Will 2026 be the year investors and governments finally take the resilience of AI’s enabling infrastructure seriously? And what may be their implications on U.S.-China AI competition?

Much has been said and predicted about the year 2026 being a critical turning point for artificial intelligence. AI is pivoting from experimental to operational mandate, or, in other words, it is no longer a “pilot project” for various industries but has become a core business component. Rather than chasing after higher AI performance metric, it will be more about solving real tasks of economic consequences, including multi-modal reasoning and multi-step problem solving. The emergence of agentic AI signifies the rise of the operational role of AI in a new generation of semi-autonomous enterprises where actions are taken and tasks are coordinated by AI models.

However, in spite of the shift to full‑scale operational integration of AI, the very resilience of AI itself is often overlooked, and the importance of AI’s enabling infrastructure, especially the undersea cable systems that underpin the global Internet connectivity, is under-discussed, under-appreciated, and under-invested. According to Telegeography, even though the global value of new undersea cables planned to enter service between 2025 and 2027 is forecasted to reach over US$13 billion, that is a drop in the bucket compared with the US$280 billion estimate for the global spending on AI investment for the year 2025 alone.

Undersea cables hardly even get noticed when the proverbial “AI stack” is mentioned. In 2023, in an early view of the AI tech stack by the venture capital firm Andreessen Horowitz, three layers were identified: applications, models, and “infrastructure vendors,” with the final group defined as “cloud platforms and hardware manufacturers that run training and inference workloads for generative AI models.” When Nvidia CEO Jensen Huang talks about AI as a “five-layer cake,” the layers are the chips, infrastructure, models, applications, and the energy layer. Even when the White House AI Action Plan mandates that the U.S. “must meet global demand for AI by exporting its full AI technology stack,” the elements of the stack includes only “hardware, models, software, applications, and standards.”

Unfortunately, all consideration for what constitutes AI infrastructure appears to terminate at the datacenters, which gets all the attention, from investment and construction to their demand on the power grids. But AI datacenters cannot run on their own as islands, separated from one another, as, simply put, AI runs on the Internet, and the Internet runs on physical infrastructure connected by cables, land or undersea. So, if, by all accounts, as much as 99% of intercontinental Internet traffic is transmitted through undersea cable systems, it can only logically follow that the same 99% of intercontinental AI transactions must be transmitted in the same way, through undersea cables.

Where AI sovereignty meets connectivity reality

But what about AI sovereignty—the trend to keep data, localized models and their training and compute within national borders? Does that mean the need for interconnection and transmission of data will be reduced?

Of course not. Hyperscale datacenters still need to be connected to one another. Training workload of large models are typically distributed for best performance, and model parameters, gradients and weights must be continuously synchronized across different compute nodes. During inference, when serving AI models to users, load balancing and redundancy across different datacenters are important. That is why AI hyperscalers are investing in datacenters around the world, and it is not only for market presence and localization.

Indeed, those cloud and social media big tech players—Google, Microsoft, Meta, Amazon—are already major investors in building new global undersea cable systems in recent years, and they also happen to be, of course, right in the center of the AI race. Pure AI players such as OpenAI may be somewhat behind in this regard, but my prediction is that they should not be far behind. It is true that AI model companies may be still prioritizing their capital investments on GPUs and datacenters, but the importance of controlling their own end-to-end destiny including undersea cables will become more evident and pressing, as their business and revenue models enter the next phase of mass industry adoption.

It should be noted that although OpenAI has not directly invested in undersea cable ownership, but Microsoft, one of its major backers, is a significant investors in this area, while its major investor Softbank has also invested in undersea cable projects such as Candle, E2A (East Asia to North America), JUPITER, and ADC (Asia Direct Cable), but they tend to be more Asia or even Japan focused only, rather than on a global scale. At some point, and possibly sooner rather than later, OpenAI may have to strive for the same level of autonomy in global cable connectivity as its AI competitors. Other major U.S. companies in the AI ecosystem to watch in this space to potentially invest directly in cable infrastructure should also include Nvidia, and XAI.

The Chinese factor for undersea cable supremacy

It may also be noted that undersea cable investment is one area that Chinese AI and Internet firms lag behind their U.S. counterparts. None of the Chinese big tech, such as Baidu, Tencent, Alibaba, and let alone smaller upstarts such as DeepSeek, have invested in undersea cables, which remains exclusively under the scope of the Chinese state-owned telecom enterprises, which are not competitive on the global scale outside of China’s borders.

Even Huawei, which had been involved in cable manufacturing and construction, sold its interests in its Huawei Marine Networks to other Chinese owners. That company became HMN Technologies, which has been put on the entity list of the U.S. Department of Commerce, and is forbidden to be a part of any project involving U.S. companies or requiring U.S. technology components.

In other words, the direct ownership and end-to-end control of the interconnection of their global datacenters may prove to be a competitive advantage that leading AI players in the U.S. It is fundamentally one layer of the AI tech stack that China will find it hard to match the U.S. on the global scale possibly for a rather substantial foreseeable future. When it comes to global adoption and customer-serving operation, this may be a significant edge by the west over their Chinese competitors.

This is why, in terms of AI industry investment strategy ,regulatory policy, and global competition, incorporating undersea cables as a key part of thinking over the AI tech stack will be critical for the global race for AI leadership.