US-China AI race: one goal, two very different approaches

BofA estimates AI-related capex will add ~0.4 percentage points to US GDP growth this year and underpins an 8% GDP forecast for Taiwan in 2026. The spending surge is creating strong demand for high-end semiconductors and infrastructure, benefiting suppliers in Taiwan, Mexico and Korea amid intensified US–China competition. Analysts caution AI's GDP contribution may fade in 2027 without additional hyperscaler investment and highlight a looming 'skills challenge' as productivity gains depend on widespread adoption. Continued capex into data centers and specialized silicon makes global supply capacity the key determinant of near-term growth outcomes.

Analysis

The incremental capital flowing into training-scale AI architectures will concentrate demand on a narrow set of suppliers and infrastructure inputs, creating supply-side power that likely outpaces broad market appreciation. Expect asymmetric pricing power for extreme-UV lithography, advanced packaging, specialty substrates, and longtime congested transport corridors; firms that control capacity (not just design IP) can sustain 20–40% revenue upside in peak years while smaller producers see margin compression. This concentration also amplifies geopolitical risk: a supply disruption in a single node (Taiwan fabs, Korean assembly lines, or a key shipping chokepoint) can produce outsized revenue volatility for the entire chain within a 30–90 day window.

Banks and capital providers are a quietly underpriced lever in this cycle — underwriting, project finance, and equipment leases become a non-linear earnings driver as capex converts to receivables and IB fees. However, credit exposure will bifurcate: commercial real estate and legacy tech capex face slower cycles while data-center-linked loans show higher utilization but also longer duration and collateral tied to specialized equipment. Regulators and export-control actions are the highest-probability catalyst for re-pricing: targeted sanctions or tightened export licensing can compress accessible supply within weeks and flip sentiment from optimistic to risk-off.

From a macro standpoint, the intermediate-story is not pure demand but infrastructure choke points — power grid upgrades, localized copper/energy shortages, and skilled-labor constraints — that create multi-year lead times and durable pricing for upstream materials. Monitor hyperscaler capex guidance cycles and sovereign funding (domestic CHIPS-like programs) as 6–24 month inflection points; should those slow, expect a 12–18 month inventory digestion that pressures equipment OEMs by 20–35%. Conversely, continued front-loading by a handful of hyperscalers implies outsized alpha in specialized suppliers and financing vehicles over the next 12–36 months.

Key Decisions for Investors

• Long ASML (ASML) or TSMC (TSM) exposure via 12–24 month call spreads: buy 24–36 month OTM calls and sell higher strikes to finance premium. Rationale: captures pricing power in high-capex cycle; target return 2x–3x vs max loss = premium (use 3–5% portfolio sizing).
• Buy bank-finance exposure (BAC) 6–12 month: go long BAC equity or receive-side structured notes sized to 3–5% portfolio. Rationale: capture incremental deal flow/fees from data-center project finance; downside: credit or rate shock could shave 10–15% from thesis, upside ~15–25% if loan growth and NIMs hold.
• Buy MSFT 24–36 month call spread (e.g., buy LEAP calls, sell higher strike) to capture platform monetization without full delta risk. Rationale: benefits from software+cloud capture of AI stacks; capped cost reduces premium exposure. Target asymmetric payoff 2:1 reward-to-risk with stop if model training spend guidance slips by >15% QoQ.
• Commodity/inputs trade: long copper (FCX or copper futures) 6–18 month with tight stop: expect sustained industrial draw from data-center electrification and packaging lines. Position size 2–4% portfolio; upside 25–50% if capex continues, downside 15–20% on demand slowdown or recession.