Can China achieve AI supremacy?

China currently has ~5 zettaFLOPS of AI compute versus the U.S. ~35 zettaFLOPS (≈15% of U.S. capacity). With >500 GW of annual power additions and the potential for nearly $1 trillion in AI data‑center capex, Bernstein estimates China could match U.S. compute by 2035—or, in a power-constrained scenario, exceed U.S. levels by >3x by 2035. Key constraints remain: domestic AI chips are roughly 25% as efficient as U.S. peers (could improve to >50% by 2035) and export controls/limited access to advanced manufacturing tools pose material downside risks.

Analysis

The strategic contest is as much about where capital flows as about raw chip performance. Expect a multiyear reallocation into grid infrastructure, large-format storage, and site-level thermal management — assets that reduce delivered cost-per-inference and shorten marginal build times. These are long-horizon, lumpy capex markets where supply bottlenecks (transformers, HV equipment, specialized copper, and high-capacity substations) create outsized returns for early suppliers and pushouts for fast followers.

Algorithmic and systems-level efficiency is a realistic counterweight to a pure “scale everything” path: model compression, sparsity techniques, and on-chip acceleration can materially slow the growth rate of demand for brute-force datacenter compute over a 3–5 year window. That means software/IP-heavy plays (inference optimizers, middleware that squeezes utilization) can earn high ROIC with far lower capital intensity than hyperscale farms, and will become prime targets for consolidation. Monitor cloud pricing and OEM utilization rates as leading indicators — a 5–10% sustained change in effective utilization typically presages a re-rating of both hardware orders and colo rents.

Geopolitics and export controls amplify optionality: equipment vendors domiciled outside sanction vectors will see elevated strategic value, but their order books will be volatile around policy shifts. The clearest near-term arbitrage is between firms that sell physical build-out capacity (grid, storage, cooling) and those whose value hinges on cutting-edge lithography — both can win, but on different timelines and risk profiles. The consensus focuses on raw compute parity; the underappreciated path to dominance is cheaper, reliably deliverable energy and the local supply chain that locks in power-to-server throughput.

Key Decisions for Investors

• Long ABB (ABB) — 18–36 months: allocate 2% NAV. Rationale: global supplier of transformers, switchgear and HVDC components that will capture outsized margins from accelerated grid buildouts. Target +40% upside; stop-loss 20% on 6-month moving-average break.
• Long Equinix (EQIX) — 12–24 months: allocate 2.5% NAV. Rationale: colo operator with diversified contracts and pricing power to capture customers migrating to low-cost-power jurisdictions. Target +30% absolute upside or 20% relative outperformance vs Digital Realty; stop-loss 15%.
• Long Lam Research (LRCX) with protective puts — 12–36 months: allocate 1.5% NAV into stock + buy 9–12 month puts (25–35% OTM) to hedge export-control shocks. Rationale: exposure to global fab tooling demand as jurisdictions localize manufacturing. Risk/reward ~3:1 net of hedge; worst-case -40% if controls materially tighten.
• Pair trade — Long EQIX / Short DLR (Digital Realty) — 12 months: equal-weight pair sized to 2% NAV gross. Rationale: preference for operators with higher interconnect density and flexible power contracting; expect 15–25% relative outperformance. Tighten or unwind if colo spot rents diverge by >10% from current term trajectories.