TSMC and Intel Face Advanced Packaging Bottleneck
Advanced packaging, the CoWoS step that bonds dies together and stacks HBM into a finished chip, has become the near-term bottleneck for AI accelerators, ahead of raw fabrication. Nvidia has reserved most of TSMC's top-tier CoWoS capacity, even as TSMC builds its first U.S. packaging plant in Arizona and adds two Taiwan sites; TSMC's Paul Rousseau told CNBC that CoWoS demand is growing at roughly an 80% compound annual rate. That squeeze is pushing other AI chip designers toward Intel's EMIB and Foveros packaging technologies, which have reportedly won billions of dollars in commitments from customers including Amazon, Cisco, SpaceX, and Tesla. For chip architects and procurement teams, packaging capacity, not logic-node access, is now the gating constraint on scaling next-generation AI accelerators.
For hardware architects and procurement teams, the real capacity constraint on next-generation AI accelerators is no longer transistor fabrication, it is the packaging step that glues dies together and stacks memory on top. Nvidia's preferential claim on TSMC's most advanced CoWoS lines means everyone else building a chiplet-based AI accelerator, from hyperscaler ASICs to smaller AI chip startups, is effectively queuing behind Nvidia, and evaluating Intel's EMIB and Foveros packaging as a second source is now a live procurement decision rather than a hedge.
What happened
Advanced packaging, the assembly step that bonds multiple dies together, integrates high-bandwidth memory, and turns a set of silicon dies into a finished GPU or ASIC, is emerging as the near-term bottleneck for AI compute, ahead of raw chip fabrication. Nvidia has reserved the bulk of TSMC's highest-end CoWoS packaging capacity, even as TSMC expands with its first U.S. packaging facility in Arizona this year and ramps two new sites in Taiwan. Paul Rousseau, TSMC's North America packaging solutions head, told CNBC that CoWoS demand is growing at roughly an 80% compound annual growth rate.
Technical context
CoWoS (Chip on Wafer on Substrate) is what lets a chipmaker combine multiple compute dies with stacked HBM into a single package, something node-level scaling alone can no longer deliver for frontier AI accelerators. Intel's competing EMIB (Embedded Multi-die Interconnect Bridge) and Foveros 3D-stacking technologies are not direct performance equivalents to CoWoS, but analysts say they are mature enough to handle AI inference-class ASICs, and Intel's packaging lines have reportedly attracted billions of dollars in customer commitments this year, with reported customers or active discussions including Amazon, Cisco, SpaceX, and Tesla.
For practitioners
Packaging capacity now directly affects lead times, cost, and architecture choices, monolithic versus chiplet, whether to include HBM, for anyone designing or procuring AI accelerators. Reserved CoWoS capacity for Nvidia-scale customers can force other projects to accept longer schedules or shift topology; Intel's ramp is a real mitigation path, but switching packaging flows means retooling IP, revalidating thermal characteristics, and shifting supply chains, not a drop-in substitution.
What to watch
Actual lead times and utilization at TSMC's CoWoS lines and the ramp schedule for its Arizona packaging site; announced customer deals and volume commitments to Intel's EMIB and Foveros lines; and how HBM supply and interposer material constraints evolve as demand keeps compounding at roughly 80% a year.
Key Points
- 1TSMC's CoWoS advanced packaging, not chip fabrication, is now the near-term bottleneck for AI accelerators, with demand growing about 80% annually.
- 2Nvidia has reserved most of TSMC's top-tier CoWoS capacity, pushing other AI chip designers toward Intel's EMIB and Foveros packaging as an alternative.
- 3Intel's packaging business has reportedly won billions in commitments from customers including Amazon, Cisco, SpaceX, and Tesla as an onshore CoWoS alternative.
Scoring Rationale
Advanced packaging capacity, now confirmed via CNBC's direct reporting and a named TSMC executive, is a well-evidenced, concrete supply-chain constraint with direct implications for AI accelerator lead times, cost, and architecture choices. It affects hardware designers, procurement, and ML infrastructure teams broadly, though it is an incremental capacity/supply story rather than a paradigm shift, keeping it in the solid-to-major band.
Sources
Public references used for this report.
View 6 more sources
- 04Intel's EMIB Challenges TSMC's CoWoS as America's Answer to the ...semiwiki.com
- 05Unveiling the Real Bottlenecks of TSMC - 36氪eu.36kr.com
- 06Apple and NVIDIA are increasingly competing directly with TSMC ...igorslab.de
- 07The CoWoS Crunch: Why TSMC's Specialized Packaging Remains ...markets.financialcontent.com
- 08Inside the AI Bottleneck: CoWoS, HBM, and 2–3nm Capacity ...info.fusionww.com
- 09Nvidia's Feynman Platform Could Clog TSMC's CoWoS Supply ...ainvest.com
Practice with real Ad Tech data
90 SQL & Python problems · 15 industry datasets
250 free problems · No credit card
See all Ad Tech problems

