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 roughly 85% of 2027 high-bandwidth memory (HBM) production, led by Nvidia at 37%, Alphabet at 36% and AMD at about 12%. Constrained EUV tool availability, HBM's roughly 3x wafer requirement versus standard DRAM, and multiyear clean-room construction timelines make memory supply a material competitive moat for AI infrastructure providers. Nvidia's $500B multiyear SK Hynix partnership, Alphabet's TPU supply arrangements, and AMD's Samsung deal and memory-efficient product designs position the companies to benefit as inference-driven AI demand accelerates.

Analysis

The investable implication is not simply accelerator scarcity; it is that supply allocation converts AI demand into a share-gain mechanism. NVDA can protect system shipment cadence and pricing while smaller GPU challengers face a higher risk of launch delays or lower-volume configurations. AMD’s allocation is strategically valuable but its unusually memory-intensive product architecture raises a less appreciated risk: each incremental unit consumes more of a constrained input, potentially capping revenue realization and diluting gross-margin upside versus NVDA even if customer demand is strong.

GOOG is the most underappreciated beneficiary because internal accelerator availability improves the unit economics of Cloud and Gemini inference before it appears as external semiconductor revenue. That supports a 6-18 month case for cloud-margin expansion and raises the competitive bar for hyperscalers that must purchase merchant GPUs. AVGO benefits from custom-silicon design wins, but its economics remain exposed to whether customers, rather than Broadcom, bear memory-cost inflation; watch ASIC gross-margin commentary rather than extrapolating TPU volumes directly into AVGO earnings.

The second-order winner is MU, provided HBM mix displaces lower-return commodity DRAM rather than merely absorbing capital expenditure. A tight market should improve contract pricing and earnings visibility over the next 1-3 quarters, though MU has historically been the most vulnerable to an abrupt inventory correction once capacity comes online. ASML’s benefit is longer dated and likely already reflected in order expectations: the relevant catalyst is sustained memory-maker EUV order conversion, not headline HBM demand. Consensus may be overpaying for announced supply reservations; these are often contingent, and the true read-through is delivered HBM bits, packaging yield, and accelerator shipment guidance.

Near-term, this is supportive for the AI complex but insufficient alone to justify chasing high-beta names after a sharp move. The thesis fails if NVDA/AMD guide to supply-constrained revenue below demand expectations, if MU reports HBM pricing or yields below plan, or if hyperscaler capex growth decelerates materially. Over 6-18 months, additional memory and advanced-packaging capacity could shift bargaining power back toward customers and compress the scarcity premium.

Key Decisions for Investors

• Maintain an overweight NVDA versus AMD through the next two earnings cycles: NVDA has the cleaner ability to monetize scarce memory through complete-rack pricing and software lock-in. Use a relative-value long NVDA / short AMD structure rather than outright beta; reassess if AMD demonstrates GPU gross-margin expansion alongside shipment growth, not just backlog.
• Build a 3-6 month long MU position on pullbacks, sized modestly given cycle risk. The upside case is HBM mix and contract repricing driving estimates higher; exit or hedge if management signals rising legacy-DRAM inventories, HBM yield issues, or materially higher capex without matching pricing commitments.
• Add GOOG on Cloud-margin weakness or post-earnings volatility for a 6-18 month horizon. The key confirmation is accelerating Cloud revenue with expanding operating margin and stable aggregate capex intensity; reduce if AI capex rises faster than monetization or search-margin pressure offsets infrastructure savings.
• Avoid treating ASML as a near-term pure-play HBM trade. Establish an alert around memory-maker EUV bookings and ASML order intake; only add on evidence that incremental memory demand is converting into equipment orders rather than being met through process optimization and existing tool capacity.
• For portfolio hedging, monitor the NVDA/AMD revenue-guide spread and MU HBM commentary each quarter. A simultaneous shortfall in accelerator supply guidance and HBM pricing would challenge the scarcity thesis and warrants reducing AI-infrastructure gross exposure.