
On November 13, Jensen Huang resigned from the top job at Nvidia, the company he co-founded in 1993 and built into the world's most valuable semiconductor maker. According to The Verge, the resignation closes one chapter of a saga that is about to widen into a larger question: how Nvidia's China dealings are handled, and how far the US government is willing to go in regulating them. Liu Gong's take: the departure matters far more to the politics of Nvidia's China exposure than to its product
The public record of the event is short. The Verge reported that Huang resigned from his position at the top of Nvidia on November 13, ending a run that began with the company's founding in 1993; the report did not name a successor, a transition timetable, or any interim management arrangement. The Verge framed the news as the end of one chapter, with a bigger story set to unfold over the coming weeks and months around Nvidia's dealings in China and the US government's willingness to intervene in them. In other words, the announcement settles a personnel question while leaving the compliance question open. Nvidia's current lineup spans H-series training parts, the downgraded H20, and L-series inference cards, and its eligibility to sell to Chinese customers still turns on export licenses from the US Commerce Department. In parallel, China's market regulator has kept its probe into
Why now: Nvidia's China business has been reshaped twice over in three years. In 2022 the US first restricted A100 and H100 exports, and Nvidia answered with A800 and H800; the October 2023 tightening swept those parts in as well. The H20, launched in 2024, became the only compliant training-class part, and in April 2025 it too was made subject to licensing, forcing billions of dollars in inventory and purchase write-downs for the quarter. In August 2025, press reports described an arrangement under which Nvidia would remit 15% of China AI-chip revenue in exchange for licensing certainty. In December 2024, China's market regulator opened a probe into whether Nvidia had breached conditions
Liu Gong's take: Huang's exit is a signal event, not a structural one. Over the next six months, Nvidia is more likely to keep shipping downgraded, compliance-first parts into China than to walk away from the market, and the variable that matters is whether export licensing and revenue-share terms harden into formal rules instead of ad hoc arrangements.
From a technical standpoint, Nvidia's moat is not any single die; it is CUDA and TensorRT together with years of kernels, libraries and customer tooling layered on top. The H20 is the case study: its dense FP16 throughput is roughly 15% of the H100's, yet its 4.0TB/s memory bandwidth runs about 19% above the H100's 3.35TB/s. That compute-for-bandwidth trade is an engineering decision aimed at preserving installed-base lock-in, and it survives a CEO change intact.
On the business side, China's share of Nvidia's data-center revenue has slid from more than a fifth in FY2022 to single digits in FY2025, a swing of over ten percentage points. China has moved from growth engine to policy exposure, and Huang was the most credible informal interlocutor on that exposure in both Washington and Beijing. Losing him costs Nvidia a seat at the table even if the roadmap never moves.
For the industry, Liu Gong's observation from working in industrial inspection: production lines overwhelmingly run A10, L4 and T4-class inference cards with TensorRT, and buyers care about supply continuity and driver-version stability far more than peak FLOPS. If US licensing tightens again, domestic inference accelerators will take those mid-range, stability-sensitive stations first — faster than training clusters, and harder to undo, because every qualified inspection model is re-validated station by station.
- China's share of Nvidia's data-center revenue fell from over 20% in FY2022 to single digits in FY2025, a drop of more than 10 percentage points.
- The H20's dense FP16 throughput is about 15% of the H100's, while its 4.0TB/s memory bandwidth is roughly 19% higher than the H100's 3.35TB/s.
- The August 2025 arrangement reported to remit 15% of China AI-chip sales means about $0.15 of every revenue dollar is diverted, directly compressing gross margin.
- Nvidia's market capitalization crossed $5 trillion in October 2025, roughly five times its level of about $1 trillion in early 2023, so the absolute cost of any single policy shock scales with the valuation.
For the industry: within three months, the clearest change is a rebuilt compliance channel, with cloud providers and server ODMs likely to hold China shipment schedules in a wait-and-see posture; the licensing pace for H20 follow-ons becomes the metric to watch. Within a year, if export terms are codified, domestic inference accelerators will genuinely gain share at mid-range compute stations, while training-cluster substitution stays constrained by interconnect and software-stack maturity.
For ordinary users: over three months, little changes, since inference pricing and availability in the cloud do not move on a CEO transition. Within a year, tighter licensing could raise inference costs
Questions worth exploring:
- How CEO turnover interacts with compliance: can a personnel change alter the substance of export licensing terms
- The engineering economics of compute-for-bandwidth: how many generations the downgraded-part strategy can last
- The substitution window for domestic inference cards: why industrial inspection falls first
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from The Verge AI (published 2026-11-13 14:47:16). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.
