Why RAM Prices Won't Come Down Anytime Soon

I've covered enough technology cycles to recognize when the market has fundamentally shifted. The current RAM pricing situation isn't a temporary squeeze that will correct itself once supply catches up. It's a permanent realignment of priorities in the hardware industry, and consumers are on the wrong side of it.

Luke Lafreniere, the technical voice behind Linus Tech Tips, laid out the mechanics of this shift in a recent discussion. His analysis confirmed what stock investors figured out months ago—when Micron's share price tripled on AI demand, the writing was on the wall.

The Enterprise Migration

The consumer hardware market is becoming a side business. Luke pointed to Crucial's exit from the RAM market as unsurprising: "If you look at how DDR5 was even designed, it was not designed from a consumer standpoint. It was designed from an enterprise standpoint. The goal has been enterprise for a long time."

This isn't speculation. Intel recently announced it had been paying too much attention to consumers and would refocus on hyperscalers. Nvidia discusses GeForce cards with what Luke describes as grudging obligation. Even small water cooling companies that seemed like scrappy consumer brands derive most of their revenue from server infrastructure.

The economics are brutal. Enterprise customers buy hardware by the pallet. They negotiate deals measured in gigawatts of power consumption—literally nuclear power plant capacity. A consumer with a problem starts a Reddit thread with 7,000 upvotes. A hyperscaler with a problem shrugs and orders another shipment.

The Collusion Problem

Luke didn't mince words about industry behavior: "Some of the most collusion in effectively any market anywhere has been in RAM. These companies love working together to price fix. They love working together to keep prices high, to not increase fab capacity too much."

This isn't conspiracy theory. The legal record is extensive. RAM manufacturers have paid billions in settlements for coordinated price manipulation. The pattern is well-documented: restrict fab capacity, coordinate on pricing, maximize profit. Luke notes it's "not even a secret at this point, let alone an open secret."

New fabrication facilities take years to come online, even when companies decide to build them. Right now, there's enough AI-driven demand that manufacturers have no incentive to flood the market. One partnership in Taiwan is working to increase capacity, but the timeline for consumer relief is measured in years, not quarters.

The DDR5 Design Problem

Here's where it gets technical in ways that matter. DDR5 wasn't engineered for the airflow conditions inside consumer desktop cases. Luke explains that server environments don't care about fan noise—they prioritize performance and can push massive airflow across components. Consumer cases with tempered glass panels and aesthetic priorities don't provide the same cooling.

The result? DDR5 modules are overheating in consumer systems. Tech communities are designing 3D-printed fan shrouds to mount 80mm fans directly over RAM sticks—solutions reminiscent of the DDR3 era's Corsair Dominator cooling brackets. Wendell from Level One Techs has documented these thermal issues extensively.

Linus Torvalds offered an even more unsettling observation during a recent appearance: many Windows stability problems might actually be RAM errors. Overheating memory performs error correction poorly, leading to application crashes, blue screens, and system instability that users blame on software.

The AI Calculation

The numbers backing AI infrastructure are staggering. OpenAI's agreements with partners reportedly include commitments measured in gigawatts—one nuclear power plant's output equals roughly one gigawatt. The combined deals with Nvidia and Oracle represent something like 16 nuclear power plants worth of energy consumption. The United States currently operates 39.

Those data centers need RAM. Lots of it. OpenAI alone has reportedly committed to purchasing up to 900,000 silicon wafers monthly. A typical modern data center consumes 100-200 megawatts. AI data centers are running 10x that—over one gigawatt for a single facility.

When you're building infrastructure at that scale, consumer memory pricing becomes irrelevant. The enterprise market can absorb higher costs because the alternative—capacity constraints on AI training—costs more.

What Doesn't Change

Luke's timeline assessment: "I don't think demand and price is just going to randomly come down. I think for the last few years we've been seeing a general shift towards enterprise, and when consumer becomes more of an afterthought, consumer gear is going to be more price-performance expensive for what you're getting."

The stock market already priced this in. Micron and Western Digital are up 300% because investors understand the demand curve. Memory manufacturers have no incentive to prioritize a market segment that generates less revenue, demands more support, and complains louder.

Could AI investment slow? Possibly. Luke notes the companies driving this expansion—Microsoft, Google, Amazon—remain hugely profitable regardless of AI outcomes. They can sustain this buildout longer than speculative players could.

The transition feels familiar because I've watched it happen before in other hardware categories. When a technology becomes infrastructure rather than product, consumer needs stop driving design decisions. We're seeing that shift now in memory, and the adjustment period will be uncomfortable for anyone building or upgrading a desktop PC.

— Bob Reynolds