1. AI semis: the “scarcity trade” has gone sovereign.
South Korea announced a $576bn AI-chip investment drive, led by Samsung and SK Hynix, including expanded DRAM/HBM capacity and packaging infrastructure. The investor debate is whether this validates multi-year AI memory scarcity, or whether governments and suppliers are now collectively building the next overcapacity cycle. Near term this is bullish for SK Hynix, Samsung, MU, ASML, AMAT, LRCX and KLAC; longer term it raises 2028–30 supply risk for memory margins.
2. China AI silicon is moving from strategic theme to valuation bubble risk.
Baidu’s Kunlunxin is reportedly targeting a $50bn Hong Kong IPO valuation, versus c.$3bn only six months ago, with Reuters Breakingviews flagging the extreme valuation and China self-reliance angle. The bull case is clear: export controls create a protected domestic accelerator market. The bear case is equally clear: this is AI scarcity being capitalised at bubble multiples. Exposed: BIDU, Huawei ecosystem, Alibaba, Tencent, ByteDance suppliers, NVDA China exposure and Asian semi equipment.
3. The first-half leadership split is brutal: hardware has crushed software.
The Guardian notes chip and storage winners have surged in 1H26, with Sandisk, Western Digital, Micron, Seagate, Samsung and SK Hynix massively outperforming, while software names such as Microsoft have lagged as investors worry about capex intensity. The debate is whether this is rational — “own the picks and shovels” — or late-cycle crowding into the only part of AI with visible orders. Read-across: positive for MU/WDC/STX/NVDA/AVGO; pressure on MSFT, CRM, NOW, ADBE and broader SaaS.
4. Qualcomm remains the most important AI-infrastructure challenger story.
The Meta data-centre CPU win, $15bn 2029 data-centre target and $3.9bn Modular deal turn QCOM from handset recovery into a credible open-AI-stack debate. The bull case is that inference/custom silicon fragments away from Nvidia CUDA over time; the bear case is that Qualcomm is late and must still prove software ecosystem depth. Exposed: QCOM, NVDA, AVGO, MRVL, ARM, AMD and META.
5. Software/cyber: selectivity remains the message, not blanket re-rating.
Reuters’ software rebound framing still matters: investors prefer AI-integrated, usage-priced and infrastructure-adjacent names, with Datadog, Palo Alto, Synopsys, Oracle and Microsoft cited as better-positioned examples. For cyber, the debate is favourable but not automatic: AI expands attack surface and identity/runtime risk, but investors still need platform KPIs, ARR attach and renewal proof. Best exposed: PANW, CRWD, ZS, FTNT, DDOG, SNPS, ORCL and MSFT.
