Question 1

Why is Nvidia restarting H200 chip sales to China?

Accepted Answer

Nvidia is restarting H200 production for Chinese customers following a shift in U.S. export policy from blanket bans to conditional sales frameworks. The new approach, formalized in the revised Bureau of Industry and Security guidance issued in early 2026, allows the sale of chips below a specified compute density threshold to verified commercial end-users in China, provided the transactions include end-use monitoring agreements and are not directed to military or surveillance applications. Jensen Huang confirmed that demand from Chinese customers has picked up significantly and that orders are already flowing. The H200, which sits below the revised compute ceiling when configured with specific firmware limitations, qualifies under the new framework. For Nvidia, this reopens a market that represented over 20% of its data center revenue before the October 2022 export controls — potentially adding $8-12 billion in annual revenue.

Question 2

How much revenue did Nvidia lose from China due to export controls?

Accepted Answer

Before the October 2022 export controls, China accounted for approximately 20-25% of Nvidia's data center revenue, or roughly $10-12 billion annually at 2024 run rates. After the initial restrictions, Nvidia attempted to serve the market with downgraded chips (the A800 and H800), but the October 2023 tightening closed those workarounds as well. By fiscal year 2025, Nvidia's reported China data center revenue had fallen to approximately $3.5-4 billion — a decline of over 60% from pre-restriction levels. The cumulative revenue impact over the restriction period (late 2022 through early 2026) is estimated at $18-24 billion in foregone sales, accounting for the explosive growth in AI infrastructure spending that Nvidia captured in every other geography during the same period.

Question 3

What is the Huawei Ascend 910C and does it compete with Nvidia's H200?

Accepted Answer

The Huawei Ascend 910C is China's most advanced domestically produced AI accelerator, manufactured on SMIC's N+2 process node (roughly equivalent to 7nm). It delivers approximately 640 TFLOPS of FP16 compute and features 96GB of HBM2e memory. In raw performance benchmarks, the Ascend 910C reaches roughly 60-70% of the H200's throughput for transformer-based training workloads and approximately 50-60% for inference. However, the software ecosystem gap is significant: Huawei's CANN framework has far fewer libraries, pre-optimized models, and developer tools than Nvidia's CUDA stack. Chinese hyperscalers like Baidu, Alibaba, and Tencent have deployed Ascend 910C clusters, but many report requiring 30-50% more engineering effort to achieve comparable model performance. The H200's return to the China market under conditional terms puts direct competitive pressure on Huawei's AI chip division at a critical moment in its scaling trajectory.

Question 4

What are the new U.S. export conditions for selling AI chips to China?

Accepted Answer

The revised U.S. export framework, updated by the Bureau of Industry and Security in early 2026, replaces the blanket compute-density ban with a tiered system of conditional sales. Chips below a specified performance-per-watt and total processing performance threshold can be sold to verified commercial entities in China, subject to several conditions: end-use monitoring agreements that include on-site audit rights, restrictions on resale to entities on the Entity List, firmware-level compute caps that prevent the chips from being clustered beyond certain scale thresholds, and quarterly reporting requirements to the Commerce Department. Military, surveillance, and certain government-adjacent end-users remain fully restricted. The policy shift reflects a growing consensus in Washington that blanket bans were accelerating China's domestic chip development without meaningfully slowing its AI capabilities, while costing U.S. companies billions in revenue that funded their own R&D advantages.

Question 5

How have Chinese AI companies adapted to the chip export restrictions?

Accepted Answer

Chinese AI companies adapted to export restrictions through four main strategies. First, they stockpiled pre-restriction Nvidia GPUs: estimates suggest Chinese entities acquired 500,000-700,000 A100 and H100-equivalent GPUs before and during the restriction windows through direct purchases and gray market channels. Second, they adopted Huawei's Ascend chips at scale, with Baidu deploying over 100,000 Ascend 910B/C units across its Ernie model training clusters. Third, they developed aggressive model efficiency techniques: DeepSeek's R1 and V3 models demonstrated frontier-class performance using significantly fewer compute resources, while Alibaba's Qwen and Baidu's Ernie achieved comparable results with training budgets 30-50% smaller than Western equivalents. Fourth, several companies including ByteDance and Tencent invested in custom ASIC designs with SMIC and other domestic foundries, though these remain 2-3 generations behind TSMC-manufactured chips in performance-per-watt.

Question 6

What impact does Nvidia's China pivot have on its stock and the semiconductor sector?

Accepted Answer

Nvidia's stock rose approximately 7% in the two trading sessions following the announcement of resumed H200 production for China, adding roughly $180 billion in market capitalization. The Philadelphia Semiconductor Index (SOX) gained 3.2% over the same period, with AMD, Broadcom, and Marvell also seeing gains as investors priced in broader easing of China chip restrictions. Analysts at Morgan Stanley raised their Nvidia price target by 12%, citing a potential $8-12 billion annual revenue uplift from China re-engagement. However, some analysts cautioned that the conditional nature of the sales introduces execution risk: the firmware-limited H200 configurations carry lower average selling prices than unrestricted versions, and the compliance overhead could reduce gross margins on China shipments by 200-400 basis points compared to sales in other markets.