Reserved Capacity Moat, Regional Supply Chain Splits, and Memory Pricing Power

Reserved Capacity Becomes the New AI Supply Chain Moat

Amazon, Alphabet, Microsoft, Meta, and Oracle lifted AI infrastructure spending to new highs this week, with combined 2026 capex guidance of roughly $790 billion to $815 billion and trailing-four-quarter capex already at $691 billion through Q3 2026, up 93% year over year. That surge is colliding with the semiconductor stack’s tightest bottlenecks: HBM, TSMC’s CoWoS packaging, and ABF substrates.

HBM4 supply is not expected to fully meet demand until mid-2027, with some suppliers reportedly tight through 2028 and effectively sold out into 2029. CoWoS remains a critical limiter, 2027 packaging capacity is projected to rise about 80% against demand near 2x, and ABF substrate utilization is near full with no meaningful additions expected until 2H 2027. The result is delayed GPU deployments, pushed-out custom-chip ramps, and data-center timing pressure layered on top of power-related build delays.

The market is shifting from spot allocation to contract-locked capacity. Nvidia has raised supply and capacity commitments to about $279 billion, including a $1.1 billion reservation advance to KYEC and a three-year push to secure about 2.5x substrate capacity. Value is moving to suppliers that can reserve scarce memory, packaging, and substrate capacity years ahead.

Export Controls Are Redirecting Demand Into Regional Supply Chains

China’s shortages of advanced AI chips and HBM are not suppressing demand; they are rerouting it across a fragmented, multi-jurisdiction enforcement regime that is reshaping routing, supply, and share. IDC said Huawei shipped 572,000 AI accelerators in H1 2026, giving it 28% of China’s market, while Nvidia shipped 678,000 units for 34%. Huawei’s gain reflects both domestic substitution and restricted access to foreign supply.

The strategic shift is that export controls are no longer just a U.S.-China trade constraint. Remote access to Blackwell systems in Europe shows compute can still reach Chinese users even when physical shipments are blocked, moving value toward offshore cloud and data-center intermediaries rather than direct chip exports. For operators and vendors, the competitive battleground is shifting from unit shipments to control of routing, hosting, and compliance layers.

AI Bottlenecks Are Repricing the Semiconductor Stack

SK Hynix’s Q3 2026 results show a split market, not a broad semiconductor upcycle: DRAM average selling prices rose in the high-20% range quarter on quarter while shipments grew only about 10%, operating margin reached roughly 80%, and HBM4 began contributing. Management said DRAM and HBM are effectively sold out through 2029, HBM demand exceeds available supply by more than 2x, and long-term agreements already cover over 50% of 2027–28 capacity.

TSMC is showing the same scarcity economics through utilization rather than pricing. Its latest results delivered 41% year-on-year revenue growth in USD and a 66.4% gross margin, while 5/4nm-and-below capacity remains fully utilized through 2026 and 2026 capex is guided to the upper end of $52–56 billion. The spillover is reaching standard memory, with conventional DRAM contract prices expected to rise 10–15% quarter on quarter in Q4 2026 as HBM absorbs capacity. Value is concentrating in AI-critical bottlenecks—HBM, advanced-node foundry, and CoWoS—while legacy segments remain tied to restocking and slower recovery.