashes, sound loss and screen glitches2Clear out junk files and repair common Windows errors3Scan for outdated or missing drivers - takes under a minuteLarge investment plans: what the numbers include
Micron’s U.S. program
In July 2026, Micron announced more than $250 billion of planned U.S. investment through 2035 and a long-term goal of producing 40% of its DRAM in the United States. The figure is an announced, multi-year plan covering a national manufacturing and development strategy; it is not spending already completed and should not be read as a near-term addition to supply. Micron’s June 2025 release separately described expanded U.S. DRAM-fab and HBM-packaging plans and said supply growth would be managed in line with market conditions. The newer $250 billion-plus figure should not be added to the earlier plan as though they were separate totals.
Rank #4
SK hynix’s Korean regional strategy
SK hynix’s June 2026 explainer described a KRW 1,100 trillion mid- to long-term Korean strategy, phased over time:
| Region | Company-stated allocation | What it represents |
|---|---|---|
| Yongin | KRW 600 trillion | Part of the planned regional semiconductor complex and related capacity |
| Cheongju | KRW 100 trillion | Planned investment in the Cheongju area |
| Southwestern region | KRW 400 trillion | Planned investment across the southwestern region |
| Total | KRW 1,100 trillion | Company-announced mid- to long-term strategy, phased rather than a single-year expenditure |
These totals cannot be ranked directly against Micron’s U.S. figure. They use different currencies, geographies, planning horizons and scopes, and neither total is the same as realized capital spending or currently operating capacity.
Why new supply takes years to reach customers
- Site and permitting work: Land preparation, buildings, power, water and environmental systems must be completed before cleanroom operations can begin.
- Tool installation: Lithography, deposition, etch, metrology and packaging equipment is installed and connected to factory-control systems.
- Process qualification: Engineers run wafers, tune recipes, measure yields and qualify reliability. HBM adds stacking, bonding, assembly and customer-package validation.
- Customer qualification and ramp: A line can produce initial samples before it reaches the consistent yields and volumes required for commercial shipments.
That sequence explains why a filing may list output in 2027 or 2028 even when construction or preparation is already underway. The date is a forecast by the company, and delays, equipment availability, yield learning or changes in customer demand can alter the ramp.
How AI is changing manufacturers’ priorities
AI demand is influencing allocation decisions as well as construction. Companies are emphasizing HBM, data-center DRAM, high-performance NAND and the packaging needed to combine memory with accelerators. Samsung’s 2026 priorities span HBM4, DDR5, SOCAMM2, GDDR7 and AI-related NAND, illustrating that the response covers multiple products rather than one universal “AI memory” category.
Best Value
At the same time, shifting capacity toward data-center products can change the mix available to PCs, phones and other markets. Micron’s filing supports describing a reallocation toward data-center and hyperscale-cloud demand during 2025; it does not quantify a permanent reduction in consumer supply or establish a specific effect on retail prices.
How to evaluate an AI-memory growth claim
- Identify the product: Ask whether the claim concerns HBM, conventional server DRAM, NAND, enterprise SSDs, packaging or a combination.
- Separate demand from capacity: A company can report strong orders while a project remains years from production.
- Check the geography and date: A U.S., Korean, Singaporean, Taiwanese or Japanese project has different schedules and incentives; a 2025 result is not a 2026 forecast.
- Read the status language: “Announced,” “under construction,” “cleanroom ready,” “initial output expected” and “mass production” describe different stages.
- Normalize investment figures: Compare currency, time horizon, included businesses and whether the number is planned spending or completed investment.
- Look for attribution: Statements that demand exceeds supply capability or that AI is driving growth are company assessments unless supported by an independent market measurement.
What this means for the semiconductor manufacturing outlook
The evidence supports a clear chain: expanding AI infrastructure is increasing demand for HBM and is also supporting server DRAM and NAND; manufacturers are responding with product-mix changes, new fabs, cleanrooms and advanced packaging; and much of the announced capacity is scheduled for 2027 or later. The chain does not support a precise industry-wide AI growth percentage, a guaranteed shortage, a price forecast or an investment-return conclusion.
For readers following the semiconductor cycle, the most useful distinction is between demand reported today and capacity that can ship after qualification. Tracking both—by product, location and expected date—gives a more accurate picture than treating every investment headline as immediate supply.
Free tools Windows power users keep installed
One-click scans. No signup required.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
