The Main Battleground Moves to Back-End Packaging — TSMC's CoWoS and Intel's Foveros/EMIB

The battleground shifts from front end to back end

AI accelerator performance can no longer be scaled by front-end miniaturization alone. As the cost and difficulty of shrinking transistors rise, how closely, widely, and densely you can "package" the compute die with high-bandwidth memory (HBM) has become the main battleground for performance and cost. At its center sits 2.5D/3D advanced packaging, epitomized by TSMC's CoWoS.

For buyers, this is not someone else's problem. AI-chip supply is rate-limited not only by the front end but by back-end packaging capacity. What TSMC and Intel add, when, and where in this space is a leading indicator of supply continuity and performance.

TSMC — CoWoS as the foundation of AI

TSMC's CoWoS (Chip-on-Wafer-on-Substrate) places multiple processor cores and HBM stacks side by side on a single interposer — a 2.5D packaging technology that has become the core of AI accelerator performance. It is precisely this structure that has made CoWoS capacity tightness translate directly into AI-chip supply tightness.

TSMC is not keeping CoWoS confined to AI; it is widening its reach. For automotive, it is advancing CoWoS-R toward AEC-Q100 Grade 2 qualification, targeting certification in Q4 2025. Advanced packaging stepping into the high-reliability automotive domain is a sign that back-end demand is spreading beyond AI. It also plans a CoWoS-based System-on-Wafer (SoW) in 2027, scaling package area to the wafer level. In addition, it has outlined a plan to integrate COUPE (Compact Universal Photonic Engine) into CoWoS as co-packaged optics (CPO) in 2026, bringing optical connections inside the package — a move to address the bandwidth and power limits of electrical wiring with light.

Intel — building "in-house packaging" with Foveros/EMIB

With TSMC ahead in back-end, Intel is countering with in-house advanced-packaging capability. Intel opened Fab 9 in New Mexico with a $3.5 billion investment, bringing an advanced-packaging volume site online. Fab 9 is co-located with Fab 11x on the same site, enabling end-to-end advanced-packaging manufacturing.

There are two technology pillars. Foveros is a 3D packaging technology that vertically stacks compute tiles, combining multiple chiplets to optimize cost and power efficiency. EMIB is an embedded bridge that connects multiple dies at low cost. Intel also unveiled new ASAT (Advanced System Assembly and Test) capabilities that provide assembly and test end to end, aiming to offer foundry customers everything through packaging. Intel positions Fab 9 as key to extending Moore's Law beyond 2030 and reaching one trillion transistors, placing packaging at the center of performance scaling.

How to read the two-company race

TSMC's and Intel's approaches are two sides of the same recognition that back-end is now the main battleground. TSMC captures scale through overwhelming CoWoS capacity and expansion (automotive, optics); Intel pursues the strength of vertical integration through in-house end-to-end manufacturing (Fab 9 + Foveros/EMIB/ASAT).

Business impact and checkpoints

From a procurement and investment view, the points to watch are: (1) whether the accelerator you adopt depends on TSMC CoWoS or Intel packaging, and whether that back-end capacity is secured; (2) whether, as with CoWoS-R spreading beyond AI into automotive, your sourced parts are exposed to back-end tightness; and (3) how next steps such as CPO (optical interconnect) and SoW (wafer scale) raise requirements for materials, thermals, and test a few years out. Advanced packaging is not merely a technology topic — it is worth tracking as a leading indicator of AI-chip supply continuity and performance.

Referenced fact cards