The future of AGI should not come at the expense of our planet

The development of modern computer science is marked by humanity’s relentless push towards the boundaries of computing efficiency: new technologies led to new applications, digital services expanded, super platforms emerged, internet traffic surged, energy consumption increased, fundamental technologies advanced, and eventually computational efficiency improved. Over the past 20 years, the global tech industry has largely been going through such a cycle.

Some argue that the path to artificial general intelligence (AGI) may break this cycle. According to the International Energy Agency’s Electricity 2024 report, after globally consuming an estimated 460 terawatt-hours (TWh) in 2022, data centers’ total electricity consumption could reach more than 1,000 TWh in 2026. This demand is roughly equivalent to Japan’s annual electricity consumption. Thus, should the future of AGI come at the expense of our planet?

I believe we cannot assess tomorrow’s challenges with today’s capabilities, but yesterday’s practices can inspire us and instill some confidence in what we do today. I started working in the tech sector around 2000, and since then, the industry has evolved through eras dominated by mainframe servers, minicomputers, distributed architecture, and cloud computing.

Each iteration of the computational infrastructure is a process of broadening the accessibility of digital services while also representing the evolution of software and hardware technologies to be integrated to improve computing efficiency. What remains constant, regardless of how business demands or computing methods evolve, is the industry’s unchanging pursuit of lower energy consumption and higher efficiency for computing tasks. Since the early days of Alipay, in our pursuit to strike a delicate balance among business expansion, continuity, and cost management, “green computing” emerged as a viable technology solution. 

Green computing is essentially a process driven by technology to sustain business viability. On the path to AGI, the exponential growth in energy consumption is now “the elephant in the room.” However, I believe the tech industry is fully aware of the severity and will, as always, drive innovations in software and hardware, as past breakthroughs have often emerged from seemingly impossible challenges.

Taking Ant Group’s experience as an example, during the “11.11 Global Shopping Festival” of 2010, our payment processing service was only four seconds away from crashing under peak loads. This compelled us to transition from relying on minicomputers to deploying distributed architectures to enhance computational efficiency. By 2021, we implemented green computing technologies at scale, using technologies like workload colocation, cloud-native time-shared scheduling, and AI-based auto scaling, which doubled our CPU server utilization compared to 2019. As a member of the technology team that made this possible, the most visible change for me was that we started from being on high alert in the project room, to eventually enjoying a can of Coke and letting the system handle the traffic peak by itself. 

In the era of intelligent computing, the energy consumption challenge is even greater, something we have experienced firsthand. Just like the 11.11 Global Shopping Festival, large-scale campaigns provide the best testing ground for new technologies. For example, each year, Alipay will launch an annual Chinese New Year campaign called Five Fortune. In 2024, for the first time, we started to pioneer green computing technologies in this AI-powered campaign at scale. 

The 2024 Alipay Five Fortune campaign introduced several AI-powered features, attracting 600 million interactions over 12 days. To scale foundation model applications while controlling computing costs, continuous optimization of hardware performance is required, along with better software hardware integration and algorithmic efficiency improvements. Currently, Ant Group has built a heterogeneous cluster of over 10,000 acceleration cards, where hardware compute efficiency (HFU) exceeds 60%, and the cluster’s effective training duration accounts for over 90% of total time. The RLHF training throughput performance is 3.59 times higher than industry-standard solutions under equivalent model effects.

During these green computing pilots, we have gained two key insights:

• In the era of intelligent computing, companies must incorporate the green computing system into strategic planning from day one. The approach to technology infrastructure is no longer about patching or making incremental improvements to the existing system, but about refactoring. The old method of “letting the business run first, then considering energy consumption” is no longer viable. Companies that realize this sooner will have a head start.

• Green computing is no longer exclusive to large corporations. Previously, it was believed that because computation costs are proportional to scale, large companies had a greater need to drive green computing initiatives, while smaller companies felt less urgency to follow suit. However, in the era of intelligent computing, the high costs of computing power make energy efficiency a critical metric for companies of all sizes and types. Given the barriers to developing green computing and the limited R&D budgets of smaller companies, we believe a green computing market will form in the future, providing products and services to businesses of all sizes.

Looking ahead, as industries undergo digital transformation, a triangular challenge emerges: ensuring the secure flow of data across entities, protecting user privacy, as well as ensuring the business is commercially sustainable. In response to these challenges, we believe the future of smart computing will evolve into the era of cryptographic computing, where data will be processed in an encrypted manner across clouds, regions, and industries. This requires more complex computations and higher computing power, thus making low-cost cryptographic computing essential for companies aiming to capitalize on digital transformation. Ultimately, low-cost cryptographic computing will unlock the value of data circulation for instantaneous use, just like turning on the tap.

From the past era of general computing, through the current era of intelligent computing, and into the future of cryptographic computing, pushing for a green and efficient computational power system has been a common pursuit for the tech industry. We believe that breakthroughs in core technologies that integrate software and hardware are crucial, ultimately achieving an optimal balance of computing power, storage, and networks. To achieve this across sectors, companies need to manage their organizations through data-driven approaches to make targeted optimizations, and actively participate in open-source projects and promote the proliferation and application of green computing technologies.

Problems that arise during technological advancement can only be solved through further technological breakthroughs. Similarly, the issue of energy consumption, exacerbated by technology, can only be fundamentally resolved by technology. When we consider what kind of earth we want to leave for future generations, and when our children ask whether technology can truly bring a better future, the endeavors of our generation in green computing take on a more profound significance.