AI demand is pushing server memory towards a tighter supply picture, with TrendForce forecasting server DRAM supply growth of about 25% in 2027 against server CPU shipment growth above 30%. The report also puts that pressure in perspective: a headline comparison between DRAM and TSMC wafer prices uses figures calculated on different bases, so it is not a like-for-like measure of profitability.
Watch Desk analysis
What happened
TrendForce analyst Hsu Chia-yuan said at an industry forecast seminar that server DRAM supply is struggling to keep pace with demand. BigGo Finance reports his forecast that server CPU shipments will grow by more than 30% year on year in 2027, while server DRAM bit supply grows by about 25%. TrendForce expects the supply squeeze to persist through 2027, with new wafer fabs not expected to come online before 2028.
The same report says server uses now account for more than 60% of global DRAM bits and over 70% of wafer starts. Hsu linked the surge to AI infrastructure, including agentic AI workloads that require more general-purpose servers, CPUs and memory alongside GPUs and high-bandwidth memory.
Why it matters
Memory is not a supporting actor when AI systems need more servers to train and run. If demand grows faster than supply, operators may face tighter availability and higher costs, while manufacturers’ choices about which products to make could squeeze other markets. Hsu also described memory prices influencing chip design, with more on-chip SRAM becoming an attractive alternative to external registered memory in some applications.
The report compares an estimated US$80,000 of revenue per DRAM wafer with TSMC’s listed prices for advanced foundry wafers. Those figures are calculated differently: one is derived from memory-module contract prices, the others are foundry list prices. The comparison is not evidence that DRAM makers are more profitable than foundries.
Our read
The useful signal is the forecast supply gap, not the eye-catching wafer-price comparison. It suggests AI infrastructure demand is reaching beyond accelerators into the less glamorous parts of the server bill, where supply cannot be conjured up by adding a line to a slide deck. These are forecasts reported by BigGo, not a guarantee of prices or shortages across every type of memory.
What to watch
- Whether server DRAM supply growth falls short of demand as forecast for 2027.
- Whether memory tightness feeds into AI server shipments, prices or product configurations.
- Whether manufacturers add capacity or redirect production between server memory and other markets.
Discussion spark: If AI server demand keeps outrunning memory supply, should manufacturers prioritise server memory, or protect supply for other markets even if that slows AI infrastructure growth?
Sources and evidence
- DRAM Wafer Revenue Tops $80,000, Outshining TSMC's 2nm Foundry Price – BigGo Finance (5 November 2026, 10:35 UTC)
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.