AI infrastructure 2026-2029: chips, network, storage, electricity
Thesis: The AI investment cycle in 2026 is no longer hope, but cash flow - Goldman sees hyperscaler capex at around 1.2 trillion in 2027. USD (+50%), S&P Global over 1.3 trillion. USD; Nvidia reported over $96 billion in revenue in Q2 FY27, TSMC reported +53% monthly revenue in August. I buy the bottlenecks along the chain: compute (NVDA, AVGO, AMD-free, TSM), network/optics (ANET, ALAB, COHR), storage (MU), and the underrated power and cooling rail (VRT, CLS, CEG, VST). Catalysts: Hyperscalers capex guidance in January/February 2027, Blackwell successor ramp, 1.6T transceiver cycle, further nuclear PPAs. Biggest risk: Fed hiked in September 2026 for the first time in three years, 10-year yields at 24-year high — if a single hyperscaler cuts capex, the whole chain falls 25-40%, storage (MU +275% YTD) is the most cyclical part. Position size: max. 12% per title, MU and ALAB max. 6%, current block together 20% as a damper. Exit: Halve position if hyperscaler capex guidance for 2027 falls below previous year or NVDA data center revenue falls sequentially for two quarters in a row.
