AI demand has made memory the unlikely star of chip economics

AI’s appetite for server memory has pushed the estimated value of DRAM per square millimetre above the nominal wafer price of TSMC’s leading-edge 2nm process, according to analysis from Kurnal Insight reported by BigGo Finance. The comparison is imperfect, but it captures a real shift: in the AI hardware race, memory is no longer merely the supporting cast beside the glamorous compute chip.

Watch Desk analysis

What happened

Kurnal Insight estimates that 1b-process DRAM is worth about $0.654 per square millimetre, using a memory price of roughly $1.50 per gigabit. BigGo Finance compares that with estimated wafer-equivalent prices of about $0.424 per square millimetre for TSMC’s 2nm process and $0.283 for 3nm.

The analysis links the change to surging demand for AI servers, particularly high-bandwidth memory and server DDR5. BigGo says spot prices for 16Gb DDR5 eTT chips reached $24.80 on 21 September, or around $1.55 per gigabit, close to Kurnal Insight’s assumption.

Why it matters

AI accelerators may get the headlines, but they cannot do much without fast memory to feed them. Higher DRAM prices and tight supply could therefore raise the cost of servers even when the headline processor is unchanged. The pressure reaches beyond GPUs, into packaging, controllers, networking and the wider infrastructure needed to keep data centres fed.

The figures also hint at why memory makers are enjoying an unusually strong moment. BigGo reports that analysts expect the gross margins of Samsung Electronics, SK Hynix and Micron to exceed 80%, with some forecasts approaching 90%. Those are reported estimates, not audited conclusions, and the comparison with TSMC is not an apples-to-apples margin contest.

Our read

This is a useful reminder that AI infrastructure is an ecosystem, not a procession of ever-fancier accelerators. Memory has become expensive because demand is intense, and the supposedly humble component is now helping set the economics of the whole machine. Silicon, it turns out, has a supporting cast with excellent timing.

The sharp limitation is that Kurnal Insight compares the potential selling value of finished DRAM with TSMC’s wafer-processing fee. The calculation excludes yield loss, packaging, dicing, defects and customer-specific contracts. It also relies partly on spot prices, while major memory suppliers earn much of their revenue through longer-term agreements. Treat this as a market signal, not proof that DRAM is simply more profitable than advanced logic.

What to watch

• Contract prices
  Whether longer-term DDR5 and HBM prices follow the reported spot-market strength.
• AI server bills
  Whether memory inflation begins showing up clearly in complete-system pricing.
• HBM supply
  Whether new capacity eases shortages without triggering another boom-and-bust cycle.
• Margins
  Whether the reported gains survive once supply expands and AI customers push back on prices.

Discussion spark: Should AI hardware buyers accept higher memory prices as the unavoidable cost of expansion, or should chip designers prioritise architectures that use less of it?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.