AI DESK · HONG KONG · WEEKLY

China's H200 Chips Are Stuck in Hong Kong

Beijing's data center math, not Washington's export desk, decided how many of the H200 chips China won this month can actually run anything.
AN

The Chips Finally Move

ByteDance and Tencent each took delivery of roughly 10,000 Nvidia H200 chips this month, the first shipments to reach Chinese buyers since the US Commerce Department's January 15, 2026 rule allowed case-by-case licensing for firms including Alibaba, Tencent, ByteDance and JD.com. The rule caps each firm at 75,000 H200 units, an older Nvidia chip design but still the fastest processor Washington's export list permits out the door. Reuters reported Beijing's original approvals in January covered more than 400,000 H200 chips combined for the three largest buyers, against Chinese import requests exceeding two million units. This month's deliveries total about 13 percent of any single firm's approved ceiling, according to Tech Times. And Beijing did not let the chips clear customs onto the mainland. It directed ByteDance, Tencent and the other approved buyers to route the purchase through Hong Kong instead, while steering fresh AI infrastructure budgets toward Huawei's domestic Ascend chip line. That routing choice, not the US license, is what actually determined whether these chips do anything this quarter.

Hong Kong's Power Math

Hong Kong operates 47 data centers with a combined capacity of about 581 megawatts, according to Hardware Busters' estimate from local grid filings. A single firm's full 75,000-unit H200 allowance would need roughly 125 megawatts to run at once, for one company's chip order alone that is more than a fifth of every watt Hong Kong has built for data centers. Ten thousand of those 75,000 licensed units per firm are now sitting in Hong Kong warehouses, unpowered, because there is no rack space or substation capacity waiting for them. Building new data center power in Hong Kong means new substations, cooling loops and grid interconnects that take years, not the weeks a licensing approval takes. The H200 units stuck in Hong Kong warehouses need local megawatts, not fresh capital, and Nvidia's financing push does not touch that constraint: this month it lined up financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to mobilize more than 500 billion dollars for AI data center buildout. Hong Kong is not part of that money. The question on a ByteDance or Tencent infrastructure planner's desk now is not how many more H200 units Washington will license, it's how many megawatts the mainland grid can commit to a Hong Kong-adjacent site before the next allocation round opens.

Whether the 13 percent delivery rate climbs from here depends on a grid timeline, not a licensing calendar. If Beijing routes the next tranche onto the mainland once local substations catch up, the Hong Kong bottleneck was a one-quarter problem. If it keeps routing through a territory that cannot power the chips, the real cap on China's Nvidia access was never Washington's rule at all. Watch the next quarterly grid capacity disclosure that would show whether new substation capacity has actually been committed to a Hong Kong-adjacent site, not the next Commerce export announcement.

Sources

PREVIOUS COLUMNS, AI DESK DESK
The Wang Report's columns are produced by AI under human editorial oversight. See our Editorial Standards.