Will Google’s new TPU orders go to Intel? Ming-Chi Kuo: Google even wants to save on MediaTek’s profits; TSMC’s yield advantage is significant.

```

Google's next-generation TPU chip packaging technology path is facing a critical test. The EMIB-T packaging technology provided by Intel for Google's new TPU (codenamed Humufish) in the second half of 2027 currently achieves a 90% technical verification yield, but still falls short of the standard required for mass production. Meanwhile, Google's tough procurement stance, aimed at cutting costs and directly challenging Nvidia, is reshaping the interests of the entire supply chain.

Famous Apple and tech supply chain analyst Ming-Chi Kuo recently revealed that the current 90% technical verification yield for EMIB-T is a positive but reasonable data point. However, Intel's mass production yield benchmark refers to the FCBGA process, whose industry standard is over 98%. The difficulty in moving from 90% to 98% far exceeds the climb from zero to 90%, especially since some specifications of Humufish have yet to be finalized, meaning there's still uncertainty in achieving mass production yields in the near to mid-term.

Meanwhile, Google has quietly approached TSMC, inquiring whether it could bypass MediaTek and directly place wafer orders for Humufish's main compute chip to cut costs from intermediary markups. This move sends a clear signal: Google is transitioning from a relaxed buyer to a meticulous cost controller. The logic is simple: to directly compete with Nvidia in the AI accelerator chip market, cost advantage is Google's core weapon, making EMIB-T's mass production yield directly relevant to Google's competitiveness.

Yield Bottleneck: The Jump from 90% to 98% Is the Hardest

Ming-Chi Kuo points out that EMIB-T's current 90% technical verification yield is a significant milestone for a technology still in development. Intel has some history in EMIB mass production, which adds credibility to this figure.

However, there is a fundamental difference between technical verification yield and final mass production yield. Intel uses FCBGA as the packaging yield benchmark, which has an industry yield above 98%. Kuo emphasizes that the leap from 90% to 98% is much harder than the climb from project initiation to 90%, especially with Humufish's specifications not all finalized, adding more variables to the yield outcome.

As a comparison, TSMC's 2026 5.5x mask size CoWoS packaging yield target also starts at 98%, meaning Google has no compromise space for its EMIB-T yield expectations. Kuo states he is positive about Intel's advanced packaging in the long run, but keeps cautious views on whether it can reach mass production standards in the near to mid-term.

Google Turns Tough: Cost Pressures Force Supply Chain Restructuring

Google has recently inquired with TSMC about directly placing wafer orders for Humufish's main compute chip (self-designed by Google) instead of using MediaTek as an intermediary as before.

Kuo analyzes that since the first-generation TPU (8t), Google and MediaTek have cooperated using a semi-COT (customer-owned tools) model. MediaTek's profit margin mainly comes from its self-designed parts, so whether Google directly orders wafers for the main compute chip doesn't critically impact MediaTek's overall profitability trajectory.

However, Google's initiative to explore cuts in wafer order transfer markups signals that its procurement strategy has fundamentally changed. The logic is clear—cost is Google's most important competitive chip in facing Nvidia in the AI chip market, elevating EMIB-T's mass production yield from a supply chain technical issue to a strategic problem Google must personally solve.

TSMC's Dilemma: Front-end Capacity Allocation Unresolved

TSMC is still evaluating how much advanced process capacity to allocate to Humufish in the second half of 2027, with two considerations at play.

First, TSMC still hopes to secure backend packaging orders for Humufish, but prospects look slim—a result of Google's deliberate planning. Second, TSMC needs to clarify EMIB-T's actual backend output capacity, to avoid mismatched allocation of scarce advanced process resources to segments that cannot fully utilize it.

Kuo points out that Humufish’s effective backend output depends simultaneously on EMIB-T packaging and substrate processes, both needing to be tracked in tandem, and any bottleneck in either will impact overall shipment pace.

In the semi-COT model, TSMC also prefers MediaTek to continue securing wafer orders for the main compute chip. Besides their long-term close partnership, the key business logic is this: MediaTek is TSMC's third largest advanced process customer for 2025. If TPU orders transfer, MediaTek’s large scale makes it a natural buffer for TSMC to rebalance wafer capacity allocation. Thus, in this supply chain game, TSMC has its own incentive to maintain the existing order.

Risk DisclaimerThe market has risks, investment needs to be cautious. This article does not constitute personal investment advice, nor does it take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular situation. Investments based on this are at their own risk. ```