A chip's global passport

GFM Editor's Note

Chips were once one of the most globally-oriented commodities.

It may be designed by an American company, manufactured in Taiwan, have its memory supplied by a South Korean company, have its servers assembled in Asia, and finally installed in data centers around the world.

The reality presented by B300 shows that this globalization chain is being recoded. Products are still produced jointly by multinational supply chains, but their flow is increasingly constrained by national security, export licenses, corporate equity, and end-use.

GFM is concerned about this issue not only because the B300 is expensive or in short supply, but also because computing power is becoming an institutional resource. Who can acquire computing power, who has the right to allocate computing power, and who can set the rules for the use of computing power will affect corporate competition, national capabilities, and the global order in the era of artificial intelligence.

(Image caption ) Close-up of NVIDIA Blackwell Ultra B300 GPU chip

Introduction

In an era where artificial intelligence is sweeping the globe, money alone can no longer guarantee a company access to the most advanced computing equipment.

When Asian companies attempt to purchase NVIDIA B300 or GB300 systems, they often face more than just a price quote. Suppliers may require buyers to specify the equity structure, source of funding, equipment installation locations, data center conditions, end users, and future customers for computing power services.

A purchase order is turning into a qualification application that is a complex mix of political, legal, commercial, and engineering conditions.

Nvidia decides on product allocation and customer lists, the US government sets export boundaries, TSMC and high-bandwidth memory suppliers affect production capacity, server manufacturers complete assembly, and data centers must prepare sufficient power, liquid cooling, and high-speed networks.

By 2026, with the acquisition of the most advanced artificial intelligence equipment, we will be close to applying for a global computing power passport.

This pass not only determines who can buy a chip, but also begins to determine which companies and countries will be able to enter the next round of the artificial intelligence competition.

(Image caption ) High-density liquid-cooled AI server systems represent a shift from single-chip to complete computing infrastructure, as well as the complexity of engineering deployment.

A computing power pass issued by multiple parties

In an era where artificial intelligence is sweeping the globe, even companies with sufficient funds may not be able to purchase the machines they need.

An Asian company preparing to build an AI data center may still need to explain to its suppliers the background of its shareholders, the source of its customers, the location of the equipment installation, and who will use the computing power in the future, even if it has already secured land, server rooms, power supply, and investors.

The procurement staff are delivering more than just an order. They must prove that the expensive servers will not be resold, shipped to restricted regions, or actually provided to end users targeted by U.S. export controls as overseas cloud services.

Even after the review is approved, the buyer still has to wait for the capacity allocation.

Wafers, high-bandwidth memory, advanced packaging, server racks, liquid cooling equipment, transformers, and data center power—any one of these components could further delay delivery dates.

This computing power pass was jointly issued by multiple parties.

Nvidia controls the product, channel, and authorized customer lists; the US government determines export borders; TSMC and memory suppliers influence production capacity; server manufacturers are responsible for assembly and delivery; and the destination country's regulations, customs, and data centers constitute the final barrier.

Chinese buyers are in the most restricted position in this network. Singapore, Malaysia, Japan, and other Asian markets, once seen by some companies as windows for overseas deployment and alternative sourcing, are now subject to stricter scrutiny.

The chip is still the same chip, but the system surrounding it is completely different.

What exactly is the B300?

The market is accustomed to referring to "B300" as a chip, but in actual transactions, this name often refers to products at different levels.

The B300 is NVIDIA's Blackwell Ultra platform data center GPU, designed for large-scale model training, inference, AI agents, scientific computing, and other high-intensity workloads. Each B300 comes with up to 288GB of HBM3e high-bandwidth memory and is typically not sold separately at retail like regular consumer graphics cards; instead, it is integrated into HGX, DGX, or other server systems.

A single DGX B300 contains eight Blackwell Ultra GPUs and approximately 2.1TB of total GPU memory. NVIDIA claims a peak FP4 Tensor Core performance of 144 PFLOPS, but this is the peak specification for the entire eight-GPU system and should not be directly equated to real-world application speeds.

GB300 NVL72 is larger in scale.

A single GB300 NVL72 rack integrates 72 Blackwell Ultra GPUs and 36 Grace CPUs, employing a fully liquid-cooled design and organizing the entire rack into a highly interconnected computing system via NVLink. NVIDIA claims that its dense FP4 computing performance is approximately 1.5 times higher than the previous generation Blackwell, and its attention computing performance is approximately 2 times higher.

Therefore, the industry's claim of "not being able to obtain a B300 visa" may encompass entirely different situations.

Some people need an eight-GPU server, some want to buy a complete GB300 NVL72 rack, and some have already obtained equipment quotas but cannot find a data center, power supply, and cooling system to support the equipment.

These differences are very important.

A GPU is a semiconductor product, while a GB300 NVL72 system is a large-scale infrastructure. In addition to preparing the equipment budget, the buyer also needs to handle rack power supply, liquid cooling, high-speed networking, backup power, installation and commissioning, software maintenance, and long-term upkeep.

Even after the equipment arrives at the warehouse, there is still a long way to go before it can form a stable and usable computing power.

(Image caption ) Behind a single B300 chip lies a multi-national artificial intelligence supply chain, connecting NVIDIA, TSMC, high-bandwidth memory manufacturers, server manufacturers, and global data centers.

Why can't I buy it even though I have money?

The shortage of B300s stems primarily from the rapid expansion of global investment in artificial intelligence.

Large cloud computing companies, internet platforms, sovereign wealth funds, research institutions, and artificial intelligence startups are all vying for a limited number of advanced GPUs.

When allocating products, Nvidia and its partners typically consider order size, payment ability, data center conditions, delivery feasibility, long-term partnerships, and compliance risks. For a newly established AI company, even with ample funding, it may not be able to rank ahead of large cloud service providers and technology groups.

Another limitation comes from manufacturing.

The supply chain behind B300 involves advanced wafer fabrication processes, HBM3e memory, and CoWoS advanced packaging. Industry estimates indicate that TSMC's monthly CoWoS capacity may expand from approximately 35,000 to 40,000 wafers by the end of 2024 to approximately 65,000 to 75,000 wafers by the end of 2025, and will continue to increase in 2026.

These are not the complete production capacity data officially released by TSMC, but they are enough to illustrate that the bottleneck of advanced GPUs is not only in wafers, but also in how to package logic chips and multiple sets of high-bandwidth memory into products that can operate stably.

Expanding HBM production capacity is equally challenging. Its manufacturing, stacking, packaging, and testing are far more complex than those of ordinary memory chips, qualified suppliers are limited, and adding new capacity takes a considerable amount of time.

After the servers are completed, high-density equipment like the GB300 requires a corresponding liquid-cooled server room. The power distribution, cooling, and floor load-bearing designs of many traditional data centers are simply not capable of directly supporting these types of server racks.

The so-called "waiting six months or longer" in the market refers to different scenarios: some people are waiting for GPUs, some are waiting for complete systems, and some are waiting for liquid cooling equipment, transformers, and available data center capacity.

The scarcity of the B300 is partly due to market supply and demand, and partly due to engineering capabilities. Chinese and some Asian buyers also face a layer of political and legal scrutiny.

(Image caption ) The imagery of combining the flags of the United States and China with chips highlights the geopolitics, export controls, and the US-China rivalry behind the B300.

The door to the Chinese market

The United States’ systematic restrictions on China’s advanced artificial intelligence chips began in 2022 and have since been revised, expanded and clarified multiple times.

Washington's basic assessment is that high-performance GPUs are not only used for commercial artificial intelligence, but may also be used for military simulation, intelligence analysis, cyber attack and defense, weapons development, and large-scale surveillance.

Advanced computing power has thus been incorporated into national security policy, and its flow is no longer entirely determined by the market.

This requires distinguishing between different generations of products.

The H20 was a downgraded version of the H20 designed by Nvidia to meet then-current US regulatory requirements for the Chinese market. In April 2025, the US government notified Nvidia that exports of the H20 to China required a license, and that this requirement would remain indefinitely. Nvidia anticipated incurring up to approximately $5.5 billion in inventory, procurement commitments, and related costs as a result.

The H200 outperforms the H20, but it still belongs to the Hopper architecture, not the newer Blackwell.

On December 8, 2025, Trump announced that Nvidia would be allowed to sell the H200 to approved Chinese customers, with a proposed arrangement for the US government to receive 25% of the related sales revenue. In January 2026, the Bureau of Industry and Security of the Ministry of Commerce officially changed the approval process for the H200, AMD MI325X, and some similar products to a case-by-case approval based on security conditions.

This arrangement does not extend to Blackwell and the next generation of Rubin products.

By May 2026, the United States had approved about ten Chinese companies, including Alibaba, Tencent, and ByteDance, to purchase H200, and approved companies such as Lenovo and Foxconn to act as distributors or service providers for some of the transactions.

However, licensing does not equate to immediate delivery of the equipment.

On July 14, 2026, U.S. Commerce Department officials told Congress that a small number of H200 aircraft had begun shipping to China, but the actual delivery volume remained very limited. Whether Chinese companies are willing to purchase them, whether Beijing will allow large-scale deployment, and U.S. permits and additional conditions could all affect the final shipment.

The door to H200 was ajar, but the passageway behind it was still very narrow.

The B300 and GB300 continue to be excluded from direct sales to China. For Chinese companies, this gap is not just a difference in chip models, but could also translate into long-term disparities in model training speed, inference costs, energy efficiency, and cluster size.

(Image caption ) Singapore, Malaysia and other Asian regions are rapidly developing AI data centers and are gradually becoming important nodes for global computing power deployment and export compliance.

Third-country routes are narrowing.

When advanced GPUs cannot be directly imported into China, some demand naturally shifts overseas.

Singapore boasts a mature financial, cloud computing, and data center industry; Johor, Malaysia, and other regions are building numerous data centers; while Japan, South Korea, and Taiwan possess deep-rooted electronics industries and server supply chains.

Chinese companies can establish subsidiaries, invest in data centers, lease cloud computing power, or cooperate with overseas companies to build computing platforms in these regions.

The fact that the equipment did not enter China does not mean that the computing power will not ultimately be used by Chinese companies. This is precisely the regulatory loophole that the United States has been trying to close in recent years.

On May 31, 2026, the Bureau of Industry and Security of the U.S. Department of Commerce issued enforcement guidance clarifying that even if a company is located in a third country, it may still need to obtain a license to export controlled advanced computing products such as 3A090 and 4A090 to countries whose headquarters or ultimate parent company is located in D:5 countries or Macau.

This requirement was not established for the first time in 2026. The BIS guidelines state that its legal basis dates back to November 2023; the 2026 document primarily clarifies market questions following changes to AI diffusion rules.

In July 2026, Nvidia further tightened its list of licensed customers in Asia.

According to the Financial Times, Nvidia has established a new customer whitelist in Singapore, Malaysia, and Japan, and the number of customers authorized to directly purchase advanced AI chips has been reduced by more than half after undergoing stricter compliance reviews.

The review process includes on-site visits to data centers, contract verification, understanding end users, confirming equipment installation locations, and determining whether the buyer's business scale matches the purchase volume.

This also explains why some Asian companies are experiencing a sudden inability to purchase their products.

The countries in which these companies operate may not be subject to a full embargo, but the companies themselves may have failed supplier risk assessments. Small size, complex equity and funding sources, lack of actual data centers, purchase volumes disproportionate to business needs, or close relationships with restricted Chinese clients can all be risk signals.

The so-called "Asian computing power blockade" is not a comprehensive embargo against all Asian countries in a legal sense.

It's more like a gradually tightening filter.

Products can still enter Asia, but each buyer must prove that they are not a transshipment agent or an agent of a restricted customer.

(Image caption ) Singapore and Asia AI Data Center. The modern, large-scale data center exterior represents its strategic role as a gateway to third countries, a hub for computing power deployment in Asia, and a key component of regional infrastructure.

A million-dollar gray market

The narrower the official channels, the higher the prices in the gray market.

In late April 2026, Reuters, citing multiple industry insiders, reported that the price of a server equipped with eight B300 GPUs in the Chinese market had risen to approximately 7 million yuan, equivalent to about 1 million US dollars, nearly double the price of about 550,000 US dollars in the US market. Monthly rental fees in the Chinese market once reached approximately 190,000 yuan.

These figures represent industry quotes for a specific period, not Nvidia's official pricing, but they reflect the scarcity premium created by these restrictions.

Market supply has tightened further, linked to a criminal case that came to light in March 2026.

U.S. prosecutors have charged Supermicro co-founder and senior executive Min-Hsiung Liao, Taiwanese sales representative Jui-Tsang Chang, and contractor Ting-Wei Sun with using a Southeast Asian transit company, falsified documents, and disguised equipment to smuggle servers containing controlled Nvidia chips to China.

The prosecution stated that the total value of the related orders was approximately $2.5 billion, of which at least $510 million worth of servers were alleged to have been successfully transferred.

Supermicro itself was not charged in the case and stated that the actions violated company policy. The case is still in legal proceedings, and all defendants are presumed innocent until convicted.

The significance of this case lies not only in the amount of money involved.

This sends a signal to the entire server supply chain: US law enforcement agencies are no longer just tracking which country the chips were originally exported to, but are also beginning to track the servers, serial numbers, contracts, end customers, and the actual location of the equipment.

The premium in the gray market price cannot be simply understood as the value of the chip itself.

This also includes transportation costs, intermediary profits, the risk of equipment seizure, criminal and civil liabilities, and the cost of losing original manufacturer services.

The higher the premium paid by the buyer, the more likely they are to become a target of fraud.

The risks of the gray market are not necessarily a fake GPU with a completely identical appearance. More common issues include fake quotas, fake purchase contracts, refurbished servers, engineering samples, disassembled equipment, mismatched motherboards and network cards, tampered serial numbers, and expired warranties.

Another approach is more covert.

The seller claims to own a B300 cluster, but in reality, it only leases computing power from other cloud platforms on a short-term basis and then subleases it to customers. Customers who believe they have acquired physical equipment or long-term dedicated computing power may end up with nothing more than a service contract that can be interrupted at any time.

GB300 is a highly complex liquid cooling and high-speed interconnect system.

Even if the equipment successfully enters the restricted market, it may not be able to operate stably in the long term without original manufacturer software updates, firmware maintenance, technical support, spare parts, and professional engineers.

For large-scale artificial intelligence systems, acquiring the hardware is only the first step. The ability to operate continuously is often more important than whether the hardware can be stored in a warehouse.

Nvidia's Dilemma

Nvidia is not the only beneficiary of these restrictions.

China was once an important market for Nvidia, and one of the world's most active markets for artificial intelligence applications and data centers. Export restrictions cost Nvidia a significant amount of potential revenue and forced the company to continuously design downgraded products that met the regulatory standards at the time.

The problem is that every product adjustment may be quickly followed by new policy changes.

Jensen Huang has repeatedly warned that if American companies completely withdraw from the Chinese market, they may push customers and developers to Huawei and other Chinese domestic platforms.

The CUDA software ecosystem is one of NVIDIA's most important competitive advantages. While chip sales can be measured in quarterly figures, the long-term impact of developers switching to a different hardware and software environment may not be apparent for many years.

On the other hand, Nvidia is an American company and must comply with US law.

If distributors, cloud service providers, or overseas customers provide controlled products to unapproved Chinese entities, Nvidia may face regulatory investigations, administrative penalties, and significant political pressure.

The company thus began to take on a task that was rarely undertaken by semiconductor companies in the past: verifying customer equity, sources of funds, data center addresses, equipment locations, and end uses.

Nvidia is no longer just selling chips.

It also implements and maintains a global access system formulated by the US government and implemented by corporate supply chains.

Transactional in Trump's policies

It is inaccurate to attribute the B300 restrictions entirely to the Trump administration's decision.

The main institutional framework for the United States' handling of advanced GPUs in China was established during the Biden administration. After Trump returned to power, he continued strict control over the most advanced products while adopting more transactional arrangements for older generations such as the H200.

Allowing specific Chinese customers to purchase H200 vehicles while demanding that the US government share 25% of the sales revenue is the most direct manifestation of this transactional nature.

On the one hand, the United States hopes to delay China's acquisition of the most advanced computing power, and on the other hand, it is unwilling to permanently hand over the entire Chinese market to Huawei.

It attempts to retain a degree of control that can be adjusted at any time: which companies can buy, how much they can buy, which generation of products to purchase, what security conditions they need to accept, and when delivery can be made.

For Chinese companies, this uncertainty itself is a significant risk.

Even if a company obtains a license today, it cannot be certain whether the next batch of equipment will be approved; a chip that meets today's export standards may be subject to new restrictions months later.

Data centers are capital-intensive projects that require years of planning, but chip policies can change in a matter of months.

The conflict between the two timescales is driving China to accelerate the establishment of its own chip, software, interconnect, and computing power cluster system.

(Image caption ) With the tightening of formal supply channels, some markets have seen high-priced resales, cross-border transshipment, and supply chain risks, with procurement costs and compliance risks rising simultaneously.

Is the blockade effective?

In the short term, U.S. export controls have indeed increased the cost for China to acquire advanced artificial intelligence computing power.

Chinese buyers are forced to use lower-performance products, wait for case-by-case approvals, or pay hefty premiums to acquire limited quantities of equipment of dubious origin that lacks manufacturer support.

Even if you choose to build a data center overseas, you will still face supplier reviews, end-use restrictions, and policy changes.

This means that the restrictions are not entirely ineffective.

It increases costs, extends deployment time, reduces the certainty of acquiring state-of-the-art hardware at scale, and forces companies to take on higher legal and supply chain risks.

However, in the longer term, export controls may also have the opposite effect.

It created a larger domestic market for Huawei and other Chinese chip companies, prompting Chinese companies to improve cluster technology, model efficiency, interconnect architecture, and hardware-software synergy.

When the best chips are not readily available, developers look for new training and inference methods to accomplish more with less computing power.

According to market data cited by Reuters, the share of Chinese AI models in global token usage rose from approximately 5% in March 2025 to approximately 32% in March 2026. While such statistics are still affected by platform definitions and sample sizes and cannot be simply equated with China achieving overall technological leadership, they at least demonstrate that hardware limitations have not prevented the rapid growth in the use of Chinese models.

Can a blockade permanently solidify the technological gap, or can it only buy a few years of time?

This is the most difficult question to answer in US policy.

If China eventually establishes a domestic computing power ecosystem sufficient to support large-scale models, American companies may lose not only some hardware orders, but also software standards, application ecosystems, and developer networks in one of the world's largest technology markets.

Asia was thrust into a choice.

Singapore and Malaysia originally hoped to leverage their data centers, land, energy, and international capital to become key nodes in Asia's artificial intelligence infrastructure.

Today, they need to manage two interests simultaneously.

On one hand, there are the security and compliance requirements of the United States; on the other hand, there is the huge computing power demand from Chinese companies and the entire Asian market.

If regulations are too lenient, relevant countries may face pressure from the United States, and companies may lose access to advanced GPUs. However, if the strictest US standards are followed, Chinese customers, capital, and some data center investments may be lost.

This conflict will gradually extend to more countries.

The flow of artificial intelligence devices is beginning to influence which customers a country can serve, which capital it can accept, and which technology ecosystem it can participate in.

In the past, Asian manufacturing could maintain relatively flexible supply relationships between different markets. With the advent of the AI computing power era, this flexibility is shrinking.

The location of a data center, the country of origin of its chips, the country of origin of its capital control, and the country of origin it serves are all beginning to take on geopolitical attributes.

(Image caption ) What is truly being redistributed is not just tens of thousands of AI servers, but also the computing power, systems, technical standards, and the allocation rights of global innovation resources in the age of artificial intelligence.

After a chip

The B300 won't always be the most sought-after GPU in the world.

Nvidia has begun to push its product roadmap toward Rubin and subsequent architectures. Today's expensive and scarce devices will become next-generation products in a few years.

But the system represented by B300 may remain.

Customer whitelists, end-use reviews, identification of overseas subsidiaries, verification of data center locations, restrictions on cloud computing power usage, and tracking of device serial numbers and service chains may extend to next-generation chips and may also extend to artificial intelligence models, training data, and other key technologies.

While chip generations change in months and years, the lifespan of a system is often much longer.

In the previous era of globalization, the passport for a commodity mainly consisted of price, contract, and customs documents.

In the age of artificial intelligence, even the most advanced computing power will be subject to scrutiny regarding nationality, equity ownership, purpose, clients, and political affiliations.

Whether a company has money or not can no longer determine whether it can stand at the starting line of the artificial intelligence competition.

The battle sparked by Nvidia's B300 is ostensibly about a production shortage, but at its core, it represents a redistribution of global technological power.

The United States wants to maintain its computing power advantage, China is trying to break through external restrictions, Asian countries are finding their place between the two, and Nvidia stands at the intersection of commercial market and national strategy.

When a chip requires a global passport, the boundaries of the market are no longer determined solely by supply and demand.

The actual number of servers allocated is not just tens of thousands.

That's the qualification for entering the next era of artificial intelligence.

Disclaimer

This article is compiled and analyzed based on publicly available information, official documents, and mainstream media reports. It is for news research and information exchange purposes only and does not constitute any investment, legal, business, or other professional advice. If there are any subsequent changes in relevant policies, market data, or corporate information, please refer to the latest official announcements.