TSMC raised its 2026 capital expenditure guidance to a range of $52 billion to $56 billion, CEO C.C. Wei said, after conducting diligence with customers on chip demand.1 The upward revision signals foundry executives see sustained order volume from AI accelerator makers well into next year.
SK Hynix, a leading memory chipmaker, listed on the Nasdaq on July 10, 2026, giving it broader access to capital markets as it expands memory production for AI servers.2 Memory supply is tightening industry-wide. At ASML Holding, memory chips accounted for 51% of net system sales in the first quarter, up from 42%, reflecting a shift in lithography-equipment demand toward high-bandwidth memory used in AI accelerators.3 That shift points to more AI accelerator capacity working its way through the supply chain over the next several quarters.
Analysts see the capex increases as an early signal of broader AI infrastructure spending. Vivek Arya said AI data center capital expenditure is expected to grow another 40% to 50% next year.4 That estimate implies foundry and memory suppliers are scaling capacity ahead of, not behind, demand from cloud providers.
The pattern raises a testable question for the next two to three quarters: will Meta, Alphabet, and Microsoft raise their own capex guidance beyond prior trend lines after TSMC and SK Hynix moved first? Foundry capacity commitments have historically preceded hyperscaler spending increases, since chipmakers set multi-year investment plans based on forward orders from cloud and AI customers.
If Meta, Alphabet, and Microsoft report capex guidance revisions that exceed their historical growth trend in upcoming quarters, it would confirm foundries and memory makers correctly read early demand signals from hyperscalers. A miss would suggest the current capacity buildout is running ahead of confirmed orders.
For now, the signal points toward looser AI hardware availability, not tighter. TSMC and SK Hynix are committing capital first, and the hyperscaler capex response has not yet been reported.