When TSMC released its fourth-quarter results, one number went largely unnoticed in the financial press.
Nvidia had reserved 60% of the world’s advanced packaging capacity at TSMC for 2026 — roughly 595,000 wafers, with 510,000 of those allocated to a single technology called CoWoS-L.
On its own, Jensen Huang’s company is absorbing more capacity than the next ten customers combined.
What should concern investors is what comes next. AMD, Broadcom, Google, Amazon — every other player in the AI race — are fighting over the remaining 40% in what one OSAT executive described to DigiTimes as a bidding war.
And total capacity, even after quadrupling in two years, still falls short.
TSMC CEO C.C. Wei wasn’t particularly diplomatic about it on the latest earnings call. CoWoS capacity, in his words, is sold out through 2025 and well into 2026. He added that supply chain constraints could affect production timelines — diplomatic phrasing for the fact that TSMC cannot deliver enough AI chips even though the wafers themselves exist.
This is where most analyses stop. The real story is somewhere else.
The misunderstanding about the GPU shortage
For two years, the industry has explained Nvidia’s supply constraints by pointing to the difficulty of manufacturing the chips themselves.
That’s an incomplete picture. Silicon is rarely the actual problem. TSMC produces 3nm wafers in adequate volumes. The bottleneck sits one step downstream: assembling the GPU with its HBM memory onto a single substrate to create a functional AI chip.
That step is called CoWoS — Chip-on-Wafer-on-Substrate. It’s a 2.5D packaging technique that TSMC has refined over a decade, and no Western competitor truly matches it at scale. Intel has its own equivalents (Foveros, EMIB), Samsung is trying to catch up, but for the AI volumes the industry actually needs today, TSMC stands more or less alone.
Paul Rousseau, who runs TSMC’s North American packaging solutions, gave a rare interview to CNBC in April in which he described 80% annual growth for this line of business. That’s faster than Nvidia itself is growing. And it isn’t enough.
In concrete terms, TSMC has gone from roughly 35,000 CoWoS wafers per month at the end of 2024 to a target of 130,000 by the end of 2026. Quadrupling capacity in 24 months on a process this complex has no precedent in the industry. Even so, the pressure is severe enough that TSMC is now subcontracting parts of the process to ASE Technology and Amkor — something that would have been unthinkable three years ago.
An entire supply chain hiding in plain sight
Packaging is just one of three simultaneous chokepoints in the semiconductor backend. The other two deserve as much attention.
HBM memory. On its fiscal Q1 2026 earnings call, Micron confirmed that its HBM capacity for calendar year 2026 was fully committed. SK Hynix, which holds 46-49% of the HBM market, wrapped up its 2026 negotiations as early as October 2025. Samsung is on the same path.
Micron’s DRAM gross margins moved from roughly 22% in fiscal 2024 to over 50% last quarter — a structural shift, not a cyclical one. According to several industry sources, Samsung is pushing through 15-22% price increases on 2026 HBM contracts. When a supplier can impose that on customers the size of Microsoft or Meta, you’re looking at a one-sided negotiation.
ABF substrates. This part of the story deserves its own paragraph because it’s both critical and oddly absent from public discussion. Every advanced processor in the world — every GPU, every high-end CPU, every Apple Silicon chip — is mounted on a substrate called ABF, short for Ajinomoto Build-up Film.
The resin is licensed by Ajinomoto, the Japanese conglomerate better known for its monosodium glutamate and bouillon cubes. Ajinomoto controls roughly 98% of the relevant intellectual property, which it licenses to a small group of East Asian manufacturers: Unimicron in Taiwan, Ibiden and Shinko in Japan, Nan Ya and Kinsus.
The United States has no domestic source of ABF. During the 2021-2022 shortage, Broadcom’s lead times stretched to 70 weeks because of this resin — more than sixteen months for a printed circuit substrate. The top five suppliers control around 70-74% of the global market, and all their high-end lines are sold out through 2026.
The least-followed oligopoly in tech
The most interesting investment opportunity, in my view, sits one level further upstream. To produce CoWoS, HBM, or any 2.5D or 3D package, you need highly specialized equipment supplied by a handful of manufacturers. The same names keep showing up in the technical literature.
Disco Corporation, based in Japan, holds about 20% of the back-end equipment market and dominates wafer thinning and dicing. Its share price has tripled in two years, but the stock remains largely absent from Western portfolios.
BE Semiconductor — Besi — based in the Netherlands, holds 42% of the die-attach market and, crucially, the global lead in hybrid bonding, the technology that should succeed CoWoS by the end of the decade.
Besi’s order backlog rose 105% year-over-year in Q4 2024, mostly driven by hybrid bonding orders. ASMPT in Singapore now derives roughly 25% of its revenue from advanced packaging. Kulicke & Soffa and Korean firm Hanmi round out the field.
Together with EV Group and SUSS MicroTec in Europe, these companies control between 60% and 65% of the global hybrid bonding market according to MarketsandMarkets data.
One event from April 2025 is worth dwelling on. Applied Materials, the American semiconductor equipment giant, took a 9% stake in Besi. For a company of Applied’s size, allocating several hundred million dollars to a Dutch mid-cap is not a passive investment. It’s a strategic signal that the future of advanced packaging hinges on sub-micron precision in copper-to-copper bonding, and that nobody else can do it at this level.
The historical parallels that should worry investors
Regular readers will recognize the pattern. When I wrote about electrical transformers in early 2025, the industry was in a similar position: structural demand exploding, physical capacity unable to keep up, and an oligopoly of specialized manufacturers that the market was failing to value correctly. The submarine cable piece told the same story: 60 ships in the world, four vertically integrated manufacturers, and an AI capex cycle that depends entirely on invisible infrastructure.
Advanced packaging follows the same logic. Building a new advanced packaging fab takes three years. A hybrid bonding tool costs tens of millions of dollars and has a lead time of 12 to 18 months. Skilled labor for these processes is rare, concentrated in Taiwan, Korea, and Japan, and difficult to train.
Meanwhile Amkor announced a 2026 capex plan of $2.5 to $3 billion — roughly three times its prior-year spending — and is investing $7 billion in a new Arizona facility tied in part to CHIPS Act contracts and “Made in America” defense chip requirements. ASE expects its advanced packaging revenues to grow from $600 million in 2024 to $1.6 billion in 2025. The company now says advanced products account for more than 80% of its total revenue.
All of that money ultimately ends up with the five or six equipment manufacturers capable of supplying the necessary tools.
Why the market is missing this
Several things explain why the sector is so poorly covered. The first is geographic. Most of the interesting companies trade in Tokyo, Taipei, Singapore, or Amsterdam. Disco has no liquid ADR. Besi is a European mid-cap. ASMPT trades in Hong Kong. For the average American fund, these tickers are administratively awkward and invisible in standard benchmarks.
The second is sectoral. These companies sit in the “semiconductor equipment” bucket in most databases, lumped together with ASML and Applied Materials. Passive ETF flows go to the front-end leaders. Few sell-side analysts cover them in depth.
The third reason is more fundamental. For a decade, OSAT and packaging were low-margin businesses, treated as cyclical commodities. The shift toward high-end advanced packaging — where TSMC’s gross margins reportedly approach 80% on certain lines — is a business-model change that many analysts haven’t fully digested yet.
Catalysts over the next 18 months
Five specific events should, by my read, materially reprice this sector.
The ramp of Nvidia’s Rubin platform, expected late 2026 and early 2027. This generation will consume most of the available CoWoS-L capacity, and the associated equipment supply contracts should be announced gradually.
The transition to HBM4. Reported pricing runs roughly 30% above HBM3E, around $500 per module according to industry estimates. The transition also reshuffles the supplier hierarchy: SK Hynix has partnered with TSMC for the 12nm base die, while Samsung is pushing its own 1c node.
Mass adoption of hybrid bonding. The relevant equipment market is projected to grow from about $152 million in 2025 to nearly $397 million in 2030, a 21% CAGR. Besi captures a disproportionate share of that growth.
The buildout of US capacity under the CHIPS Act. The new Arizona fabs will need to be equipped by the same five-to-seven global suppliers — there’s no domestic American alternative for most of these tools.
And one tail risk worth flagging: Nvidia is currently testing an alternative technology called CoWoP, or Chip-on-Wafer-on-Platform, which would mount the interposer directly on PCB and bypass the ABF substrate entirely. If that approach is validated for the Grace Rubin 150 platform expected late 2026, it could restructure the entire ABF ecosystem.
What premium subscribers get
I spent six weeks mapping this supply chain using the same methodology as the submarine cable and transformer pieces.
The full report contains my 9-position portfolio organized in three categories: backend equipment makers, substrate and memory suppliers, and OSATs. It includes a non-consensus view on Besi that diverges from the mainstream sell-side take, plus a European mid-cap directly exposed to CHIPS Act contracts that few analysts cover.
The report also breaks down real margins by technology (CoWoS-S vs CoWoS-L, hybrid bonding D2W vs W2W), the unit economics of Amkor’s new Arizona capacity, and the sensitivity of each name to a CoWoP scenario. The short thesis covers an ABF supplier whose valuation assumes the current architecture remains dominant indefinitely — a bet that could age poorly if Nvidia shifts direction in 2027.
The catalyst calendar lays out earnings dates, HBM4 qualification milestones at Nvidia, and the two strategic decisions from Jensen Huang that could reprice the entire chain by Q3 2026.