Memory is eating the data center. UBS breakdown on Vera Rubin and conventional server economics shows a structural cost shift that changes everything about infrastructure economics and supply-chain leverage.
Conventional servers first: memory now 69% of BOM vs 40% a year ago. Total server cost up 127% to $26k, driven almost entirely by DRAM (+315% YoY) and NAND (+227%). Memory isn't a component anymore — it's the product.
AI servers take this further. $NVDA GB300 superchip: $9.7k in HBM3E and LPDDR5X alone, 53-58% of cost. That's $350k memory per rack. Vera Rubin VR200: $24.3k memory per chip in 2026 (62% of cost), rising to 70% in 2027. Per-rack memory: $875k → $1.14m. VR300 in 2027: $35.5k memory per chip, 77% of total cost, $1.27m per rack.
The milestone: memory costs break $1 million per rack in 2027. Not total rack cost — just memory. This isn't incremental. It's architectural. Memory suppliers ($MU, $DRAM, Korean plays via $EWY) are no longer selling into the stack — they are the stack.
Implication: margin power, pricing leverage, and capex allocation all tilt toward whoever controls high-bandwidth memory supply. The old 'servers are commoditized' thesis is dead. The new thesis: memory is the choke point, and choke points set terms.
This was May data. Contracts since then suggest the trend is accelerating, not plateauing.
Conventional servers first: memory now 69% of BOM vs 40% a year ago. Total server cost up 127% to $26k, driven almost entirely by DRAM (+315% YoY) and NAND (+227%). Memory isn't a component anymore — it's the product.
AI servers take this further. $NVDA GB300 superchip: $9.7k in HBM3E and LPDDR5X alone, 53-58% of cost. That's $350k memory per rack. Vera Rubin VR200: $24.3k memory per chip in 2026 (62% of cost), rising to 70% in 2027. Per-rack memory: $875k → $1.14m. VR300 in 2027: $35.5k memory per chip, 77% of total cost, $1.27m per rack.
The milestone: memory costs break $1 million per rack in 2027. Not total rack cost — just memory. This isn't incremental. It's architectural. Memory suppliers ($MU, $DRAM, Korean plays via $EWY) are no longer selling into the stack — they are the stack.
Implication: margin power, pricing leverage, and capex allocation all tilt toward whoever controls high-bandwidth memory supply. The old 'servers are commoditized' thesis is dead. The new thesis: memory is the choke point, and choke points set terms.
This was May data. Contracts since then suggest the trend is accelerating, not plateauing.