Why TSMC Is Letting Go of CoWoS—and Where Its Real Second Moat Lies

Hi everyone,

I once traveled to Silicon Valley to attend Intel’s annual media event and had the privilege of interviewing legendary chip architect Jim Keller, who was then in charge of Intel’s CPU business.

He noted that every major cloud provider at the time wanted to develop its own server chips, so Intel had to stay vigilant. “If you end up like Qualcomm, with Apple above you [using its own smartphone chips] and MediaTek below you, squeezing you in the middle, that’s not a good place to be,” he said.

Nvidia now finds itself in exactly that “not a good place to be.” I believe two major developments over the past few weeks are both related to this.

Last week, Google underwent a major leadership shakeup. Demis Hassabis, formerly CEO of Google DeepMind, stepped away from the front lines to become chairman and chief scientist, while Google Chief Scientist Jeff Dean left to launch a startup with three star researchers.

On the well-known All-In podcast, several Silicon Valley venture capitalists who host the show offered this explanation: The market for renting out computing capacity has become so hot and so lucrative that it is crowding out resources inside Google.

Budgets are flowing toward building data centers and buying Google’s internally designed TPUs, rather than developing the most cutting-edge AI foundation models, prompting some of the company’s AI luminaries to leave in frustration.

From the perspective of Taiwan’s hardware industry, what this episode highlights is that Google is pouring astronomical resources into its next-generation TPUs, developed with assistance from Broadcom and MediaTek（聯發科）. The outlook is promising.

Some research firms predict that within two years, Google could even catch up with Nvidia to become TSMC’s largest customer.

When your biggest customer becomes your biggest competitor, Nvidia is in an extremely unfavorable position. And it is not just Google. Every major cloud provider—including AWS, Microsoft, Meta, and even China’s ByteDance—is aggressively adopting internally developed AI chips.

Jensen Huang（黃仁勳） has little choice but to counter with an unconventional strategy: finance small and midsize customers outside the major cloud providers so they can buy more Nvidia chips.

That brings us to another major piece of news from a few days ago. Nvidia announced that it would join forces with six financial giants, including Blackstone, BlackRock, and Goldman Sachs, to form a US$500 billion lending platform that will provide financing to AI startups and emerging cloud providers to purchase Nvidia GPUs. The platform will provide “financing support of up to 25% of project opportunities,” Huang wrote in a post on X.

Back to this issue of the newsletter.

The aforementioned next-generation Google TPU, which has triggered a series of ripple effects across the industry, has attracted the most attention in Taiwan because one version boldly adopts Intel’s EMIB-T packaging technology, marking the first breach of TSMC’s dominance of the 2.5D packaging market for AI chips.

This is seen as an opening salvo in Intel’s revival and counterattack against TSMC. But when TSMC President C.C. Wei was asked at last month’s earnings call what he thought about major customer Google turning to EMIB-T, his answer was: “Welcome!”

What is going on?

After spending some time investigating, I found that it has to do with an emerging strategic direction at TSMC.

Read on in this issue of the newsletter.

Subscribe now

Read more