Google Awards 3M+ TPU Packaging Orders to Intel Foundry, Breaking TSMC's CoWoS Monopoly
Summary
Key Takeaways
According to Morgan Stanley, Google has placed an order for over 3 million next-gen TPU advanced packaging units with Intel Foundry, set for 2028 production. The order focuses exclusively on packaging using Intel's EMIB (Embedded Multi-die Interconnect Bridge) technology to integrate Google's custom TPU dies with HBM (High Bandwidth Memory) into a single module. This is Google's first diversion of TPU packaging from TSMC's CoWoS (Chip-on-Wafer-on-Substrate) ecosystem to a third-party supplier.
Google conducted months of validation on Intel's packaging, covering yield, thermal management, signal integrity, and mass production consistency. TSMC's persistent CoWoS capacity shortage is the primary driver. Separately, NVIDIA is performing early technical validation of Intel's 18A process for 4-GPU die packaging compatible with its 2028 Feynman architecture, but has not placed a formal order. Intel expects 18A mass production in 2027 and 14A in 2028, and this order significantly boosts Intel Foundry's credibility in AI chip packaging.
Why It Matters
Ostensibly a supply chain diversification move, this is Intel Foundry's strategic encirclement of TSMC's AI chip packaging ecosystem. By securing Google's TPU packaging with EMIB, Intel aims to decouple Google from TSMC's CoWoS and eventually extend to 18A/14A wafer manufacturing, creating a full-stack lock-in.
Intel downplays EMIB's engineering limitations: compared to TSMC's mature CoWoS with >95% yield, Intel's multi-die + HBM stacking faces thermal management and signal integrity risks at scale. The 2028 production timeline exposes Google to 2+ years of technology obsolescence risk, as TSMC may introduce advanced CoWoS-L or SoIC by then.
The hidden trap: once Google's TPU interconnect is optimized for Intel's bridge specifications, it loses compatibility with TSMC's 3D Fabric ecosystem, creating architectural vendor lock-in. Intel is defending against TSMC and NVIDIA, while leveraging this order to court AMD, Broadcom, Marvell as future packaging customers.
PRO Decision
【Vendors (Competitors)】TSMC must accelerate CoWoS-L and SoIC capacity expansion and yield improvement, offering Google competitive pricing and 3D Fabric compatibility guarantees to prevent defection. Proactively demonstrate CoWoS's roadmap advantages (higher bandwidth, lower power) to NVIDIA, AMD to undermine Intel's EMIB appeal.
【Enterprises (CIOs/Architects)】For enterprises relying on Google TPU cloud services, monitor supply stability gains but demand independent benchmarks comparing EMIB vs CoWoS packaged TPUs on inference latency and power efficiency. If developing custom AI chips, avoid early lock-in to Intel packaging; maintain interoperability with TSMC's 3D Fabric.
【Investors】This order confirms AI chip packaging supply chain restructuring, but Intel's 2028 production limits near-term revenue. Focus on Intel 18A process customer adoption, not packaging PR. TSMC's CoWoS capacity doubling in 2025 remains the primary battlefield; Intel's breakthrough does not yet shift competitive dynamics.
Get 3-5 key AI infrastructure signals weekly →
💬 Comments (0)