ANALYSIS: H200, the battle for access between Washington, Beijing, and Hong Kong

Twenty thousand fleas, three capitals

The number that opens the door

Approximately 10,000 H200 chips for ByteDance. Approximately 10,000 for Tencent. The Financial Times, as reported by several media outlets on August 19, attributes these shipments to sources familiar with the matter. Beijing reportedly authorized the entry of small batches into mainland China while encouraging that the bulk of the capacity remain in Hong Kong.

These figures do not come from a public Chinese registry. They must therefore be treated as estimates. But they describe a tangible shift after months of U.S. licenses, Chinese hesitation, and shipments that were announced but then delayed.

Twenty thousand chips are not just a purchase. They represent the bottleneck where Washington controls the outflow, Beijing controls the inflow, and Hong Kong bears the brunt of their contradictions.

The War Over Access

The word “sale” oversimplifies matters. A U.S. license is required. Chinese approval is required. An accepted customer is required. Servers, electricity, cooling, a network, and a location to run the computing clusters are required.

Each step becomes a political lever. The manufacturer can produce. The customer can pay. Yet capacity only truly exists when all the doors open in the right order.

This is the war for access: no one alone owns the chain, but anyone can block it.

In artificial intelligence, power belongs less to the buyer than to the one who manages to plug in.

The H200 is no longer the pinnacle

A powerful machine that’s intentionally outdated

The H200 is part of Nvidia’s Hopper architecture. The manufacturer highlights 141 gigabytes of HBM3e memory and 4.8 terabytes per second of memory bandwidth. These specifications make it a very powerful accelerator for training and running artificial intelligence models.

But it is no longer Nvidia’s most advanced generation. The Blackwell products—and subsequent platforms—have raised the bar. This is precisely why Washington is willing to consider certain H200 exports while maintaining restrictions on the highest-performance chips.

America isn’t selling the technological frontier to China. It’s selling a lower tier—high enough to be profitable and, it hopes, low enough to preserve the gap.

The Lag That Is Actually an Acceleration

Being two generations behind does not mean being useless. A large quantity of well-integrated H200 chips can drive models, accelerate inference, and relieve the burden on teams working with more limited capabilities.

The relevant comparison, therefore, is not just H200 versus Blackwell. It’s also H200 versus what the Chinese company might have deployed without access to U.S. technology, at the same time, given the constraints of its own hardware fleet.

A chip that’s outdated by global standards can still transform the room for maneuver of whoever receives it.

Technological lag becomes a weapon when deciding who has the right to catch up.

Washington Opens a Door Under Scrutiny

The case-by-case rule

The Bureau of Industry and Security announced on January 13, 2026, that license applications for the H200, the AMD MI325X, and comparable products would be reviewed on a case-by-case basis, subject to security conditions. Applicants must, in particular, demonstrate that the exports do not reduce the production capacity available to U.S. customers.

Chinese buyers must also adopt compliance and customer screening procedures. The products must undergo independent testing in the United States to verify their performance and safety.

This is not a lifting of controls. It is a transformation of them: Washington is replacing a simple ban with a framework in which each authorization becomes a thread that the government can extend, monitor, or sever.

Trade Becomes Conditional

This policy allows Nvidia to regain a portion of a major market, but it places the company at the heart of a national strategy. Every sale now comes with an authorized geographic region, a verified customer, and the risk of review.

For Chinese companies, the U.S. license offers no certainty. It merely opens the first barrier. Beijing then retains its own approval authority.

The globalization of semiconductors isn’t going away. It’s becoming a series of revocable permissions.

The market promised free flow; strategic rivalry has restored its border checkpoints.

Beijing holds the second key

Authorizing Without Liberalizing

Available information points to a selective easing of restrictions by China. ByteDance and Tencent are reported to have received small batches, while other companies may obtain similar volumes. Purchases would remain subject to approval by the Chinese state planning authority.

Beijing is pursuing two conflicting objectives. Its AI champions want more Nvidia computing power. Its domestic semiconductor industry wants time, orders, and a protected market to grow.

China is therefore not opening the door; it is rationing the oxygen. Enough for its labs to breathe, but not enough for foreign manufacturers to take all the air.

Rationing as Industrial Policy

A small quota does not meet the entire demand. It forces companies to choose which projects, teams, and models get access to the highest-performing hardware legally available.

This selection process can become an industrial policy tool as powerful as a subsidy. The government does not merely fund a sector; it decides who can use a scarce imported resource.

Chinese control thus complements U.S. control rather than simply being subject to it. Washington limits the capacity that can leave the country. Beijing limits the capacity that can enter and where it can be located.

Two rival governments are vying for the same chip, each controlling a different half of its journey.

Hong Kong Becomes the Waiting Room

Outside the mainland customs border

Hong Kong occupies a unique position. The territory is distinct from the customs border of mainland China, while remaining under Chinese sovereignty. According to reports, companies are being encouraged to set up a large portion of the H200 chips authorized by Washington there.

This solution serves Beijing’s interests: it provides access to advanced computing without fully opening the mainland market to American hardware. It also serves Chinese companies, which can leverage computing power close to their teams and networks.

Hong Kong is becoming a buffer zone: Chinese enough to serve Beijing’s leading companies, yet separate enough that importing hardware does not appear to be an industrial capitulation on the mainland.

Regulatory Geography

A chip has no nationality of its own once installed. Yet its location changes its status, its controls, its access, and the way political authorities frame its use.

A few kilometers of administrative border can thus matter more than a generation of architecture. Computing is global in its effects, but extraordinarily local in its power outlets.

The Sino-American rivalry, therefore, does not merely draw a map of manufacturers. It draws a map of the places where artificial intelligence can be legally trained.

The most strategic territory is sometimes not the one that designs the chip, but the one that has the right to power it on.

The wall is no longer the license

Megawatts Before Models

Installing thousands of accelerators requires much more than a warehouse. It requires data centers capable of housing high-density servers, providing continuous power, dissipating heat, and connecting clusters with sufficiently low latency.

Articles reporting on the Hong Kong plan highlight a shortage of local capacity. Precise figures on future megawatts vary depending on the projects and calculation methods; the constraint, however, remains consistent: a permit without infrastructure results in a tied-up or underutilized resource.

Geopolitics can sign a permit in a single day. It cannot build a substation, a cooling system, or a high-density data center by the mere force of a decree.

The calculation has substance

We talk about AI as if it were a cloud. The H200 reminds us that it has a physical weight: circuit boards, servers, cables, cooling water or air, workers, land, and power grids.

This physicality turns the narrative on its head. The question is no longer just whether China can buy an American chip, but where it can deploy tens of thousands of units without hitting the physical limits of its territory.

Hong Kong may be a regulatory compromise, but it remains an industrial bottleneck.

The cloud always ends up somewhere inside a building that generates heat.

Engineers are becoming a queue

Who gets the computing hours

Behind every quota are teams. A lab with a large fleet can test more hypotheses, run simulations longer, serve more users, and start over after a failure without sacrificing as many weeks.

A limited fleet forces trade-offs. The priority project moves forward. The secondary project waits. One team gets the cluster; another must scale back, rent elsewhere, or work on a less demanding architecture.

The chip war trickles down to the engineer’s schedule: the power of a nation becomes a computing hour granted, postponed, or denied to someone who wanted to try.

Scarcity shapes research

Constraints can stimulate efficiency. Chinese teams have already invested in software optimization, energy-efficient architectures, and the use of domestic hardware. But it would be naive to turn every shortage into an automatic virtue.

Less capacity also means fewer experiments, more compromises, and greater dependence on the players who control access. Ingenuity reduces costs; it doesn’t abolish the laws of physics.

The true impact of these shipments will therefore be measured not only by the number of chips, but by how they change the workflow of the teams receiving them.

Scarce computing power doesn’t just slow down machines; it selects which ideas will have time to thrive.

Nvidia walks a fine line

Selling Without Deciding

Nvidia designs the product, organizes its supply chain, and seeks out customers. But in the Chinese market, the company no longer decides on sales on its own. Washington determines whether exports are allowed. Beijing determines whether imports are allowed. Infrastructure determines whether the products can be used.

In its quarterly report filed with the Securities and Exchange Commission, Nvidia describes export controls as a significant risk to its sales and competitive position. The company must plan under rules that are subject to change.

The global AI leader can manufacture the most coveted product on the planet and still be unable to guarantee that it will cross the two political borders standing in its way.

Inventory Held Hostage

Reports have mentioned a large stockpile of H200 chips, some of which are intended for the Chinese market. The exact volume and its allocation may change. The dilemma, however, is clear: producing before authorization exposes the company to excess inventory; waiting for authorization risks losing the customer or missing the technological window.

Every month counts because the product line is evolving. An unsold chip doesn’t remain static in the product hierarchy; it becomes obsolete as new products emerge.

Nvidia wins when both capitals simultaneously open their doors. It loses when one closes after the other has already given its approval.

In this war, even the dominant supplier is operating against a clock it doesn’t control.

The American Paradox

National Security and Revenue

Washington claims it wants to protect its technological lead and prevent the most powerful capabilities from strengthening a strategic rival. At the same time, it authorizes certain sales of the H200 to approved customers, under certain conditions.

The official rationale is to sell a lower-generation product while maintaining the most advanced frontier. This approach supports the revenue of U.S. companies, keeps Chinese customers within the Nvidia ecosystem, and provides the government with levers for compliance.

The contradiction is no secret: the U.S. wants to slow China down without stopping sales to it, and to protect its lead without handing the market over to Chinese competitors for free.

The Risk of Funding China’s Catch-Up

Even though it is less advanced, the H200 remains useful. Critics may therefore argue that these sales strengthen China’s AI capabilities, with commercial, scientific, and potentially military applications.

Supporters of the policy counter that a total embargo would accelerate domestic substitution, deprive U.S. companies of revenue, and reduce Washington’s visibility into users.

There is no risk-free solution. There is only a trade-off between the capabilities transferred today and the influence retained for tomorrow.

Selling a lower rung is still selling a rung; the question is who gets to climb the ladder thanks to it.

The Chinese Paradox

The Independence Nvidia Seeks

Beijing is pushing its companies to use more domestic semiconductors. This policy aims for strategic autonomy, which is understandable given that Washington can cut off access through regulations.

But the immediate needs of its AI champions do not disappear while the local industry catches up. ByteDance, Tencent, Alibaba, and others want to train, deploy, and improve models now—not just once the domestic ecosystem has reached every level of performance and availability.

China wants to prove it can do without Nvidia while granting its top companies just enough Nvidia to avoid paying too high a price for this demonstration.

Buying Time, Protecting the Market

Small batches can be seen as a compromise: they buy time for research labs without entirely handing the market over to American products. Localization restrictions in Hong Kong, if they prove to be long-term, reinforce this logic.

This compromise comes at a cost. It complicates deployment, fragments infrastructure, and can reduce operational efficiency. But Beijing seems willing to accept a portion of this cost to preserve its industrial policy.

Autonomy is therefore not a clean break. It is a managed transition, filled with dependencies that the state chooses to tolerate without acknowledging them as permanent.

Technological independence often begins with deciding which dependencies one is still willing to tolerate for a while.

Hong Kong Under Pressure

Infrastructure as Politics

If a significant portion of licensed chips is to remain in Hong Kong, the territory will have to absorb new demand for data centers, energy, and connectivity. This could boost investment, but it could also expose the limitations of a densely populated area where every megawatt competes with other uses.

The announced projects do not immediately result in operational capacity. Between the allocation of land and the training of a model lie construction, connectivity, equipment, and permits.

Hong Kong is thus entrusted with a strategic mission without any change to its geography: to serve as a compromise between two superpowers that have shifted their confrontation right into its server rooms.

Remote access does not eliminate the border

Engineers located elsewhere can access clusters installed in Hong Kong. But remote access raises its own questions: data security, bandwidth, latency, operational control, and user compliance.

Computing data can traverse the network more easily than cargo passes through customs. Yet it is not thereby free from all regulations. Governments can monitor the provider, the data center, the client, and the data flow.

The physical territory therefore remains central, even when the user experience appears entirely remote.

The border does not disappear into the network; it merely shifts the location from which it exercises control.

Southeast Asia: A Detour or an Outlet

Computing Power Leased Elsewhere

Even before the relaxation of restrictions on H200s, reports described how Chinese groups were using data centers in Southeast Asia to access advanced accelerators. Leasing computing power abroad allows certain physical or commercial constraints to be circumvented without directly importing the hardware.

This approach is no magic solution. It depends on the U.S. regulations applicable to the provider, end-user controls, local capacity, and the political will of the host country.

When the chip cannot come to the engineer, the engineer sends his work to the chip. It is this shift that future controls will inevitably seek to track.

The Spread of Controls

Singapore, Malaysia, and other regional economies may attract investment in data centers. They may also face increasing pressure to demonstrate that their infrastructure is not becoming a backdoor to restricted users.

The Sino-American rivalry thus exports its compliance costs. Companies that wanted to sell hosting services must now know the identity, usage, and sometimes the actual economic nationality of their customers.

Control over chips becomes control over the cloud, and then control over the lease agreement.

Every successful workaround teaches governments where to set up the next border checkpoint.

The Myth of the Decisive Chip

Hardware Does Not Replace the Ecosystem

Twenty thousand H200s represent significant capacity. They guarantee neither a superior model, nor a scientific breakthrough, nor a sustainable advantage. Performance also depends on data, algorithms, teams, networks, software, training time, and the ability to deploy to users.

An obsession with the number of chips can therefore be misleading. It provides a concrete metric in a competition that is difficult to measure, but it does not automatically reveal what companies will actually do with the hardware.

A chip accelerates what is possible. It does not choose the right idea, the right data, or the team capable of transforming computation into sustainable capability.

What the sources do not yet reveal

We do not know what proportion of the shipments has been installed, where each unit is operating, which models they are assigned to, or what measurable benefit results from them. Nor do we know whether Beijing will authorize much larger volumes.

The figures of 10,000 are based on anonymous sources cited by the Financial Times. They are plausible and widely reported, but they should not be treated as an official tally.

Honesty therefore requires us to distinguish between reported shipments and actual operational capacity.

A box crosses a border in a day; power, however, must still be built.

The Seven Signs of the Coming Shift

What would prove the opening

First sign: public confirmations from Nvidia or purchasing companies. Second: Chinese approvals for other groups. Third: mainland volumes exceeding the initial small batches. Fourth: Hong Kong data centers capable of absorbing the demand.

Fifth: a U.S. modification of licensing terms. Sixth: Chinese data center revenue clearly reappearing in Nvidia’s financial results. Seventh: evidence of large-scale use in specific models or services.

Until these confirmations are in place, the opening remains a crack, not a door. It allows hardware to pass through; it guarantees neither continuity, nor volume, nor a sustainable strategy.

What would signal the closure

Conversely, a tightening of restrictions by Washington, a Chinese suspension, a customs block, an inability to host services in Hong Kong, or an accelerated shift to domestic chips could bring this phase to a close.

Semiconductor policy changes faster than infrastructure. A company may obtain a license under one set of rules and plan its deployment under another.

That is why announcements should be viewed as provisional statements, never as a guaranteed new normal.

In this industry, the future can be overturned by a regulation published before the data center is even completed.

Twenty thousand doesn’t lock anything down

The number resurfaces in a different context

Twenty thousand H200s reported. The figure resurfaces, but its meaning has changed. At first, it seemed to herald Nvidia’s return to China. In the end, it mainly demonstrates the precision with which three jurisdictions can carve up access to the same capacity.

Washington chooses the generation and the conditions. Beijing chooses the companies, the volumes, and the location. Hong Kong must provide the physical embodiment of the compromise. Nvidia provides the product, but not the freedom to sell it. Engineers receive the processing power, but not equally.

The most strategic chip isn’t the one that exists. It’s the one that an authorized engineer can actually use, for long enough, in a facility capable of powering it.

The war lies in the permissions

There are no trenches around an H200. There are forms, user lists, tests, customs borders, data centers, and megawatts. This administrative banality does not make the conflict any less fierce. It makes it deeper.

States have realized that AI is governed by bottlenecks. Whoever controls one step can slow down the entire chain without banning the word “intelligence” or seizing a single model.

Twenty thousand chips may have begun to circulate. The freedom to compute, however, remains locked away in a series of keys.

When three powers each hold a key, who really owns the machine?

Signed, Maxime Marquette, columnist

Columnist’s Transparency Statement

Editorial Positioning

I am not a journalist, but a columnist and analyst. My expertise lies in observing and analyzing the geopolitical, economic, and strategic dynamics that shape our world. My work consists of dissecting political strategies, understanding global economic trends, contextualizing the decisions of international actors, and offering analytical perspectives on the transformations that are redefining our societies.

I do not claim to possess the cold objectivity of traditional journalism, which is limited to factual reporting. I strive for analytical clarity, rigorous interpretation, and a deep understanding of the complex issues that affect us all. My role is to make sense of the facts, situate them within their historical and strategic context, and offer a critical analysis of events.

Methodology and Sources

This text respects the fundamental distinction between verified facts and interpretive analyses. The methodological rule is consistent: factual information is published only if it is supported by a verifiable source, and the sources actually used in this article are listed under “Sources,” never here.

Categories of primary sources used by the publication, when applicable: official press releases from governments and international institutions, public statements by political leaders, reports from intergovernmental organizations, and dispatches from recognized international news agencies.

Types of secondary sources: specialized publications, internationally recognized news media, analyses from established research institutions, and reports from sector-specific organizations.

When an article cites statistical, economic, or geopolitical data, it comes from data-producing institutions (intergovernmental organizations, central banks, national statistical institutes), and the specific institution is listed under “Sources.”

Nature of the Analysis

The analyses, interpretations, and perspectives presented in the analytical sections of this article constitute a critical and contextual synthesis based on available information, observed trends, and expert commentary cited in the sources consulted.

My role is to interpret these facts, contextualize them within the framework of contemporary geopolitical and economic dynamics, and give them coherent meaning within the broader narrative of the transformations shaping our era. These analyses reflect expertise developed through continuous observation of international affairs and an understanding of the strategic mechanisms that drive global actors.

This article describes a situation documented as of its publication date, not a prediction: subsequent developments may alter these perspectives. No updates are promised in advance; when an article is corrected or supplemented, the change is dated within the text.

ANALYSIS: H200, the battle for access between Washington, Beijing, and Hong Kong