Tariff Talks Cause Technology Stocks and Markets to Dip Further
By Robert Castellano and George Gilder
The April 15 announcement by NVIDIA (NASDAQ: NVDA) that it will need to take a $5.5 billion charge to its earnings due to President Trump sanctioning the company’s H2O AI GPU chips shows the vulnerability of even market-leading companies to trade disputes.
The H2O AI GPU chips previously has been allowed to be shipped to China under President Biden’s administration. A quick search of recent posts by President Trump on his TruthSocial page shows that tariff policy is affecting nations around the world.
“Had a very productive call with the President of Mexico yesterday,” President Trump posted on his TruthSocial website. “Likewise, I met with the highest level Japanese Trade Representatives. It was a very productive meeting. Every Nation, including China, wants to meet! Today, Italy!”
Japan is among many nations seeking to strike a deal during President Trump’s 90-day pause to so-called “reciprocal tariffs.” After meeting with a trade delegation from Japan, Trump posted: Big Progress!
But President Trump expressed dissatisfaction with Fed Chairman Jerome Powell for not cutting interest rates. The traditional independence of the Fed is not outside the scope of President Trump to voice his opinions.
Even though the European Central Bank (ECB) is expected to cut interest rates for the seventh time, Trump posted that Chairman Powell is always “TOO LATE AND WRONG.” Oil prices are down, groceries — even eggs — are down, he added.
“Powell’s termination cannot come fast enough!” President Trump posted.
Technology stocks are deeply affected by these trade barriers, as shown by Nvidia. It likely will not be the only technology company to report the need to take charge to account for tariff fallout or reduced earnings guidance.
On the bright side, Taiwan Semiconductor Manufacturing Co Ltd. (NYSE: TSM), the largest contract chip maker in the world, posted a bigger-than-expected first-quarter profit increase. At least that important semiconductor company was able to avoid tariff tumult.
U.S.–China Microchip Tariff Theater
Gaining new currency in Silicon Valley are the epithets “Bull in a China Shop” and “monkey-wrench in the works,” as President Trump’s “art of the deal” and “beautiful tariffs” confront the most complex and intricate industrial systems and supply chains by far in the history of human endeavor.
Think of all the world’s Trump Towers and all of Europe’s gothic cathedrals and all our great cities’ filigreed skyscraper gantries, together with their lighting and heating systems and underground power fabrics and sources. Imagine these huge structures all rendered in miniature gallimaufries of features trillionths of a meter in size. That’s nano and pico-scale.
Think of an apparatus egregiously too small even to see that can be fabricated only on huge, elaborate machines sourced in intricate supply chains that circle the globe. Consider that these systems are dependent on unique skills commanded by geniuses of specialization entrenched in a few special locations, often in Taiwan and Israel, or in Silicon Valley redoubts heavily manned by Asian talent (often suspected by Washington paranoids of “spying”) —that’s the world electronics industry. American lawyers and politicians need not apply.
With all industry and commerce increasingly riddled and ratcheted with artificial intelligence (AI) systems, fraught with graphics processors (GPUs) performing trillions of floating-point operations per second, high-bandwidth memories (HBM) in intricate three-dimensional peta-flop towers, and new AI accelerators mounting in Brobdingnagian complexity amid Gigawatt power supplies and nanoscale fragility, that’s what we’re dealing with. Let’s hope we navigate the artfully.
Yankee doodle it with “on-shoring” jabberwocky, and it will dissolve into silicon-smithereens. Ironically, the structure will remain intact chiefly in China with its millions of engineers, who we are bullying into a semblance of self-sufficiency.
Let’s hope that Trump’s theatrical negotiations grasp this reality.
Nvidia’s Blackwell GPUs, uniquely fabricated in Taiwan, lead the market for AI training, where giant databases are parsed in massively parallel arrays for “large language models” (LLMs) with trillions of “tokens.”
At the same time, the rise of hyperscaler application specific processors (ASICs) from Google (NASDAQ: GOOG), Amazon (NASDAQ: AMZN) and others challenges Nvidia’s presence in AI inference, where these trained arrays become “agents” and are applied to specific applications. Enabling these feats is an awesomely intricate transition to the fourth generation of high bandwidth memory (HBM4) from SK Hynix and Micron, setting the stage for a coming leap in computing performance.
The shifts toward “hyperscale” special purpose AI chips at cloud giants Google, Amazon, and Microsoft cast shadows on Nvidia’s 80% share in AI inference from its older or legacy chips, but are no competition to Nvidia’s 92% share in training. Nvidia’s GPUs remain the backbone of AI infrastructure for Microsoft, Google, and OpenAI. In a continuing tidal surge of capability, Nvidia’s GB200 (the dual-chip Blackwell GPU) began limited deployment to these so called “hyperscalers” in late 2024 and is ramping up in 2025, delivering 40,000 FP8 TFLOPS (that’s 40 thousand trillion floating point operations per second, folks) and 16 TB/s (trillion bytes) of HBM bandwidth.
Table 1 provides a comparison of Nvidia’s AI GPUs, highlighting the exponential growth in floating-point operations per second (FLOPS) and HBM bandwidth over the past few years.
The Blackwell GPU’s dominance, however, does not come without challenges. AI compute bottlenecks are shifting from raw GPU FLOPS to memory bandwidth and interconnect efficiency. As individual GPUs scale beyond 1,000 watts of power consumption, signifying ultimate gigawatts per rack, companies must develop solutions that optimize energy efficiency while maintaining performance gains. The upcoming Rubin architecture in 2026, expected to be fabricated on TSMC’s 3nm or 2nm process, will address these concerns by compounding core density and interconnect bandwidth.