Direct:
AI-assisted parallel reporting
Nvidia's HBM allocation is now the clearest fault line between SK Hynix and Samsung, and both Korean majors keep sinking fresh money into aging China NAND fabs they know could lose Western tool access at any annual license renewal. SK Hynix also went public with more detail on the CPO roadmap this run touched on last night.
DRAM / HBM / HBF
Nvidia's memory orders are pulling SK Hynix and Samsung apart
Nvidia's AI memory demand now explains most of the gap between the two Korean memory makers. SK Hynix pulled KRW17.6 trillion from Nvidia in the first half of 2026, making Nvidia its single largest customer at 13.35% of revenue, down from roughly 24% of full-year 2025 revenue as SK Hynix diversifies toward Big Tech's in-house ASIC programs. Samsung's numbers run the other way: Nvidia doesn't appear anywhere in Samsung's top five customers for the period, which instead reads Alphabet, Amazon, Apple, Hong Kong Techtronics, and Supreme Electronics, together good for about a quarter of Samsung's revenue. The split tracks the DRAM share data too, with Samsung holding roughly 38.5% of the market and a gap to SK Hynix that BigGo Finance pegs at 9.7 percentage points. It isn't necessarily locked in: SK Hynix's Nvidia share could recover in the second half as HBM4 ramps alongside Nvidia's Vera Rubin platform, but right now Samsung is winning volume everywhere except the one customer relationship that sets AI-era margin.
*Sources: DigiTimes (Aug 25) · TrendForce (Aug 17) · BigGo Finance
Flash & SSDs
Samsung and SK Hynix double down on aging China NAND fabs
Samsung and SK Hynix are betting their older China NAND lines are worth defending, even as Washington's export rules get less predictable with each annual license cycle. SK Hynix restarted construction and equipment installation at Dalian Fab 2 in August, targeting first-half 2027 mass production and roughly a 50% increase in China NAND capacity; the company's combined 2026 investment in its Wuxi DRAM and Dalian NAND plants topped 1 trillion won, with Dalian spending alone up 52% year over year. Samsung has already finished converting its Xi'an fab from 128-layer V6 to 236-layer V8 NAND and is now laying groundwork for 286-layer V9, targeting full operation in 2027. Korea JoongAng Daily's framing is blunter than DigiTimes': both companies are pouring fresh money into factories they know could face tighter US tool restrictions at any renewal, because walking away from existing China capacity now costs more than the risk of a future clampdown.
*Sources: DigiTimes (Aug 25) · TrendForce (Aug 12) · Korea JoongAng Daily
Optics & AI infrastructure
Update: SK Hynix's CPO roadmap gets sharper, and DigiTimes now frames it against Nvidia directly
Last night's run flagged SK Hynix's Nature Electronics CPO paper only in passing, as background to its Hot Chips packaging disclosures. DigiTimes has since run a dedicated piece with more detail worth logging: the paper, co-authored with University of Virginia researchers and published Aug 20, sets specific targets, over 100 Tb/s of bandwidth per node, sub-1 pJ/bit energy efficiency, and chip-to-chip latency under 10 nanoseconds, reached through a progression from 2D packaging to 2.5D interposers to full 3D heterogeneous integration. The endpoint, a photonic interposer linking memory directly to processors so multiple AI accelerators can share one large memory pool, is explicitly a post-2030 target. The new angle is DigiTimes' framing that this puts SK Hynix on a different track than Nvidia, whose own CPO push centers nearer-term on switch-side pluggable optics rather than memory-side integration. Worth repeating: this is a research roadmap, not a product announcement, so treat the timeline as SK Hynix's stated ambition rather than a committed schedule.
*Sources: DigiTimes (Aug 25) · TrendForce (Aug 20)
Quick hits
- SK Hynix and Alibaba are reportedly working the “value per chip” problem together, as complex chip configurations bump against the physical and economic limits of scaling, though DigiTimes gives few specifics behind the paywall: DigiTimes (Aug 24)
- Zhongji Innolight's 1.6T transceiver surge is DigiTimes' latest evidence that optical networking, not just GPUs, memory, and power, is becoming the next hyperscaler bottleneck: DigiTimes (Aug 25)
- DigiTimes' own Research Insight desk expects the three largest upstream memory makers to more than triple their combined 2026 DRAM and NAND revenue, with pricing staying firm until 2027 capacity arrives. Four days old now and not logged in last night's run, but it's the macro backdrop the supplier-specific stories above sit inside: DigiTimes (Aug 21)
Reddit: https://ift.tt/gEchTQW