Roots · Memory supply chain · 9 min read
The RAM crisis, explained
The memory in your laptop is not competing with another laptop. It is competing with an AI data center. That sounds abstract until the price of a normal DDR5 kit doubles, a phone stops getting a RAM upgrade, or a cheap PC simply disappears from the shelf.
This is not just a temporary shortage. AI servers are pulling supply toward high-bandwidth memory (HBM) and server DRAM, while new fabs take years to build. The result is a structural squeeze: less ordinary memory available, more expensive devices, and a market where enterprise buyers can outbid consumers.
Memory became the bottleneck
An AI accelerator is not useful by itself. It needs enormous amounts of data delivered quickly, so the accelerator is paired with HBM: stacked memory placed close to the compute package. Chatbots, image generators, recommendation systems, and training clusters all turn memory bandwidth into a core infrastructure requirement.
The ordinary memory market was already concentrated. Samsung, SK hynix, and Micron make the large majority of DRAM. When those companies direct more cleanroom space and investment toward higher-margin server products, PC and phone makers cannot instantly find a replacement supplier.
The HBM squeeze is a capacity problem
HBM does not simply use the same memory in a different box. It requires advanced manufacturing, stacking, packaging, and testing. Industry estimates often describe HBM as consuming several times more wafer area per useful bit than commodity DRAM. The exact multiple depends on the generation and the accounting method; the direction is not controversial: one wafer moved into HBM is a wafer not making ordinary memory.
Gartner's projected combined DRAM and SSD cost increase by the end of 2026, versus early 2025.
A widely cited 2026 estimate for the share of memory output routed to AI data centers and hyperscale infrastructure.
The rough window analysts give for new capacity to catch up, not a promise that prices normalize then.
These are estimates and forecasts, not a single official scoreboard. We use them to show the shape of the problem, not to pretend the future is precisely known.
Who gets the memory first?
The answer is usually the customer with the largest order, the longest contract, and the most expensive product. Hyperscalers can commit years ahead for AI clusters. A PC builder buying memory for a low-margin $500 machine has much less leverage. That is allocation, not a conspiracy: suppliers are following the economics of a market with only a few major producers.
Micron’s December 2025 announcement that it would exit the Crucial consumer business made the trade-off visible. Micron said the move would improve supply and support for its larger strategic customers. The company did not say that every consumer module disappeared overnight. It did show where the priority sits.
Why the price moves so fast
Memory is a commodity market. When a manufacturer sells the same basic type of chip to many buyers, a relatively small supply change can move the market price for everyone. Buyers then compete for inventory, distributors reprice kits, and retailers pass along the increase.
That is why a headline like “RAM is 400% more expensive” needs a follow-up question. Is it a spot price, a contract price, one capacity, one generation, or one unusually scarce module? A large increase can be real without applying to every stick of RAM. The defensible takeaway is simpler: ordinary memory has become materially more expensive and more volatile.
What it means for your devices
PCs feel the squeeze in the bill of materials. Memory can become one of the largest component costs, especially when an “AI PC” is expected to ship with 16GB or 32GB as a baseline. Manufacturers have three choices: raise the price, reduce another component, or accept a smaller margin.
Phones face the same math. More RAM helps on-device AI and multitasking, but the memory cost is harder to hide in a mid-range phone. A flagship may hold at 12GB instead of moving to 16GB. A budget model may keep its specification and become more expensive, or keep its price and ship with less elsewhere.
Storage is connected too. NAND flash is a different memory technology, but it shares the same broad semiconductor investment cycle and is also being pulled by data-center demand. SSDs can rise alongside DRAM even though an SSD is not “made of RAM.”
What you can do about it
- If your current machine works, keep it. A shortage is a reason to avoid an unnecessary upgrade, not a reason to panic-buy.
- If you need a device, compare the complete configuration. A cheap laptop with soldered 8GB RAM may be a worse long-term value than a slightly pricier upgradeable model.
- Treat capacity as a workload decision. More RAM helps when you run many apps, large projects, virtual machines, or local AI models; it does not make every task faster.
- Check the actual memory type and configuration. DDR4, DDR5, LPDDR5X, HBM, and NAND are not interchangeable products.
How to read the claims
Watch for precise-looking numbers without a defined baseline: “400% higher,” “70% of supply,” or “three times more wafer.” Ask what period, product, region, and measurement the number covers. Separate facts already observed from forecasts. A company announcement can tell you what a company chose to do; it cannot prove the whole market will behave the same way.
Also keep the scale in view. “AI uses most of the new capacity” does not mean AI consumes most of every memory chip already installed in every device. The shortage is about marginal supply, allocation, and price. Those are enough to hurt without turning every dramatic claim into a literal one.
A useful warning label
Forecasts for 2026–2030 move quickly, and some figures circulating online combine analyst estimates, trade reporting, and unattributed supply-chain claims. For this explainer, the most concrete anchor is Micron’s own Crucial announcement. Market context comes from TrendForce’s DRAM research coverage and reporting that attributes the 2026 price and allocation estimates to Gartner. Check the original release or research note before treating any forecast as settled fact.
The point is not to memorize a percentage. It is to notice who is buying capacity, what kind of memory they need, and who is left negotiating after them.