Google's selection of packaging technology for its next-generation TPU chip is undergoing a critical test. The EMIB-T packaging technology provided by Intel for Google's new TPU (codenamed Humufish), scheduled for the second half of 2027, has currently achieved a 90% yield in technical validation. However, this still falls short of the standards required for mass production. Google's aggressive procurement stance, aimed at cost reduction to directly compete with NVIDIA, is reshaping the profit dynamics across the supply chain.

A prominent analyst specializing in Apple and technology supply chains has disclosed that the current 90% technical validation yield for EMIB-T is a positive yet reasonable data point. However, Intel's benchmark for mass production yield is based on the FCBGA process, which commonly achieves yields above 98% across the industry. The difficulty of improving from 90% to 98% far exceeds the challenge of progressing from zero to 90%. Furthermore, with some specifications for Humufish not yet finalized, uncertainty remains regarding whether the mass production yield can meet targets in the near to medium term.

Concurrently, Google has quietly approached TSMC to inquire about the possibility of placing wafer orders directly for the main compute chip of Humufish, bypassing MediaTek to eliminate markups from intermediate steps. This move sends a clear signal: Google is transforming from a lenient buyer into a meticulous cost-controller. The underlying logic is that to challenge NVIDIA head-on in the AI accelerator chip market, cost advantage is Google's core weapon. Consequently, the mass production yield of EMIB-T has escalated from a supply chain technical issue to a strategic problem that Google must address directly.

**Yield Bottleneck: The Leap from 90% to 98% is the Most Difficult** The analyst pointed out that the current 90% technical validation yield for EMIB-T is a positive milestone for a technology still under development. Intel's historical experience in EMIB mass production lends credibility to this figure. However, there is a fundamental difference between technical validation yield and final mass production yield. Intel's reference benchmark for packaging yield is FCBGA, for which current industry yields are generally above 98%. The analyst emphasized that the leap from 90% to 98% is significantly more difficult than the climb from project initiation to 90%, especially given that some Humufish specifications are not yet finalized, introducing more variables into the final yield outcome. For reference, TSMC's yield target for its 5.5x reticle size CoWoS packaging in 2026 also starts from 98%, indicating that Google has little room for compromise on its yield expectations for EMIB-T. The analyst expressed a positive long-term view on the direction of Intel's advanced packaging development but maintains a cautious stance regarding its ability to meet mass production standards in the near to medium term.

**Google's Shift to Aggressiveness: Cost Pressure Forces Supply Chain Restructuring** Google recently inquired with TSMC about the possibility of placing wafer orders directly for the main Humufish compute chip (designed internally by Google), rather than continuing the previous model of routing orders through MediaTek. The analyst's analysis suggests that Google and MediaTek have collaborated using a semi-COT model since the first-generation TPU. MediaTek's profit margin is primarily concentrated on the parts it designs itself, so whether Google places wafer orders directly for the main compute chip is not a critical variable affecting MediaTek's overall profit trajectory. Nevertheless, Google's proactive inquiry about eliminating markups on wafer orders itself conveys an important message: Google's procurement strategy has undergone a fundamental shift. The driver for this change is straightforward – to compete directly with NVIDIA in the AI chip market, cost is Google's most important competitive advantage. This also elevates the mass production yield of EMIB-T from a supply chain technical issue to a strategic problem that Google must personally resolve.

**TSMC's Dilemma: Front-End Capacity Allocation Remains Undecided** TSMC is still evaluating how much advanced process capacity to allocate for Humufish in the second half of 2027, considering two main factors. First, TSMC still hopes to secure the back-end packaging orders for Humufish, but currently this appears unlikely – a result of Google's deliberate strategy. Second, TSMC needs to understand the actual back-end output capability of EMIB-T to avoid misallocating scarce advanced process capacity to a stage that cannot fully utilize it. The analyst pointed out that the effective back-end output for Humufish depends on both the EMIB-T packaging and the substrate环节; these two need to be tracked in coordination, as a bottleneck in either will impact the overall shipment schedule. Under the semi-COT model, TSMC also prefers that MediaTek continue to handle the wafer orders for the main compute chip. Beyond the long-standing close cooperation between the two companies, a more critical business rationale is that MediaTek is TSMC's third-largest customer for advanced processes in 2025. If TPU orders were to shift, MediaTek's scale makes it a natural buffer for TSMC to rebalance its wafer capacity allocation structure. This means that TSMC has its own inherent motivation to maintain the existing order within this supply chain博弈.