Semiconductors: $200B of CapEx Won't Matter if These 4 Companies Aren't Getting Paid

Hyperscaler spend is only useful when it becomes supplier revenue

The headline number is huge. The investable number is smaller.

The signal is revenue, not capex guidance

The four largest U.S. hyperscalers have laid out roughly $720–745 billion of 2026 capex guidance, and analysts attribute about $250–490 billion of that to AI infrastructure. That is still an enormous demand engine. But hyperscaler spending is not the same thing as supplier revenue. A large share of that money can be absorbed by construction, networking, power, leasing, and multi-tier supply-chain costs before it reaches chip makers. The cleaner signal is where cash becomes reported revenue.

Why four choke points matter more than the whole stack

What can actually be traced is supplier revenue, and the available framework says that flow concentrates in four layers: accelerators, leading-edge foundry, high-bandwidth memory, and fab equipment. For investors, those are the clearest places to look for payout rather than for the broadest narrative.

Bulls see the largest corporate capex cycle in recent memory. Bears point to ROI questions and capex fatigue. Both can be true. The near-term question, though, is simpler: who gets paid first, and who gets paid most?

TSMC shows how semiconductor capex is concentrating

The key shift is not that AI demand exists. It is that spending is funnelling through fewer hands.

TSMC is becoming the main foundry bottleneck

Semiconductor capex is rising from $166 billion in 2025 to $200 billion in 2026, while TSMC alone was 25% of total spending in 2025 and now projects $52 billion to $56 billion in 2026 capex. Management says 5G, AI, and HPC are driving the increase, while most other foundries are guiding flat to lower spending. That is what concentration looks like: the ecosystem expands, but the number of major cash collectors stays small.

TSMC's own financial position shows the commitment is real. In Q2, it generated nearly TWD 783 billion in operating cash, spent TWD 496 billion on capex, and still ended the quarter with TWD 3.1 trillion in cash. It also raised its full-year budget to $60 billion to $64 billion, with roughly 70% to 80% going to advanced processes and 10% to 20% going to advanced packaging and related areas. When one foundry is fronting so much of the expansion, the bottleneck is becoming the collector.

AI demand is already shaping TSMC's mix

HPC reached 66% of TSMC revenue, up from 58% a year ago, and TSMC is directing 70% to 80% advanced-process allocation from its raised capital budget. That suggests AI-related demand is not just a future story; it is already influencing capacity planning.

The risk is that this strength depends on a narrow demand pool. If hyperscaler spending stays firm, TSMC keeps acting as a central toll booth in the AI stack. If that demand weakens, the market may quickly punish the company that looks like the clearest proof the cycle is real.

Memory is the second payout lane, with HBM at the centre

That payout funnel does not stop at the foundry.

Memory already dominates the next capex cycle

The memory segment is expected to grow 30% in 2026, and memory companies are projected to account for 45% of 2026 semiconductor capex. That makes memory a central part of the AI buildout, not a side story.

The mechanism is straightforward: AI training and inference systems require more memory and storage per server. As compute capacity expands, the memory wall matters more, which is why HBM is becoming such an important payout zone.

Why SK Hynix and Micron deserve attention

The capex filings are becoming more telling. Micron Technology and SK Hynix each should increase CapEx by over 40% in 2026, which shows these companies are committing capital where demand is already visible rather than waiting for a more distant AI economy.

SK Hynix also has the clearer HBM positioning in the current outlook. Its 2026 market commentary names it as the primary anchor of this shift because of its ability to deliver HBM3E and next-generation HBM4 at scale. For investors, that matters because HBM leadership can translate into revenue durability once suppliers are qualified with major AI customers.

Samsung is still a giant, but the payout is less clean because a large part of its spending is spread across more businesses. The higher-growth AI memory exposure sits more directly with Micron and SK Hynix.

The bear case is about timing, not direction

Bears are not wrong to flag future pricing pressure. One early signal is that IDM CapEx should decline again in 2026 by about 9% even as memory firms spend more. That suggests the market sees a near-term shortage in AI memory, but it also points to the risk of a longer-term margin fight if supply expands too quickly.

For now, the clearest watch items are capex filings, HBM supply announcements, and whether scarcity is still converting into margin.

The smart-money map: focus on the toll booths

The smart money is not buying the highway. It is buying the toll booths. Hyperscalers may be pointing to roughly $720–745 billion of 2026 capex, but that trail only matters where it becomes supplier revenue: accelerators, leading-edge foundry, HBM, and fab equipment. TSMC's raised US$60 billion to US$64 billion budget is the clearest signal that the bottleneck still has to expand.

Catalysts to watch

• TSMC: updates on its HPC segment mix and 3nm contribution to wafer revenue. If advanced nodes keep gaining share, the foundry toll-road thesis remains intact.
• SK hynix and Micron: signs of stronger HBM3E and next-gen HBM4 traction and more aggressive capex deployment into AI memory demand. SK hynix is already seen as the primary anchor of this shift in memory.

Red flags

• TSMC raises spending, but the revenue mix stops leaning further into AI-led demand.
• Pricing pressure appears as advanced capacity expands.
• Customer concentration starts to matter more than demand strength.

The invalidation test is simple: if TSMC keeps the raised budget but HPC momentum stalls, pricing slips, or customer concentration becomes the bigger overhang, the thesis weakens.

Capex is the rumor. Revenue receipt is the proof.