A U.S. Commerce official told Congress on Tuesday that shipments of Nvidia's H200 chips to China amount to a "very small quantity," according to CNBC. In May, the Commerce Department authorized approximately 10 Chinese companies to purchase Nvidia's H200, the company's second-most powerful artificial intelligence chip, Reuters reported. Later in the hearing, Under Secretary Jeffrey Kessler called the volume "trivial." The remark confirms that months after licenses were issued, the geopolitical tug-of-war over AI chips has produced almost no actual hardware crossing the Pacific—even as the rest of the world fights over every wafer that does.
The H200 drama is a sideshow. The real story is unfolding in Arizona fabs, Idaho packaging plants, and the balance sheets of hyperscalers who have committed approximately $725 billion to AI infrastructure in 2026 alone —nearly double last year's spending, according to analysis from A.L. Capital Advisory. The five largest hyperscalers—Amazon, Microsoft, Alphabet, Meta, and Oracle—are collectively pouring between $660 billion and $725 billion into AI infrastructure this year, Intellectia reported. That figure exceeds the GDP of Switzerland. And it is running headlong into constraints that money alone cannot solve.
Can Fabs Keep Pace With a $700 Billion Bet?
TSMC's 3nm monthly output stood at around 120,000 to 130,000 wafers at the end of 2025 and is now projected to reach about 180,000 wafers by the end of 2026, representing a year-on-year increase of more than 40%, TrendForce reported in April. By early 2026, TSMC's 2nm capacity has reached approximately 90,000 to 100,000 wafers per month across its Taiwan facilities, and both plants are completely sold out for the year, according to Tech Insider. TSMC's three advanced-packaging backend fabs are sold out through 2027 with lead times of 52 to 78 weeks, and Nvidia alone holds an estimated 60% to 70% of total CoWoS capacity, Silicon Analysts tracking data shows.
The packaging bottleneck is less visible than chip shortages but just as binding. CoWoS has become as difficult and capital-intensive as wafer fabrication itself, in part because the newest variant must manage warpage—the physical bending of large packages during heat cycles—and preserve signal integrity across enormous interposers, TechTimes noted. Broadcom has warned that constrained foundry capacity at TSMC will limit chip supply into 2026, as demand for AI accelerators continues to outpace manufacturing expansion, Astute Group reported in March.
Nvidia maintains its stranglehold. Nvidia holds roughly 86% of the AI GPU market, according to industry estimates cited by Companies History. NVIDIA leads with $215.9 billion in fiscal 2026 revenue, far ahead of TSMC ($122.4 billion), Broadcom ($63.9 billion), Intel ($52.9 billion), and AMD ($34.6 billion), TrendX Insights reported. AMD is gaining ground— shares reached record highs above $558 in mid-June 2026 and pushed the company's market capitalization above $900 billion for the first time —but the gap remains vast.
Where Does the Memory Come From?
Nowhere, for now. Micron's high-bandwidth memory capacity is sold out through calendar year 2026, the company disclosed in its fiscal Q1 2026 earnings call, according to Introl. TrendForce expects average DRAM memory prices to rise between 50% and 55% this quarter versus the fourth quarter of 2025, CNBC reported. TrendForce analyst Tom Hsu told CNBC that type of increase for memory prices was "unprecedented."
The culprit is high-bandwidth memory—HBM—the stacked DRAM modules that sit beside every AI accelerator. NVIDIA's B300 GPU requires eight HBM chips, each containing 12 individual DRAM dies. That means a single B300 GPU consumes 96 DRAM dies—and a fully configured DGX B300 system with eight GPUs requires 768 DRAM dies just for the HBM modules alone, Tech Insider calculated. Samsung, SK Hynix, and Micron have all been aggressively converting production lines to HBM, as the revenue per wafer for HBM is estimated to be three to five times higher than conventional DDR5, according to industry analysis.
The reallocation is crushing consumer electronics. Data centers now consume an estimated 70% of all memory chips produced worldwide, IDC reported. NVIDIA plans to slash RTX 50-series GPU production by 30-40% in H1 2026 due to GDDR7 shortages, Introl noted. Unlike the 2020–2023 global chip shortage, which stemmed primarily from pandemic-related supply chain disruptions, this shortage is driven by the intentional reallocation of manufacturing capacity toward highly profitable products intended for AI data center development. According to a 2026 Kearney's PERLab analysis, the shortage is expected to last at least until 2030, Wikipedia's tracking of the crisis shows.
Samsung has stumbled. Samsung Electronics has struggled to meet NVIDIA's qualification standards for its 12-layer HBM3E chips due to yield and performance issues, relegating the world's largest memory maker to a tertiary position and tightening the bottleneck further, EnkiAI reported. That leaves SK Hynix and Micron splitting the spoils. In October, SK Hynix said it had secured demand for its entire 2026 RAM production capacity, CNBC noted.
