Some of the newest reports on hardware in tech media today have to do with efforts to diversify away from Taiwan Semiconductor Manufacturing, abbreviated TSMC, which has held fab foundry customers in its palm for decades.
The context: World chip makers are trying to rival the Taiwanese firm, but they’re moving slowly: recent American investments in companies like Intel have not fully diversified this part of the market, especially for AI-specific chips, and as for Europe, Africa, or other Asian countries, none of these quarters have produced a robust TSMC competitor. Taiwan, it seems, just has too much of an edge.
To be fair, it’s not that other countries and other firms don’t have foundry production. It’s more that the customers, the designers of chips which have been put onto another track than the manufacturers, want the best, and TSMA is the best, leading to estimates that TSMC still fabricates 90% of the market for high-end AI chips, for customers like Nvidia.
“TSMC pulled away from Samsung and Intel at 7nm, then 5nm, then 3nm, and the company became the only place on earth that could reliably manufacture leading-edge silicon at scale,” wrote Alex Sirois about a week ago at 24/7 Wall Street, providing somewhat of a summary, showing how the stock 10xed for buy and holders. “The AI boom poured gasoline on that lead. Today TSMC controls roughly 72% of the global foundry market and prints chips for NVIDIA, AMD, Apple, and Broadcom on advanced nodes that are running near capacity.”
But now, news from Google might signify that change will happen in the future.
Google Icefish
This week, The Information reports that Google is “turning to Samsung” for production of a new TPU code named “Icefish.” A closer look, though, suggests that although Samsung may be involved, TSMC is still involved, too.
Look at this coverage by Berry Zwets at TechZine.
This is how Zwets phrases it:
“In the current design, TSMC remains responsible for the most demanding part of the chip: the computing engine, manufactured using 1.4nm technology. Samsung would produce the component that connects the chip to the memory.”
Think about this. This isn’t an end run around a manufacturer with a virtual monopoly in fab. This is a handoff to a secondary supplier of a very limited part of production. This doesn’t seem like an effort that would substantially deal with that above 90% number.
Here’s just one of many reports on why U.S. chip making efforts are going slowly. This is even stranger, because one of the major chip makers setting up production on U.S. soil is none other than TSMC.
“The United States invested $30 billion in the semiconductor sector to restore its industrial glory,” summarizes the author identified by his first name, Jerry, at Circuit Digest, in a piece partially titled “Chip Turmoil in the U.S.” that attempts to explicate. “However, the effort has encountered significant obstacles rather than a victorious resurgence.”
Intel hit problems. TSMC hit problems with its U.S. production. Why the U.S. would want TSMC or any non-U.S. firm to own domestic production facilities probably has to do with that very wide, very deep moat: TSMC is the best. But people not close to the industry should be forgiven for wondering: if the whole goal is to diversify into U.S. production, should TSMC really be the owner?
Ongoing Dominance
In further characterizing the market, Zwets continues:
“Samsung is expanding its 2nm production capacity, but its share of the advanced AI chip market remains limited. TSMC completely dominates this segment. The deal with Google would represent a significant step toward Samsung’s ambitions to attract more major AI chip customers for its foundry division.”
Here’s another review from December of last year. While one would think that the tech world has finally woken up to the need for diversification, again, it will take a while.
What is Icefish?
Citations online are scant, but here’s how NIK (@ns123abc) puts it on X, above a meme of a bulked-up fellow that is supposed to be Samsung:
“Google is in talks with SAMSUNG to make part of its next-gen TPU ‘Icefish’ because TSMC is too overbooked by Nvidia to handle it.”
Is that what’s happening?
Anyway, Google has more formally announced two chips, a TPU 8t optimized for training, and a TPU 8i optimized for inference. Is Icefish a TPU8? I guess we’ll find out.
Meanwhile, let’s see if the world market can adapt to changes that shake the dominance of TSMC in fabrication.
Stay tuned.
