China's LineShine hit 2.198 exaflops — quintillions of calculations per second — on a system built almost entirely from conventional CPUs, the general purpose chips the AI era's supercomputer buildout has largely abandoned for specialized graphics processors.
China's LineShine has taken the No. 1 spot on the TOP500 list of the world's most powerful computers by doing something no other system in the ranking's history has done: sustaining more than 2 exaflops of double-precision performance using only conventional CPUs, the general-purpose chips the AI era's supercomputer buildout has largely abandoned for specialized graphics processors.
LineShine, housed at the National Supercomputing Center in Shenzhen, displaced the US Department of Energy's El Capitan system at Lawrence Livermore National Laboratory. It is the first Chinese system to top the semi-annual TOP500 list since 2017 and, according to Tom's Hardware's coverage of the release, the first CPU-only machine ever to clear the 2-exaflop mark.
The TOP500 ranks systems by how fast they solve the dense linear-algebra problems that underpin classical scientific computing: simulating nuclear reactions, modeling climate, running computational fluid dynamics, or screening new drugs. LineShine hit 2.198 exaflops, more than 2 quintillion 64-bit floating-point operations per second, by stringing together a very large fleet of general-purpose CPUs. El Capitan, the machine it dethroned, runs on AMD's specialized GPU accelerators, the same class of chip that has powered both the prior generation of exascale systems and the AI training buildout. So do nearly all of the other flagship systems announced during the AI era, including the next two US Department of Energy machines in the new ranking.
For the better part of a decade, the conventional wisdom in high-performance computing has been that the path to peak performance runs through specialized accelerators. China's researchers bet on a different route: scale out the general-purpose CPU instead of swapping in a GPU. On the TOP500's chosen benchmark, that bet paid off, and the Glenn K. Lockwood system profile for LineShine treats it as a genuine inflection point in the GPU-versus-CPU debate rather than a one-off scoreboard result.
The headline ranking also obscures the size of the field. Only five publicly verified exascale machines now exist worldwide, and the new top 10 includes systems from Italy, Switzerland, and Japan alongside the US, China, and Germany, a reminder that the leading edge of supercomputing is multinational, even when headlines focus on bilateral competition. The United Kingdom, for context, has 11 entries in the full 500-system list. The University of Bristol's Isambard-AI, built from 5,400 Nvidia superchips, is its highest-ranked machine at No. 11.
The energy cost is real. LineShine draws roughly 42.2 megawatts at full tilt, comparable to a small city's continuous power demand. The TOP500 also measures a single slice of the compute stack, dense double-precision arithmetic, and the Guardian's coverage notes that it says little about how well a system trains large language models or runs the sparse, low-precision workloads that dominate commercial AI. The United States still leads on those AI-training workloads, and the next two US national-lab systems, Frontier at Oak Ridge and Aurora at Argonne, both GPU-based, sit just below El Capitan in the new ranking.
What LineShine at No. 1 shows is that the AI chip race and the classical supercomputing race are turning out to be different competitions, and that China has chosen to compete hard on the second. The European Union's recently announced EUR 20 billion AI gigafactory plan, which calls for facilities with more than 100,000 AI processors each, is a recognition that governments on both sides of the Atlantic still see AI training as the prize that matters most. LineShine's score at the top of the TOP500 does not change that bet, but it does complicate the assumption that the only road to the world's most powerful computer runs through the AI chip.
What to watch: the November 2026 TOP500 cycle, which will show whether any GPU-based system can match or exceed LineShine's 2.198-exaflop result, and whether Shenzhen publishes a fuller technical accounting of how a CPU-only design reached a level that until this year was the exclusive territory of GPU-accelerated machines.