1. Semiconductors have entered a technical bear market despite exceptional operating data — the debate has decisively shifted from demand to duration, valuation and capital-cycle risk.
The Philadelphia Semiconductor Index is now more than 20% below its 22 June peak after another broad sell-off on Friday, with Nvidia, Broadcom, Micron, Intel and equipment names under pressure; Reuters estimates the index is down roughly 24% from its record, even though it remains materially higher year to date. What changed is market psychology: investors are no longer rewarding evidence that 2026 AI demand is strong, because that is already embedded in estimates; they are questioning whether hyperscaler capex, memory pricing and leading-edge utilisation can remain exceptional through 2027–29. Bulls view the decline as a positioning reset after an extreme Q2 rally, pointing to durable AI infrastructure demand and a wafer-fab-equipment market that could reach $250bn by 2028. Bears see the familiar semiconductor reflexivity problem — record margins drive record capex, which eventually destroys scarcity. Most exposed remain NVDA, MU, AVGO, AMD, TSMC, ASML, AMAT, LRCX and KLAC; relative beneficiaries from a sustained rotation could be cloud platforms and software-infrastructure vendors that monetise installed compute rather than sell incremental capacity.
2. Moonshot AI’s Kimi K3 release has created another “DeepSeek moment”, intensifying fears that model efficiency and open source can weaken the link between AI capability and infrastructure spend.
Moonshot released a nearly 3tn-parameter open-source model that it claims competes with or exceeds leading OpenAI and Anthropic systems on selected coding benchmarks, adding fuel to Friday’s global chip sell-off. The key investor debate is not whether the benchmark claims prove that fewer chips will be required — larger and cheaper models can stimulate far more inference demand — but whether frontier-model economics are commoditising faster than the industry can monetise its infrastructure. The bull case for semis is Jevons’ paradox: lower model costs broaden adoption and ultimately expand aggregate compute. The bear case is that cheaper Chinese and open models reduce premium pricing, weaken proprietary-model moats and allow enterprises to achieve acceptable performance without continually purchasing the most expensive accelerators. Second-order winners could include enterprises, cloud customers, inference software and orchestration vendors; pressure points include NVDA’s scarcity premium, frontier-lab economics, and hyperscalers carrying very large depreciation burdens. Most exposed: NVDA, AMD, AVGO, TSMC, MSFT, GOOGL, AMZN, META and private model providers OpenAI and Anthropic.
3. TSMC’s post-results weakness is the clearest signal that semiconductor equities now require proof of customer returns, not simply supplier execution.
TSMC delivered Q2 net income growth of 77% yoy to approximately $22bn, raised full-year revenue growth expectations to more than 40%, and lifted 2026 capex to $60–64bn, yet its shares weakened alongside the wider complex. TSMC remains strategically advantaged because it captures Nvidia GPUs, Broadcom-designed ASICs, hyperscaler custom silicon and advanced packaging regardless of architecture; however, the market is increasingly interpreting higher capex in two ways. Bulls see capacity responding to contracted, multi-year demand and packaging bottlenecks. Bears see rising depreciation, synchronised foundry and memory investment, and a growing probability that returns accrue to AI users once supply normalises. The second-order read-through is still positive for ASML, AMAT, LRCX and KLAC because spending is committed, but increasingly ambiguous for chip designers and memory suppliers whose earnings rely on persistent scarcity and pricing power.
4. Cybersecurity’s demand thesis strengthened again as real-world incidents spread across healthcare and critical industries, making AI-security spending harder to treat as discretionary.
Abbott disclosed two incidents involving unauthorised access to internal systems, while Clover Health identified unusual login activity affecting employee accounts with access to certain member information; Ecopetrol separately reported theft of data linked to roughly 3,300 accounts. These disclosures follow the White House’s creation of an AI and cybersecurity coordination group bringing together frontier-model developers and critical-infrastructure operators to manage vulnerabilities discovered by advanced AI systems. The debate is whether this urgency drives incremental sector growth or merely reallocates budgets towards a smaller number of platforms. Our read is that the incidents support spend across identity, endpoint, data security, exposure management and recovery, but consolidation should favour vendors controlling telemetry and enforcement rather than every point product. Most exposed: PANW, CRWD, ZS, OKTA, CYBR, FTNT, RBRK, CVLT, TENB and QLYS.
5. The software tape is increasingly confirming a structural budget and valuation split: cyber and infrastructure software are holding up better while conventional application SaaS remains trapped between AI disruption and weak organic growth.
On Friday, Palo Alto Networks and CrowdStrike rose modestly even as Salesforce and SAP declined, extending the pattern in which investors favour compulsory security, telemetry and infrastructure control points over seat-based application software. The central debate is whether traditional SaaS is merely oversold or structurally impaired. Bulls argue that embedded workflows, proprietary data and switching costs will allow leading vendors to migrate towards consumption and agent pricing. Bears point to IBM’s warning that enterprise budgets are being redirected towards servers, storage and memory, while AI agents threaten seat growth before vendors have established meaningful incremental revenue. The second-order implication is that low valuation multiples may remain poor catalysts without evidence of core reacceleration: PANW, CRWD, DDOG, SNOW and selected workflow/control-plane vendors appear better positioned, while CRM, WDAY, ADBE, TEAM, HUBS and SAP remain the principal battlegrounds.
