SK Hynix's $26.5B Listing Bets on China's H200 Backlog

One Week, Two Levers

On July 7, China's Ministry of Commerce, the agency that issues the country's export and technology licenses, met with Alibaba, ByteDance and the startup Z.ai to discuss a new rule: restrict how much of China's most advanced AI software foreign users can reach, treating its top models as a strategic asset to be kept under domestic control rather than software any foreign user can access. Reuters reported the proposal splits into three tiers. Basic open source tools, the kind any developer can already download and modify freely, would only need a filing. Mid-tier systems would need a security review. The top tier, models still unreleased or newly open-weighted, would be domestic use only.

Two days later, on July 9, Washington moved the opposite direction on a different asset. The Commerce Department authorized three Chinese firms, Alibaba, ByteDance and DeepSeek, to buy the H200, a version of Nvidia's flagship AI training chip modified to comply with existing US export limits. Research firm TrendForce puts total approvals under 200,000 chips split across every buyer, with no single customer allowed more than 75,000, less than half what companies asked for.

Same week, opposite direction. Beijing tightening exports of its own model software while Washington loosens, on its own terms, imports of the chip that runs it. Two governments have stopped treating AI as one thing to control and started treating it as two separate assets, a model and a chip, each gated differently.

Where The Memory Sits

Beijing's model export clampdown is not new logic, Washington got there first. On June 13, a US directive forced Anthropic to cut off foreign access, including Anthropic's own overseas staff, to its two most capable models, Claude Fable 5 and Mythos 5, after a reported jailbreak tied to cyberattack risk. Partial access returned June 26, but only for organizations Washington had pre-approved. That is the same tiered logic Beijing is now writing into its own rules: the model itself, not just the chip that trains it, is now something both governments treat as a controlled national asset, not commercial software you download like an app.

High bandwidth memory is the actual bottleneck, not the chip. Even if every one of those 200,000 H200 chips clears customs, running them requires high bandwidth memory, the fast memory stacked directly onto the chip package that feeds it data quickly enough to actually train a large model. Without enough of it, the chip sits idle. SK Hynix, Samsung and Micron built nearly all of that supply. SK Hynix alone holds 58 percent of global high bandwidth memory revenue, per Counterpoint Research, and just raised $26.5 billion in a July 10 Nasdaq listing, the largest ever US IPO by a foreign company, earmarked for a new memory fab in Korea. None of that balance sheet sits in Beijing.