On October 27, 2026, Taiwan’s Commercial Times reported that TSMC was considering another round of price increases for its CoWoS advanced packaging services, potentially as much as 20%, citing supply chain sources. These remain reported pricing discussions. But they raise a question worth asking: as AI chips grow larger and more complex, why should packaging prices stay the same?
The commercial case for higher CoWoS prices is strong. Manufacturing is becoming harder. Consistent yields are becoming more valuable. Expanding capacity requires spending money and accepting risk well before the resulting products ship. For customers, those capabilities determine whether expensive compute chips and memory become finished products on schedule.
Packaging’s role in AI has changed. Customers want more compute dies, more memory, and greater bandwidth in each package. That changes the work the packaging supplier must perform. Thinking only in terms of a fee per chip overlooks the additional engineering behind each product.
It also explains how capacity can keep growing while supply remains tight. Larger packages mean fewer units fit within a given processing area. Additional process steps and inspections consume more equipment time. Monthly wafer capacity tells only part of the story: what matters is how many working products each wafer yields and which processing stages those products must pass through.
Expansion may be chasing a moving target. More equipment does not necessarily translate into a proportional increase in shipments.
Yield becomes more valuable as complexity rises.
Kazutoshi Iwata, general manager of Taiwanese equipment maker Eternal Precision Mechanics, illustrated this with substrate lamination: a 98.5% pass rate per cycle leaves roughly 86% passing after 10 cycles; at 95% per cycle, that falls to about 60%. Assuming a constant pass rate and no rework, repeated processing magnifies a small quality gap. In complex packaging, customers must look beyond the quoted fee to the number of usable products delivered.
By the time compute dies and high-bandwidth memory (HBM) reach final integration, substantial manufacturing costs have already been incurred. A failure late in the process can put more accumulated value at risk. Customers therefore have to weigh packaging fees alongside usable output and the supplier’s ability to resolve engineering problems.
Part of TSMC’s pricing power comes from coordinating those processes well enough to deliver consistently. That capability takes time to build and depends on the wider supply chain.
Asking TSMC to expand CoWoS quickly means asking it to line up facilities, equipment, materials, and engineers in advance, then bear the risks of investment and production ramp-up before orders become revenue. If future demand disappoints, depreciation, maintenance, and payroll still have to be paid. A reasonable return must account for that lead time and uncertainty.
For U.S. chip designers and cloud companies, another easily overlooked cost is waiting.
An AI product with a completed design and fabricated compute dies can still be held up by packaging. Until it ships, the investment already made cannot produce the planned returns. Customers deploying computing capacity may also have to delay service launches or revise equipment deployment plans. On-time delivery has measurable economic value.
The purchasing decision is therefore a comparison of total costs. Is it cheaper to pay more for packaging, or to accept delayed shipments, qualify an alternative, and take on yield uncertainty? If CoWoS delivers sufficient advantages on those measures, a higher price can still be the better deal.
A 20% increase in packaging fees also does not translate into a 20% increase in the cost of an AI server. The effect depends on packaging’s share of total cost and the performance, yield, and delivery benefits the customer receives. The price change for one manufacturing step cannot, by itself, tell us whether the economics of the entire system have worsened.
NVIDIA CEO Jensen Huang has already expressed this logic publicly. On June 5, 2024, when asked about TSMC raising chip manufacturing prices, he acknowledged its contribution to the industry and said higher prices could reflect the value it provided.
His response matters because customers can recognize that a supplier’s technology deserves an appropriate return. NVIDIA needs its manufacturing partners to keep investing so that new AI chips can become products delivered at scale. Purchase prices and supply capabilities belong in the same calculation. Squeezing procurement costs today could prove expensive if it weakens future capacity expansion and technology investment.
TSMC Chairman C.C. Wei has addressed the other side of that relationship: how far to raise prices. At the company’s annual shareholders’ meeting on June 4, 2026, he discussed sharply rising memory prices. TSMC, he said, wanted wafer prices to reflect its value and generate profits, while emphasizing that it would not suddenly impose the kind of fourfold increase he cited in memory. The company’s priority was sustainable operation over the long term.
Together, these views point to two conditions for reasonable pricing. Suppliers need returns that support the next round of investment. Customers need enough economic room to keep launching products and expanding their markets. CoWoS price adjustments that reflect manufacturing difficulty, yield, and delivery capabilities fit that relationship. Considering customers’ ability to absorb higher costs, even during a shortage, can also make future cooperation more predictable.
If the market is willing to pay a premium for AI chip architectures and software ecosystems, the manufacturing and integration technologies that make those products possible deserve an appropriate return as well. Expecting suppliers to invest ahead of demand in next-generation capabilities while keeping prices tied to the previous generation is difficult to sustain.
Competition still sets limits. If an increase exceeds the benefits customers gain from yield, performance, and delivery, it strengthens their incentive to redesign products or qualify alternatives. Today’s pricing power still depends on tomorrow’s technical progress.
With product requirements rising, expansion taking time, and customers placing a high value on computing capacity they can actually deploy, higher CoWoS prices make perfect sense. They bring pricing closer to the work, risk, and value involved, while strengthening the economics of the next round of expansion.
Sources
- Zhang Jiarui, “TSMC’s CoWoS Prices Reportedly Set to Rise by Up to 20%,” Commercial Times, October 27, 2026.
- Jules Insight, “Why Is AI Packaging Turning to the Toyota Group Supply Chain? Eternal Precision Mechanics’ Cross-Industry Expansion Experiment.”
- FTV Finance Desk, “TSMC Considers Raising Prices; NVIDIA’s Jensen Huang Says They Reflect the Value Provided,” FTV News, June 5, 2024.
- Chang Chien-chung, “C.C. Wei’s Remarks at TSMC’s Annual Shareholders’ Meeting: Highlights from Competition to Employee Bonuses,” Central News Agency, June 4, 2026.