Memory costs more per millimetre than logic. The patent war over DDR5 and HBM

Kernel Insight, a Korean analytics firm, ran the numbers. One square millimetre of DRAM wafer in the 1b node costs 0.654 dollars. One square millimetre of wafer in TSMC's N2 process costs 0.424 dollars. Memory is the simpler product to build: no multiple patterning, no extreme ultraviolet lithography. Yet it costs 54 percent more per unit of area than the most advanced logic. The usual relationship is flipped.

The cause is structural, not technical. Demand for AI servers is draining manufacturing capacity. As the Chinese outlet 钛媒体 reports, one AI server uses 8 to 10 times more DRAM than a traditional server, and more than three times as much NAND. Samsung, SK hynix and Micron have moved more than 70 percent of their new capacity in advanced processes to higher-margin products, meaning HBM and server memory, and away from consumer DDR5. The 1b node also makes the base die of HBM3E stacks for AI accelerators.

Prices are the result, and they are not coming back to normal. The CEO of SK hynix keeps saying the AI industry does not exist without memory. Acer chairman Chen Wei forecasts that prices will keep rising until mid-2027, though he expects the shortage to be over before 2030.

On the other side sits demand that can vanish faster than it appeared. At the launch of V4.1-Flash, DeepSeek announced that its new architecture and its compression of the context cache cut HBM requirements to a quarter and SSD requirements to an eighth of earlier values. For memory suppliers that is a warning: part of current demand comes from inefficiency in the models, not from the number of queries.

Law is entering the game as well. On 23 September the US International Trade Commission opened investigation 337-TA-1523 into DRAM memory and products containing it. Netlist filed the complaint, naming Micron, Hewlett Packard Enterprise, Lenovo and Supermicro. The company has a record of wins. SK hynix paid it 40 million dollars in 2021. A jury awarded 303 million dollars against Samsung in 2023 and 118 million a year later, and 445 million from Micron in 2024. Netlist had already sued over DDR5 and HBM modules, and since July investigation 337-TA-1511 has been running, with Nvidia, Broadcom and Google among the defendants. Each of those cases is a cost that sooner or later lands in the price of memory.

The patent war is the second, less visible mechanism pushing prices up. Settlements and verdicts, counted in hundreds of millions of dollars, are booked as operating costs, and those sooner or later reach the memory price list for end customers. Combined with a tight production balance, this means the availability of modules depends not only on demand from data centres but also on legal risk. For a buyer of AI servers that is one more argument to sign long-term contracts and secure supply before prices climb again.