AI Power Crunch Squeezes Taiwan Wafer Capacity, Chip Prices Expected to Keep Rising Through H2 2026
Nashnova编辑部
Surging demand for power semiconductors in AI data centers is outpacing Taiwan's foundry expansion, stretching lead times by one to six months. This means → upward pressure on chip prices will persist into at least H2 2026, even as consumer-electronics demand stays soft.
How many chips does one AI rack consume?
AI servers are shifting to high-voltage DC architecture — delivering power at higher voltage to cut conversion losses. A single rack needs roughly 200 power devices.
One 8-inch wafer yields enough chips for only two to three racks. This means → deploying 1,000 AI racks burns through 300-plus wafers, and the supply gap opens from there.
Infineon has announced a major 12-inch capacity expansion, equivalent to roughly 70,000 additional 8-inch wafers, but suppliers still expect demand to remain tight.
Why can't weak consumer electronics cool down foundry pricing?
Power MOSFETs — basic transistors that switch current on and off — are also needed for cooling fans and telecom gear. Production stays on 6-inch and 8-inch wafers, where product variety is high and batch sizes small, limiting capacity flexibility.
European and American chipmakers outsource production to Taiwan, mainly to UMC, Powerchip (PSMC), and Vanguard International Semiconductor. When large offshore orders land, they absorb available capacity fast, crowding out other customers.
This means → even with soft end-consumer demand, foundry prices are expected to rise further in H2 2026. The bottleneck sits on the supply side, not the demand side.
How far have lead times stretched?
Power-semiconductor delivery times have extended by roughly one to one-and-a-half months across the board; some products now take up to six months.
Some suppliers have begun negotiating two-to-three-year capacity agreements with foundries. The last time the industry did this was during the chip shortage in H2 2021.
In plain terms = the 2021 crunch was a short-term, pandemic-driven shock. This cycle is structural, driven by AI — it runs longer and won't self-correct when consumer electronics recover.
How is the AI infrastructure supply chain spreading beyond Taiwan?
Taiwan remains the core manufacturing base: TSMC supplies advanced AI chips, ASE and SPIL provide packaging, Delta Electronics and Lite-On supply power and thermal products.
But production is expanding into the U.S.: Wistron's Fort Worth plant has started building NVIDIA GB300 Grace Blackwell Ultra superchips, and Foxconn is scaling AI server capacity in Houston.
This reflects a geographic risk-diversification push — Arizona is developing into an advanced manufacturing hub, while Texas becomes a server-integration center.
Why is power supply becoming the next bottleneck?
Taiwan's five main AI data centers have a combined capacity of roughly 660 MW. By comparison, South Korea's 11 major facilities total about 4.2 GW — more than six times as much.
South Korea has proposed investing roughly $397.8 billion by 2029 to support 8.4 GW of data-center capacity, with Samsung, SK Hynix, and Naver folding data centers, semiconductors, and renewables into a single infrastructure strategy.
Taiwan projects annual power-demand growth of about 2.5% from 2026 to 2035; four gas-fired units totaling roughly 5.2 GW are expected online by end-2026. But global hyperscalers now demand more than just "available power" — clean energy, long-term PPAs, grid capacity, and supply certainty are becoming prerequisites for site selection.
In plain terms = whether Taiwan keeps attracting AI infrastructure investment no longer hinges on how well it makes chips — it hinges on whether it has enough power, and whether that power is green enough.
Content is for reference only, not financial advice.