Memory Chip Shortage Leaves Hyperscalers Behind as Hardware Costs Squeeze AI Spending

A tightening supply of high-bandwidth memory chips is emerging as the defining constraint on the AI buildout, dragging down the stocks of Amazon, Alphabet, Microsoft, and Meta Platforms while a basket of memory-related equities has surged 41% over the past month.

All four hyperscalers have seen their shares decline over that same period, even as the Nasdaq has gained nearly 1%, underscoring a growing divergence between the companies spending most heavily on AI infrastructure and the suppliers profiting most directly from that spending.

The bottleneck centers on high-bandwidth memory, or HBM — a specialized form of dynamic random access memory that is central to AI computing workloads. The HBM market is concentrated among three players: SK Hynix holds roughly 60% share, with Samsung and Micron each accounting for approximately 20%.

That concentration has given memory makers significant pricing power. Microsoft and Meta both cited higher component costs on recent earnings calls as a factor driving their capital expenditure figures higher, reflecting how deeply the shortage has penetrated hyperscaler cost structures.

The ripple effects have extended beyond AI-focused memory into long-term data storage. Shares of Sandisk, Western Digital, and Seagate have also climbed sharply, reflecting expectations that demand across the storage stack will remain elevated.

The equipment companies that manufacture the machinery used to produce these chips — Applied Materials, Lam Research, and KLA Corp — occupy what some analysts now describe as the most critical position in the AI supply chain. Applied Materials CEO Gary Dickerson said last month that the company has "unprecedented visibility" from customers given the strength of demand, suggesting near-term earnings estimates are unlikely to come under pressure.

New fabrication plants, known as fabs, were expected to ease the HBM shortage, but capacity additions have not come online quickly enough to relieve the constraint. The shortage has also prompted the hyperscalers to pursue custom chip strategies as an alternative to relying on Nvidia. Amazon, Alphabet, Microsoft, and Meta have partnered with Marvell Technology and Broadcom to co-design proprietary AI accelerators.

Amazon has said that if its internal chip business were a standalone entity, it would carry a $50 billion annual revenue run rate. Nvidia CEO Jensen Huang, whose company remains the dominant supplier of AI accelerators, made a $2 billion investment in Marvell in March and earlier this month described it as the next trillion-dollar company — even as Marvell works with Amazon on chips designed to reduce dependence on Nvidia hardware.

Broadcom, which has partnered with Alphabet's Google on a similar custom chip effort, has seen its stock fall sharply. Shares stood at $479 before the company's June 3 earnings report and ended Thursday at $411 — still down roughly 14% from that pre-earnings level after recovering a portion of a 22% post-earnings slide.

Meta faces an additional structural challenge. Unlike the other three hyperscalers, Meta's revenue is almost entirely derived from advertising, leaving it without a cloud services business that could more directly monetize its AI infrastructure investments. Meta's stock is down 12.55% year to date.

SK Hynix, the dominant HBM supplier, is planning a New York stock listing in an effort to broaden its investor base and raise its profile among international institutional investors.

The current dynamic signals a meaningful shift in where AI-related market value is accruing. Chips and the equipment used to make them — once considered commodity-like businesses vulnerable to cyclical price declines — have become supply-constrained enough that pricing power has moved decisively to the manufacturers, leaving the hyperscalers in a more dependent position than the earlier phase of the AI buildout suggested they would be.

Disclaimer