Thesis Labv0.2.0
T002 active ●●● 12-24 months created 2026-05-18 · updated 2026-05-18

The semi capex cycle is being underwritten by hyperscalers, not consumers — and equipment is the cleanest expression

Claim. TSMC's $52-56B 2026 capex and hyperscaler $600-700B AI spend together create a multi-year demand floor for advanced-node capacity. ASML and TSMC capture the spend with less competitive risk than GPU vendors; AVGO captures custom-silicon dispersion as hyperscalers diversify away from a single GPU supplier.

The thesis

There are three layers to express AI silicon: GPUs (NVDA), foundry (TSM), and equipment (ASML). NVDA has the highest expected revenue beta but also faces custom-silicon substitution from every major hyperscaler — Google's TPU on Broadcom, AWS Trainium, Meta MTIA, Microsoft Maia. TSM wins regardless of which design wins, because they all run on N3/N2. ASML wins regardless of which foundry wins, because High-NA EUV is sole-source. The cleanest 'shovels' positioning is ASML + TSM with a smaller NVDA position as exposure to the leader and AVGO as the custom-silicon hedge.

Candidate tickers

  • ASML core — Monopoly on EUV. Mkt cap broke $500B post-TSMC capex raise. Order book visible 18-24mo out.
  • TSM core — Foundry duopoly winner. 63-65% gross margin guidance signals pricing power. Geopolitical risk is the main offset.
  • NVDA core — Leader but priced for it. Hold as participation, not as conviction overweight.
  • AVGO core — Custom ASIC for Google/Meta. Wins if hyperscalers continue diversifying off NVDA. Lower multiple than NVDA at writing.
  • AMD watching — MI series gaining real share but inconsistent. Wait for evidence of sustained hyperscaler wins.
  • MU watching — HBM beneficiary but memory cycles are brutal. Skip until cycle clarity.

Evidence

Falsifiers — what would change my mind

  • TSMC cuts capex guidance >15% in any quarter.
  • ASML order book book-to-bill <0.9 for two consecutive quarters.
  • Hyperscaler aggregate capex guidance revised down >15%.
  • Evidence of training compute demand plateauing (e.g., a major lab reports diminishing returns on scaling).