Hi everyone,
I recently came across an eye-catching story: a little-known smart TV brand called Onn, virtually unheard of in Taiwan, is now poised to overtake Samsung as the top-selling TV brand in the U.S. this year, thanks to its acquisition of the more familiar Vizio.
I happened to be familiar with Onn. It’s Walmart’s in-house electronics brand. Two years ago, I stumbled upon a 55-inch Onn TV at a Walmart in Hawaii, priced at just $248. I was stunned.
“Electronics in the U.S. are unbelievably cheap,” I thought. Of course, bringing a TV back to Taiwan wasn’t exactly practical, so I settled for a photo instead.
What I didn’t realize at the time was how Onn could sell at such rock-bottom prices. The answer? It doesn’t make money from the TVs themselves. Onn is part of Walmart’s growing advertising business, and the real revenue comes from highly targeted ads delivered through the smart TV platform.
The ads are precise because they’re powered by data from Walmart, the world’s largest brick-and-mortar retailer and the second-largest e-commerce player in the U.S.—drawing from the purchase behavior of tens of millions of members.
It creates a closed-loop system where marketing and retail converge, allowing Walmart to track exactly what people buy after seeing an ad.
Walmart’s ad business has quietly become one of the retailer’s most promising new revenue streams. Last year, it grew 30% and accounted for 10% of the company’s operating income.
But the real growth engine is e-commerce, which crossed $100 billion in annual revenue for the first time last year.
Yes, Walmart is still a distant second behind Amazon. But it’s growing much faster—over 20% annually—while Amazon’s growth has slowed.
The difference? Fresh food delivery. Walmart’s ability to tap into its nearly 5,000 stores and warehouse infrastructure gives it an edge in speed and service quality that Amazon can’t easily match.
This is why, despite assumptions that Walmart—long seen as a “Made in China” proxy stock—would be among the hardest hit in a U.S.-China trade war, its stock has remained strong.
In fact, Walmart is trading at a lofty 41x P/E ratio, higher than even Amazon, which sits at 37x and has the AI halo effect.
If there’s a model for a legacy retailer reinventing itself, Walmart may just be it.
Now, back to this week’s newsletter.
The biggest headline from the past two weeks? Nvidia’s H20 GPU is now cleared for sale to China, despite earlier export restrictions.
I’ve obtained an exclusive full-length research report from a major international investment bank analyzing China’s AI chip sector. It sheds light on why Jensen Huang(黃仁勳) was able to persuade Donald Trump to reverse course—and what exactly the “Godfather of AI” is so worried about.
Let’s dive into this week’s newsletter.
Over the past few years, Taiwan has come to embrace Nvidia founder Jensen Huang as a local icon—equal parts Silicon Valley visionary and approachable “uncle next door.”
Whether he’s walking the streets of Taipei in a bomber jacket, cracking jokes in Taiwanese at COMPUTEX, or pausing for selfies with fans, Huang has cultivated a uniquely affable image in his ancestral homeland.
So when he appeared at a supply chain conference in Beijing this July, clad in a traditional Chinese Tang suit and referring to himself as an “American Chinese,” it marked a striking shift in tone for many Taiwanese observers.
His trademark warmth remained intact: Huang patiently engaged with a local reporter who struggled to phrase her question in English, gently asking her to repeat it multiple times until he understood.
But the real headline from Beijing was buried in an unexpected comment about Nvidia’s proprietary CUDA software stack—the foundation of the company’s AI dominance.
“If you build a CUDA-compatible product,” Huang said, “I won’t be angry.”
Then, in heavily accented Mandarin, he added, “No problem.” He even took a veiled swipe at Intel: “If it were x86, they’d be mad.”
The off-the-cuff remark may signal a significant policy shift: a softening of Nvidia’s previous stance barring CUDA compatibility with third-party GPUs.
It also fits a broader pattern of Huang extending an olive branch to Beijing—this was his fourth visit to China this year alone, during which he has publicly emphasized his Chinese heritage and donned culturally symbolic attire.
Crucially, Huang’s lobbying efforts also appear to have swayed President Donald Trump. On July 14, it was announced that the “China-specific GPU” H20, which the U.S. Commerce Department had banned from export to China in April, would once again be permitted for sale.
Major U.S. media outlets, including The New York Times and The Wall Street Journal, took notice. Once seen as a political outsider, Huang had somehow managed to convince the notoriously difficult Trump to rescind an executive order.
This achievement has positioned the Taiwan-born Huang, who moved to the U.S. at age seven, as Trump’s new favorite entrepreneur, following in the footsteps of Elon Musk and Tim Cook.
Huang’s core argument, which reportedly persuaded Trump, is that for the U.S. to win the U.S.-China tech race, American AI chip technology must become the global standard. Ceding the massive Chinese market to local competitors, he argued, would be a grave error.
Stanford professor H.-S. Philip Wong(黃漢森) echoed the sentiment in a June interview with TechTaiwan, noting that China’s domestic alternatives to U.S. chips once struggled to gain traction—until Washington’s export bans “effectively created a market for them.”
Just how big is this burgeoning market?
A July 17 report from Bernstein Research titled “China AI: Assessing Domestic Chip Supply and Demand” projects that 2025 will be a breakout year for Chinese AI semiconductors.
Domestic players are projected to capture 42% of the market, a significant jump from 29% in 2024.
Bernstein further predicts that even with the H20’s re-export to China, the “import substitution” trend by local players is irreversible. By 2026, domestic companies are expected to surpass NVIDIA and AMD in market share for the first time, reaching 53%.
Bernstein’s performance rankings of top Chinese AI chips versus Nvidia’s latest offerings underscore the shift.
To stay within the U.S. export control limits, Nvidia engineered the H20—a deliberately downgraded version of its flagship H100, banned from being shipped to China in 2023.
With a Total Processing Performance (TPP) score of just 2,368, the H20 delivers exactly 15% of the H100’s power, narrowly clearing the regulatory bar set by the Commerce Department. Designed specifically for the China market, the chip reflects Nvidia’s strategy of compliance through technical constraint.
The ranking reveals that 15 Chinese AI chips now surpass the H20 in performance, including multiple offerings from Huawei, Cambricon, and Hygon. Huawei’s Ascend 910C, designed to rival NVIDIA’s GB200, achieved the highest score, boasting 4.3 times the performance of the H20.
Why a “Weaker” GPU Is China’s Hottest Commodity
On paper, Nvidia’s H20 looks hopelessly outgunned in China. Yet, immediately following the news of its unbanning, Chinese tech giants like Tencent and ByteDance reportedly clamored to acquire this “politically incorrect” American product.
There are three key reasons.
1. Supply shortage. China’s domestic fabs, like SMIC, are stuck at 7nm processes due to U.S. restrictions on EUV lithography equipment. SMIC must rely on multi-patterning workarounds with extremely limited output.
According to Bernstein, China’s entire monthly 12-inch wafer capacity (15,000 to 20,000 wafers) is largely consumed by Huawei’s smartphone business.
By late 2025, SMIC’s 7nm capacity is expected to double to 30,000–40,000 wafers per month, finally freeing up some capacity for AI chips. However, AI chip yields remain constrained, especially since large AI chips suffer from low yield rates, estimated at just 20%–40%.
2. High-Bandwidth Memory (HBM) Advantage. Starting late 2024, the U.S. will restrict exports of high-bandwidth memory (HBM2 and above) to China—a blow to AI model training and inference.
Yet Nvidia’s H20 comes prepackaged with 96GB of HBM3 memory, delivering 4TB/s of bandwidth—twice that of the H100.
It’s like a racing league that caps engine size at 1,000cc—only for a clever automaker to add twin turbos and top-tier tires, producing a car that nearly matches the performance of a 2,000cc engine.
Taiwanese computer science professor Shih-Hao Hung(洪士灝) notes that despite its weaker core, the H20’s memory configuration makes it ideal for running large language models that fit within 96GB. “It might actually perform inference tasks faster than the H100,” he said.
Chinese rivals lag behind here: Huawei’s 910B chip only offers 1TB/s, while others are stuck using standard DDR6 memory.
3. Nvidia’s Unparalleled Ecosystem. As many have surmised, the third reason is NVIDIA’s formidable “moat”: the CUDA ecosystem and NVLink high-speed interconnect technology.
China’s engineers, especially at cloud giants like Alibaba, are deeply entrenched in Nvidia’s CUDA development environment. Switching ecosystems is painful.
Furthermore, NVLink, which allows “multiple GPUs to be used as a single one,” is essentially indispensable for training large models.
Consequently, Chinese companies aiming to build domestic large models are left with only two choices: NVIDIA and Huawei. Among Chinese vendors, only Huawei possesses a technology similar to NVLink, known as “CloudMatrix.”
The Chinese Supercomputer That Spooked Washington
In May, Huawei unveiled its CloudMatrix 384, a GB200 NVL72 equivalent for China that links 384 Ascend 910C chips via optical connections across 16 cabinets.
The system reportedly stunned U.S. officials, including David Sacks, the White House’s “AI czar” and former PayPal COO. According to the New York Times, it helped solidify Huang’s case that abandoning the Chinese market only accelerates Huawei’s rise.
Ren Zhengfei, Huawei’s founder, told People’s Daily in June that China’s chip disadvantage could be offset through “cluster computing,” delivering practical performance despite lagging per-chip power.
However, a closer look reveals a significant gap in “cluster computing” technology between the U.S. and China.
This AI system, which sprawls across 16 cabinets and connects 384 910C chips via optical fiber, achieves double the computing power of NVIDIA’s “hundred-million-dollar cabinet” but consumes 4.1 times the power.
In the Western world, where power consumption and data center space are meticulously scrutinized, these specifications are utterly uncompetitive.
It also suffers from a lack of CUDA compatibility. Bernstein notes that outside Huawei, only one Chinese cloud provider is even testing the system.
Still, Huang himself acknowledged the system’s utility during a June press conference in France, suggesting that China’s abundant energy resources could enable “brute-force” solutions that are less feasible in the U.S.
Given enough time to iterate and scale within the Chinese market, Huawei could, over time, evolve into Nvidia’s most formidable domestic challenger.
According to The New York Times, this was also a key part of Jensen Huang’s pitch to Trump: casting Huawei as a looming “bogeyman” that the U.S. must contain.
But from another angle, this so-called bogeyman may not be quite as fearsome as it seems.
Look beyond data centers, and cracks begin to show. Huawei recently launched the MateBook Fold—its first laptop running HarmonyOS—with a SMIC-manufactured 7nm processor.
This flagship device, priced at roughly NT$100,000 (approx. USD $3,000), still utilizes a 7nm processor manufactured by SMIC, according to consultancy TechInsights. This is the same process node as the Mate 60 Pro smartphone processor, which stunned the Western world two years ago.
This “indicates that U.S. tech sanctions will continue to affect the foundry’s capabilities in mobile phones, computers, and AI/cloud products.”
This aligns with expectations within the Taiwanese industry. Without access to EUV lithography machines, China will remain stuck at the 7nm process for a considerable period, as it represents the last generation of process technology that can be mass-produced using traditional optical methods.
Meanwhile, TSMC is on the cusp of mass-producing 2nm process technology, effectively widening the gap to three generations.
A foreign analyst put it bluntly: “Washington’s H20 exception doesn’t signal a broad easing of semiconductor restrictions. It’s a four-generations-old product. Nvidia’s access is limited, and the broader tech chokehold remains firmly in place.”
In other words, as long as export restrictions on EUV tools and HBM memory remain in place, the hardware gap between China and the U.S. is likely to widen in the foreseeable future—creating a “one world, two systems” divide in AI infrastructure, with growing disparity between the haves and have-nots.
Read more here
Nvidia to Resume H20 AI Chip Sales in China
Exclusive|How Taiwan Glass Became Nvidia’s Unlikely Savior for GB200