AI Chip Makers Face Memory Crunch That Threatens Computing Speed

Micron Technology this week became the first U.S. memory-chip company to briefly cross $1 trillion in market capitalization, a milestone that reframes the AI hardware race from a GPU story into a memory story. The bottleneck slowing the world's most expensive AI clusters is not compute power. It is the speed at which data reaches that compute, and Micron, SK Hynix, and Samsung are the three companies that control the answer.

Micron's market cap crossing the $1 trillion threshold, however briefly, places it alongside Nvidia in a class of AI infrastructure companies whose valuations reflect not current earnings alone but structural scarcity. The company's HBM4 chips, designed specifically for Nvidia's next-generation Vera Rubin GPUs, deliver bandwidth exceeding 2.8 terabytes per second, according to Micron's own specifications cited by Scientific American . That figure is the operating constraint for every frontier AI training run and inference deployment happening at scale today.

Keren Bergman, an electrical engineering professor at Columbia University, described the demand clearly: "The reason HBMs are in such high demand is that they have pretty good storage, and they're extremely, extremely fast." Bergman added that because of the growing size of AI models, the available memory capacity close to the processor is "one or two orders of magnitude less than what you need." That gap between supply and demand, measured in orders of magnitude, is the investment thesis in a single sentence.

The stakes extend beyond corporate earnings. Hadi Esmaeilzadeh, a computer architecture researcher at UC San Diego, told Scientific American: "It's in the national security interest that we bring chip manufacturing back to the United States. Our dependence on AI systems is growing, and our supply chain is somewhere else." With SK Hynix and Samsung headquartered in South Korea, Micron's Boise-based operations represent the only significant North American HBM production capacity in an era when AI infrastructure has become a geopolitical asset.

The Physics of the Bottleneck: Why You Cannot Simply Build Faster

The architecture of high-bandwidth memory is the reason supply cannot scale at the same pace as demand. Standard memory chips spread horizontally across a circuit board. HBM stacks memory layers vertically, up to 12 or 16 layers high, connected through the silicon itself using structures called through-silicon vias, or TSVs. This vertical stacking places memory physically adjacent to the processor rather than at the end of a long electrical highway.

Esmaeilzadeh framed it with precision: "Now there's higher connectivity between the two, providing higher bandwidth. It's like adding more lanes on highways." The analogy is useful, but the constraint it implies is more important than the innovation. Engineers cannot add lanes indefinitely. TSV fabrication requires precision at scales that limit yield and manufacturing throughput. Every additional stacked layer compounds the complexity.

This is not a problem that capital expenditure alone resolves on a short timeline. The fabrication processes for HBM are distinct from those for logic chips or standard DRAM. Building new HBM capacity requires specialized equipment, trained process engineers, and multi-year lead times. The demand surge from AI model training and inference is measured in months. The supply response is measured in years.

Two Demand Drivers, One Supply Constraint

The demand for HBM does not come from a single point in the AI stack. It comes from two distinct and compounding sources.

Training large AI models requires clusters of thousands of accelerators operating simultaneously. Each accelerator is only as productive as the data it receives. A GPU waiting for memory is idle compute, and idle compute in a cluster costing hundreds of millions of dollars is a capital efficiency problem. The pressure on memory bandwidth during training is sustained and intensive.

Inference, the process of running a trained model to serve actual users in chatbots, coding tools, or AI agents, creates a second and structurally different demand profile. Inference workloads repeat continuously, cycling through the same memory-intensive operations billions of times per day across global deployments. As AI usage scales from early adopters to mass deployment, inference demand grows faster than training demand because user volume multiplies while model count stays relatively stable.

The parallel trend toward local AI deployment, documented in Sam Witteveen's analysis of AMD's Ryzen Threadripper 9980X and Radeon AI Pro R9 700 hardware, adds a third vector . AMD's Radeon AI Pro R9 700 carries 32GB of VRAM, a figure that underscores how memory capacity has become the differentiating specification in hardware purchasing decisions at every tier of the market, from hyperscaler data centers to on-premise enterprise deployments. Terms on AMD's HBM supply arrangements were not disclosed publicly.

Competitive Landscape: Three Companies, One Chokepoint

The global HBM market concentrates around three producers: SK Hynix, Samsung, and Micron. SK Hynix and Samsung are based in South Korea. Micron is the sole U.S.-headquartered manufacturer at scale. Market share figures by producer for 2026 were not disclosed in the source material, but Micron's $1 trillion valuation milestone signals that capital markets are pricing Micron as a primary beneficiary of AI infrastructure spending regardless of its position relative to Korean competitors .

The geopolitical dimension of this concentration is no longer theoretical. U.S. export controls on advanced semiconductors to China, the concentration of leading-edge fabrication in Taiwan, and the South Korean dominance of HBM supply have each entered national security discussions at the policy level. Esmaeilzadeh's comment to Scientific American on national security interests reflects a view that has moved from academic to legislative in the past 18 months .

For institutional investors, the supply concentration creates a durable pricing dynamic. When three producers control a component that is one to two orders of magnitude undersupplied relative to demand, pricing power is structural rather than cyclical. The question is not whether HBM margins compress but when, and that answer depends on fabrication capacity timelines that no announced investment has yet resolved.

Bubble Risk Is Real, But Memory Is the Last Line to Cut

Sundar Pichai at Google and Sam Altman at OpenAI have each warned publicly of a potential AI bubble, and Scientific American noted that data center construction has stalled while banks have grown cautious about the debt financing backing AI infrastructure buildout . These are legitimate systemic risks.

Our view: memory sits at the base of the AI infrastructure stack in a way that differentiates it from the higher-risk layers of the investment thesis. Data centers can be delayed. Software licenses can be cancelled. GPU orders can be deferred. But AI models already in production, serving hundreds of millions of users, require memory to function. That existing installed base creates a floor under HBM demand that speculative infrastructure spending does not.

The distinction matters for portfolio construction. An investor long on AI data center REITs is exposed to the construction and debt dynamics Pichai and Altman flagged. An investor long on HBM producers is exposed to a different risk profile: the risk that AI model sizes plateau, reducing the bandwidth and capacity requirements per accelerator. Bergman's "one or two orders of magnitude" gap makes that plateau scenario difficult to construct near-term .

The Plocamium View

The market has correctly identified Micron as an AI infrastructure beneficiary. It has not yet fully priced the second-order consequence of the memory bottleneck, which is that it shifts negotiating leverage across the entire AI value chain.

When memory is the binding constraint, every company building AI products faces a supplier with structural pricing power. Nvidia can design the world's most advanced GPU, but if it ships without sufficient HBM, it is a partial product. Hyperscalers can commit to $100 billion capital expenditure programs, but if memory allocation falls short of model requirements, training timelines extend and inference capacity lags user demand. The constraint is architectural, not cyclical.

Plocamium's thesis is this: HBM producers are the toll road, not the destination. The AI application layer will see winner-take-most dynamics, intense competition, and significant capital destruction among also-rans. The memory layer will see sustained pricing power, limited new entrants given fabrication complexity, and a geopolitical tailwind from U.S. and allied governments seeking to secure domestic supply chains.

The analogous historical precedent is ASML's position in EUV lithography. One or two companies controlled a process no one else could replicate at scale, and they extracted durable margins across an entire technology supercycle. HBM does not have ASML's complete monopoly, but the three-player concentration in a market with structural undersupply is the closest current analog in AI hardware.

For PE and institutional allocators, the implication is a rotation within AI infrastructure exposure: away from the application and software layer, where competition is intensifying, and toward the component layer where physics, manufacturing complexity, and geopolitics combine to create barriers that capital alone cannot replicate on any short timeline.

The forward question is not whether demand for HBM grows. Bergman said it plainly: "It's very clear that we're not even close to meeting the compute demand that's out there." The forward question is which producer scales fastest, which government subsidizes most effectively, and whether Micron's $1 trillion moment is a ceiling for this cycle or the floor for the next one. Plocamium's position: it is the floor.

The Bottom Line

Micron's $1 trillion milestone is the AI supply chain's most important data point of May 2026. It tells institutional capital that the memory layer, long treated as a commodity input, has repriced as strategic infrastructure. Three producers control a component that is orders of magnitude undersupplied relative to AI's requirements. Physics limits how fast new capacity can come online. Geopolitics is adding policy-driven demand for domestic supply. Investors still underweight the memory layer relative to the GPU layer are positioned for the last cycle, not the current one.