AI supercomputer competition has long been described by speed: how many operations per second, or FLOPS, a system can deliver. In 2026, however, a second axis has moved alongside it: how much power is needed to deliver that performance, meaning power and efficiency. Even if compute performance rises, capacity cannot grow if the available power supply and grid do not keep up. Power is becoming the ceiling on compute capacity. This shift anticipates, at the front line of supercomputing, the broader data center constraint moving from semiconductors to power infrastructure.
LineShine won the top spot through efficiency
The symbolic change in the June 2026 TOP500 list was the new number one. China's LineShine took first place with 2.198 exaflops. What stands out is how it achieved that: 42.2 MW of power consumption, 52.07 Gigaflops/Watt of efficiency, and a CPU-only design with no GPUs. LineShine's HPL-MxP performance was only 3.6 times its HPL performance, supporting the absence of GPUs, because GPU systems tend to show a much larger ratio.
The contrast is the second-place system. The U.S. El Capitan delivered 1.809 EF using AMD EPYC and Instinct MI300A. That is the mainstream answer: load large numbers of GPUs and push absolute performance. LineShine instead showed an approach that competes on power per unit of performance, or efficiency. If power becomes scarce, absolute performance in FLOPS is not enough. Competitiveness also depends on how much compute can be produced per watt. Efficiency is not only an environmental metric; it is a practical way to get more computation from a limited power envelope.
Scale is moving toward the superfactory: Microsoft Fairwater
At the same time, absolute scale keeps expanding. Microsoft Fairwater Atlanta, scheduled to begin operation in October 2025, aims to become a distributed AI superfactory that completes AI training jobs in weeks. The scale is different from conventional sites. The Fairwater network connects hundreds of thousands of GPUs, exabyte-scale storage, and millions of CPU cores, while Atlanta adopts NVIDIA GB200 NVL72 with Blackwell. Sites are linked through 120,000 miles of dedicated fiber deployed for the AI WAN. The facilities use two-story designs to raise GPU density and lower latency.
As scale rises, resource constraints such as power, cooling, and water move to the foreground. Symbolically, Fairwater Atlanta's initial cooling water volume was described as equivalent to the annual consumption of 20 homes. Keeping systems with hundreds of thousands of GPUs running is itself a problem of securing power and infrastructure. The ceiling on compute capacity is starting to be set less by the number of GPUs procured and more by the power, cooling, land, and grid connections that support them.
The competition is running globally
This is not a single-country story. In the TOP500 list, exascale systems appeared simultaneously in Asia, North America, and Europe for the first time. Europe also reached the threshold with its first exascale machine, JUPITER Booster, at exactly 1.000 exaflops. On the policy side, the EU is building 19 AI factories and 13 antennas, providing free access to SMEs and startups. New AI factories are being added in six countries: Czechia, Lithuania, the Netherlands, Spain, Poland, and Romania, while antennas include non-EU countries such as Switzerland and Serbia. AI factories are large-scale model development platforms centered on AI-optimized supercomputers, and each region is adding compute capacity while securing power and infrastructure.
From FLOPS to power and efficiency: the numbers behind the shift
01
Top rank won by efficiency
China's LineShine ranked first at 2.198 EF, using 42.2 MW and a CPU-only design to reach 52.07 GF/W. HPL-MxP at 3.6 times HPL supports the no-GPU design.
02
The mainstream answer is GPU scale
Second-place El Capitan in the U.S. uses EPYC plus MI300A for 1.809 EF. It contrasts with LineShine, which competes through efficiency.
03
Scale becomes superfactory
Microsoft Fairwater connects hundreds of thousands of GPUs, exabyte-class storage, and millions of CPUs. Initial cooling water equals about 20 homes per year, making scale a resource constraint.
04
Global parallel buildout
Exascale machines appeared at the same time across Asia, North America, and Europe. JUPITER in Europe reached 1.0 EF, while the EU is building 19 AI factories and 13 antennas.
Business impact and checkpoints
For AI data centers and high-performance computing, the checks are: 1. whether plans are evaluated not only by performance in FLOPS but also by power capacity and efficiency, meaning power per unit of performance; 2. whether power, cooling, water, land, and grid connection can support hundreds of thousands of GPUs; 3. whether procurement and design are optimized around efficiency, including power conversion, cooling, high-voltage distribution, and SiC/GaN adoption; and 4. how to access regional AI infrastructure such as AI factories. AI compute capacity is no longer decided by GPU count alone. Available power and the efficiency of using it are becoming the next competitive axis. Related articles cover power architecture and cooling design.
