The HBM Bottleneck: 3 Tech Companies Have Locked Up 85% of the Supply of AI's Scarcest Asset

Morgan Stanley estimates Nvidia, Alphabet, and AMD have pre-secured 85% of 2027 high-bandwidth memory (HBM) production—37%, 36%, and 12%, respectively—creating a major supply-chain advantage in AI infrastructure. HBM remains constrained by limited EUV lithography availability, its roughly 3x wafer requirement versus conventional DRAM, and multiyear clean-room construction timelines. The secured supply supports Nvidia's GPU leadership, Alphabet's TPU expansion, and AMD's ramp of HBM-intensive MI455X GPUs as AI inference demand accelerates.

Analysis

The relevant signal is not accelerator demand but allocation economics: scarce memory and packaging convert vendor roadmaps into a quota system. NVDA’s secured inputs should protect shipment cadence and gross-margin durability over the next 2-4 quarters, whereas AMD’s higher memory intensity creates a less obvious execution risk—each incremental system consumes more of the constrained component, potentially limiting unit growth or forcing lower-margin supply sourcing. This favors NVDA over AMD unless AMD demonstrates that customer pricing offsets its materially higher memory content.

GOOG and AVGO are the more underappreciated read-throughs. Captive TPU capacity can lower Google Cloud’s unit compute cost and make price-performance competition against AWS/Azure more aggressive without sacrificing economics; AVGO benefits only if custom-chip volume converts into recognized semiconductor revenue rather than merely long-dated capacity reservations. MU is a tactical beneficiary of HBM mix, but the market should distinguish HBM pricing from the broader DRAM cycle: incremental HBM wafers displace conventional DRAM output, which can tighten commodity DRAM and amplify earnings upside if PC/mobile demand is merely stable.

Consensus may be over-crediting reservations as delivered supply. The critical missing disclosures are take-or-pay terms, HBM yield progression, CoWoS/substrate availability, and whether demand is tied to end-customer deployments rather than preemptive inventory building. A meaningful improvement in HBM yields or an AI capex pause would unwind scarcity premiums first in memory and custom-silicon suppliers, while NVDA’s software moat makes its downside relatively more cushioned.

Key Decisions for Investors

• Initiate/maintain a 3-6 month long NVDA / short AMD pair, sized beta-neutral. Thesis is superior supply conversion and lower component-intensity risk; reassess if AMD’s next earnings call shows MI-series revenue growth and gross margin above guidance without a material inventory build, or if the pair underperforms by 15%.
• Accumulate MU on 1-3 month weakness ahead of memory pricing and earnings updates, but keep position tactical rather than structural. Upside requires HBM mix to tighten conventional DRAM as well; exit if management indicates HBM yield gains are expanding supply faster than demand or if DRAM contract pricing turns negative.
• Watch for a long AVGO catalyst around confirmed custom-ASIC shipment/ramp commentary over the next two earnings cycles. Do not underwrite reported partnership values as revenue; require backlog conversion, semiconductor segment margin stability, and customer concentration disclosure before establishing a full-size position.
• Use GOOG as a 6-18 month relative long versus hyperscaler peers with less captive accelerator capacity, contingent on Cloud margin expansion and sustained capex productivity. Falsification is accelerating capex with no corresponding Cloud operating-margin improvement over two reporting periods.
• Avoid chasing ASML solely on this bottleneck narrative. The nearer-term economics are governed by tool delivery timing and customer capex budgets; upgrade only if order intake/bookings demonstrate that memory-fab EUV demand is incremental to, rather than cannibalizing, logic-fab demand.