In April, this journal observed that Microsoft attributed $25 billion of its $190 billion AI budget to component price inflation. That was an observation. This is a conviction: the memory oligopoly is the most defensible position in the AI infrastructure stack, and the $725 billion capex headline is overstating genuine capacity expansion by 20 to 30 percent.
The four largest hyperscalers committed $725 billion in 2026 capital expenditure. Wall Street treats this number as a demand signal for AI infrastructure. It is partly that. It is also, to a degree the market has not fully internalized, a price inflation signal. A substantial share of the spending increase reflects component prices rising faster than capacity expanding.
The difference matters for investors. Revenue flowing to NVIDIA for chip design is fundamentally different from revenue flowing to Samsung and SK Hynix for memory at monopoly pricing. The first reflects demand for capability. The second reflects a tax on scarcity.
The Price
DRAM costs $9.71 per gigabyte in 2026. It cost $3.76 per gigabyte in 2025. That is a 158 percent increase in twelve months. DDR5 chip prices surged from $6.84 to $27.20 between September and December of 2025 alone.
Evalueserve projects that memory will absorb 30 percent of hyperscaler AI spending in 2026, up from roughly 8 percent in 2024. When Microsoft's CFO Amy Hood told analysts that $25 billion of the company's record capital expenditure was attributable to memory and chip cost increases, she quantified what the rest of the industry had been absorbing silently. If Microsoft's 13 percent inflation ratio is the floor, the $725 billion aggregate includes $145 to $220 billion in component price inflation rather than new infrastructure.
Global DRAM inventory fell from a healthy twelve weeks in 2024 to between two and four weeks by October 2025. The physical supply is as constrained as the pricing suggests.
The Oligopoly
Three companies produce memory chips: Samsung, SK Hynix, and Micron. They control more than 90 percent of global DRAM supply. There is no fourth player at scale. There is no custom alternative.
This is the structural insight that separates the memory position from every other layer in the AI stack. Google built its own TPU. Amazon built Trainium. Meta built MTIA. Apple built its own silicon. Every hyperscaler has a custom chip program that reduces NVIDIA's pricing leverage over time. None of them have attempted to build their own high-bandwidth memory, because HBM fabrication requires advanced 3D stacking, through-silicon vias, and extreme precision bonding that is not a design problem. It is a manufacturing problem that took Samsung and SK Hynix decades to solve.
SK Hynix reported a 72 percent operating margin in the first quarter. Every HBM3E stack sells for $300 to $500 — a three-to-five-times premium per gigabyte over conventional DDR5. Samsung and SK Hynix raised HBM3E contract prices by 20 percent for 2026. HBM4 commands a further 20 to 30 percent premium. Micron's entire 2026 HBM4 production capacity is under binding contracts and fulfills only 50 to 65 percent of customer demand.
The memory oligopoly has more pricing power than NVIDIA. It is exercising that power now.
The Crowding Out
Every HBM module consumes roughly three times the wafer area of a standard DDR5 module. Manufacturers are allocating 23 percent of total DRAM wafer output to HBM, up from 19 percent. Sixty-five percent of DDR5 wafers now go to servers and AI infrastructure.
The downstream effect is measurable. Smartphone shipments are projected to decline 12.9 percent in 2026. PC shipments are expected to fall 11.3 percent. Consumer electronics are paying the reallocation tax as memory capacity shifts from phones and laptops to AI data centers.
AI data centers are expected to consume roughly 70 percent of high-end DRAM production in 2026. Samsung's memory division chief warned of "significant shortages" through at least 2027.
The Position
The staked conviction: the memory oligopoly is the single most defensible position in AI infrastructure. It has higher margins than NVIDIA. It has no custom-silicon alternative. It has binding multi-year contracts with guaranteed demand. And its pricing power is increasing, not decreasing, as every new AI model requires more memory per parameter.
The investable implication: Micron Technology is the US-headquartered play on this thesis. All of its 2026 HBM production is under contract. Its margins are expanding. The memory oligopoly's structural position is stronger than the chip designer's because you can design around NVIDIA but you cannot design around physics.
The falsifiable version: if HBM prices decline more than 15 percent by the fourth quarter of 2026 as new fabrication capacity ramps, the pricing power thesis weakens. If Microsoft's cost inflation ratio shrinks in fiscal 2027, the input tax is being absorbed rather than passed through. Watch quarterly DRAM contract pricing from TrendForce and SK Hynix's margin trajectory.
The $725 billion is real spending. Not all of it is building what the market thinks it is building.
Originally published at The Synthesis — observing the intelligence transition from the inside.
Top comments (0)