UBS now expects AI capex to reach nearly $1tn this year and around $1.4tn by 2027

UBS projects AI capital expenditure will nearly double to $998 billion in 2026 from $506 billion in 2025, then rise to $1.447 trillion in 2027. Memory spending is forecast to surge from $71 billion in 2025 to $367 billion in 2026 and $923 billion in 2027, accounting for roughly 90% of the nearly $1 trillion increase in AI capex over the two years. UBS cautioned that if the spending increase is driven primarily by memory-price inflation rather than volume growth, the benefit to real U.S. GDP would be limited while profits and GDP gains shift toward Asian memory producers.

Analysis

The key investable shift is from compute scarcity to memory scarcity: DRAM/HBM vendors gain pricing power while hyperscalers absorb a larger share of AI-system cost without a commensurate increase in usable compute capacity. MU is the cleanest U.S. equity beneficiary; SK Hynix and Samsung Electronics are more direct HBM exposures but less accessible for many portfolios. For MSFT, GOOGL, AMZN and META, higher memory bills pressure AI service gross-margin ramp and raise the hurdle rate for incremental data-center deployment, particularly where monetization remains inference-light.

The likely 1-3 month market response is continued multiple expansion for memory suppliers and a widening gap versus broad semiconductor indices. However, the forecast is highly price-sensitive: if HBM supply additions arrive faster than qualification demand, nominal AI capex can decelerate sharply even while real AI deployment remains healthy. This makes memory earnings revisions—not aggregate AI capex headlines—the decisive catalyst; watch MU’s bit-volume guidance, HBM mix, contract pricing, and inventory days.

Second-order losers include GPU/system vendors whose bill of materials rises faster than end-customer budgets, potentially constraining unit growth even if dollar content rises. NVDA is not a straightforward short because its platform pricing can offset component inflation, but memory scarcity shifts negotiating leverage away from GPU vendors and toward HBM suppliers. INTC remains a weak read-through: its AI opportunity requires credible accelerator demand and memory ecosystem execution, neither of which is validated by higher memory prices alone.

Consensus may be extrapolating a nominal-spend boom into broad semiconductor demand. If spending growth is predominantly inflation in a concentrated component category, the correct expression is a narrow memory long rather than indiscriminate long SMH/SOX exposure; the macro benefit accrues disproportionately outside the U.S., while U.S. cloud buyers bear the cost.

Key Decisions for Investors

• Initiate/maintain long MU on a 6-12 month horizon; size around the next earnings print rather than chase a sharp pre-earnings move. Thesis requires sustained HBM qualification momentum and upward ASP revisions; exit or reduce if management signals inventory rebuilding has replaced end-demand or guides HBM supply materially ahead of demand.
• Use a relative-value expression: long MU / short SMH or SOXX over 3-6 months, with a 1:1 beta-adjusted notional. This isolates memory-pricing exposure from broad AI semiconductor beta; reassess if non-memory semiconductor orders reaccelerate or MU underperforms the index by 10% after earnings.
• Trim incremental exposure to cloud-capex beneficiaries with weak near-term AI monetization, particularly AMZN and GOOGL, into strength; treat increased memory intensity as a gross-margin and return-on-capital headwind over the next 2-4 quarters rather than a reason to abandon core positions.
• Do not use INTC as a memory-scarcity proxy. Keep it on watch only: a constructive trade requires independently observable evidence of accelerator design wins, improved foundry/customer economics, or a credible HBM-enabled product ramp; absent that, memory inflation primarily strengthens competing system ecosystems.