High AI Capex Demand a 'Multi-Year' Cycle, Says Rudina Seseri

Glasswing Ventures founder Rudina Seseri says demand for AI infrastructure remains strong and she does not expect AI capex to fall as companies pursue vertical integration. She also expects growth in specialized industry data sets rather than broad, general-purpose data. The outlook is supportive for AI infrastructure spending, though the piece is commentary without new company financials.

Analysis

The investable takeaway is not simply “AI capex stays high”; it is that capex is becoming more inelastic as customers internalize the model stack. That shifts profit pool toward the picks-and-shovels layer: accelerators, networking, memory, power, cooling, and data-center capacity. In practice, the best second-order beneficiaries are the vendors that solve bottlenecks inside the buildout path—NVDA, ANET, AVGO, VRT, DLR, EQIX—because vertical integration usually increases complexity, not just spend.

The underappreciated loser is the long-duration AI software cohort that depends on generalized model access and paid distribution. As large platforms move upstream, they can bundle AI features into existing products and squeeze standalone pricing; that is a margin and multiple problem for public software names without proprietary data rights. By contrast, vertical AI winners will be incumbents with scarce datasets and workflow control in healthcare, industrial, and financial services, where the value is in conversion of proprietary records into decision advantage, not in raw model quality.

The catalyst path is mostly 1-3 months: hyperscaler capex commentary, enterprise cloud spend, and supply-chain lead-time checks. The thesis breaks if top cloud vendors signal utilization gains and slower incremental capex, or if power/interconnect constraints force project delays rather than pricing power. Over 6-18 months, the key variable is whether AI economics broaden from experimentation to measurable operating savings; if not, today’s infrastructure optimism can turn into a value trap for anything not directly monetizing the buildout.

Key Decisions for Investors

• Long NVDA / ANET into the next hyperscaler earnings window; use a 5-8% pullback as entry. Risk-reward favors infrastructure because incremental spend is still being pulled forward by vertical integration rather than displaced.
• Pair long AVGO vs short IGV on any AI-fueled software bounce. Thesis: infrastructure captures the spend while generic software faces bundling pressure and multiple compression if model access becomes commoditized.
• Add EQIX or DLR on weakness as a slower-moving beneficiary of AI localization and data gravity. Falsify if hyperscaler capex guides down or vacancy/lease spreads soften for two consecutive quarters.
• Use ARKK as a hedge/short overlay rather than a direct AI expression. If the market starts questioning ROI on AI software, duration-heavy venture-style names should underperform before the hardware complex rolls over.