Memory: DRAM 1b costs 54% more per square millimetre than TSMC's N2

Making DRAM ought to be simpler than making a leading-edge processor, and therefore cheaper. The figures from the South Korean firm Kernel Insight, reported by the Chinese site IT之家, say the opposite. A DRAM 1b wafer costs 0.654 dollar per square millimetre. A wafer on TSMC's N2 process costs 0.424 dollar. That is a gap of 54%.

Two technologies, two levels of complexity

The 1b node is a generation of DRAM below 10 nanometres, with a density of around 0.436 gigabit per square millimetre. These wafers go into DDR5 modules: desktop sticks of 16 to 64 gigabytes, sometimes 128 or even 512, memory for laptops, and RDIMM and LRDIMM server modules of up to 256 gigabytes per stick. Machines built on them can reach 12 terabytes of memory. Part of this output also feeds the base dies for the stacked memory used in AI accelerators.

TSMC's N2 process requires double exposure and EUV lithography. In theory, a simpler product should cost less to make. The opposite is happening, and the cause is cyclical: the general shortage of RAM driven by the build-out of AI servers.

The effect of AI demand

Large AI server projects absorb both high-bandwidth memory and high-capacity DDR5 modules. Capacity expansion at the three big memory makers lags behind demand. The head of SK Hynix has said publicly that the AI industry cannot exist without memory, and that his group is speeding up the rollout of its next-generation production. The president of Acer, quoted by IT之家, expects the shortage not to last until 2030, even though prices should keep climbing before easing in the middle of 2027.

This price effect runs through the whole chain. In Italy, industry observers note that AI is sending chip costs soaring and that the bill eventually lands with businesses and consumers alike.

A component that is simpler to manufacture can cost more: when AI demand saturates capacity, price no longer depends on technical difficulty but on scarcity.