1. AI semis: yesterday’s panic has flipped into “the orders are still real”.
Micron and Qualcomm have reset the AI tape after the prior semiconductor sell-off. Reuters reported Micron disclosed $22bn of multi-year customer commitments for memory chips, while Qualcomm guided to $15bn of data-centre chip sales by 2029; Nasdaq futures rose c.2% on the read-across. The investor debate is whether this proves AI capex is demand-backed rather than speculative, or whether take-or-pay contracts and large customer commitments are simply pulling forward an increasingly crowded memory/ASIC cycle. Biggest exposed: MU, QCOM, NVDA, AVGO, MRVL, AMD, SK Hynix, Samsung and ASML.
2. Memory/HBM is becoming the cleanest “AI scarcity” trade.
SK Hynix is reportedly targeting a $29.4bn Nasdaq ADR listing, while Micron’s update reinforces the idea that premium memory has moved from cyclical commodity to strategic AI bottleneck. The bull case is that HBM supply remains tight through 2027, supporting pricing and margins; the bear case is classic memory cyclicality returning once capex catches up. Second-order winners are ASML, AMAT, LRCX and KLAC, because HBM scarcity still requires more wafer, packaging and equipment intensity.
3. Qualcomm is no longer just a handset recovery story — it is trying to buy and guide its way into AI infrastructure.
Qualcomm’s $15bn 2029 data-centre target, Meta customer disclosure, and reported $3.9bn Modular acquisition put it directly into the custom AI silicon/software stack debate. The key question is whether Qualcomm can build a credible open alternative to Nvidia CUDA and compete with Broadcom/Marvell custom silicon, or whether this remains a late entrant with execution risk. Exposed: QCOM positively; AVGO/MRVL competitively; ARM, NVDA and AMD depending on how much workload shifts to custom inference.
4. Cybersecurity is getting a new AI-infrastructure angle: firewalls for data centres.
Fortinet is being framed as a beneficiary of AI data-centre buildout, with its Q1 revenue at $1.85bn, product revenue up 41%, and Q2 revenue guidance of $1.83–1.93bn. The debate is whether this is a genuine acceleration in network security demand from AI clusters, or a temporary high-end appliance refresh. The more interesting implication is that AI infrastructure is not just GPUs and power; it also raises demand for high-throughput security, segmentation, SASE and hybrid network control. Most exposed: FTNT, PANW, CHKP, CSCO, NET and ZS.
5. Software/SaaS still needs proof that AI is monetisation, not margin pressure.
The contrast with semis is stark: chip names are producing hard orders, customer commitments and multi-year revenue targets, while software is still largely arguing narrative — agent productivity, workflow redesign, governance and consumption pricing. For SaaS, the buy-side bar remains: show AI drives attach, usage, retention or margin expansion, not seat compression. Cyber remains better positioned than horizontal SaaS because AI increases attack surface, identity sprawl and governance needs. Key exposed names: PANW, CRWD, ZS, DDOG, SNOW, MDB, NOW, CRM, ADBE and PLTR.
