The demand for AI infrastructure is extending pressure from GPUs and HBM memory to a much broader part of the semiconductor industry. Taiwanese foundries face a second half of 2026 with increasingly tight capacity and new price hikes, both in advanced processes and in certain mature technologies used for power management, interconnects, photonics, and other components essential for AI servers.
The key factors of AI pressure on Taiwanese factories in 20 seconds
- AI is occupying more capacity in both advanced processes and mature nodes.
- The average utilization of global 8-inch fabs could reach 90% in the second half of 2026.
- Prices for these processes have already increased by between 5% and 15%.
- TSMC is particularly tightening its 3 nm nodes.
- TrendForce expects continued pressure on capacity and prices during 2027.
This phenomenon helps explain why the current AI investment cycle is having much broader effects than the focused attention on NVIDIA and large accelerators might suggest. Building an AI server requires numerous additional semiconductors: power management circuits, controllers, network components, interposers, optical devices, and specialized chips.
A significant portion continues to be manufactured with mature technologies.
And here appears a second bottleneck.
TrendForce estimates that the average utilization of 8-inch fabs among the ten largest foundries will reach around 90% in the second half of 2026, compared to the 88% average forecast for the entire year.
The demand increase coincides also with a structural reduction in certain older lines. TSMC and Samsung have been reallocating investments from mature nodes toward advanced processes and advanced packaging, precisely the technologies where AI customers are willing to pay a premium.
The result is paradoxical: while billions are being invested in new factories, certain types of available capacity are decreasing at the same time.
From 3 nm to power chips: AI needs many more wafers
The most evident pressure continues to be in advanced processes.
TSMC remains the leading contract manufacturer globally and one of the biggest beneficiaries of AI spending. In Q2 2026, it recorded $40.2 billion in revenue, with expectations of between $44.6 billion and $45.8 billion for Q3.
Demand comes from GPUs, internally designed accelerators by major cloud providers, and other advanced circuits.
Counterpoint Research estimates that the revenue of the expanded global foundry market grew by 23% year-over-year during Q1 2026, reaching $86 billion, mainly driven by GPU and AI ASICs. TSMC’s revenues increased by 41% year-over-year during that period.
But extraordinary demand doesn’t allow capacity to grow quickly.
TSMC President and CEO C.C. Wei reminded earlier this year that building a new fab takes two to three years. Even a sharp immediate increase in investment has limited effect on wafers produced in 2026 and only begins to contribute partially in 2027.
This gap between the pace of AI demand growth and the time needed to build factories is keeping supply tight.
Supply chain estimates from June placed TSMC’s monthly capacity for 3 nm wafers between 160,000 and 175,000 in Q2, which was not enough to fully meet demand. Some industry sources then pointed to possible price increases of up to 15% during the second half of the year.
Counterpoint also notes a similar situation: their data indicates an accumulated increase of approximately 8% to 15% for TSMC’s N3 process during 2026, with N3 and N2 expected to remain without significant relief in the short term.
However, pressure is no longer limited to the most modern nodes.
8-inch fabs are back at the center of the market
One of the less visible effects of AI growth is emerging in facilities that use 8-inch (200 mm) wafers.
These fabs do not produce the most advanced GPUs, but they remain critical for numerous analog and power components.
AI servers require large quantities of power management integrated circuits (PMICs) and discrete power devices. As the energy density of racks increases, so does the electronics needed to distribute, convert, and control that electricity.
TrendForce notes that this demand increase coincides with production cuts of 8-inch wafers carried out by major manufacturers like TSMC and Samsung.
Prices for foundry services in these lines rose between 5% and 15% from Q1 to Q2 2026, depending on the manufacturer and process. The sector is also preparing a third round of adjustments that could start during the second half of the year and extend into 2027.
This tension directly affects specialized Taiwanese foundries.
Vanguard International Semiconductor, for example, has acknowledged that AI-related demand is growing beyond available capacity.
Its President, Fang Leuh, considers this a structural growth and expects shortages to be especially evident in 2027. The company plans to invest between 60,000 and 70,000 million Taiwanese dollars, approximately $1.85-$2.16 billion, mainly in its new 12-inch factory in Singapore.
Vanguard is also negotiating new prices with customers for 2027 and anticipates that price increases next year will not be lower than those in 2026.
UMC, Vanguard, and PSMC also feel the new demand
The shift is gradually affecting nearly all major Taiwanese foundries.
Latest data from TrendForce show that TSMC, UMC, Vanguard, and PSMC increased their revenues in Q2 2026, supported by a combination of high demand and higher prices.
United Microelectronics Corporation (UMC) has also increased investments to expand capacity, especially in Singapore and Taiwan.
This movement is significant because AI demand is reaching technologies previously considered peripheral to large accelerators.
TrendForce highlights rising demand for power devices of 55 nm or larger, silicon interposers of 65/55 nm, and FPGAs manufactured in 40/28 nm processes. Technologies related to silicon photonics, integrated capacitors, and components used in advanced packaging are also occupying capacity.
PSMC provides another example of how production is reorganizing.
The logical capacity of 12-inch lines is gradually decreasing because certain HBM-related processes are occupying part of these resources. According to TrendForce, PSMC made an initial price hike in Q1 and was preparing another for Q3.
Therefore, the pressure does not only stem from companies needing to produce more chips.
There is also internal competition among the factories themselves.
Manufacturers are dedicating more space to high-demand, higher-margin AI-related components while reducing or containing production of less profitable products.
Pricing pressures could extend far beyond GPUs
This shift has consequences for almost the entire electronics industry.
When a foundry allocates capacity to AI chips, that same line becomes unavailable for other products. If old factories are gradually phased out, supply could tighten just as new needs emerge.
Hence, price pressures may end up affecting products that seemingly have little relation to ChatGPT, large language models, or data centers.
Automobiles, industrial equipment, telecommunications, computers, consumer electronics, and energy systems all use many of the same mature processes.
The situation is also benefiting manufacturers outside Taiwan.
Samsung has recently increased prices for certain foundry services after its 4 nm capacity in Pyeongtaek reached high utilization levels. Simultaneously, difficulties in obtaining capacity at TSMC are prompting some clients to explore alternatives.
Therefore, the competition isn’t necessarily leading to a price war. With several lines close to full capacity, manufacturers have more margin to pass costs and investments on to their customers.
And the problem doesn’t seem limited to 2026.
TrendForce expects an ongoing upward trend in prices for mature nodes through 2027, and Counterpoint does not anticipate significant relief in the N3 and N2 advanced processes either.
The industry is responding with new factories, expansions, and increased investment. But there’s an inevitable lag between deciding to build a production line and obtaining the first commercial wafers.
While this new supply is being developed, Taiwan will continue to hold a critical share of global capacity and a major part of the tension caused by AI growth.
The AI bottleneck can no longer be measured only in GPUs. It’s also increasingly measured in wafers, production lines, and the years required to build new factories.
Frequently Asked Questions
Why is AI driving up chip manufacturing costs?
AI servers require advanced GPUs or ASICs, but also numerous power, communications, interconnect, and control chips. This additional demand is occupying capacity across both advanced nodes and certain mature factories.
How much have 8-inch fab prices increased?
TrendForce estimates average increases of between 5% and 15% from Q1 to Q2 2026, with new hikes planned for the second half of the year and into 2027.
Is TSMC having trouble meeting demand?
Demand for advanced processes remains very high. Market estimates indicate that 3 nm capacity is still tight despite expansion efforts, as TSMC continues increasing investments and building new facilities.
When might supply improve?
There is no specific date, as it depends on the process. Analyses from TrendForce and Counterpoint suggest that some constraints and price pressures will persist through 2027, especially in technology areas directly related to AI infrastructure.
