Nvidia Restarts H200 Chip Production for China After 10-Month Freeze

Nvidia CEO Jensen Huang confirmed at GTC 2026 that the company is restarting H200 manufacturing for China after a 10-month freeze. Export licenses are secured, purchase orders from Chinese customers are in hand, and production is ramping. ByteDance, Alibaba, and Tencent have collectively been approved to buy more than 400,000 units. The AI chip freeze that reshaped Chinese AI development for nearly a year is over — with conditions.

What Jensen Huang Actually Said

Huang's statement at GTC was direct: Nvidia has received export licenses for multiple Chinese customers, has purchase orders in hand, and is in the process of restarting H200 manufacturing. He said this publicly at the company's largest annual developer conference, which means it is not a rumour or leak — it is official corporate disclosure.

The H200 is Nvidia's current-generation data center GPU, built on the Hopper architecture. It is not the flagship Blackwell (B200, GB200) — those remain on the restricted list for China. But the H200 is a serious training and inference chip, and for Chinese AI labs that have spent 10 months working around its absence using Huawei Ascend 910C alternatives and domestic alternatives, its return is significant.

China once accounted for approximately 25% of Nvidia's data center revenue. Since the export restrictions tightened in May 2025, that share collapsed to a small fraction. Even a partial reopening — capped, licensed, and inspected — meaningfully changes Nvidia's revenue picture and China's AI development trajectory.

The 75,000 Chip Cap Per Customer

The US government is considering capping H200 shipments at 75,000 units per Chinese customer. For context: a large-scale AI training cluster for a frontier model requires 10,000-50,000 GPUs. A 75,000-unit cap per company means ByteDance, Alibaba, and Tencent can each build one major training cluster — not an unlimited fleet.

The cap is also designed to prevent aggregation: Chinese companies cannot simply buy from each other or pool chips beyond the licensed allocation without triggering additional scrutiny. US officials are also considering total China-wide H200 shipment caps approaching 1 million units across all customers.

The approval structure requires US government inspection of shipments and imposes a 25% duty on H200 chips entering China. Beijing approved the sales in January, with ByteDance, Alibaba, and Tencent specifically cleared for over 400,000 units combined — meaning the Chinese regulatory side cleared first, and the US licensing followed in March.

The Groq Chip: Nvidia's China-Specific Move

Alongside the H200 restart, Nvidia is preparing a China-specific version of its Groq inference chip for the Chinese market, expected to launch by May 2026.

Groq is Nvidia's inference accelerator — a chip optimised for running trained models at production scale, not training from scratch. The China version is not a downgraded export-controlled variant; it is adapted to work with Chinese AI frameworks and ecosystem tooling. This matters because inference is the high-volume, high-frequency workload: every time a user queries a model like Baidu's Ernie or Alibaba's Qwen, that is an inference call.

The Groq chip for China targets the inference market specifically because inference is where China has the most commercial demand. Chinese AI labs have trained competitive models (DeepSeek, Qwen, Doubao) using Huawei Ascend clusters and older Nvidia A100s they stockpiled before export restrictions. What they need now is cost-efficient inference at scale — and that is what Groq addresses.

Why 10 Months Without H200 Mattered

Between May 2025 and March 2026, China's major AI labs operated without access to Nvidia's current-generation training hardware. The impact was not what many predicted:

DeepSeek released R1 and V3 during the freeze, demonstrating that training efficiency improvements could partially compensate for hardware constraints. Alibaba's Qwen 2.5 and Tencent's Hunyuan models also advanced significantly. Huawei's Ascend 910C became the de facto domestic alternative, though it remains roughly 30-40% less energy-efficient than H200 for comparable training workloads.

The freeze accelerated China's domestic chip ecosystem in ways that will outlast the H200 restart. Huawei now has enterprise relationships with Chinese AI labs that it did not have two years ago. Those relationships do not disappear when H200s become available again. China's AI development is now running on a dual-track hardware strategy — Nvidia for frontier workloads, Huawei for domestic compliance and supply security.

What the H200 Restart Means for US-China AI Competition

The restart is a concession with conditions. Washington is allowing H200 sales because:

First, Chinese labs have already demonstrated they can advance without Nvidia hardware. The strategic objective of slowing Chinese AI development via export controls has partially failed, and the cost — lost Nvidia revenue, weakened US-China trade relationships — is becoming politically significant.

Second, H200 is now one generation behind Blackwell. Nvidia's B200 and GB200 NVL72 systems are what frontier US labs are running. Letting China buy H200s does not give them access to the current frontier — it gives them access to last year's frontier.

Third, the 75,000-unit cap and US inspection requirement give the US government visibility into which Chinese companies are building what scale of AI infrastructure. That intelligence has value.

The concession is real, but it is managed and monitored. China gets H200s. It does not get Blackwell. It does not get uncapped volume. And every shipment gets inspected.

What Developers Should Watch

Three things follow from the H200 restart:

Nvidia's China revenue recovery: Even at 75,000 units per major customer, if ByteDance, Alibaba, Tencent, Baidu, and a dozen smaller Chinese AI companies each buy at or near the cap, that is millions of units and tens of billions in revenue. Nvidia's guidance calls will reflect this.

Huawei Ascend trajectory: Huawei has been moving fast on Ascend 910C and the upcoming 910D. If H200s return to China, does Huawei's enterprise AI customer base stabilise or shrink? Watch Huawei's domestic market share in cloud inference infrastructure over the next two quarters.

US export control policy revision: The H200 restart signals that the US is moving toward a managed access model rather than a blanket freeze. Watch for whether Blackwell export restrictions follow the same path in 12-18 months, as Blackwell becomes the new generation behind the next Nvidia architecture (Rubin).

Key Takeaways

• Nvidia is restarting H200 manufacturing for China — Jensen Huang confirmed at GTC with licenses secured and purchase orders in hand, ending a 10-month freeze
• ByteDance, Alibaba, and Tencent approved for 400K+ units combined — Beijing cleared the sales in January, US licenses followed in March
• 75,000-unit cap per Chinese customer under consideration — enough for one major training cluster per company, not an uncapped fleet
• 25% duty and US inspection required on all H200 shipments to China — this is managed access, not a free market reopening
• Groq inference chip coming to China by May 2026 — not a downgrade, adapted for Chinese AI frameworks and the high-volume inference market
• China's 10-month H200 freeze accelerated Huawei Ascend adoption — those relationships persist even after H200 availability returns
• H200 is now one generation behind Blackwell — the US is letting China access last year's frontier, not the current one