Which Buyer Breaks First
Where the memory-price surge actually ends, and it isn't the household budget
Steve Williams · babnews.org
The question everyone is asking about memory prices is whether the industry can keep raising them, or whether there is a point where the market can no longer support the increases. It is a reasonable question, and it has a precise answer that most coverage misses: the market already hit that point, on one side, and the price kept rising anyway. Understanding how that is possible tells you where the surge actually ends, and it is not where you would guess.
The consumer ceiling has already been reached
Start with the data from this quarter, because it looks like relief and is the opposite. After roughly 60 percent quarter-over-quarter jumps earlier in 2026, conventional DRAM contract prices decelerated to 13-to-18 percent in Q3, and are projected to slow further to 8-to-13 percent in Q4. NAND followed a similar curve. The instinct is to read decelerating increases as the beginning of the end of the shortage.
The analysts are explicit that it is not. TrendForce states plainly that the slowdown is driven by consumer-electronics manufacturers' unwillingness and inability to absorb higher costs, not by any improvement in supply. Memory is still just as scarce. What changed is that the buyers stopped paying. That is the signature of demand destruction, the blunt economic mechanism where a price rises until it drives its own demand away, and on the consumer side it is now binding.
You can see the ceiling in behavior rather than just in the numbers. Notebook shipments dropped sharply as component costs flowed through to retail prices. Retail flash drives and memory cards went sluggish because the upstream cost can no longer be passed to consumers. PC makers accumulated inventory in the first half of the year specifically to avoid accepting another round of increases. Smartphone and notebook brands began downgrading specs and cutting their cheapest models, concentrating what volume remains among a few leading names. Every one of those is a market telling you it found the wall.
So on the consumer side, the answer to "is there a point the market can't support" is yes, and we are at it.
Yet the price keeps rising
Here is the fact that reframes the whole question. Despite the consumer market hitting its ceiling, memory prices are still climbing. Slower, but up, quarter after quarter, with real relief pushed to mid-2027 or later. The consumer ceiling did not stop the increases. It simply removed the consumer as the buyer who sets the price.
That is only possible because memory is not one market. It is two, with two entirely different buyers and two entirely different ceilings, and only one of them has broken.
The AI ceiling is far higher, and it now sets the price
While consumer memory decelerates, enterprise server memory is going the other way. A Citi research note projected 64GB DDR5 server modules climbing from around $873 in Q1 2026 to roughly $1,586 by Q4, an 80-percent-plus increase over three quarters, at the same time consumer prices were slowing. The two halves of the same industry are moving in opposite directions.
The reason they diverge is the reason the surge persists. Hyperscalers building AI infrastructure behave nothing like a PC buyer. They will pay more, they lock in supply years in advance on multi-year contracts, and they pass the memory cost straight into cloud bills that their customers expense. A consumer buying a laptop can do none of that. So when the memory makers can fill only 55-to-60 percent of total demand, as Micron disclosed, they allocate the scarce bits to whoever pays the most and commits the earliest, which is never the consumer. The AI buyer's ceiling is not affordability, because it can afford almost anything and pass it along. Its ceiling is something else entirely, which I will come to.
The result is that the AI market has become the price-setter for the entire industry. It bids up the memory that both markets draw from, and the consumer market either matches the price or loses the allocation. Since the consumer can no longer match, it loses the allocation, and the price keeps rising on the demand that remains.
How the industry "supports" increases the consumer can't
This is the mechanism that answers the original question, and it is worth stating plainly because it is uncomfortable. The industry does not support ever-higher prices by everyone paying them. It supports them by culling the buyers who cannot.
Each price increase prices out the least-able-to-pay tier: the sub-$100 phone, the budget SSD, the entry-level laptop. That demand does not pay the higher price. It exits the market. And here is the perverse part, exiting demand technically relieves the shortage at the margin, because there is now slightly less demand chasing the scarce supply, which lets the price keep climbing on the tiers that remain. The market does not hit a wall and stop. It narrows, progressively, from the bottom up, shedding its cheapest, most price-sensitive segments and concentrating on the buyers who can still pay. Fewer units, higher value, and a whole bottom rung of the market quietly deleted.
That cull is regressive, and it is the same pattern I have traced through the price of a phone: the cost of the AI buildout rolls downhill until it reaches someone who cannot pass it on, and that someone gets removed from the market rather than relieved. The affordability ceiling is real, but it does not stop the price. It just decides who is still in the room when the price goes up again.
The real ceiling is the AI capex cycle
Which brings us to where the actual ceiling sits, and it is not in memory at all. As long as hyperscalers keep spending on AI infrastructure and keep passing the memory cost into cloud bills their customers pay, the top of the memory market has room to keep rising, and it drags the entire price structure up behind it. The binding constraint on how high memory can go is therefore the durability of the AI capex cycle, not the affordability of any device.
That reframes what a "break" would even look like. The break will not come from a consumer who can no longer afford a phone; that break already happened, and the price ignored it. It will come from the AI side, and in one of two ways. Either the returns on the enormous buildout disappoint and hyperscalers slow their spending, or the financing that underwrites all those pre-committed, multi-year supply contracts gets more expensive and the commitments soften. In either case, the deep-pocketed demand holding up the top of the market cracks, and only then does the price structure have room to fall. That is precisely why the analysts tie real relief to 2027 and beyond, and specifically to supply catching up rather than to a demand wall. On the side of the market that sets the price, the demand wall is nowhere close.
The honest bounds
A few caveats keep this from being over-read. The consumer ceiling is real but not absolute; some consumer demand is inelastic enough to keep paying, and essential replacements still happen at higher prices, so the consumer market shrinks and re-prices rather than vanishing. Supply will also eventually catch up, new capacity and the natural cycle mean this is a multi-year condition, not a permanent one, and when supply arrives the dynamic reverses. And the AI-capex ceiling, while far higher, is not infinite: memory is becoming a large enough share of AI system cost that even hyperscalers have a limit, and a sharp enough macro or financing shock could find it faster than a gradual consumer squeeze ever would.
But the shape of the answer is clear, and it is more precise than "how high can it go." The consumer ceiling has been reached and it did not stop the price. The industry supports continued increases by culling demand from the bottom, not by everyone absorbing them. And the ceiling that actually matters, the one whose break would let memory prices fall, is the AI capex cycle, not the household budget.
You do not remove a bottleneck. You relocate it. The memory price ceiling moved off the consumer's affordability and onto the durability of the AI buildout. Memory does not fall until the buildout does, and the buyer whose limit finally stops the price is not the one buying a phone. It is the one building the data center.