The consensus narrative is that AI demand is pulling the entire semiconductor complex higher. Worldchip sales in April were up 93.9% year-over-year. AMD is up more than 130% year-to-date. SK Hynix, Samsung, and Micron have all breached the $1 trillion market capitalization threshold within the same week in late May 2026. The story feels like a single megatrend.
That story is incomplete. The semiconductor complex is not being driven by one demand wave. It is being restructured by supply-side cannibalization. The recovery is not broad. It is concentrated, and the concentration is the point.
The HBM diversion
High-bandwidth memory - the specialized DRAM stacked directly onto AI processors - now accounts for roughly 20% to 25% of total global DRAM wafer production. A single gigabyte of HBM consumes more wafer area, more packaging capacity, and more process steps than an equivalent amount of commodity DRAM used in smartphones and PCs. That means every megawatt of new AI data center capacity does not just add demand. It subtracts from the supply available for everything else.
The result is a demand cannibalization event that has pushed conventional DRAM prices up approximately 60% in 2025, with another 30% to 40% increase forecast through 2026. HSBC estimates an 8% HBM supply deficit for 2026. Morgan Stanley raised its 2026 DRAM price increase forecast from 26% to 30% in October 2025 as the squeeze tightened. This is not a cycle where demand pulled supply along behind it. This is a cycle where HBM is consuming the manufacturing capacity that commodity segments used to share.
The implication is straightforward. SK Hynix, Samsung, and Micron are not sitting on $1 trillion valuations because they are growing volume into new end markets. They are sitting there because they control the allocation decision - which wafers become HBM at premium pricing, and which become commodity DRAM at now-inflated pricing. Pricing power, not unit growth, is carrying the recovery.
The capex concentration
The same concentration dynamic is visible in capital spending. Total semiconductor capital expenditure for 2026 is estimated at approximately $200 billion, up roughly 20% from 2025. Five companies alone - TSMC, Samsung, SK Hynix, Intel, and Micron - are committing approximately $180 billion of that total. That is 90% of the industry's capex flowing through five balance sheets.
TSMC is guiding $52 billion to $56 billion in capex for 2026, up 27% to 37% year-over-year. It now invests approximately 60% of its revenue in capital expenditure. Samsung is adding roughly 1 billion DRAM bits of capacity at its M15X fab and accelerating 1c node conversions at M14 and M16 for HBM and server memory production. Meanwhile, the three largest mature-node foundries - GlobalFoundries, UMC, and SMIC - are spending billions to expand capacity at 22nm and above, chasing a completely different set of customers.
This is the two-market split. The advanced node market, dominated by TSMC and Samsung, is absorbing the vast majority of incremental capex, building out 2nm and beyond alongside advanced packaging capacity. The mature node market is running a parallel expansion at trailing-edge processes, serving automotive, industrial, and IoT chips that have no path to 2nm but still require capacity.
The structural divergence was confirmed in May 2026 when GlobalFoundries announced its first-ever dividend - a signal that the mature-node foundry is shifting from pure growth capex toward cash returns, even as TSMC deepens the investment gap.
What the market is misattributing
AMD's first quarter 2026 results illustrate the demand-side narrative that the market defaults to. Revenue of $10.3 billion, up 38% year-over-year. Data center revenue of $5.8 billion, up 57%. The stock surged 16% on the earnings print. The consensus read: AI chip demand is broadening beyond Nvidia.
That reading is not wrong, but it misses the supply-side constraint that is amplifying the move. AMD's data center growth is not happening in a market with infinite foundry and memory capacity. It is happening in a market where TSMC's advanced packaging capacity is allocated, where HBM supply has an estimated 8% deficit, and where the foundry bottleneck determines who ships and who waits. AMD's revenue growth is partly a function of having secured capacity in a constrained system - not purely a function of superior product demand.
The same applies to the semiconductor equipment sector. Wafer fabrication equipment sales are projected to expand 9% in 2026, with total WFE (wafer fabrication equipment) segment value exceeding $112 billion. But the spending is concentrated in the same five companies that dominate capex. The equipment vendors whose portfolios align with advanced node and packaging expansion capture the growth. The rest follow a much flatter trajectory.
Investor Takeaway
The semiconductor recovery is not being driven by a broad-based surge in chip demand. It is being driven by a structural reallocation of wafer capacity and capital expenditure, where HBM production diverts capacity from commodity segments, pricing power concentrates in the top three memory makers, and capex flows almost entirely through five balance sheets. The question is not whether demand is strong. The question is how long the supply constraint holds.
If SK Hynix, Samsung, and Micron maintain their current capex discipline - prioritizing HBM yield improvement and node migration over broad capacity expansion - the pricing environment sustains. If they respond to $1 trillion valuations by accelerating general-purpose capacity, the commodity DRAM and NAND segments will see the oversupply dynamics of the last cycle return.
The key issue is not whether AI demand remains healthy. The more important question is whether the companies that control wafer allocation keep treating supply as the scarce resource. If they do, the current structure holds. If they treat the rally as license to build, it does not.