HYPERSCALER CAPEX vs. BUYBACKS — THE PARADIGM SHIFT
Combined AMZN/GOOGL/META/MSFT CapEx set to hit $610B in 2026 (+70% YoY), nearly 3x the $217B spent in 2024. Meanwhile, buybacks collapsing from ~$110B (2024) to an estimated $50-70B. The CapEx-to-Buyback ratio explodes from ~1.3x (2020-2023 average) to 9-12x in 2026.
This is not visionary spending — it’s catch-up. A decade of prioritising buybacks over infrastructure ($1.1T in repurchases 2020-2025 by 6 major tech cos) created a CapEx deficit now being compressed into 2-3 years at peak cost. Same dynamic as European defence: 20 years of underinvestment forcing a simultaneous rush at crisis premium.
It's a return to building real infrastructure — reminiscent of the telecom buildout of the late 90s or the railroad boom of the 1850s, as Benzinga recently compared it
Key data points:
• AMZN: $200B CapEx, zero buybacks since Q2 2022 — Bezos DNA, never played the buyback game
• GOOGL: $180B CapEx, FCF projected to collapse 90% (Pivotal Research). Was biggest buyback spender ($62B in 2024)
• META: $125B CapEx (>50% of revenue). CFO explicitly deprioritised buybacks. Off-BS debt via Blue Owl ($27B)
• MSFT: $105B CapEx but core AI dependency on OpenAI — spending on infrastructure around a partner’s technology
The market needs to re-learn how to value these companies. We’re shifting from asset-light multiples (P/E, buyback yield) to capital-intensive metrics (ROIC, depreciation, asset utilisation). Companies will now be judged on real strategy and execution, not financial engineering for investor appeasement.
Winners: AMZN (CapEx-native, custom chips, 15yr infrastructure DNA), GOOGL (DeepMind, TPUs, but execution risk). Losers in relative terms: MSFT (rented AI strategy), META (highest leverage, narrowest monetisation path).
Underappreciated: $88B in bonds issued in 3 months (Sep-Nov 2025). JPM projects $1.5T in AI data centre bonds over 5 years. Fortress balance sheets becoming leveraged infrastructure plays.
CONCLUSION — THE REAL PLAY
If $610B in CapEx is coming regardless of which hyperscaler wins the AI race, the best risk-adjusted play is the infrastructure layer. As we wrote last week: the picks and shovels always get paid. Data centre REITs (EQIX, DLR), power infrastructure (Vertiv, Eaton, Siemens Energy), cooling systems, and networking vendors (Broadcom) benefit from the spending regardless of who captures the AI application revenue. This is CapEx that cannot be cancelled — contracts are signed, land is secured, power is committed. The hyperscalers are competing on who builds faster. The infra names get paid either way. That’s the asymmetry