NVIDIA's H200 chips destined for China have moved from the approval-review stage to the stage where physical units are reaching select companies. The Financial Times reported on August 19 that ByteDance and Tencent each received approximately 10,000 H200 units. In July, a senior US Commerce Department official had confirmed that shipments had begun, but described the volume only as "very few," without disclosing arrival volumes by buyer. Approval from both the US and China, customer orders, and actual arrivals proceed on separate tracks. This latest report reveals how far the arrival stage has actually progressed.
The H200 is a GPU from NVIDIA's Hopper generation. While not the company's current flagship product, the H200 SXM features 141GB of HBM3e memory and 4.8TB/s of memory bandwidth, according to NVIDIA. For Chinese AI companies, it represents one option for filling gaps in computing resources. At the same time, the volumes confirmed so far don't reveal how much this will change China's model development or inference capabilities.
About 10,000 Units Each: Approval Becomes Reality
The volume reported by the Financial Times totals roughly 20,000 units combined for ByteDance and Tencent. According to the newspaper, additional companies may also receive H200 shipments. However, since approval from China's National Development and Reform Commission (NDRC) is also required, US export licenses alone don't determine whether the chips actually arrive.
On July 14, Jeffrey Kessler, Under Secretary at the US Bureau of Industry and Security (BIS), acknowledged before Congress that H200 shipments to China had begun, but described the volume as "very few." As of May, the US had reportedly issued licenses to about 10 companies. Unlike July's report, which didn't specify actual arrival volumes by company, August's report shows that supply has reached at least two companies.
The roughly 20,000 units here should not be equated with future purchase volumes or approved license amounts. This figure represents arrival volumes reported by the Financial Times based on anonymous sources; as of August 19, the Chinese government, NDRC, and NVIDIA had not disclosed company-specific arrival volumes.
Two Gates Controlled by the US and China
On the US side, shipment approval comes with multiple conditions. The BIS rule that took effect on January 15, 2026, shifted certain advanced computing products, including the H200, from a default-denial policy for exports to China to case-by-case review. Products covered include those with TPP below 21,000 and total DRAM bandwidth below 6,500GB/s.
The rule also caps total shipments to China and Macau at no more than 50% of shipments to the US domestic market. It requires that domestic US supply not be delayed, that production capacity for more advanced products not be diverted, and that know-your-customer (KYC) checks and independent third-party inspections within the US be conducted. The 50% figure is not a per-company cap but a condition on total shipments to China and Macau relative to total domestic US shipments.
Beyond this, approval for import and purchase must also be obtained from the Chinese side. In its Form 10-Q filed in May 2026, NVIDIA disclosed that it had obtained licenses since February for small quantities of H200 destined for specific Chinese customers, but had recorded no sales as of the filing date, with China's approval of imports remaining uncertain. The company also disclosed a 25% tariff applicable when products are temporarily imported into the US for inspection purposes.
What the H200 Fills: Gaps in Memory and Supply
According to NVIDIA's product page, the H200 offers nearly double the memory capacity and 1.4 times the memory bandwidth compared to the H100. In training and inference for large-scale models, memory capacity for storing models and data, along with read/write bandwidth, shapes processing performance alongside raw GPU computational power. NVIDIA states that the 141GB of HBM3e and 4.8TB/s bandwidth accelerate processing for generative AI and large language models.
The H200 is a distinct product from the H20, which was designed with reduced performance specifically for the Chinese market; this shipment concerns the H200. Additionally, the H200 belongs to the Hopper generation and is not NVIDIA's current flagship product. How much this limited supply will alleviate the computing resource shortage among Chinese AI companies cannot be quantified unless allocation destinations, operational purposes, and additional supply volumes are disclosed.
The US restricts total supply volume and inspection requirements as conditions; China selects which companies receive shipments. The arrival of roughly 10,000 units each represents the start of limited supply operating under this dual constraint.
400,000 Approved Units, 20,000 Delivered
On January 28, Reuters reported that ByteDance, Alibaba, and Tencent had received approval to purchase a combined total of over 400,000 H200 units. However, given the strict conditions, customers had not yet moved to place orders at that time. This figure represents early reporting on purchase approvals and differs in both timing and meaning from the roughly 20,000 units in confirmed actual arrivals as of August.
The same gap appeared in NVIDIA's 10-Q filing from May. While the company had obtained licenses for small quantities, it had not recorded any sales of H200 units to China as of the filing date. Even with licenses issued, approval from the Chinese side, followed by ordering, shipping, and inspection, must all be completed before the chips translate into sales or usable computing resources.
The Financial Times report has now made visible the scale of the first actual arrivals: approximately 20,000 units. The number of shipments reaching additional companies, along with H200 sales figures to China that NVIDIA discloses, will serve as the key indicators for measuring how far this limited supply expands.