China Orders About $70 Billion in Nvidia Chips, but Near Term Deliveries May Cover Just 1%

Six million chips ordered, with only a fraction expected soon

Alibaba, Tencent, ByteDance and other Chinese technology companies have ordered about six million Nvidia RTX Pro 5500 chips worth roughly $70 billion, yet confirmed delivery plans would cover only about 1% of that total over the next four months, according to people familiar with the purchases.

The companies expect several thousand chips before the end of 2026, followed by tens of thousands around February 2027. Large Chinese cloud providers anticipate receiving about 700,000 by the end of 2027 as Nvidia increases production, although that estimate could change with manufacturing allocations.

At the center of the shortage is a professional graphics processor that Chinese buyers can obtain under current US export rules. Washington restricts access to Nvidia's more powerful products, making the RTX Pro 5500 one of the strongest available choices for companies expanding artificial intelligence services.

Nvidia faces more demand than its planned production can meet. The company is also directing manufacturing resources toward its most advanced and profitable products, used by American AI businesses including OpenAI and Anthropic. Chinese customers therefore face a practical constraint even when a chip meets export requirements: permission to buy cannot create additional factory capacity.

The reported orders are priced at more than $11,000 per chip, with some buyers quoted about $11,400. Nvidia has declined to comment on shipment numbers, and Nvidia, Alibaba, Tencent and ByteDance have not publicly confirmed the full order totals. The figures describe purchases and delivery expectations disclosed by people familiar with the business.

What the $35 billion and $70 billion figures cover

ByteDance, Alibaba and Tencent account for more than three million cards valued at over $35 billion. The larger estimate of about six million cards, worth roughly $70 billion, includes other Chinese AI and cloud computing customers. The two figures describe different groups of buyers.

At approximately $11,400 each, six million cards would cost $68.4 billion, close to the reported $70 billion total. Three million at the same price would cost $34.2 billion. The published totals are approximate, and individual contracts may carry different prices.

Nvidia has reportedly indicated that it could manufacture no more than three million RTX Pro 5500 cards worldwide in 2027. Chinese demand alone would therefore amount to about twice that potential annual output, before any cards were allocated to customers elsewhere.

Some partners have been told they may receive only 5% to 10% of their requested quantities. Those estimates concern allocations to particular partners, while the roughly 1% figure concerns confirmed deliveries over the next four months. Buyers consequently face both a slow start and uncertainty over how much of their orders will eventually arrive.

Why a workstation card has become an AI priority

The RTX Pro 5500 is based on Nvidia's Blackwell architecture and was launched in September 2026. It was designed for powerful professional computers, known as workstations, rather than the enormous computing clusters built around Nvidia's leading data center processors.

The card has 84 gigabytes of GDDR7 memory and the same number of CUDA computing cores as Nvidia's RTX 5090 gaming processor. CUDA cores perform many calculations at once, which is useful when AI software must process large amounts of information. Memory holds the model and the information needed while those calculations run.

Chinese companies mainly want these cards for inference, the work an AI model performs after it has been trained. When a chatbot answers a question or software generates a response to a document, it is carrying out inference. Training is the earlier process of learning patterns from large collections of information.

Inference generally requires less computing power than training a major model from scratch, making workstation hardware useful for some services. The RTX Pro 5500 gives buyers a combination of computing capacity and substantial memory at a time when access to more powerful Nvidia products is limited.

Customers are also responding to a shrinking set of alternatives within Nvidia's range. The RTX 6000D, introduced for China in 2025, reportedly offered similar performance but attracted limited initial interest as buyers hoped stronger chips would become available. That product has since been discontinued, and its Chinese market price reportedly doubled.

When buyers expect the cards to arrive

Nvidia's initial delivery schedule stretches from small batches in late 2026 to larger commercial shipments in early 2027. Samples allow customers to evaluate hardware and prepare systems before wider deployment, while commercial deliveries supply the quantities needed to expand operating services.

The reported timetable brings together purchasing decisions, government discussions and expected shipments:

• September 2026: Nvidia launches the RTX Pro 5500.
• Late September 2026: China's industry ministry asks companies about intended purchases and their planned uses, and reportedly signals that it intends to approve orders.
• October 13, 2026: Tansun Technology discloses an agreement to purchase 20,000 cards.
• December 2026: Early samples and the first batches of several thousand cards are expected.
• First quarter of 2027: Commercial deliveries are expected to begin, with tens of thousands of cards anticipated around February.
• End of 2027: Large Chinese cloud providers expect about 700,000 cards, subject to production changes.

About 1% of six million cards would be 60,000. The longer range expectation of 700,000 would cover roughly 12% of the same order total. Even that larger allocation would leave most of the reported demand unfilled.

Tansun Technology, a company listed in Shenzhen, illustrates how the shortage extends beyond China's largest internet groups. Its October 13 disclosure warned that supply shortages, demand and foreign technology controls could delay deliveries or reduce volumes under its agreement.

Factory priorities limit how quickly supply can grow

Nvidia's production choices are shaped by demand for several generations of processors at once. Manufacturing more RTX Pro 5500 cards means finding capacity while the company is supplying advanced AI systems and preparing newer products.

From 2027, Nvidia is expected to devote much of its capacity at Taiwan Semiconductor Manufacturing Co. to Rubin, its next platform for large data centers. TSMC declined to comment on those production plans. Nvidia has also reportedly halted RTX 5090 production as it shifts resources toward its enterprise AI business.

Chinese buyers are therefore competing for factory allocation as well as finished cards. Even a large purchase commitment can leave a customer waiting when a supplier has already assigned production to other products.

Nvidia encountered a related mismatch with its H200 processor. Chinese customers previously placed orders exceeding two million against reported inventory of about 700,000 chips. That prompted discussions about additional manufacturing, while sales also depended on decisions in Washington and Beijing.

The RTX Pro 5500 offers a different route to computing capacity because it meets current US export thresholds. Its reported backlog nevertheless shows that changing products does not automatically solve the shortage of available hardware.

Two governments shape the purchasing window

US rules restrict Nvidia's most powerful AI products from reaching China. The Blackwell B300 cannot be exported there under the rules described in the reports, while H200 sales require decisions on individual cases and licensing conditions. The RTX Pro 5500 falls below the relevant restriction thresholds.

Beijing also influences whether Chinese companies can proceed with purchases. In late September, China's industry ministry reportedly asked ByteDance, Alibaba and other businesses how many cards they wanted and how they intended to use them. The ministry indicated that it intended to approve the orders. ByteDance alone was considering about one million cards.

Nvidia argues that restrictions on both sides constrain its business. A company spokesperson described the pressures facing American suppliers:

U.S. firms continue to be restricted by a combination of outdated U.S. export controls, which cover gaming products released nearly half a decade ago, and China's own limits on U.S. imports.

Chinese authorities are simultaneously encouraging domestic chip purchases. For cloud providers, buying Nvidia hardware and investing in Chinese alternatives can therefore happen together: each offers a possible source of computing capacity when another route is constrained.

Chinese alternatives face their own supply pressures

Huawei, Cambricon and Alibaba's chip design subsidiary T-Head are potential alternatives for inference workloads. Nvidia's limited deliveries give Chinese cloud companies another reason to use those suppliers rather than wait for their entire foreign order to arrive.

Domestic producers face restrictions of their own. US controls on advanced chip manufacturing equipment and related services limit access to tools needed to expand production. A prolonged shortage of high bandwidth memory, which moves information rapidly alongside powerful AI processors, also constrains output and pushes up prices.

The RTX Pro 5500 uses GDDR7 rather than high bandwidth memory. Chinese manufacturers building other AI processors must secure the memory suited to their own designs, so a shortage in one product category can redirect demand toward another without removing pressure across the supply chain.

Huawei had already narrowed Nvidia's lead in Chinese AI accelerator shipments. IDC figures described in October put Huawei's market share gains at 28% in the first half of 2026, compared with Nvidia's 34%. Nvidia shipped 678,000 accelerators during that period, against Huawei's 572,000.

Those shipment figures show an expanding domestic supply base alongside continued demand for Nvidia cards. The reported six million RTX Pro 5500 orders suggest Chinese customers still want substantial quantities of Nvidia hardware even as local manufacturers win more business.

Cloud expansion depends on more than processors

Alibaba, Tencent and ByteDance are also trying to secure the memory needed to operate more servers. Their pursuit of Korean memory supply agreements includes requests to Samsung Electronics and SK hynix for commitments extending through 2035, much longer than the usual three to five years.

Those negotiations cover server DDR and low power DDR memory. They serve different roles from the memory attached to a graphics processor, but all are needed to build functioning computing systems. A cloud company must coordinate processors, memory and other equipment before additional hardware becomes usable capacity.

Samsung and SK hynix have already negotiated substantial commitments with major customers. Industry estimates suggest agreements with global technology companies could absorb 70% to 80% of the manufacturers' planned memory output for the following year. Chinese customers are seeking more supply in a market where much future production is already reserved.

For companies expanding AI services, the immediate consequence is a gap between purchasing plans and deployable computing power. The next milestones are the first RTX Pro 5500 batches expected in December 2026 and broader deliveries in early 2027. Their size will determine how quickly buyers can begin turning these large orders into operating systems.

Key Points

• Chinese companies have reportedly ordered about six million RTX Pro 5500 cards worth roughly $70 billion.
• Confirmed deliveries over the next four months would cover about 1% of those orders.
• Nvidia could reportedly produce no more than three million cards worldwide in 2027.
• The cards meet current US export thresholds and are sought mainly for running trained AI models.
• Large Chinese cloud providers expect about 700,000 cards by the end of 2027, subject to production changes.
• Huawei, Cambricon and T-Head offer alternatives, while manufacturing and memory shortages also constrain domestic supply.