AI Infrastructure Investment Needs to Go Beyond GPUs to Networks, Lawmakers Forum Says

As competition in AI intensifies, calls are growing to expand the scope of AI infrastructure investment to include networks. The essential infrastructure elements are also widening to include GPU, data centers, power, and wired networks. Industry players are arguing that AI infrastructure investment should go beyond computing resources to encompass networks as well.

Against this backdrop, the "Global Forum for a Viable Digital Ecosystem" was held on the 22nd at the initiative of lawmaker Lee Jung-heon. At the forum, network investment in the AI era, network usage fees, and regulatory policy were discussed.

At the event, speakers explained that network importance is rising as the generative AI industry spreads. Kim Myeong-su, a professor at Kangwon National University, said networks are a key factor determining the quality of AI services. He especially noted that AI service performance is affected by whether large volumes of data can be transmitted and received stably.

The network segments cited in this context included FTTH, backbone networks, edge networks, and data center interconnection networks. Kim Myeong-su, a professor at Kangwon National University, said the overall quality of the network can determine AI service performance.

Omdia said the spread of AI is expected to sharply increase both total mobile traffic and the share of AI traffic. Omdia projected that AI will drive mobile traffic to grow 2.4 times from 2025 to 2030.

Omdia forecast that AI traffic will account for 52% of total traffic in 2030. It also projected that the share would rise further to 81% in 2035.

The basis for these projections was a shift in traffic structure. Traditional traffic consisted mainly of one-way content delivery centered on the web and video, but the recent spread of interactive AI services is now driving change. As a result, future traffic is expected to feature a larger share of real-time request, computation, and response feedback loops.

The rise in AI traffic is also highlighting the importance of telecom infrastructure investment. However, telecom carriers' CAPEX is shrinking. CAPEX for the country's three mobile carriers fell from KRW 9.6 trillion to KRW 6 trillion over the past 6 years.

The industry sees network investment demand rising without a proportional increase in telecom carriers' revenue. Mani Manimohan and the GSMA explained the "decoupling" phenomenon, in which network traffic and telecom carrier revenue diverge after traffic growth.

Under the traditional structure of the telecom market, users pay telecom fees, and carriers have used that revenue to build and maintain networks. However, as the influence of large content providers, AI companies, and cloud providers over traffic volume and delivery methods has expanded, and as the quality required of telecom networks has risen, there are calls for discussion on whether continued investment is possible under the existing structure alone.

There were also calls for an AI infrastructure strategy. The scope of AI infrastructure was described as computing, power, and networks, and the tasks identified included securing investment funds and improving regulatory predictability. The purpose of improving regulatory predictability is to expand private investment.

Another necessary task cited was network upgrading. Speakers said that if network upgrading improves AI service productivity, a virtuous cycle is needed so that those gains can lead to reinvestment.

Along with these discussions, a key issue raised was the imbalance in network cost structures between major overseas CPs and domestic CPs. According to domestic wireless traffic shares announced by the Ministry of Science and ICT in December 2023, Google accounted for 30.6%, Netflix for 6.9%, and Meta for 5.1%. Among domestic companies, Naver accounted for 2.9% and Kakao for 1.1%.

In South Korea, there are differences in the burden of network usage fees by provider relative to traffic contribution. Google accounts for a high share of total domestic wireless traffic through YouTube and other services, but does not directly pay network usage fees to telecom carriers. Naver, meanwhile, paid KRW 73.4 billion in 2016, and Kakao paid about KRW 30 billion in 2016.

In this regard, experts said that voluntary commercial negotiations should take precedence over the government's uniform setting of network usage fees. They also said that if negotiations become deadlocked, mechanisms for mediation, arbitration, and dispute resolution need to be established.

Experts said discussions over network usage fees need to move beyond whether a specific company bears costs. They argued that the debate should be expanded to the question of how ecosystem participants share the cost of massive network infrastructure investment in the AI era.

A telecom industry official said Big Tech generates significant traffic in South Korea but remains relatively reluctant to shoulder corresponding network usage fees. The official said the European Union is facing a similar problem, and added that the EU is pursuing the Digital Network Act (DNA), which would allow regulators to mediate if voluntary negotiations fail.

Professor Kim suggested that costs need to be shared fairly, saying that this would allow telecom carriers to reinvest in networks and, as a result, better support AI services. He added that a virtuous cycle needs to be established in which cost burdens, network reinvestment, and AI service support repeat.

Source: IT DAILY · Seong Won-young
Original: https://www.itdaily.kr/news/articleView.html?idxno=241800

References

• https://www.itdaily.kr/news/articleView.html?idxno=241800