The U.S. Department of Commerce has moved to close what experts describe as a significant gap in America’s AI export control regime, issuing new guidance that could restrict Chinese companies from acquiring the world’s most advanced artificial intelligence chips through overseas subsidiaries.

The action targets a potential weakness that emerged after the Trump administration abandoned enforcement of the Biden-era AI Diffusion Rule in May 2025. Industry observers say the decision may have inadvertently allowed Chinese firms operating outside mainland China to purchase cutting-edge processors from companies such as Nvidia and AMD without the licenses typically required for exports to China.

The new guidance, released on Sunday, clarifies that advanced AI chips exported to entities headquartered in China will remain subject to U.S. licensing requirements regardless of where those entities are physically located.

The move represents the latest effort by Washington to tighten restrictions on technologies viewed as critical to the global race for artificial intelligence supremacy and national security.

At the center of the issue are Nvidia’s Blackwell and Rubin processors and AMD’s MI350X accelerators, chips considered among the most powerful AI computing products currently available. These processors are essential for training and deploying frontier AI models and have become a focal point of U.S. efforts to limit China’s access to advanced computing capabilities.

The Commerce Department’s clarification suggests concerns inside Washington that Chinese firms may have been exploiting overseas operations in countries such as Malaysia and other Southeast Asian technology hubs to continue acquiring advanced U.S. hardware.

While the government has not disclosed the scale of the activity, one industry source familiar with semiconductor supply chains estimated that hundreds of thousands of advanced chips may have been shipped during the period in question, according to Reuters.

If accurate, such volumes would represent a substantial flow of AI computing power at a time when U.S. policymakers have been attempting to constrain China’s ability to develop next-generation artificial intelligence systems.

Technology analyst and former U.S. State Department official Chris McGuire described the issue as a major vulnerability in America’s export control framework.

“This is a HUGE problem,” McGuire wrote on social media, arguing that the gap effectively allowed overseas subsidiaries of Chinese firms to purchase Nvidia Blackwell processors without obtaining export licenses.

“Chinese companies have been buying these chips, very likely at scale,” he added.

While restrictions can prohibit direct shipments into China, multinational corporations often operate extensive networks of subsidiaries, affiliates, and data center operations across multiple jurisdictions. As a result, regulators have increasingly focused not only on where advanced chips are shipped, but also on who ultimately controls and uses them.

The issue has become more pressing as Southeast Asia emerges as a critical hub for AI infrastructure investment. Countries such as Malaysia, Singapore, and Indonesia have attracted billions of dollars in data center spending from global technology companies seeking access to power, land, and connectivity needed for AI computing.

Those same locations have also become areas of concern for U.S. officials seeking to prevent advanced semiconductors from indirectly reaching restricted Chinese entities.

The guidance reflects a broader shift in U.S. policy from controlling physical exports to monitoring beneficial ownership and end-use access. Rather than focusing solely on geography, regulators are increasingly scrutinizing whether Chinese companies can gain effective access to advanced AI computing resources regardless of where those resources are located.

Importantly, the Commerce Department stopped short of requiring existing facilities to shut down operations or disconnect servers already equipped with advanced processors. That distinction may prove significant for cloud providers, data center operators, and infrastructure investors that have built AI computing facilities serving multinational customers.

The guidance appears aimed primarily at preventing future transactions rather than disrupting existing deployments.

The decision is likely to create new compliance challenges for semiconductor manufacturers, cloud providers, and data center operators, which may now face heightened scrutiny over customer ownership structures and cross-border corporate relationships.

For Nvidia and AMD, the clarification adds a fresh challenge to an already evolving regulatory landscape. Both companies have repeatedly found themselves caught between growing global demand for AI computing power and increasingly stringent U.S. export restrictions.

The latest action also underscores a broader reality of the U.S.-China technology rivalry: as export controls become more sophisticated, companies and governments continue searching for new pathways to access critical technologies.

Washington’s latest move exposes policymakers’ belief that controlling advanced AI capabilities requires monitoring not just where chips are shipped, but who ultimately benefits from their computing power.

The episode also raises questions about how much advanced AI hardware may already be operating within the global networks of Chinese technology firms. While the new guidance closes a potential pathway for future acquisitions, it remains unclear how much computing capacity may have already been accumulated during the period when the regulatory ambiguity existed.

SoftBank Group plans to invest 45 billion euros ($53 billion) over the next five years in France, marking one of the most ambitious artificial intelligence infrastructure commitments ever announced in Europe.

The Japanese technology conglomerate said the investment forms part of a broader 75-billion-euro initiative aimed at deploying 5 gigawatts of AI data-center capacity across France. The project places France at the center of Europe’s effort to narrow the widening AI infrastructure gap with the United States and China, both of which have raced ahead in building the computing capacity required for next-generation AI systems.

SoftBank said the first phase will focus on constructing 3.1 gigawatts of AI data-center capacity in northern France by 2031, with major facilities planned in Dunkirk, Bosquel, and Bouchain. The company described the initiative as its largest AI infrastructure investment in Europe.

“The commitment marks SoftBank Group’s largest AI infrastructure investments in Europe,” the company said. “It is designed to support the rapid growth of artificial intelligence by expanding access to high-performance compute capacity in France.”

The scale of the announcement highlights a growing realization across governments and technology companies that access to computing power has become one of the most important strategic assets in the AI economy. Training and operating advanced AI models now requires vast networks of data centers packed with specialized processors, memory systems, and networking equipment, creating an arms race for compute capacity among nations.

Speaking alongside French President Emmanuel Macron, SoftBank founder and CEO Masayoshi Son suggested the initiative could ultimately represent a much larger economic undertaking than the headline figure implies.

“It’s a massive size of investment coming,” Son said. “We are doing that in the U.S. already, we are expanding a lot in the U.S., so we have the momentum, which we can make France the center of Europe, and Europe needs this kind of AI technology.”

“There’s no choice. U.S. is going fast, China is going fast, Europe, Japan, Asia have to also go fast, not to be left out,” he added.

The project reflects Son’s increasingly aggressive strategy of positioning SoftBank at the heart of the global AI ecosystem. Over the past two years, the company has transformed itself from a technology investor into one of the world’s largest backers of AI infrastructure.

Its exposure spans multiple layers of the AI value chain. SoftBank controls a significant stake in semiconductor designer Arm Holdings, whose processor architectures are widely used in AI servers and data centers. It has also emerged as one of the largest investors in OpenAI, committing tens of billions of dollars to the company behind ChatGPT.

The French initiative underpins how AI infrastructure is increasingly becoming a preferred investment theme among global technology investors. While much attention has focused on AI models and applications, industry leaders are viewing data centers, power systems, semiconductor supply chains, and networking infrastructure as the critical bottlenecks that will determine who wins the next phase of the AI race.

France has aggressively positioned itself as Europe’s preferred destination for large-scale AI investment. The country benefits from a substantial nuclear energy base, which provides relatively stable electricity supplies compared with some neighboring countries. Access to reliable power is becoming a decisive factor for AI infrastructure projects because advanced data centers consume enormous amounts of electricity.

That advantage could prove increasingly important as Europe grapples with rising energy costs. The continent’s ambitions to compete with the United States and China in artificial intelligence have been complicated by higher electricity prices and concerns over energy security. Those challenges have intensified amid disruptions linked to the U.S.-Iran conflict, which has driven volatility in global energy markets.

The economics of AI infrastructure are increasingly tied to access to affordable power. Industry analysts have warned that regions unable to provide reliable and competitively priced electricity may struggle to attract major AI investments, potentially creating a divide between European countries that can support hyperscale data centers and those that cannot.

SoftBank’s choice of northern France reflects those realities. The region offers industrial infrastructure, access to power networks, and proximity to major European markets, making it an attractive location for AI computing hubs. The company said it will partner with Schneider Electric to develop a large-scale industrial production cluster in Dunkirk, further embedding the project within France’s broader industrial strategy.

Investors welcomed the announcement, sending SoftBank shares sharply higher. The stock has already surged more than 70% this year as markets increasingly view the company as one of the biggest beneficiaries of the AI boom.

The investment also signals growing confidence that demand for AI computing capacity will continue to expand rapidly over the coming decade. Technology companies are pouring unprecedented sums into data centers as they compete to train larger models, deploy AI agents, and support enterprise adoption of generative AI.

Accountants call it marginal cost. Economists often discuss it through the lens of unit economics. Whatever discipline you choose, the underlying question is the same: what happens to the cost of producing one more unit as scale increases?

Traditional software companies enjoyed one of the most beautiful economic characteristics ever discovered in business. Once the software has been built, the cost of serving an additional customer approaches zero. As users increased, the marginal cost curve moved closer and closer toward zero, creating a near-asymptotic relationship. This economic structure enabled extraordinary operating leverage and helped create some of the most valuable companies in history. In practical terms, adding more customers often made the business stronger, more profitable, and more efficient.

That is why software produced something Adam Smith never truly experienced in his era: accelerating returns. The fixed asset, the software platform, could continue serving increasing numbers of users while variable costs grew only marginally. The result was a business model where scale itself became a competitive advantage.

Artificial Intelligence changes that equation. Unlike traditional software, AI often behaves more like a classical industrial enterprise. Every new user may trigger additional inference costs, computing costs, storage costs, model-serving costs, and infrastructure expenses. As usage scales, costs do not naturally collapse toward zero. In many cases, they rise alongside demand. The economic profile begins to resemble manufacturing more than software.

This introduces the old economic reality of diminishing returns. If not carefully managed, each additional customer may contribute less value than the previous one. In extreme situations, growth itself can become expensive. Yes, more customers can actually push an AI company toward financial distress if the unit economics are poorly designed!

This challenge explains why nearly every serious AI company is attempting to build proprietary inference infrastructure, optimize models, develop custom chips, or reduce dependence on third-party providers. The battle is no longer merely about intelligence; it is increasingly about economics.

Simply, without solving the inference-cost problem, AI businesses may follow the economic trajectory of traditional industrial companies rather than the trajectory of software legends like Facebook. Put differently, without strong inference economics, AI begins to look more like a cement factory than a social network or software operating system.

And that is why the race to reduce inference costs may become one of the most important competitions of the AI age. It is also one of the reasons Tekedia Capital invested in Piris Lab. We believe the future of AI will not be determined solely by who builds the smartest models, but also by who can run those models most efficiently. Intelligence without economical delivery remains a constrained opportunity. And before AI models can evolve, hardware must have emerged to power them. That conviction is what led us to write the cheque for Piris Lab.

Piris Lab is developing a next-generation photonic computing system designed to perform AI inference at the speed of light. By leveraging photons rather than relying solely on traditional electronic architectures, the company seeks to dramatically reduce latency, improve performance, and lower the cost of deploying AI at scale.

Good People, if AI is to become truly ubiquitous, powering everything from personal assistants and autonomous systems to healthcare, manufacturing, and scientific discovery, the economics must improve. The industry cannot sustainably scale if every additional user significantly increases computational costs.

We see photonics as one of the most promising pathways to solving that challenge. In many ways, future AI winners may emerge not only from advances in algorithms, but also from breakthroughs in the physical infrastructure that makes intelligence affordable and accessible.

We believe the company is helping build the foundational infrastructure required to advance the next phase of the AI revolution.