Much has been said and written about artificial intelligence (AI) and its future impact on productivity, the issue of job losses and the resulting inequalities within and between countries. Some studies indicate that AI has the potential to eliminate millions of jobs in a variety of industries, including manufacturing, inventory management, sales, logistics, and services such as software, legal and accounting services, insurance, and more. Just to name a few.
“Creating a 21st century AI superpower requires four key ingredients: rich data, tenacious entrepreneurs, well-trained AI scientists, and a supportive policy environment,” said Dr. Kai Fu Lee, Chairman and CEO of Sinovation Ventures, a leading technology-savvy investment firm.
He said that while hundreds of companies are pouring resources into AI research, a few giants such as Google, Facebook, Amazon, Microsoft, Baidu, Alibaba and Tencent are investing resources to become global leaders. There is a fierce competition between the United States and China to become the world leader in AI. According to Dr. Li, in the current situation, China has an advantage in entrepreneurship, data and government support, and is rapidly catching up with the United States in expertise. Dr. Lee believes that the real potential threat posed by AI is that it could cause massive social upheaval and political collapse due to widespread unemployment and rising inequality.
PricewaterhouseCoopers estimates that the adoption of AI will increase global GDP by USD 15.7 trillion by 2030. China is expected to take home US$7 trillion of this, nearly double North America’s US$3.7 trillion. The estimates indicate that the balance of economic power will tilt in favor of China, and that China will strengthen its political influence and soft power.
There is now a race between battery-powered “giants” and “startups” to develop AI products and train algorithms for specific tasks such as medical diagnostics, mortgage lending, insurance, and logistics.
The ‘grid’ approach is followed by the ‘Big Seven’ who have harnessed the power of machine learning and built it into salable, standardized services. Today’s AI research relies on specialized high-performance computing (HPC) systems, often referred to as AI supercomputers. These machines differ from traditional supercomputers used for scientific simulations because they are optimized for massive parallel processing, primarily using graphics processing units (GPUs) and AI accelerators such as the NVIDIA H100/B200, AMD MI300X, and Google’s Tensor Processing Units (TPUs).
Power requirements from the grid
AI supercomputers are extremely power hungry, consuming far more power per rack and per facility than traditional data centers.
•A single NVIDIA H100 GPU consumes approximately 700 watts of power.
•A rack of 8 GPUs uses approximately 5.6 kilowatts, and the compute power alone for an entire training hall with 1,000 racks can exceed 5 to 6 megawatts (MW).
•A large-scale AI supercomputer with around 50,000 GPUs can require 200 to 500 MW continuously, which is comparable to the output of a medium-sized coal-fired or nuclear power plant.
•Overall usage for AI data centers ranges from 20 MW to 1 GW, depending on size.
•The largest “gigawatt-class” supercomputing data centers (2026-class facilities by Amazon, Google, and Microsoft) each require about the same power as a nuclear reactor, about 1 GW, to support computing and cooling loads.
Supercomputers used for AI research consume large amounts of energy. GPU-accelerated systems require hundreds of megawatts of consistent power and exabytes of storage. The biggest challenge is not the computing hardware, but the availability of grid power, which defines the limits of AI infrastructure expansion. Sustainable operations increasingly rely on direct sourcing of renewable energy, on-site generation, and strategic geographic deployment close to power plants.
While power availability is one of the challenges in AI development, an equally important concomitant variable is the availability of high-performance chips that are central to a variety of bespoke applications.
Next-generation AI chips are specialized processors designed to handle artificial intelligence workloads such as machine learning, deep learning, and real-time decision-making.
Unlike traditional CPUs, modern AI systems use:
GPUs (graphics processing units) are used for massively parallel processing, making them ideal for training large AI models. It has the ability to conduct massive parallel competition. However, the power consumption is very high. GPUs are primarily used in data centers, research, and large-scale AI systems.
TPUs (Tensor Processing Units) are used for AI-specific high-speed chips. These are used for deep learning and neural networks. These are very efficient and are used for matrix operations. These chips are less flexible and designed specifically for AI workloads such as language processing, speech recognition, and real-time translation.
NPU (Neural Processing Unit) is used in smartphones, surveillance cameras, surveillance systems, etc., and is ideal for on-device AI. Power consumption is lower. NPU and TPU chips are ideal for training large models.
ASIC – These chips are intended for specific tasks and have customized AI hardware.
FPGA – These chips are reprogrammable and flexible. They are used in specialized systems such as defense, finance, and autonomous machines. These chips are optimized for speed, efficiency, and low latency, which are important for real-world AI applications.
No single chip can efficiently handle AI tasks. The future lies in combining different AI processors. Combining NPUs, GPUs, TPUs, ASICs, and FPGAs enables advanced applications such as self-driving cars, smart factories, and defense systems.
(The author is a former CBIC member and former Bank of Japan DG member)
