A chip that was never allowed to be fast became the single most fought-over piece of silicon in China's AI market. The Nvidia H20 is not a flagship. It is a compromise — and the fight over it tells you everything about how compute access in China is now decided in Washington, not in a data center.
What the H20 actually is
The H20 is a cut-down version of Nvidia's Hopper architecture, created after earlier US rules blocked the full H100 and H800 from China. Reports describe it as carrying 96GB of HBM3 memory, with FP16 throughput around 15% of the H100, at roughly 400 watts. The design goal was narrow on purpose: enough to run inference for large models (大模型), not enough to train them efficiently.
When it first went on sale in early 2024, distributors quoted each card at roughly ¥86,000–110,000 (≈ US$12,100–15,400 / HK$94,500–121,000), and a server with eight of them at about ¥1.4 million (≈ US$197,000 / HK$1.54 million).
The 2025 ban-and-unban rollercoaster
Few products have reversed direction this many times in a single year.
- October 2022: the US bans exports of A100 and H100 to China.
- October 2023: tighter rules also catch cut-down versions, so Nvidia launches the H20 to stay compliant.
- February 2024: distributors begin taking H20 orders; by early 2025, demand jumps as Chinese frontier models such as DeepSeek-R1 drive inference demand.
- April 9–10, 2025: the US Commerce Department tells Nvidia that H20 exports to China now require a license. Nvidia discloses a charge of about US$5.5 billion against H20 inventory and purchase obligations, and effectively halts shipments.
- July 14–15, 2025: a reversal. Jensen Huang, in Beijing for the China International Supply Chain Expo, announces the US has approved H20 licenses and sales will resume. Nvidia also unveils a separate China-specific RTX Pro GPU.
- Nvidia had reportedly built a stockpile of 600,000–700,000 H20 units and ordered roughly 300,000 more from TSMC ahead of the restart, according to Reuters reporting.
The 15% "license premium"
The un-ban came with a price tag that surprised even seasoned observers. In August 2025, the Financial Times reported that Nvidia and AMD agreed to remit 15% of their China chip revenue to the US government in exchange for export licenses. It was the first time a US company had accepted a revenue share as the condition for a license — effectively a tax on selling into China.
What changed in 2026
The story did not end with the restart.
- December 2025: the US permits limited H200 sales to China, attached to a revenue condition reported at 25%.
- January 13, 2026: a Commerce final rule moves the H200 and AMD's MI325X to case-by-case review rather than presumptive denial, defined by a performance ceiling (total processing performance below 21,000, memory bandwidth below 6,500 GB/s). It caps China and Macau shipments at no more than 50% of the units sold to US customers.
- May 31, 2026: the rules close a loophole — overseas subsidiaries of companies headquartered in China or Macau now also need licenses to buy advanced chips.
Huang has repeatedly framed the stakes plainly: China's AI chip market could reach US$500 billion in the coming years, and losing it would be a major blow to Nvidia.
Why this is really about inference
Strip away the drama and the pattern is consistent. The US has allowed China to buy just enough compute to serve models, while blocking the hardware needed to train frontier ones at scale. The H20 is the physical embodiment of that line. Chinese labs train on domestic accelerators or older stock; they lean on H20 where Nvidia's software stack and mature ecosystem still win.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the H20 episode shows that US export controls now shape not just which chips China can buy, but how its entire split between AI training and inference is organized.
Honest limitations
- The specs are Nvidia's and secondary. Exact H20 throughput and memory figures come from reports and teardowns, not a China-specific Nvidia datasheet we can independently verify. Treat them as widely reported, not audited.
- The 15% figure is a media report. It originates from the Financial Times. Nvidia has not published the terms; we present it as reported, not confirmed by filing.
- The lane keeps moving. A "resumption" in July 2025 did not mean stability. New rules in late 2025 and 2026 keep redrawing what is permissible, so any procurement plan built today can be outdated within a quarter.
- It is not a training engine. Buyers who need to train large models (大模型) should not read "H20 is back" as "the constraint is gone."
What readers can do now
- If you procure compute in China, do not treat the H20 as a long-term fixed supply. Build contracts around available inventory and license status, and keep a domestic-accelerator fallback for training.
- If you run inference at scale, benchmark the H20 against domestic options on your own models. Nvidia's software maturity is real, but the price-per-token gap has narrowed.
- If you follow policy, watch the case-by-case review thresholds. The bandwidth and performance ceilings, not the brand names, are what will decide what is legal to import next.
- If you invest, price in regulatory churn. A chip that was banned and un-banned twice in a year is a reminder that China-bound compute revenue carries political risk no datacenter metric captures.