Direct:
When Foxconn (2317.TW), Quanta (2382.TW), Inventec (2356.TW), and Wiwynn (6669.TW) assemble an AI rack, the GPU and memory bill of materials dwarfs the assembly fee. As Nvidia GPU ASPs climb and DRAM contract prices rip higher, the revenue line inflates without the gross profit line keeping pace. The percentage margin gets squeezed mechanically, even if absolute dollars per rack hold steady.
Quanta's numbers show the trajectory most clearly. Gross margin stair-stepped down from 8.58% in Q2 2024 to 7.92% in Q1 2025, 7.05% in Q2 2025, then 6.85% in Q3 2025, with management pointing to AI server mix shift as the structural pressure (the Q3 step also caught FX and geopolitical headwinds, but the secular direction is what it is). Wiwynn's Q1 2026 gross margin of 7.6% on NT$276.5 billion in revenue (up 62% YoY) was actually a step up from 7.2% in Q4 2025, so the pattern isn't uniform across every ODM and every quarter. But the underlying squeeze is the same one everyone is talking about.
Wiwynn's response is the most interesting. Starting in April, they moved certain customers to a “procurement agency model” where memory is purchased on behalf of the hyperscaler and excluded from reported revenue entirely. April revenue dropped 16% MoM as a direct accounting consequence, but rack shipment volume actually rose and margins improved. Translation: they're getting the inflated DRAM cost off their reported revenue so the percentages stop looking ugly. Wistron (3231.TW) is reportedly exploring similar sales model shifts.
It's a working capital and optics problem caused by absolutely brutal memory pricing. TrendForce initially forecast 55-60% QoQ for Q1 2026 conventional DRAM, then revised the same quarter to 90-95% QoQ in early February. Server DRAM specifically was pegged at roughly 90% QoQ, enterprise SSD at 53-58%, NAND Flash overall at 55-60%. Q2 2026 conventional DRAM is forecast another 58-63% on top. HBM is a separate conversation (its pricing is set through long-term agreements with the GPU vendors, not the spot/quarterly contract cycle), but it competes for the same wafer capacity, which is the deeper reason everything else got pulled tight. When component costs move that hard, anyone carrying inventory between purchase and assembly absorbs the working-capital load and the margin-percentage distortion, even when per-rack profitability holds.
Memory suppliers are reallocating wafers toward server DDR5 and enterprise SSDs because that's where CSPs are signing LTAs at any price. The capacity that would otherwise smooth out client SSD and DDR5 module pricing is structurally locked up through at least 2027.
Sources used:
- Wiwynn Q1 2026 financial results press release (May 7, 2026)
- Taipei Times coverage of Quanta Q2 2025 and Q3 2025 earnings (Aug 2025, Nov 2025)
- Alpha Spread Quanta Q1 2025 earnings call summary
- TrendForce, “Memory Price Outlook for 1Q26 Sharply Upgraded” (Feb 2, 2026)
- TrendForce, “AI Server Demand to Drive Memory Contract Price Increases in 2Q26” (Mar 31, 2026)
- DigiTimes, “Wiwynn shifts memory purchases as component costs pressure ODM margins” (May 9, 2026)
- DigiTimes, “Commentary: AI server ODMs face margin squeeze” (May 13, 2026)
Reddit: https://ift.tt/lxz6vTL