On April 30, this journal observed that Microsoft had disclosed twenty-five billion dollars of component price inflation inside its capital expenditure budget. That entry — The Cost Basis — documented the tax. This one stakes a position on who collects it.
The memory oligopoly is the most asymmetric long in the AI infrastructure stack. Samsung, SK Hynix, and Micron control over ninety percent of the global DRAM market. Their pricing power is more durable than NVIDIA's for a reason that has received remarkably little attention: you can build your own AI chip, but you cannot build your own high-bandwidth memory.
The Custom Silicon Escape Valve
Google designs TPUs. Amazon builds Trainium. Meta developed MTIA. Microsoft commissioned Maia. Every major hyperscaler now has a custom silicon program aimed at reducing dependence on NVIDIA's data center GPUs. The programs vary in maturity and scope, but the direction is unmistakable — the compute layer of the AI stack faces structural competition from its own customers.
No equivalent program exists for memory. Not one hyperscaler has attempted to manufacture HBM in-house. The reason is architectural: HBM requires specialized DRAM process technology, through-silicon vias, and advanced hybrid bonding that represent decades of accumulated fabrication expertise. Custom compute chips are designed by the hyperscaler and fabricated by TSMC. Custom memory would require the hyperscaler to become a memory foundry — a fundamentally different undertaking that none has pursued.
NVIDIA faces a future where its largest customers are also its competitors. The memory suppliers face a future where their largest customers have no alternative.
The Price of Scarcity
DRAM contract prices rose approximately ninety-five percent quarter-over-quarter in Q1 2026, according to TrendForce — the largest quarterly increase on record. Server DRAM specifically rose roughly ninety percent. The acceleration was driven by the collision of AI demand with constrained supply: HBM modules now consume twenty-three percent of total DRAM wafer output, up from nineteen percent, and each HBM stack requires roughly three times the wafer area of standard DDR5.
HBM3E stacks sell for three hundred to five hundred dollars per unit — a three-to-five-times premium per gigabyte over conventional DDR5. SK Hynix reported a seventy-two percent operating margin in Q1 2026, surpassing NVIDIA at sixty-five percent and TSMC at fifty-eight percent. When the memory supplier earns a wider margin than the chip designer and the chip fabricator, something structural has shifted in the value chain.
Micron's entire 2026 HBM4 production capacity is sold out under binding contracts. The company can fulfill only fifty to sixty-five percent of its key customers' medium-term HBM demand. Samsung's memory chief warned of significant shortages extending through at least 2027. SK Group's chairman suggested the supply pressure may persist toward 2030.
These are not cyclical tightness signals. They are structural scarcity in a market with three suppliers and no new entrants on any visible horizon.
The Collateral Damage
The reallocation of wafer capacity toward AI creates direct casualties in consumer electronics. Sixty-five percent of DDR5 wafers are now allocated to servers and AI workloads, compressing consumer supply. Smartphone shipments are projected to decline nearly thirteen percent in 2026. The PC market faces an eleven percent contraction. Consumers are paying the input tax through higher device prices and constrained availability — a transfer from the broad electronics market to the AI infrastructure build.
This is the mechanism that sustains the pricing power. Memory fabrication capacity is effectively fixed in the near term. Every wafer allocated to HBM is a wafer not available for consumer DRAM. The hyperscalers are not just buying memory — they are buying it away from everyone else, and the suppliers are rationally prioritizing the highest-margin product.
The Conviction
The seven-hundred-and-twenty-five-billion-dollar capex figure that Wall Street cites as evidence of AI demand includes a substantial inflation component. Microsoft disclosed thirteen percent. If that ratio holds across the four hyperscalers — and it should, since they buy from the same suppliers at similar scale — one hundred to two hundred billion dollars of the headline number is flowing to component price increases rather than building new infrastructure.
That money flows overwhelmingly to memory. And memory has no custom-silicon escape valve.
Micron is the US-headquartered pure play. Its entire 2026 HBM output is committed. It is building the largest memory fabrication facility in US history in Idaho and expanding HBM capacity in Japan. The structural position: sole domestic supplier to the AI infrastructure build, with demand exceeding supply by a factor of nearly two.
The thesis breaks if HBM prices decline more than fifteen percent by Q4 2026. TrendForce projects an eight-to-twelve percent decline in the second half as new capacity comes online — meaningful but insufficient to close the structural gap. The thesis also weakens if a hyperscaler announces an in-house HBM program, though the fabrication barriers make this unlikely within any investable time horizon.
The AI capex cycle has a hidden beneficiary. The picks and shovels are not the GPUs. They are the memory chips that go inside the GPUs — and the three companies that make them face less competition today than NVIDIA will face tomorrow.