🔍 Read the full analysis: How To Prepare For AI-Driven Memory Costs In IT Budgets on ThorstenMeyerAI.com
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
SK hynix said on October 28 that memory demand is likely to exceed supply through 2030, while Samsung and Micron have also described tight supply. AI data centers are increasing demand for high-bandwidth memory made from the same wafer supply as conventional memory, adding cost pressure for IT buyers. The duration and severity of shortages remain forecasts, and suppliers’ financial interests are part of the debate.
SK hynix expects memory demand to exceed supply through 2030, the South Korean chipmaker said in its October 28 outlook, raising the prospect of continued cost pressure for organizations buying servers, PCs and phones. The forecast reflects demand from AI infrastructure and the long lead times for adding production capacity; it is a company projection, not a guarantee of prices or supply levels through that date.
In its official release, SK hynix said supply was not keeping pace with market growth. It cited the technical difficulty of leading-edge manufacturing and the time required to build new production facilities as reasons it expects demand to outstrip supply through 2030. On the company’s earnings call, Chief Financial Officer Kim Woo-Hyun said higher memory costs and shortages were expected to reduce PC and mobile-device shipment volumes, while AI infrastructure demand would more than offset that decline.
The pressure reaches beyond the specialized memory used in AI accelerators. High-bandwidth memory (HBM) is made from DRAM, as are familiar memory products used in computers and phones. TrendForce estimates that HBM production uses a larger share of wafer capacity than its share of memory-bit output: by the end of 2026, it expects HBM to account for about 22% of wafers started by the three largest producers and about 9% of memory bits. Those are market estimates, not a direct measure of retail prices.
The supplied report says a 32-gigabyte DDR5-6000 kit in the United States costs at least $447, compared with a recorded low of $72 in Tom’s Hardware’s price index. It also cites TrendForce estimates that ordinary memory contract prices rose 93% to 98% in the first quarter of 2026. These figures describe different measures—retail kit pricing and chip contract prices—and should not be treated as the same price series.
Infographic 01 · Where the memory goes
AI data centers get first call on memory.
High-bandwidth memory (HBM) for AI chips and the memory in your phone come out of the same factories. HBM needs far more wafer per gigabit, and its buyers sign years ahead.
The bakery problem
HBM share of wafers vs. share of the memory bits they produce
Same factories, same wafers: every wafer used for AI memory is missing from phone and PC memory.
Who gets first call
Long contracts lock up supply years ahead
- SK hynix: deals with 14 major customers, usually five years, with deposits; more than 50% of 2027–2028 capacity
- Micron: more than 75% of fiscal 2027 shipments already committed
- HBM demand: customers want “more than double” what SK hynix can supply today
- Cloud budgets: memory ≈ 47% of big cloud providers’ capex in 2026, 68% in 2027 (TrendForce estimates)
Three makers, three horizons, one customer’s view
- SK hynixThrough 2030Demand above supply (company forecast)
- SamsungInto 2028Tight supply, based on market research
- MicronNo “line of sight”Tighter in 2027 and 2028 than in 2026
- Acer (customer)Easing from 2H 2027Says suppliers talk up the shortage
HBM is a wedding cake, phone memory is bread. Same ovens. The cake customers pay more and order years ahead.Thorsten Meyer · my analogy
Sources: TrendForce, 2 Jun (wafer and bit shares, estimates); TrendForce, 25 Aug (capex shares, estimates); SK hynix, 28 Oct 2026; SK hynix call, 28 Oct; Micron call, 30 Sep; The Register, 1 Oct; Samsung call, 29 Oct; Acer via Tom's Hardware / DigiTimes. Wafer-per-gigabit ratio: my arithmetic. All horizons are forecasts.
Powered by Thorsten Meyer AI
Budget Exposure Across IT Purchases
For IT departments, memory costs can affect both large infrastructure projects and routine device refreshes. Server configurations with more memory may become more expensive, while higher component costs can flow into laptops and phones or lead manufacturers to offer less memory at a given price. That can complicate procurement plans, refresh schedules and per-user device budgets.
The impact will vary by organization. Buyers with fixed-price contracts, existing inventory or flexibility over configuration may face different costs from those purchasing at spot prices or replacing equipment immediately. SK hynix also said long-term agreements with 14 major customers cover more than half of its production capacity for 2027 and 2028. That is evidence of advance commitments at one supplier, but does not establish how much capacity is available to other buyers or what their contract terms will be.
Planning teams should treat the forecast as a risk scenario rather than a settled price path. The source report notes that a major customer, Acer, has accused suppliers of talking up shortages to protect margins. Suppliers have an interest in the outlook they present, while customers have an interest in lower prices. Neither the accusation nor the suppliers’ projections independently settles how tight the market will be.
Products Worth Considering
eufy Security eufyCam C37 4-Camera Kit with 360° Pan and Tilt
Dyson Ball Animal 3 Corded Upright Vacuum
Dyson V8 Cyclone Cordless Vacuum
As an affiliate, we earn on qualifying purchases.
Why AI Memory Competes for Wafers
DRAM holds data a device is actively using; NAND flash stores files and applications when the device is off. HBM is a form of DRAM built by stacking memory chips and placing them close to AI processors, enabling high data-transfer rates. AI systems use HBM to move data quickly, and chipmakers are directing more manufacturing resources toward this higher-value product.
The supplied report cites TrendForce estimates that HBM will use about 30% of the three largest manufacturers’ wafer starts by the end of 2027, while providing about 13% of their memory-bit output. This illustrates the trade-off in wafer allocation, but the figures are forecasts and do not by themselves predict the price or availability of a particular PC or phone component.
Other suppliers have described a similar, though not identical, outlook. The source says Samsung expects tight supply into 2028, while Micron has said it has no “line of sight” to a balance between supply and demand. Differences in company forecasts matter: the 2030 date is SK hynix’s projection, not a consensus timetable established by the figures provided.
“Supply is not keeping pace with the growth of the market.”
— SK hynix, in its October 28 earnings release
Forecasts, Prices and Supply Gaps
The 2030 outlook is a forecast from SK hynix, based on its assessment of demand and production constraints. The supplied material does not establish that shortages will persist at the same level every year, or that all memory types and buyers will face identical conditions. Samsung’s cited outlook runs into 2028, and the source gives no common industry forecast resolving the difference.
It is also unclear how much of the projected pressure will reach end users, and when. Contract pricing, retail component prices and the prices of complete devices can move on different schedules. The figures supplied do not provide a full, independently verified breakdown of production, inventory, supplier margins or the price effects by product category. Acer’s allegation about supplier messaging is an attributed claim, not proof that forecasts are inaccurate.
Track Procurement and Supplier Updates
IT budget owners can monitor supplier earnings calls, contract-renewal terms and vendor pricing for changes to lead times or memory configurations. Before committing to a purchase plan, teams can ask suppliers to specify whether quoted prices are fixed, how long they remain valid, and what substitutions or delivery delays could apply. Comparing configurations may also help identify where a lower-memory option would meet actual workload needs, without assuming that every organization can reduce capacity safely.
The next useful signals will be updated supply and demand forecasts from SK hynix, Samsung, Micron and market researchers such as TrendForce, alongside shipment and pricing data. Until those arrive, buyers should distinguish confirmed current quotes from projected market trends and avoid treating any single company’s long-term forecast as a settled outcome.
Key Questions
What did SK hynix announce?
On October 28, SK hynix said it expects memory demand to exceed supply through 2030. It attributed the outlook to market growth, manufacturing challenges and the time needed to build new facilities.
Why can AI demand affect ordinary PC and phone memory?
HBM used alongside AI processors is made from DRAM, using wafer capacity that can also support other memory products. As manufacturers allocate more capacity to HBM, the supply available for conventional memory can face pressure.
Does the forecast mean all devices will become more expensive?
No. The forecast does not establish the price of every device or component. Costs can depend on product design, supplier contracts, inventory and the timing of manufacturing and retail price changes.
How should IT teams use the forecast in their budgets?
They can treat it as a procurement risk to test against, while using current supplier quotes and contract terms for actual budget decisions. The supplied information does not support a single price increase that every organization should apply.
Is the shortage forecast independently confirmed through 2030?
No. Through 2030 is SK hynix’s forecast. The supplied report cites other suppliers describing tight conditions on different timelines, but it does not establish an independent industry consensus for that date.
Source: ThorstenMeyerAI.com
