Nvidia Seeks More H200 Production as China Orders Top 2 Million Chips

China orders exceed Nvidia inventory by more than 1.3 million chips

Chinese technology companies have ordered more than 2 million Nvidia H200 artificial intelligence chips for 2026, against reported inventory of about 700,000 units. Nvidia has approached Taiwan Semiconductor Manufacturing Co. about additional production, with manufacturing expected to begin in the second quarter of 2026, according to reports published on December 31, 2025.

The figures put the immediate inventory gap at more than 1.3 million chips. Existing stock would cover less than 35% of the reported orders, even if every available unit were allocated to those buyers. The volume of any additional TSMC order has not been disclosed, so the production discussions do not yet establish how much of that gap Nvidia could close.

The commercial opportunity depends on decisions in both Washington and Beijing. President Donald Trump announced in December that authorized Chinese customers could buy H200 chips subject to a 25% fee. Chinese authorities had not approved imports when the production discussions were reported.

That uncertainty sits alongside a strong appetite for computing capacity. Major Chinese internet companies reportedly account for most of the orders, viewing the H200 as a substantial upgrade over hardware currently available to them. Beijing, meanwhile, is promoting domestic chips and reportedly restricting foreign processors in new data centers receiving state funding.

The result is a market where demand, permission to import and permission to deploy are separate issues. Orders indicate what customers want to buy. They do not establish that the chips can enter China or be used in every planned facility.

Two chip variants and a production timetable

Nvidia's reported inventory includes approximately 100,000 GH200 Grace Hopper superchips and about 600,000 standalone H200 chips. Both variants are expected to be offered to Chinese customers. The GH200 combines Nvidia's Grace central processing unit with a Hopper graphics processor, bringing two types of computing hardware together in one product.

The H200 belongs to Nvidia's Hopper generation and uses TSMC's 4 nanometer manufacturing process. Nvidia is seeking additional output while also expanding its newer Blackwell products and preparing its upcoming Rubin line. This would extend production of an older generation because it remains attractive to customers whose access to newer hardware is restricted.

Initial deliveries are expected to come from existing inventory, with the first batch targeted to arrive before the Lunar New Year holiday in mid February 2026. Additional manufacturing beginning in the second quarter would not meet that first shipment window. A production start also does not establish a delivery date for finished chips.

The sequence reported so far is:

• December 2025: Trump announces permission for H200 sales to authorized Chinese customers with a 25% fee.
• December 31, 2025: Reports describe more than 2 million orders for 2026 and Nvidia's discussions with TSMC.
• Before Lunar New Year in mid February 2026: Nvidia expects initial shipments from existing inventory, subject to approvals.
• Second quarter of 2026: Additional TSMC production is expected to begin, with the volume unspecified.

Why buyers see the H200 as an attractive purchase

Nvidia has reportedly indicated a price of around $27,000 per chip to Chinese customers, although the actual price would vary with order size and customer arrangements. An eight chip module is expected to cost about 1.5 million yuan, compared with approximately 1.2 million yuan for the previously available H20 module.

Those module prices imply a 25% increase. The reported performance gain is much larger: the H200 is described as delivering roughly six times the performance of the H20, a reduced capability product Nvidia designed for China that was subsequently blocked by Beijing from entering the country.

The sixfold figure should not be read as a promise that every AI application will run six times faster. Actual results depend on the software, the task and the rest of the computing system. Still, the reported difference explains why buyers may accept a higher purchase price to obtain more useful computing capacity.

The expected module price is also described as roughly 15% below alternatives sold through the gray market, where prices exceed 1.75 million yuan. Comparing 1.5 million yuan with 1.75 million yuan gives a discount of about 14.3%, consistent with that estimate because the reported gray market price is above the comparison figure.

The chip price in dollars and the module price in yuan describe different purchasing arrangements. Neither establishes a universal final bill, and the reports do not explain exactly how the announced 25% fee would be reflected in each customer's contract.

ByteDance plans a larger Nvidia budget

ByteDance reportedly plans to spend about 100 billion yuan on Nvidia chips in 2026, provided China allows H200 sales. That compares with roughly 85 billion yuan in 2025, an increase of 15 billion yuan, or about 17.6%.

The budget is evidence of the scale of intended spending, but it is not a confirmed H200 purchase total. It covers Nvidia chips, and the reported plan remains conditional on Chinese approval. Dividing the entire budget by a single quoted chip price would therefore give a misleading estimate of how many H200 units ByteDance intends to buy.

The broader order figures suggest that access to more capable processors remains valuable to China's large internet businesses despite the push toward domestic alternatives. The sources of demand are concentrated among companies with substantial AI computing needs, rather than described as a general increase across every part of the economy.

Tencent combines capital spending with overseas capacity

Tencent is also described as pursuing additional funding and computing capacity. It reportedly plans a $5 billion bond issuance as part of its AI infrastructure expansion. Its reported second quarter capital expenditure reached 52.8 billion yuan, up 176% from a year earlier, with much of the spending directed toward AI computing and data centers.

A 176% increase means that spending was about 2.76 times the comparable earlier level. That capital expenditure figure measures a broader infrastructure commitment, not a disclosed order for H200 chips. The account does not specify how much of it went to processors rather than other data center equipment.

Tencent also reportedly signed a five year agreement worth about $7 billion with Oracle to obtain access to approximately 100,000 advanced AI chips through data centers in Southeast Asia. Access through overseas facilities is a different route from importing processors for installation inside China.

The distinction matters for interpreting demand. A company can expand its computing capacity through a service or lease arrangement without receiving the underlying chips at a domestic site. The reported Oracle agreement does not identify those roughly 100,000 processors as H200s, so they should not be added to China's reported H200 order total.

Nvidia says Chinese sales will not reduce US supply

The proposed production expansion raises a practical allocation issue: Nvidia would be serving renewed Chinese demand while continuing to supply other customers and ramp up newer products. The reported inventory gap shows why existing stock alone cannot cover all the Chinese orders.

A Nvidia spokesperson, responding to questions about the proposed sales and supply, said authorized exports would not affect the company's ability to serve customers in the United States:

Licensed sales of the H200 to authorized customers in China will have no impact on our ability to supply customers in the United States.

That assurance does not disclose the size of an additional TSMC order, a delivery schedule for the full Chinese backlog or the manufacturing capacity assigned to each product generation. TSMC declined to comment on the reported discussions, leaving the production plan without public confirmation from the manufacturer.

Nvidia also argues that excluding American chips from China benefits foreign competitors. That position is part of the company's case for reopening sales, rather than proof that regulators will approve imports or that domestic purchasing restrictions will be relaxed.

US export permission is only one approval

Trump's December decision reopened a possible route for H200 exports that had been barred under the previous US administration's controls on advanced AI chips. The announced arrangement applies to authorized customers and includes a 25% fee. It is not unrestricted access to Nvidia's entire product range.

Beijing was still considering whether to allow H200 imports at the time of the December 31 reports. Chinese officials were described as concerned that access to advanced foreign chips could slow the development of domestic AI semiconductor producers, although they had not signaled immediate opposition to H200 shipments.

One proposal reportedly under consideration would require buyers to purchase a specified proportion of domestically produced chips alongside each H200 purchase. No ratio has been disclosed, and the proposal should not be treated as an adopted requirement.

Such an arrangement would address a different policy objective from the US fee. Washington's announced condition concerns access to American technology. A Chinese purchasing condition would seek to preserve demand for local producers while allowing companies to acquire foreign processors. Neither the final Chinese import decision nor a deadline for that decision has been announced in these accounts.

State funding may restrict where foreign chips can be used

A separate reported Chinese directive requires new data center projects receiving state funding to use domestic AI chips. Projects less than 30% complete were reportedly told to remove foreign chips already installed or cancel planned purchases. Projects further along would be assessed individually.

The reported scope needs careful qualification. The issuing regulators were not identified, and it was unclear whether the guidance applied nationwide or only in certain provinces. It would therefore be premature to describe it as a publicly established national ban covering every new data center.

The directive reportedly covers Nvidia's H20, H200 and B200 processors. Reports describing H200 and B200 availability through gray market channels concern an earlier period when US export controls prohibited their sale to China. Trump's later H200 decision changes the potential export route for that product; it does not establish that B200 exports are also permitted.

The financial reach of the reported rules could be broad. A review of government tender documents put state support for Chinese AI data center projects at more than $100 billion since 2021. Many facilities receive some form of public funding, although the number affected by the directive is unknown.

At least one planned project in northwestern China, developed by a private technology company with state backing and intended to use Nvidia chips, was reportedly put on hold before construction. Private ownership alone would therefore not necessarily place a project outside restrictions tied to public funding.

Beijing points to domestic progress without settling imports

China's National Development and Reform Commission offered an official account of domestic computing development on December 31. In the commission's press briefing transcript, spokesperson Li Chao said demand for different types of computing chips was growing rapidly under China's AI Plus initiative, and domestic products were being adapted to a range of applications.

Li highlighted cluster interconnection technologies, including supernodes, as an opportunity for domestic computing capacity to approach leading international levels. These technologies connect groups of processors so they can work together on large tasks. They concern the performance of a computing system, rather than simply the capability of one chip.

Li also said closer cooperation across the supply chain would continue to improve domestic computing capacity. The statement supports Beijing's focus on local development, but it does not announce an H200 import decision, confirm the reported data center directive or provide comparative benchmark results.

Reports describe Chinese products as having reached performance comparable to the H20, while saying no domestic equivalent to the H200 is yet available. Li's comments about progress through interconnected clusters address a different measure. Stronger systems built from multiple processors do not, by themselves, establish that an individual domestic chip matches the H200.

The central unresolved issue is how these policies will meet commercial demand. Nvidia has a reported order backlog larger than its inventory, customers have substantial spending plans, and Beijing is promoting local suppliers. The next concrete milestones are import approval, any final purchasing conditions and confirmation of the additional manufacturing volume and delivery schedule.

Key Points

• Chinese companies reportedly ordered more than 2 million H200 chips for 2026, compared with Nvidia inventory of about 700,000 units.
• Nvidia has approached TSMC about additional production expected to begin in the second quarter of 2026; the volume is unknown.
• Initial shipments are targeted for before Lunar New Year in mid February, subject to approval.
• Reported pricing is around $27,000 per chip or 1.5 million yuan for an eight chip module, with terms varying by customer.
• ByteDance reportedly plans a 100 billion yuan Nvidia chip budget for 2026 if China permits H200 sales.
• US permission includes a 25% fee, while China's import decision and a possible domestic chip purchasing ratio remain unresolved.
• Reported restrictions on data centers receiving state funding could limit deployment, but their geographic scope is unclear.
• China's development commission says domestic chips and cluster technologies are advancing, without announcing an H200 import ruling.