Advanced Packaging Shortage Becomes the Binding Constraint on AI Chip Supply

The tightest point in the artificial intelligence hardware supply chain has moved away from wafer fabrication and settled in the back end of the process. Michael Mertin, chief executive of Austrian substrate maker AT&S Austria Technologie & Systemtechnik AG, told Nikkei Asia on July 27, 2026 during a visit to Malaysia that the bottleneck is no longer chip fabrication but packaging. His comment matches what buyers of AI accelerators have been reporting since the start of the year, namely that finished dies are available faster than the capacity needed to turn them into shippable products.

Modern AI accelerators are not single pieces of silicon. Nvidia Blackwell and Rubin processors, AMD Instinct parts and the custom chips designed by Google, Amazon and Meta are assemblies of several small dies placed next to stacks of high bandwidth memory. Connecting them requires TSMC advanced packaging platform known as CoWoS, short for Chip on Wafer on Substrate. The logic die and the memory stacks sit on a passive silicon bridge called an interposer, and the whole assembly is then mounted on an organic substrate that links it to the server board. Without that step, a fully working 3 nanometer wafer cannot become a usable accelerator.

TSMC has confirmed the squeeze repeatedly. At the annual shareholders meeting held on June 4, 2026, chief executive C.C. Wei described CoWoS capacity as extremely tight and sold out through 2026. Independent allocation trackers put lead times at 52 to 78 weeks, with bookings already running into 2027. The company has been expanding aggressively, moving from roughly 35,000 wafers per month at the end of 2024 to about 75,000 at the end of 2025, with a target between 125,000 and 130,000 by the close of 2026. Even that near fourfold increase falls short. Demand for 2026 is estimated at close to one million wafers while supply is running at around 80 percent of that level.

The second quarter results published on July 16, 2026 reinforced the picture rather than easing it. TSMC reported record revenue of 40.2 billion dollars, up 36 percent year on year, with net profit rising 77.4 percent and gross margin reaching 67.7 percent. Management raised the full year outlook for AI related growth above 40 percent, guided third quarter revenue to between 44.6 and 45.8 billion dollars, and confirmed that advanced packaging remains fully booked with lead times above one year. Executives also pointed to a roadmap for packages larger than 14 reticle sizes and to early work on glass carriers, both of which are multi year projects rather than near term relief.

Allocation is highly concentrated. Nvidia is estimated to hold close to 60 percent of global CoWoS capacity for 2026, equivalent to roughly 595,000 wafers, drawn from TSMC lines and from additional capacity at outsourced assembly firms Amkor and ASE. Together with Broadcom and AMD, the top three customers absorb an estimated 85 percent or more of what is available. Smaller designers, including well funded startups and the internal chip teams at cloud operators, are left to compete for the remainder. Prices reflect the imbalance, with packaging costs rising two to four times faster than the underlying wafers.

A second and structurally different problem sits underneath the packaging crunch. The insulating film used inside every high performance chip substrate, known as Ajinomoto Build-up Film or ABF, comes almost entirely from one source. Ajinomoto Fine-Techno, a subsidiary of the Japanese food and chemical group best known for monosodium glutamate, holds about 95 percent of the world market. Capacity constraints in packaging can be solved with time and capital. A single supplier holding almost all output of an irreplaceable material is a dependency that cannot be engineered away on the same timescale.

Ajinomoto notified substrate manufacturers in May 2026 that it would raise ABF film prices by approximately 30 percent, with the new terms taking effect during the third quarter of 2026. Those prices are now coming into force. Taiwanese substrate makers confirmed receipt of the notice and have been negotiating terms since. Combined with comparable increases on copper clad laminate materials from Resonac and Mitsubishi Gas Chemical, the change is pushing costs across the whole substrate chain. For substrate manufacturers, a 30 percent increase in film cost translates into a 3 to 6 percent rise in finished substrate prices. Spot prices for ABF substrates have already climbed by more than 30 percent, while contract prices are moving up 5 to 10 percent in the second half of 2026.

The volume gap is widening rather than closing. Industry projections place the ABF supply shortfall at around 10 percent in the second half of 2026, close to 21 percent in 2027 and potentially above 40 percent by 2028. The driver is not only unit growth. As AI chip packaging moves from three layer configurations toward eleven layer designs, and toward thirteen layer designs after 2030, the amount of film consumed per chip multiplies even without extra volume. Lead times for finished ABF substrates are running 16 to 24 weeks and are expected to stay there through 2027.

Capacity responses are underway but slow. Samsung Electro-Mechanics has committed 1.2 billion dollars to new ABF substrate lines at its plants in Sejong, South Korea and Thai Nguyen, Vietnam, with volume production not expected before the third quarter of 2027. Its chief executive told shareholders in March 2026 that demand for high end substrates already exceeded the company capacity by more than 50 percent. Ajinomoto has bought land in Gifu Prefecture for a third film plant, a project valued at 1.2 billion yen for the site alone, with construction due to start in 2028 and mass production targeted for 2032. Taiwan three leading substrate fabricators, Unimicron, Kinsus Interconnect and Nan Ya PCB, are running at full utilisation and are expected to remain there through the end of the year, and all have raised capital spending plans for 2026.

AT&S is positioning itself as the western alternative. On June 15, 2026 the company announced an expansion of its Kulim site in Malaysia worth up to 2 billion euros, backed by long term commitments from AMD and a second technology customer it has not named publicly. Alongside the announcement it lifted guidance for the 2026 and 2027 financial year to revenue growth of 45 to 55 percent at constant currency, an EBITDA margin between 32 and 37 percent, and capital spending of 1 to 1.2 billion euros. The group reported revenue of 1.8 billion euros for its previous financial year, up 21 percent on a currency adjusted basis. Its shares, worth about 13 euros in early 2025, reached a record near 200 euros in June 2026.

Mertin argues the crisis has also changed how the industry has to work. Substrates are no longer passive carriers, because power delivery and signal density inside the package have become primary design constraints. Suppliers now need to be involved in chip design from the earliest stages, sharing roadmaps with their customers rather than simply building to a specification. That raises the barrier for any new entrant, since manufacturing scale alone is not enough to win qualified work.

What separates this shortage from the semiconductor glut that followed the 2021 and 2022 crunch is the absence of inventory. That earlier episode corrected itself as excess parts worked through distribution channels. Film, specialty glass cloth and advanced substrates are consumed during production and never accumulate as stock. At the same time, spending by cloud operators is committed years in advance, so there is no consumer slowdown available to soften demand. New capacity in film, substrates and packaging takes two to four years to move from decision to volume output.

As of late July 2026 the situation remains active and is expected to persist. Buyers planning AI infrastructure should assume that packaging and substrate supply adds at least one to two years to delivery schedules on top of any wafer lead time, which makes commitments signed now the determining factor for deployments planned in 2027 and 2028.