The memory bill, from a factory in Korea to your invoice

Five numbers already on this site, put in order. Together they follow one price rise from a trade report in Seoul to what you pay a model to answer a question.

8 September 2026 · 6 min read

The boring part of a computer became the expensive part

For thirty years memory was the part nobody argued about. You bought a computer, it came with memory, and the interesting decisions were about the processor. AI broke that. An AI processor is fast enough that it spends most of its life waiting for data to arrive, so the thing that decides how much work it can do is not the processor at all. It is the memory sitting next to it and how quickly that memory can feed it.

There are three companies in the world making that kind of memory at scale. Demand for it went up faster than any of them can build new factories, and factories take years. What follows is what that shortage looks like as it travels.

Follow it, one step at a time

Each box is a figure we have already published with its own source. The arrows are the part you have to judge for yourself, because a chain of true numbers is not the same thing as proof that one caused the next. Two of these steps are stated by the company itself, which is as close to proof as this subject gets. The others are our reading.

A worked example

Suppose you run a small support team and you send a million words a day through a frontier model. In July you paid the old rate. In September, after OpenAI raised its frontier output price by two thirds, the same work costs you a bit under double what it did, for a model doing the same job.

Nobody at OpenAI decided to charge you more because memory got expensive. The path is longer than that. Memory got scarce, so the company that makes AI processors told investors its profit margin would fall. The companies buying those processors raised their spending plans, and Amazon named memory as the reason for the whole of its increase, from about 200 billion dollars to about 220 billion. Compute became something to ration rather than something to hand out. And when compute is rationed, the price of the most capable model is the first thing to move.

What would end it

More memory factories, and better ones. Samsung has said it plans to use the newest generation of chipmaking machines for full scale memory production by 2028. That is the fix, and the date is the point: it is two years away. Nothing anyone announces this year changes what memory costs next year, because the machines have to be built, installed and tuned first.

So the honest expectation is that this stays expensive for a while, and that the cost keeps moving up the stack rather than being absorbed at the bottom.

What we do not know

We do not know how much of Korea's 209 percent export jump is price and how much is volume. The trade report gives one figure for both. We do not know what any of these companies pay for memory, because those contracts are private. And we cannot see whether a cloud provider is passing its higher costs through to customers or absorbing them to keep market share, because none of them publish that.

Those three gaps are why this piece is a chain and not a calculation. Anyone who gives you a number for how much memory adds to your AI bill is guessing.

Sources

Everything above is written from these. Each line says what that document proves.

1. motir.go.kr: Korea's Ministry of Trade, Industry and Resources, August trade report. Carries the $46.65bn and the 209.0 percent.
2. fool.com: Nvidia's Q2 FY2027 earnings call. Carries the 71 to 72 percent margin floor and the memory attribution.
3. investing.com: Amazon's Q2 2026 earnings call. Carries the $220bn and the sentence blaming the rise on memory.
4. openai.com: OpenAI's GPT-6 Astra launch post. Carries the input and output prices, up from $5 and $30 in July.
5. news.samsung.com: Samsung and ASML. Carries the plan for High NA in DRAM mass production by 2028.