Jen-Hsun Huang comes to Taiwan to respond to Google's TPU competition: Huida GPU is more versatile and Nvidia is not afraid of falling behind in the market

Jen-Hsun Huang came to Taiwan for the fifth time this year, relying on the "versatility" theory to maintain GPU leadership, and strive to expand TSMC's 3nm production capacity by 50% to counter the challenge of Google's TPU and the supply chain pressure of the Trump administration.
(Preliminary summary: Huida’s Q3 earnings report shattered the AI bubble theory, Huang Renxun: Blackwell demand is so good, Nvidia rose 5% after the market)
(Background supplement: Huang Renxun had a fried chicken dinner with the heads of Samsung and Hyundai, setting off South Korea’s “Chicken Meme Stock” carnival)

Contents of this article

Taipei 11 The moonlit night is filled with humid heat, but Nvidia CEO Jen-Hsun Huang’s pace is obviously more urgent. On November 27, he set foot in Taiwan for the fifth time this year. In addition to the rumored visit to the sick TSMC founder Zhang Zhongmou, the outside world is not paying attention to his appearance in a classic leather jacket, but to the next round of the chip war: how he can maintain the core voice of AI for Huida under the pressure of Google TPU and the US manufacturing policy.

He (Zhang Zhongmou) is very good, his condition is very good...and I will leave soon, I will go home today.

"Universal" theory: GPU value

Investors have recently become obsessed with the efficiency of Google's self-developed ASICs, questioning whether GPUs can still maintain a premium. Huang Jen-Hsun fired back during an interview in Taipei, reiterating that fungibility is the insurance purchased by companies. Citing the monthly update of AI models as an example, he pointed out that ASICs built for a single task may quickly become obsolete after algorithm refurbishment; in contrast, GPUs are like cash in the pocket of a data center and can be scheduled to the latest models at any time, covering everything from training to inference.

This statement is not just a slogan. According to Spyglass analysis, the Huida B200 single-chip computing power can reach 9,000 TFLOPs, exceeding the Google TPU by 4,614 TFLOPs. Huang Renxun emphasized that the performance gap is important, but the CUDA ecosystem and software tool chain are "an entire generation" ahead. In other words, what customers buy is not a piece of silicon, but an insurance policy against the unknown.

Huida GPU is extremely versatile and can run anywhere and in the cloud; Huida is the only system in the world that can run all AI models... No matter what your needs are, Huida can help. Huida's position is very solid and unique, but we still have to work hard every day.

Capacity decisive battle: TSMC’s 3nm monthly production capacity increases by 50%, Huida locks in 35,000 pieces

When the demand for AI chips jumps from millions to tens of millions, design is no longer the only winner. Huang Renxun had dinner with TSMC founder Zhang Zhongmou during this trip, which ultimately resulted in a more specific production capacity guarantee. According to TechNode, Nanke’s Fab 18B 3nm monthly production capacity will increase from 100,000 to 160,000 wafers, an increase of 50%, of which 35,000 wafers will be reserved exclusively for Huida, corresponding to the 20 million AI chip shipment plan in the next five quarters.

For Jen-Hsun Huang, this is not just a contract, but also a declaration to the capital market that "there is no innovation without a supply chain." He bluntly stated in front of the media:

"Without TSMC, there would be no Huida."

To supply chain experts, this sentence is a cold reality: TSMC is the only one in the world that can expand production of advanced processes and packaging at the same time.

Trump Stress Test: Hybrid Supply Chain Under the Slogan of Made in America

When Trump returns to the White House in 2025, "Made in America" once again becomes politically correct. Huida made a high-profile announcement that Arizona Fab 21 has produced the first batch of Blackwell chips, delivering a compliance report card to Washington. But Huang Renxun also revealed the reality in Taipei: Even if the wafer is completed in the United States, advanced packaging such as CoWoS still relies on Taiwan, and server system assembly is also inseparable from partners such as Hon Hai and Wistron. , this hybrid model of "wafers made in the United States and packaging made in Taiwan" may become the best balance point for Huida to compete with international politics in the next four years.

From Silicon Valley to Washington, and back to Taipei, Huang Renxun responded to the same question every step of the way: Amidst the dual risks of rapid model evolution and geopolitical tug-of-war, can Huida still sit firmly at the core of AI? He used general arguments to convince customers, and he also used TSMC’s capacity commitments to appease investors. In the face of vertical integration pushed by rivals like Google, he's betting the market will still be willing to pay for flexibility. TSMC has become the key bookmaker in this gamble - without it, any strategy is just a blueprint on a whiteboard.