The Refresh Replaced the Roadmap — The Week Memory Backlog Pushed Non-AI Silicon Off the 2027 Calendar

The quietest yet most consequential story of the week didn't come from the US — it came from Taipei. On June 22, Technews Taiwan reported that Qualcomm is shelving its next-generation Snapdragon X3 PC SoC launch and replacing it with a "refresh" of the current X2 line (Kalambo / Mahua / Glymur Refresh). The stated reason: memory supply crunch. Read as a single-product slip, this is a footnote. Read alongside the events that hit Korea, Taiwan, Japan and the US in the same week, it is the first visible signal that AI memory and packaging allocation has begun crowding non-AI silicon off the 2027 roadmap.

Korea — the supply side gets locked up

On June 22, SK Hynix's market cap surpassed Samsung Electronics for the first time ever (KRW 2,080T vs 2,051T, +5.61% close). The number that matters isn't the share price — it's Hanwha Investment & Securities raising its SK Hynix target to KRW 4.3M the same day, with the explicit reasoning that "long-term agreements (LTAs) smooth earnings volatility." Translation: hyperscalers are pre-buying 2027–2029 HBM volume. Korea's May semi exports hit $29.4B (+154% YoY), April +158%, March +138%. The first 20 days of June printed a record $62B. DDR5 16Gb spot reached $46.17, and the industry now sees a memory supercycle running through 2027 with another ~15% contract uplift implied. Reporting on Samsung's expanding long-term HBM deals confirms the same structure on the second-largest supplier. The point is structural — what Korean memory makers "have left to sell" is already pre-allocated.

US — demand silos harden

The same week, NVIDIA began shipping its next-generation Vera CPU systems to OpenAI, Anthropic, Oracle and xAI. Foxconn pegged a single 1GW Vera Rubin-based AI datacenter at ~$47B capex with $1.3B annual power — one system pulls tens of times the HBM of a standard server rack. AMD and Korean startup MangoBoost rolled out a heterogeneous-compute stack claiming up to 90% AI server TCO reduction — but the stack still consumes HBM. Intel telegraphed an emerging CPU shortage as its turnaround completes, while the Intel-NVIDIA co-developed PC processor was pushed one more notch out to CES 2028. When PC and inference compete for the same memory pool, the loser is consistently PC. Qualcomm's X3 shelving is the first hard evidence.

Taiwan — same fab, different priority

Taiwan's TAIEX exploded +2.75% / 1,276 points to a record 47,741.51 on June 22. TSMC hit NT$2,510 (market cap NT$65T / ~US$2.1T). The same day, TSMC installed its first CoPoS panel-level advanced-packaging pilot at Chipbond's Longtan fab; ASE locked limit-up to a NT$3T market cap. ABF substrate maker Unimicron (3037) announced ~NT$14B in new capex to scale Vera Rubin substrate production, targeting 45% share. MediaTek separately won the exclusive contract for Google's TPU v9 'Triggerfish' inference chip (per Ming-Chi Kuo). In other words, TSMC's advanced-node, CoWoS and CoPoS lines are being preferentially allocated to Vera Rubin, HBM4 stacks and Google TPU silicon — bumping PC-class SoCs that share the same fabs and packaging lines further back in the queue. If memory shortage is the first mechanism behind Qualcomm's X3 pause, packaging-capacity lock-up is the second.