The Amnesia Period

We spent three years building machines that can do almost anything except remember. Then, to fix that, we started buying up all the memory.

The first amnesia is the one everyone talks about. Today's most capable AI systems have no persistent memory. The skillset is fully intact, the reasoning is real, and none of it carries from one session to the next. Think of Jason Bourne pulled from the water: elite, fluent, deadly, and unable to tell you who he is. Capability arrived before continuity, and the whole industry is now spending enormous sums trying to bolt memory onto systems that were built without it.

But there is a second amnesia, and it is the more literal one. The cure for the first, feeding these models the memory they lack, has created an appetite for physical memory so large that it is now starving everything else. The machines that cannot remember are eating the memory that the rest of the electronics industry runs on. That is the story worth telling, because it is already showing up in prices, in delayed launches, and in the margins of businesses three rungs removed from any data center.

It is a reallocation, not a shortage

The word "shortage" makes it sound like demand simply outran supply and everyone waits for the fabs to catch up. The mechanism is sharper than that, and understanding it changes what you watch.

Inside every DRAM fab, the same wafer can become high-bandwidth memory bound for an AI accelerator or commodity DDR5 bound for a laptop. In 2026 the makers are overwhelmingly choosing the first, and the reason is margin, not malice: HBM commands several times the profit of conventional DRAM. The catch is the conversion rate. HBM carries roughly a three-to-one consumption ratio against DDR5, so every wafer redirected to high-bandwidth memory removes about three times as much equivalent commodity memory from everyone else's market.

You can see the reallocation in a single pair of numbers. Total DRAM wafer capacity is growing around 14% this year. Commodity DRAM capacity is growing about 10%. The difference is not a supply failure. It is a decision, repeated across three companies, about who gets served first.

So the laptop maker and the appliance maker were not outbid. They were deprioritized by design. This is the old line doing new work: you do not add memory, you relocate who gets it.

The flattered number sits on the supply side

Every celebrated figure in this story belongs to the winners. Samsung, SK Hynix, and Micron are posting record memory margins. Late last year those margins ran ahead of TSMC's foundry business, which is a sentence that would have read as a typo a few years ago. SK Hynix's AI-memory revenue has more than tripled since 2024. HBM capacity for all of 2026 sold out before 2026 began.

Read those numbers as the whole story and you would conclude the industry is thriving. Anchor there and you miss the other side of the ledger entirely. The health of three suppliers is being measured, dollar for dollar, in the condition of a thousand buyers downstream. And the buyers do not report on the same page, so the pain is real but diffuse, and easy to leave out of the headline.

The oligopoly is why this lasts. With only three makers setting allocation, there is no fourth competitor with an incentive to divert capacity back to low-margin commodity parts to win share. When all three tout record margins from the same choice, the old cyclical equilibrium is gone. This is not a normal upcycle that corrects when supply catches up. The demand driver is structural and the incentive to serve it is permanent, which is exactly why every major supplier now warns the imbalance runs to 2027 and beyond.

The tax lands unevenly

Start with the prices, because they are the visible part. DRAM contract prices rose on the order of 90 to 110% quarter over quarter in early 2026, climbed another roughly 60% the next quarter, and are still rising double digits into the third. The recent cool-down is not relief. Suppliers are explicit that it reflects consumer manufacturers hitting the ceiling of what they can absorb, not any easing of supply. Cumulatively, DRAM and NAND have moved in multiples, not percentages.

The tax then lands by two variables: how much memory a product carries, and how little room it has to pass the cost along.

Memory-heavy products feel it first. Gaming hardware carries more DRAM and faster storage than an average laptop, so it led. The PlayStation 5 sits at $649 after a second increase, and the Xbox Series X reached about $800 from a $500 launch. HP told investors that memory now accounts for around 35% of a PC's bill of materials, roughly double where it sat a quarter earlier. When a single input doubles its share of your cost structure in ninety days, nothing about your pricing survives intact.

Price-sensitive products feel it worst. IDC projects the global smartphone market shrinks nearly 13% in 2026, the sharpest drop on record, with PCs down around 5%. Gartner expects the sub-$500 laptop to become financially unviable within a couple of years. A device that cannot raise its price without losing its only advantage, and cannot cut its memory without losing its function, has nowhere to put the cost. It simply becomes a worse business, or stops being one.

The integrated giants weather it. Apple and Samsung have pricing power, product mix, and in some cases their own fabs, so they absorb and reprice and move on. The single-product firm and the thin-margin firm do not have those levers. Same shortage, opposite outcomes, sorted almost entirely by where a company sits on the chain. E Ink is the cleanest illustration: it cut its 2026 growth guidance as rising memory prices softened demand for e-readers, with its consumer segment down double digits while its shelf-label business held. One company, one shortage, split down the middle by which customers could pass the cost on.

The cost you cannot see is erasure

Here is the part that appears in no price index, and it is the sharpest cost of the three.

The deepest damage from the memory shortage is not the products that got more expensive. It is the products that never shipped. Memory-module makers pulled planned launches into the following year to wait out pricing. Valve delayed its next Steam Machine outright and still has no confirmed price or date for its VR headset, both tied to sourcing. Sony and Nintendo have warned that tighter component supply could delay future releases. Even a photomask supplier like Photronics, several rungs upstream of any consumer, told investors its revenue took a hit from delayed chip design releases linked partly to memory-supply allocation, and watched gross margin fall from the high 30s to the low 30s year over year.

None of that shows up as a line item. It shows up as a spec that quietly dropped from 16GB to 8GB, a price tier that disappeared from the catalog, a launch that slid a year and took its roadmap with it. The memory shortage does not just tax products. It deletes them.

Which is the second meaning of amnesia, and the more exact one. The machines forget between sessions. The market is now being made to forget alongside them: forget the budget laptop, forget the on-time launch, forget the spec that was printed on the box.

What the period actually is

So the period has a shape. We built a technology that cannot hold a memory, and its hunger for memory is erasing capability everywhere the cost cannot be passed along. The winners are three companies. The payers are most of the rest of the electronics industry, and they are paying in three currencies at once: higher prices, thinner margins, and the products they will never get to make.

If you run a downstream business, the useful question is not whether memory got expensive. Everyone knows it did. The question is which kind of payer you are. Can you pass the cost on, or is your memory line the quiet thing that deletes your next product before it reaches a shelf?

You do not remove a bottleneck. You relocate it. This one relocated the forgetting from the machine to the market.