Hyper-Dense AI: The Grip on China and the Geopolitical Race - Report
- Gabriele Iuvinale
- 30 mag
- Tempo di lettura: 8 min
Executive Summary
The Artificial Intelligence (AI) landscape is undergoing a radical transformation, driven by power density, that is, the ability to concentrate immense computational capabilities in tiny spaces. This trend, extensively analyzed by the Goldman Sachs Global Institute, is not only a technological phenomenon, but has profound geopolitical, economic and military implications, especially for China.
The progression is staggering: a single modern server cabinet can outperform a supercomputer of just a few years ago by five times. This compression is vital for advanced AI models that require interconnected GPUs and ultra-fast communications. However, this creates enormous infrastructure challenges: power consumption equivalent to hundreds of homes per cabinet, the need for sophisticated liquid cooling systems, and the obsolescence of traditional data centers, requiring new hyperspecialized infrastructure. These massive investments are crucial for AI cost-effectiveness and for unlocking entirely new AI capabilities, today only theoretical, promising transformative innovations.
China finds itself at the center of this geopolitical dynamic. Export restrictions imposed by the United States from October 2022, aimed at limiting access not only to advanced chips but, crucially, to high-speed interconnect technologies, are creating an ever-tightening stranglehold. Goldman Sachs' analysis points out that with new architectures integrating hundreds of interconnected GPUs, the technology gap is becoming virtually unbridgeable for China, which will no longer be able to import comparable systems. This accelerates Beijing's push toward domestic chip and interconnect production, fueling an increasingly pronounced divide in the global technology landscape.
A key element in this scenario is the recent withdrawal of a U.S. regulation. The U.S. Department of Commerce (BIS) announced on May 13, 2025, the repeal of the "Artificial Intelligence Diffusion Framework," a rule that imposed global licensing requirements. Despite the repeal, BIS issued three guidance documents clarifying enforcement priorities toward China, with warnings on the use of U.S. AI chips for Chinese models, recommendations against diversion tactics, and a warning against the use of advanced chips developed or manufactured in China. Currently, China is pushing hard toward self-sufficiency, with manufacturers such as Loongson introducing all-Chinese servers and Inspur building domestic clouds. At the same time, Beijing is investing in the development of advanced AI models, with startups such as DeepSeek competing with Western giants. However, restrictions on high-speed interconnections remain a critical challenge, creating a gap that is difficult to close in the short term.
In the military, China risks a technology gap in the new generation of AI-based weapon systems (autonomous, strategic, cyber warfare), with lower performance due to limitations in interconnections. This can lead to difficulties in the development of strategic and decision-making AI, persistent dependence on foreign technologies and high costs for domestic development, as well as the risk of a "two-speed" military. On the economic and technological front, China faces an extremely costly path to self-sufficiency, with massive investments that could divert resources. The current performance of Chinese products still does not match that of global leaders, limiting their international competitiveness. U.S. attempts to deter collaborations with Chinese entities risk a fragmentation of the global AI market, isolating China and hindering its access to partnerships and talent. The significant and persistent lag in technological development due to limitations on interconnections threatens Beijing's coveted technological leadership.
In sum, while China is determined to achieve AI self-sufficiency, the "Achilles' heel" of high-density interconnections, exacerbated by U.S. sanctions and external technological acceleration, poses a real risk to its ability to compete at the same level as the United States both in next-generation AI and in maintaining its economic growth and military security.
The Artificial Intelligence (AI) landscape is undergoing a radical transformation driven by power density – the ability to concentrate enormous computing power into increasingly smaller spaces. This shift, extensively analyzed by the Goldman Sachs Global Institute in its report "1,000 power homes in an archive: the growing power density disrupts AI infrastructure," is not merely a technological phenomenon but has profound geopolitical, economic, and military implications, especially for China.
The Battle for Hyper-Dense Connections and US Restrictions
China faces a crucial obstacle in its AI development: hyper-dense connections. These ultra-fast "digital highways" are essential for efficiently communicating a vast number of processors (GPUs) within advanced AI systems. Starting in October 2022, the United States imposed severe export restrictions specifically targeting these interconnection technologies, aiming to limit China's ability to build cutting-edge AI with potential strategic or military implications.
These measures have forced companies like NVIDIA to create "de-powered" versions of their chips (such as the H800 or H20) for the Chinese market. While these chips retain significant raw computing power, their inter-chip data transfer speed is drastically reduced. Initially, for smaller AI clusters, this limitation could be managed. However, as highlighted by the Goldman Sachs analysis, next-generation AI architectures (like NVIDIA's "Oberon" systems or the future "Kyber," which integrate hundreds of GPUs) require near-instantaneous coordination. Without rapid interconnections, these architectures become ineffective, creating "bottlenecks" that waste computing resources. The consequence is that China can no longer simply import systems minimally comparable to those available elsewhere, creating a virtually insurmountable technological gap in the short term. Faced with this chokehold, China is accelerating its massive investments in domestic chip and interconnection technology production, aiming for self-sufficiency. This scenario is inexorably leading to a sharper division in the global technological landscape, with two separate AI ecosystems competing.
Shifts in US AI Chip Export Policy
A key element in this scenario is the recent withdrawal of a US regulation. The U.S. Department of Commerce's Bureau of Industry and Security (BIS) announced on May 13, 2025, the repeal of the "Artificial Intelligence Diffusion Framework," a Biden-era rule that imposed global licensing requirements for advanced semiconductors and computing products.
This withdrawal, pending a new regulation within the next four to six weeks, reflects the Trump administration's objections regarding stifling US innovation and undermining diplomatic efforts. Despite the repeal, the BIS has published three guidance documents clarifying its enforcement priorities, particularly concerning China: an advisory on potential consequences from allowing the use of US AI chips for training and inference of Chinese AI models; recommendations to protect supply chains from diversion tactics; and an explicit warning against the use of advanced integrated circuits (ICs) developed, designed, or manufactured in China.
The Status Quo of Chinese AI: Progress and Obstacles
Currently, China is aggressively pursuing self-sufficiency in AI and high-density servers to overcome US-imposed limitations. Among the notable developments, CPU manufacturer Loongson has unveiled the first entirely "Made in China" server, leveraging proprietary processors and interconnection technologies. This marks a significant step towards independence, although its performance does not yet match global leaders. In parallel, Inspur is building its cloud without Western components, while Huawei is working to defend its AI chips globally against US pressure. Despite these hardware efforts, Beijing is investing massively in developing advanced AI models, with startups like DeepSeek directly competing with Western giants and the emergence of ambitious projects like the "general agent" AI Manus (though still in experimental phase). China is also investing in AI education from primary school to cultivate future generations of tech talent. However, US restrictions on high-speed interconnections remain a critical challenge: they prevent China from importing systems with the necessary computing density for the most complex AI models, creating a short-term technological gap that is difficult to bridge.
The Silent Revolution: Hyper-Dense AI Reshaping the World
The world of Artificial Intelligence is experiencing a rapid transformation driven by "power density." As the Goldman Sachs Global Institute's analysis highlights, we are moving from traditional data centers to true "power archives." While the "Summit" supercomputer occupied a basketball-court-sized area in 2018, today a single NVIDIA server cabinet, the NVL72, is five times more powerful yet occupies only one three-hundredth of that space. This incredible ability to concentrate power is fundamental for modern AI, as models require their GPUs to communicate at light speed.
This power compression brings monumental challenges:
Unstoppable Thirst for Energy: A single NVIDIA "Oberon" server cabinet consumes ten times more power than an AI server from a few years ago. By 2027, the "Kyber" system, with 576 GPUs in a single cabinet, is projected to require enough energy to power 500 homes. We're talking about 50 times higher power density than five years ago, and the trend is still upwards.
An Inferno of Heat: Such immense power generates incredible heat, necessitating liquid cooling systems with complex infrastructure.
Goodbye Old Data Centers: Current data center structures are unsuitable; next-generation AI demands new, purpose-built data centers designed for these extreme requirements.
Investing heavily in these new infrastructures makes precise sense: it makes AI more economical and efficient by reducing operational costs for training and inference, and crucially, it unlocks the possibility of entirely new AI capabilities that are currently only theoretical, leading to revolutionary innovations.
The Big Bet: Specialization and Future Risk
Building these new AI infrastructures is an enormous gamble. These are not versatile structures but highly specialized "AI cathedrals." As the Goldman Sachs analysis points out, if AI demand were to decline, these data centers would not be easily convertible, making financing more complex and their future value less certain. Yet, for companies with true AI ambitions, immobility poses a greater risk. Infrastructure decisions made today will define their competitive positioning for the next decade. With construction times exceeding 18 months and systems like "Kyber" set to revolutionize power density by 2027, the time to act is now.
China's Military Risks in Hyper-Dense AI: The Critical Role of Interconnections
In the context of AI and hyper-dense server development, China faces significant challenges in its military advancements, primarily due to limitations on crucial high-speed interconnections. These restrictions, highlighted by the Goldman Sachs analysis, create several concrete risks:
Technological Gap in Next-Gen AI-Driven Weapon Systems: China risks developing a gap in the next generation of AI-based weaponry (autonomous systems, advanced strategic planning, cyber warfare, large-scale surveillance), which demand immense computing power and efficient chip communication.
Inferior Performance in Advanced Systems: Even with many AI chips, limitations in interconnections reduce their effectiveness when working in networks, leading to higher latency and less real-time processing, making systems less responsive and potentially inferior.
Difficulties in Strategic and Decision-Making AI: Delays in developing AI for strategic planning, logistics, and war scenario simulations, which require massively interconnected clusters, could compromise China's ability to achieve strategic superiority.
Reliance and Vulnerabilities: Despite self-sufficiency efforts, a persistent gap in interconnections could maintain reliance on external solutions, creating a strategic vulnerability in case of further sanctions.
High Costs and Domestic Inefficiency: The push for self-sufficiency incurs enormous costs and long timelines, diverting resources and potentially leading to less efficient systems.
Risk of a "Two-Speed" Military: A disparity might emerge between military units with access to more advanced AI technologies and those limited by less capable hardware, impacting overall cohesion and effectiveness.
China's Economic and Technological Risks in the Hyper-Dense AI Era
In its bid to become a global leader in artificial intelligence, China faces significant economic and technological risks, primarily stemming from US restrictions on key hyper-dense interconnection technologies.
Cost of Self-Sufficiency: China is embarking on an extremely expensive path towards technological self-sufficiency, with massive investments and years of R&D that could divert resources from other vital economic sectors.
Limited Competitiveness: The current performance of Chinese products, though improving, does not yet match global leaders, limiting the competitiveness of Chinese companies in the international AI market.
Global Market Fragmentation: US attempts to discourage collaborations with Chinese AI entities risk fragmenting the market, isolating China and hindering its access to crucial partnerships, talent, and supply chains.
Persistent Technological Lag: Limitations on high-speed interconnections prevent China from effectively building and scaling the hyper-dense servers needed for training complex AI models, creating a short-term gap that is hard to bridge. This slows China's ability to develop next-generation AI for both civilian and military applications, jeopardizing its coveted technological leadership and the expected economic value from AI expansion.
Comments