Political risk / Jul 9, 2026 / 4 min
Beijing Meters the Chips Trump Licensed
On July 8, The Information reported Beijing will let Alibaba, ByteDance, and DeepSeek buy fewer than 200,000 Nvidia H200s under four use-case restrictions — seven months after Washington approved the exports and Nvidia booked zero China revenue from them.
Beijing finally answered the question Nvidia's CFO has been ducking since May: yes, some H200s may enter China — but only through a quota system that treats American silicon like a strategic reserve, not a market reopening.
What's new: The Information reported July 8 — and Reuters, the South China Morning Post, and regional outlets confirmed July 9 — that Chinese officials have told Alibaba, ByteDance, and DeepSeek they may soon receive permission to buy limited H200 chips. Firms must document how many they need and why. The total cap across all buyers may stay below 200,000 — less than half what the companies requested earlier this year, per The Information. Policy is still being finalized; nothing has shipped.
Why it matters: For seven months the standoff was surreal: Trump cleared H200 exports in December 2025 and licensed roughly ten Chinese buyers. Beijing withheld import approval to protect Huawei Ascend and domestic self-sufficiency. Nvidia CFO Colette Kress said on the May 21 earnings call the company had generated zero H200 revenue in China and excluded the market from its outlook entirely. July's rationing plan breaks the logjam — but on Beijing's terms, not Nvidia's.
The four conditions:
- Huawei Ascend and domestic processors keep priority for inference workloads at scale
- H200s may be used for training on public data — not proprietary inference pipelines
- Each firm must justify requested chip counts to regulators
- Total authorized volume reportedly caps below 200,000 units across all participants
The numbers:
- 200,000 — reported ceiling for all firms combined; one hyperscale data center can deploy more than 400,000 Blackwell GPUs alone, per Bloomberg relaying The Information
- 10 — Chinese firms Washington has licensed to purchase H200s, per Reuters
- $0 — H200 revenue Nvidia recorded in China through May, per CFO Colette Kress
- +3.7% — Nvidia's July 8 session, its biggest one-day move in over a month, on the report (Straits Times)
- 20% — China once represented roughly that share of Nvidia data center revenue before export controls tightened
The quotes:
- Kress, May 21 earnings call: "We have yet to generate any revenue, and we are uncertain whether any imports will be allowed into the country."
- Chinese embassy spokesperson Liu Chang, July 9: "We advocate that China and the United States achieve mutual benefit and win-win outcomes through cooperation, and we oppose the politicisation, instrumentalisation, and weaponisation of technological and economic issues."
- SCMP analyst framing, July 9: Beijing's move is a "middle-ground solution" to "temporarily ease the frontier training bottleneck" while domestic chipmakers catch up.
- Shi Shenchang, export-controls lawyer at Co-Effort Law Firm, to SCMP: domestic chips are "unlikely to fill the country's computing-power gap in the near term."
The backstory: This is not Washington meeting Beijing halfway — it is both governments governing compute from opposite sides. The U.S. sells down-tier Hopper silicon with a 25% surcharge while blocking Blackwell and Rubin. China meters even that downgrade: enough foreign training capacity to keep Qwen, Doubao, and DeepSeek competitive, not enough to rebuild Nvidia's Chinese franchise. DeepSeek's R2 training failure on Huawei Ascend — reported by the Financial Times in August 2025 — left little mystery why labs need foreign trainers even as they chase domestic inference silicon.
What we cannot verify: The Information cited two unidentified people; Alibaba, ByteDance, DeepSeek, Nvidia, and China's commerce ministry did not comment. Four use-case restrictions come from trade-press analysis of the reported structure, not a published Chinese government order. Chip counts, pricing, delivery timelines, and the 25% U.S. surcharge mechanics remain unsettled. Kress's zero-revenue disclosure is audited; July's import plan is still deliberation.
Convina's view: Beijing is not letting Nvidia back in — it is letting just enough Nvidia in to close one gap while Ascend keeps the crown jewel. Inference touches users and hoards proprietary behavior; training on public data is the expendable layer Beijing will rent from Santa Clara. That architecture outlasts any single chip generation, and it tells every investor watching the July 8 rally the same thing: Nvidia gets a purchase order, not a market. The AI sovereignty race now runs on quota ledgers, not product launches.