The Next AI Debate Is About Geopolitics, says the headline of an insightful Foreign Policy piece of 28 October by Jared Cohen, president of global affairs at Goldman Sachs and co-head of the Goldman Sachs Global Institute. He reminds us that data is sometimes called the "new oil". If so, as outlined below, data centres are critical refineries.
According to Juniper Networks, an AI network platforam, there are three phases of developing an AI model:
- Phase 1: Data preparation–Gathering and curating data sets to be fed into the AI model.
- Phase 2: AI training–Teaching an AI model to perform a specific task by exposing it to large amounts of data. During this phase, the AI model learns patterns and relationships within the training data to develop virtual synapses to mimic intelligence.
- Phase 3: AI inference–Operating in a real-world environment to make predictions or decisions based on new, unseen data.
Phase 3 is generally supported with existing data center and cloud networks.
According to Digital Realty, a datacenter platform, "modern companies are running hugely demanding workloads on data center infrastructure. Sticking with the ChatGPT example, a recent report has predicted that training and implementing generative AI models will cost $76 billion by 2028. This figure is more than double the annual cost of Amazon Web Services (AWS), the world’s largest public cloud provider."
As Cohen explains in his Foreign Policy think-piece, "Data centers are the factories of AI, turning energy and data into intelligence. Industry leaders estimate that a few major U.S. technology companies alone are expected to invest more than $600 billion in AI infrastructure, particularly in data centers, between 2023 and 2026."
Cohen points out that AI computation is highly energy intensive, increasingly served by energy-intensive graphics processing units (GPUs) rather than central processing units (CPUs). The International Energy Agency expects that global data center electricity consumption could double as soon as 2026, driven largely by AI.
AI workloads are ultra-high-density, requiring concentrated power supplies. Data centers must operate 24/7, requiring land with access to abundant, affordable, and reliable power. Many developed markets’ electrical grids are unprepared and unused to adapting rapidly and at scale—in the United States, the lead time to build upscale electricity grid assets is up to 10 years.
The United States’ data centers are primarily based in Silicon Valley and Northern Virginia. The country needs far more power and far more differentiated data centers. America's abundant energy reserves do not translate into transmission, connectivity, or the geographical spread of data center requireements.
As Cohen points out, China is executing its own strategy to lead in AI infrastructure, including dozens of nuclear reactors planned or under construction, one-third of clean energy investments worldwide. and a national data center initiative launched in 2022 called “Eastern Data, Western Computing”, with $6.1 billion investment in eight major data center hubs.
What is more, chips and cables that connect GPUs are made of critical rare earths such as germanium and gallium largely controlled by China. What do China’s curbs on Gallium and Germanium mean for the World?
There are currently around 8,000 data centers globally built by a diverse group of companies, demonstrating a wide array of potential partners for data center buildouts. Indeed, every country will face hard choices about where their AI workloads will run.
Cohen points out that Government and industry are working to improve the efficiency of chips, with denser circuits and new architectures already reducing semiconductor energy needs at remarkable scale. Meanwhile, large-language model workloads require less bandwidth than traditional internet content such as images and videos, and AI applications often have different latency requirements than traditional cloud services. While the debate is far from settled, innovation may enable future data centers to be built farther away from customers.
Cohen concludes that to win in today’s high-stakes geopolitical competition, the United States will need to enlist its asymmetric advantage of global alliances and partnerships, both in the public and private sectors. Such partnerships include close allies or friendlier countries which are relatively powerful in the data center market placce. Coming to mind are countries like Canada, the Nordic countries in Europe, Japan, South Korea, India, Brazil, Vietnam, the Philippines, and the Gulf states (with subsea fibre optic cables in the Red Sea and the Persian Gulf carrying 90 percent of Europe-Asia data traffic.)
What Cohen omits to mention is that global data transmission is closely related to international trade and economics, which have become much more closely interwined and interdependent. As the world's largest trader and manufacturer par excellence, China is deeply embedded in the global supply and value chain, encompassing data generation and data flows, which remain the backnone of AI development, including data centre networks, "de-coupling" or 'de-risking" notwithstanding.
As the inter-connected digital age of the 4th and 5th Industrial Revolution continues to unfold, the battle for AI data centres, like US-China great power rivalry, cannot be framed in Manichean terms.
A PowerPoint presentation follows. Download The geopolitics of AI data centres
