AI交易风险焦点：过度工程化复杂性成热议话题

IndustryTech Hardware Q2-26 preview: Carry AI bear spray Robert Sanders AI trade risks in focus with over-engineering complexity the topic du jour Research Analyst+44-20-754-58394 There are so many potential AI bear cases that can be seized on at any one time,whichgives bears ample material.On the demand side,these includesmonetization/ROI including from sluggish enterprise adoption, open-source modelproliferation, and hardware obsolescence. On the supply side, there are risks ofdouble ordering amidst severe component shortages. Related physical risksinclude power grid constraints. On the financial side, core risks include hyperscalerFCF exhaustion and related funding risks, private debt and asset-backed over-leverage, and a potential neocloud pricing war in the future. Lastly, there arestrategic/technology risks include over-engineering risks, HBM being replaced byLPDDR6, hyperscalers reverting back to their core "swimming lanes" and rentingout rather than building, or conversely some powerful companies (eg Google) whomay choose to flood the market. Looming over all of this are geopolitical risks.Recently, it is over-engineering that has entered the limelight given that Nvidiaappears to be facing risks around the Kyber NVL144 rack relating to the 78-layerPCB mid-plane (1.5m!) on top of CPO delays, both relating to thermals and signalintegrity. Google/Broadcom too appears to be having issues with its TPU v9 chipset. Johannes SchallerAssociate Director of Equity Research - Germany+49-69-910-31731 George BrownResearch Associate+44-20-754-74893 Michael Kuhn Research Analyst+49-69-910-46002 Yet there remains grounds for optimism in the medium-term Despite hardware development challenges borne of complexity, we still seereasons for optimism in the medium-term, seeing several pushbacks as validagainst the bear cases. For example, those worrying about AI monetisation mustacknowledge AI efficiency gains and Anthropic's rapid progress (finance +software/internet now half of its ARR?), those worrying about the grid mustacknowledge BTM power. As for the financial side, we would argue the stakes aretoo high for the hyperscalers' core profit centers to dial back capex growthsignificantly, even if they are keen to drive down capex to below $20bn/GW (vsNvidia at $40bn/GW and rising). In short, hyperscaler cash burn could persist, evenif the peculiar tax of persistent memory margin stacking is at risk. Overall then, withthe chip industry constrained and with supply air gaps from Nvidia delays not likelyto alter the direction of travel with Google filling the gap, we remain constructive.Demand destruction though in non-AI categories from memory cost and poorsupply though merits more attention, given that this could bring nasty surprises. We value stocks on a P/E, EV/EBIT(DA) andDCF basis, benchmarking multiples vs.peers and/or expected growth rates. Keyrisks include macro, FX, and share shifts. Semicaps: Prefer sub-suppliers over OEMs into numbers After a period of lagging from February to April, EU semicaps moved up in May and 10 July 2026SemiconductorsTech Hardware This report is being provided solely on thebasis that it, in whole or in part, will not beused in any manner contrary to laws andregulations of the United States, includingbut not limited to Executive Order 14032,by a United States citizen, a permanentresident alien, an entity organized underthe laws of the United States or anyjurisdiction within the United States(including foreign branches), or any personin the United States. June helped by greater optimism on more upside to semi capex in 2027/28(supported notably by Intel) and greater willingness to value on 2028E multiples.However, the hard realities of clean room constraints (eg Intel Ohio not ready until2030) mean that there is a danger that WFE expectations step ahead of physicalreality (we think they already have for 2027). We prefer sub-suppliers as we seegreater scope for pricing and volume upside as OEMs build up stocks. As for theOEMs, we see ASML as likely to accelerate its suppliers to scale up its low NA EUVcapacity to more than 100 low NA tools during 2028. However, we suspect ASMLmay prefer to wait until the Q3 results before providing an update on EUV capacity,at which point we expect some moderation on high NA expectations. Our target onASML is increased to €1,800 (from €1,600) with our 2028E EPS of ~€60 nowbroadly in line with likely buyside views. As for Besi, unfavourable developments onHB adoption in HBM are credible, but likely to be offset by resilient 2.5D-leddemand. For the sub-suppliers, we favour Comet and Inficon. Broad-based semis: We still like Infineon and ST. Soitec still wins from NPO. In terms of broad-based semis, availability continues to trump product quality. Asa result, those companies not facing allocation challenges look set to enjoy greatergross margin improvement, while the pricing backdrop remains positive for mostas lead times continue to stretch