Tencent, Alibaba, and ByteDance Snap Up NVIDIA RTX PRO 5500; Potential Orders Exceed $35 Billion

Wen | Singularity Research Society, Author | Xingzhi, Editor | Dr. Qi

AI Agents are getting more capable, but users' quotas may burn out faster.

On October 24, Moonshot AI released Kimi K3.1 Pro. This is the first model in the K3.1 series, with a parameter scale reaching the 3-trillion level and support for a 1-million-token context. The company highlighted the model's capabilities in programming, 3D creation, and long-horizon Agent tasks. But after the new model launched, some paying users found that their quotas were nearly exhausted before their tasks were even finished.

On October 26, Moonshot AI issued a public apology, acknowledging that Kimi members using K3.1 Pro experienced rate limiting and "unnecessary quota consumption," and compensated subscribers within the validity period with a quota equivalent to 25% of their monthly total.

Where did the problem come from? When a user sends a single instruction, the backend may have already launched multiple sub-Agents, each responsible for breaking down the task, parallel retrieval, analyzing materials, checking results, and summarizing output. What the user sees is one task, but what the model actually executes may be a whole chain of calls. The more complex the task, the more tokens and computing resources are consumed in the backend.

Similar complaints have also appeared in the Claude user community.

Some users posted on Reddit that when Claude Opus 5.5 organized 87 text materials, it automatically launched 12 peer-level sub-Agents and exhausted the entire 5-hour usage quota in 5 minutes; other users reported that using Opus 5.5 for code tasks for less than 30 minutes triggered the 5-hour quota limit.

These individual cases may not be universal, but complaints about quotas being consumed too quickly have brought a real problem with Agent products to the forefront: how much computing cost do backend calls add, and how much actual benefit do they bring to users?

Yesterday, Anthropic submitted an IPO application to the U.S. Securities and Exchange Commission. In the prospectus, it included matters such as obtaining sufficient computing resources and continuously investing in computing infrastructure in the company's operating and risk framework, and disclosed the company's arrangements for using various computing resources such as AWS Trainium, Google TPU, Nvidia GPUs, and AMD GPUs.

Users care about whether their quotas are sufficient, while model companies need to consider how much computing power is required to support these complex tasks and whether they can obtain sufficient commercial returns from them.

A potential procurement order recently disclosed by Reuters has pushed this supply-demand pressure to a larger industrial scale.

According to Reuters, citing people familiar with the matter, ByteDance, Alibaba, and Tencent together have order demand for more than 3 million Nvidia RTX PRO 5500 cards. Based on a list price of about $11,400 per card, the potential order scale exceeds $35 billion. At the same time, Nvidia cannot guarantee that it will meet all demand, and actual delivery quantities may be far lower than the order intentions.

A potential $35 billion order and users complaining that quotas burn too fast appear to be two different problems, but both are related to growing inference demand. As AI moves from answering questions to executing complex tasks, the consumption of computing resources increases accordingly, and the pressure extends from the user side to model companies and the chip supply chain.

How much has demand for inference computing power actually grown? Why are Chinese internet giants still concentrating purchases of an Nvidia graphics card whose performance is inferior to top data center chips? And how will this rush to snap up cards affect the commercialization process of domestic chips?

Even workstation cards are being snapped up. What are the big companies anxious about?

If we go back a year, the product route represented by the RTX PRO 5500 might not necessarily have attracted such great attention.

According to previous Reuters reports, when Nvidia launched restricted products for the Chinese market such as the RTX 6000D, some large technology companies were not very interested. They believed that the price and performance of such chips did not match, and they also hoped that the U.S. government would allow more powerful products to re-enter the Chinese market.

That waiting did not bring a definite result. U.S. chip export restrictions on China continue to exist. Exports of products such as the H200 need to meet specific licensing conditions, while the more powerful B300 cannot be sold to China under current restrictions. At the same time, AI models and applications are still expanding, and enterprises need to continuously deploy new models and expand service scale, but computing power supply may not be able to keep up with the pace of business.

In such a supply-demand environment, computing power that can be bought and deployed is more important than ideal chip specifications.

What exactly can the RTX PRO 5500 card do? From its product positioning, the RTX PRO 5500 is a professional workstation GPU under Nvidia's Blackwell architecture: 84GB GDDR7 memory, memory bandwidth of about 1398GB/s, and maximum power consumption of 600W. Nvidia positions it for professional workloads such as Agentic AI, physical AI, simulation, and graphics processing.

84GB of memory and the corresponding computing power give it the conditions to handle some AI inference tasks. Compared with consumer-grade graphics cards, professional workstation GPUs also have designs and support capabilities for professional scenarios.

But do not confuse it with the H200 and B300. The latter two are aimed at large-scale data center clusters, serving high-intensity AI training and inference, and relying on large-scale interconnection to handle heavier model workloads. The RTX PRO 5500 is closer to a workstation GPU that can handle professional computing and some AI inference.

Whether a model can run efficiently depends on model scale, quantization method, parallel deployment, software adaptation, and actual throughput.

It can meet part of the computing demand, but it cannot be pushed as a low-cost substitute for high-end data center GPUs on the strength of 84GB of VRAM alone. Reuters, citing people familiar with the matter, said the Chinese companies purchasing these graphics cards are mainly doing so for AI inference.

Inference happens after model training is complete: a single user question, a single code generation, a single enterprise knowledge base query, or even an Agent calling multiple models in succession while executing a task, all generate inference computing demand.

Training is like concentrated construction, while inference is like daily water supply. Model training usually concentrates a large amount of computing power to complete parameter updates in a relatively concentrated period of time; inference follows product usage, and the more users there are, the more frequent the calls, and the more complex the tasks, the more continuously model services must supply computing power.

After Agents move from "answering questions" to "executing tasks," inference volume will be amplified layer by layer. A single user request may be split into several steps: retrieving materials, breaking down the task, calling tools, generating results, and checking results. The question enterprises care about has also shifted from "are there enough cards to train models" to "can models be run continuously, stably, and economically."

This explains why a product with inferior performance to top-tier data center chips is still being snapped up by major companies.

What exactly are these cards used for? Regarding the purchase quantity, deployment scenarios, and alternatives, we sought confirmation from Tencent and ByteDance. Tencent said it was not aware, and ByteDance and Alibaba had not responded as of publication.

The available information is not yet sufficient to break this order down into the specific business plans of the three companies: it cannot be concluded that ByteDance is definitely doing this for Doubao, or that Alibaba is definitely doing this for Qwen. How much each has ordered, where they will be deployed, and whether they have been confirmed for use alongside domestic chips all remain to be confirmed.

What can be confirmed is the direction of industrial application. In theory, this type of graphics card can handle online services for small and medium-scale models, some enterprise internal AI applications, Agent task execution, and professional computing workloads with certain VRAM requirements. When it comes to specific tasks, parameter scale, quantization schemes, concurrent requests, and latency requirements still need to be evaluated. Some businesses may also adopt multi-card deployment, but multi-card efficiency and cost cannot be directly inferred from single-card specifications.

Inference also does not mean training is completely ruled out. Workstation GPUs can be used for development, testing, and some lightweight fine-tuning; but in this purchase, Reuters points to AI inference as the main use. Without more evidence, it should not be understood as major companies sweeping up top-tier pre-training computing power.

As for why they have not completely switched to domestic chips, the answer must be viewed in terms of software ecosystems, existing clusters, migration costs, and deployment efficiency. Buying Nvidia does not mean the three companies do not use domestic chips; it may simply be a supplement to the existing computing power system. What is more worth asking is how much enterprises are willing to pay for obtainable inference computing power, and how large they judge future demand will grow.

Moreover, we must also be clear that a $35 billion order scale does not mean Nvidia has already secured a confirmed revenue of the same amount.

According to Reuters, related demand from Chinese customers exceeds 6 million cards; Nvidia expects global production of the RTX PRO 5500 in 2027 may not exceed 3 million cards. The company also has to balance different product lines and allocate more capacity to the next-generation Rubin platform.

Even if all purchase intentions are fulfilled, delivery will be constrained by production capacity. For buyers, locking in supply in advance is about securing future resources. Even if the final delivered quantity differs greatly from the initial demand, as long as they obtain part of it, it may still supplement computing power, support expansion, or adjust the load of existing clusters.

How much locking in supply is worth still depends on contract terms, the final transaction price, the specific delivery quantity, and whether the products can be deployed as planned. For the capital market, these are the pieces of information that must continue to be tracked to judge the value of the order.

For Nvidia, this order shows that even if some high-end products are subject to export restrictions, the Chinese market still needs its product ecosystem and computing capabilities. As for how much of the demand can be converted into actual revenue, that will also depend on subsequent deliveries.

The $35 billion purchase intention is enough to reflect major companies' expectations for future computing power demand, but it cannot directly prove that China's AI inference is already comprehensively in short supply.

IDC data shows that in the first half of 2026, China's AI accelerator shipments were about 2.008 million units. Among them, Nvidia shipped about 678,000 units, accounting for 34% of the market; Huawei shipped about 572,000 units, accounting for 28%. Alibaba's T-Head accounted for 7%, Cambricon 5%, Baidu Kunlunxin 4%, and Hygon 3%.

During the same period, China's GPU shipments fell 26% year on year to about 991,000 units, while shipments of other types of AI accelerators grew nearly 80% to about 1.016 million units, already surpassing GPUs. IDC expects China's overall AI accelerator shipments to grow 19% in 2026, while GPU shipments may decline 9%.

Therefore, this wave of card snapping cannot simply be understood as a shortage of all types of computing power. What enterprises truly need is computing resources that can run their target models, meet concurrency and latency requirements, and be deployed in a timely manner.

A chip's VRAM, bandwidth, computing capability, and software support affect whether a model can run and how efficiently it runs; once it enters actual business operations, servers, networking, scheduling, and operations and maintenance must also be considered. Even if two chips have similar theoretical computing power, actual throughput, response latency, and total cost may differ greatly.

For enterprises that need to quickly launch model services, procurement timing is equally important. After the chips arrive, model adaptation, system deployment, and business testing still require investment. If migrating existing systems requires extensive engineering changes, both time and cost will affect the final choice.

This is also one of the reasons Nvidia products remain attractive: Nvidia's long-accumulated CUDA ecosystem, operator libraries, and development tools bring engineering convenience to its products. Enterprises can reuse some existing development and operations experience, reducing uncertainty during migration.

However, even if the chips arrive smoothly, enterprises still need to consider whether the additional computing power can generate sufficient business revenue.

Assuming the entire batch of chips is ultimately purchased for US$35 billion, amortizing only the equipment procurement amount evenly over five years would correspond to an equipment cost of about US$7 billion per year. This calculation only illustrates the scale of the investment and does not represent the three companies' actual procurement contracts, payment schedules, or accounting treatment.

Moreover, chips are only part of the cost of AI infrastructure; servers, network equipment, data centers, power, cooling, and operations and maintenance all generate corresponding expenses. After the equipment is in hand, if utilization is insufficient, the unit computing cost may be relatively high; if model service fees are low and the user payment rate is low, the additional computing power may not necessarily generate sufficient revenue.

Therefore, judging whether this procurement is cost-effective requires looking at several operating indicators: the comprehensive cost per million tokens, the effective throughput per card, computing power utilization, the volume of paid AI service calls, and whether related business revenue can cover depreciation and operations and maintenance expenses.

Different enterprises also have different return paths. Cloud vendors can directly generate revenue by selling computing power and model services; platform companies with consumer AI products may generate returns indirectly by improving user retention and driving advertising or transaction conversion. The latter's commercial value may not be directly reflected in AI product revenue, but it likewise needs to be verified through actual operating data.

Only when the chain from procurement intention to chip delivery, and then to business revenue and profit, is completed does the US$35 billion take on clearer commercial significance.

Amid the scramble for chips, how can the computing power dividend of domestic substitution be realized?

This order occurred during a period of continuous adjustment in US chip export policy toward China. Sales restrictions differ across products. Only by understanding these rules can one understand why the major companies are scrambling for chips and what competitive environment domestic chips face.

(Note: As of October 2026, specific sales and deliveries remain subject to relevant licenses, supply arrangements, and import policies.)

In its 2026 regulatory filings, Nvidia disclosed that although the H200-related license was approved, the Chinese government's import restrictions still affected actual sales of some products, and the company was unable to sell all licensed products.

This means whether chips can enter the Chinese market depends not only on whether the United States permits exports, but also on buyer qualifications, supply arrangements, and China's import policies. For purchasing enterprises, complying with export restrictions is only the starting point; whether the chips can arrive on time and be put into business use is equally important.

If Nvidia can stably supply more compliant products in the future, some enterprises may readjust their procurement plans. For customers that have already deeply used Nvidia's software ecosystem, continuing to purchase compatible products can reduce migration and adaptation costs; domestic chip manufacturers, meanwhile, will face more direct competition.

But the recovery of imported supply does not mean domestic substitution will lose room. According to IDC data cited by Reuters, in the first half of 2026, Huawei's shipment share in China's AI accelerator market reached 28%; domestic suppliers such as T-Head, Cambricon, Kunlunxin, and Hygon have also occupied certain shares. Domestic manufacturers are expanding shipments and using actual deployments to accumulate software adaptation and cluster operation experience.

For major companies such as Tencent, Alibaba, and ByteDance, the stability of computing power supply is also becoming increasingly important. Training, inference, and cloud services all require continuous investment in computing resources, and policy or capacity risks facing a single supply channel may affect the pace of business expansion. Therefore, enterprises are motivated to evaluate different chip solutions and allocate computing power according to task requirements.

Restricted Nvidia supply has opened a market window for domestic chips, and whenever news of large computing power procurement is released, the capital market looks for potential beneficiary directions. But from procurement intention to listed company performance, there are still multiple links that need to be verified.

Server and system integrators may benefit from subsequent deployment, but the RTX PRO 5500 is a workstation GPU, and its deployment method differs from large data center GPU clusters. Based on procurement quantity alone, it is impossible to calculate server demand, let alone directly determine the corresponding rack, network, and cooling solutions.

Power supply, cooling, and data center infrastructure companies may also gain incremental demand, with the specific scale depending on chip deployment methods, power configurations, and whether existing facilities can accommodate them. For high-speed interconnection links such as optical modules, judgment needs to be more cautious. Large AI clusters may drive demand for high-speed networking, but this round of procurement does not equal data center interconnection orders of the same scale. Whether related companies benefit still depends on whether specific projects and equipment procurement can be implemented.

Domestic chips and their supporting ecosystems also have opportunities to take on part of the new demand, but from customer testing to large-scale deployment, and then to revenue recognition by listed companies, time is still needed. To judge the degree to which specific companies benefit, their product capabilities, customer orders, mass production deliveries, and financial contributions should be further verified.

Procurement demand, chip delivery, infrastructure construction, listed company orders, and final revenue are five different links. The computing power concept is related to this order, but that does not mean the relevant companies will necessarily benefit; only by finding specific business connections can the investment logic be supported.

Conclusion

The US$35 billion procurement intention reflects leading companies' expectations for future inference demand, and also exposes real problems such as restricted imported chip supply and domestic solutions still requiring continuous verification.

For purchasing enterprises, securing computing power resources is a prerequisite for business expansion; subsequently, they must complete deployment, control inference costs, and make the additional computing power serve actual business operations.

For domestic chip manufacturers, the supply gap provides opportunities, but long-term competition still depends on products, software ecosystems, and delivery capabilities. For the capital market, order scale cannot replace judgment about actual projects and performance.

How much industrial change this round of chip scrambling can ultimately bring still comes down to who can stably obtain computing power and who can turn it into lower costs, better services, and sustainable revenue.