AI could take capex to $1Tn in CY27

BofA now sees global hyperscale capex reaching $800 billion in 2026 and $1 trillion in 2027, driven by accelerating AI demand and tight supply for high-end compute hardware. Microsoft raised 2026 capex guidance to $190 billion, Alphabet to $185 billion, Meta to $135 billion, and Amazon maintained $200 billion, underscoring a powerful investment cycle in AI infrastructure. The report points to continued pricing power for semiconductor and hardware suppliers such as Nvidia as hyperscalers absorb higher component costs.

Analysis

The market is still underestimating how self-reinforcing this capex cycle is: once hyperscalers commit to multi-year buildouts, procurement becomes less about demand discovery and more about securing scarce supply, which locks in vendor pricing power through 2026. That shifts margin leverage away from cloud platforms and toward the hardware stack with the tightest bottlenecks — advanced packaging, high-bandwidth memory, optics, and power delivery — where incremental supply is slowest to respond.

The second-order effect is that the capex surge is not just bullish for NVDA; it likely creates a broader beneficiary set in the picks-and-shovels layer while pressuring lower-tier component suppliers that lack allocation priority. If component inflation is already being absorbed by the hyperscalers, then the next margin squeeze may show up in cloud unit economics, especially for non-AI workloads that cannot absorb the same pricing changes. That sets up a divergence: AI monetization leaders can justify spend, but slower-growth cloud lines may see returns on invested capital trough before they recover.

The biggest risk is duration mismatch. Investors are pricing this as a 12-month earnings tailwind, but the supply tightness and capex commitments imply a multi-year trade, while any slowdown in enterprise AI adoption would hit sentiment faster than the physical supply chain can adjust. A reversal likely needs one of three catalysts: a sharp fall in AI inference usage, a delay in model monetization, or a policy/supply shock that loosens hardware availability and compresses vendor pricing power.

Consensus is too focused on 'AI winners' as a monolith. The better setup is relative value: the hyperscalers are increasingly buying growth at the expense of current FCF, while the durable upside likely sits in the infrastructure layer with pricing discipline and constrained supply. The trade is not to chase every AI-linked name, but to own the scarce inputs and fade the parts of the stack where incremental capex simply accelerates depreciation without near-term revenue conversion.

Key Decisions for Investors

• Long NVDA vs short a basket of lower-quality AI infrastructure suppliers for the next 6-12 months; the thesis is allocation-driven pricing power, not broad-based hardware beta. Use a 1:1 notional pair and take profits if GPU lead times begin normalizing or gross margins compress by >200 bps.
• Add selective exposure to the 'tight-supply' layer via long AVGO or MRVL on dips over the next 2-4 weeks; these names should benefit from the custom-silicon and networking mix shift. Risk/reward is best if bought after any post-earnings pullback tied to capex fatigue, with a 3-6 month horizon.
• Consider a relative short on MSFT or AMZN vs long NVDA if the market starts questioning hyperscaler FCF conversion over the next 1-2 quarters. The trade monetizes the spread between monetized AI infrastructure demand and the cost of funding it.
• Buy 6-9 month call spreads on NVDA rather than outright calls; implied vol likely remains elevated, and the spread captures continued capex upside while limiting downside if spending growth normalizes in late 2026.
• Set a watchlist for memory and optics names on any guidance raises from hyperscaler suppliers; if pricing passes through faster than expected, rotate into those laggards before consensus catches up.