Supercomputing is splitting into two tracks: publicly funded exascale systems ranked by TOP500/HPL FLOPS and hyperscaler-built AI campuses measured by megawatt–gigawatt capacity.
Supercomputing is having another 'super' year. Record-breaking buildouts – fueled by AI demand and hype – are pushing performance and scale in two very different directions.
Two Tracks, Two Metrics #
The action is unfolding on two tracks. One is publicly funded and measured by the twice-yearly TOP500 list. The other is hyperscaler-funded and measured in gigawatts of capacity.
That split produced a rare double on June 23, 2026: Microsoft claimed its new Wisconsin campus hosts the “world’s most powerful supercomputer,” while the TOP500 list named a Chinese system, LineShine, No. 1. Both were right – for their respective metrics. These systems do different jobs and don’t run the same workloads.
Public labs continue to build exascale machines and submit High-Performance Linpack (HPL) scores to the TOP500. Private companies keep building AI training campuses that dwarf those machines in raw power, cooling, and footprint, even if they never run HPL.
Exascale Facilities Keep Growing #
On the government-funded track, national programs continue to deploy exascale-class systems on budgets that now represent a fraction of what a single hyperscaler spends on a multi-site AI campus.
LineShine (China). The system debuted at No. 1 on the
June 2026 TOP500 listwith 2.198 exaflops on HPL – the first Chinese system to lead since 2017. It runs 20,480 nodes of Chinese-built LX2 Arm processors (about 14 million cores), with no GPUs, and draws 42.2 MW. The CPU-only design reflects US export restrictions.Eni HPC7 (Italy). The energy company’s HPC7 entered at No. 6 in June 2026, the only privately built system in the current Top 10.
Lux (US). Oak Ridge National Laboratory's
new AI cluster– built with AMD and Hewlett Packard Enterprise – uses AMD Instinct MI355X GPUs and AMD EPYC CPUs. Deployed in early 2026 under a public-private model, Lux supports fusion, materials science, and national security research, according to the US Department of Energy.Jupiter (Germany). In May 2026, researchers at Jülich and Nvidia used Jupiter’s GH200 Superchips to fully simulate a 50-qubit universal quantum computer for the first time, beating the previous 48-qubit record. At roughly exascale performance, it ranks among the most energy-efficient systems in its class.
Alice Recoque (France). Europe’s second exascale site pairs AMD Instinct MI430X GPUs with SiPearl's Rhea2 processors. Installation began in 2026 at the CEA's Very Large Computing Center under a EuroHPC JU
contractwith Eviden totaling €354.8 million.
Hyperscalers Build at a Different Scale #
On the private track, hyperscalers are standing up AI training campuses measured in hundreds of megawatts to gigawatts, funded and operated outside the national lab model.
Colossus 2 (xAI). Spanning Memphis, Tennessee, and neighboring Southaven, Mississippi, the site generates power on-site rather than waiting for utility interconnection –
an approachthat can compress typical multi-year data center timelines into weeks. Target capacity is 2 GW, with roughly 555,000 Nvidia GPUs.Fairwater (Microsoft). The Wisconsin site went live in June 2026and is linked to an earlier Atlanta campus via dedicated fiber. Both sites use closed-loop liquid cooling to eliminate operational water consumption. Each exceeds 350 MW and is scaling toward multi-GW.
Stargate (OpenAI/Oracle). In March 2026, OpenAI and Oracle scrapped a
planned 600 MW expansionof the Abilene, Texas, campus after financing talks failed, though the existing build remains on schedule. Committed capacity stands at 1.2 GW.Prometheus (Meta). In January 2026, Meta signed
nuclear power agreementswith TerraPower, Oklo, and Vistra to supply its New Albany, Ohio, site, targeting 1 GW of operational capacity when it comes online later in 2026.
These facilities don’t appear on the TOP500 because they aren’t designed to run HPL, and owners don’t submit scores. For data center planners, the comparison that matters is power capacity, cooling architecture, and build timeline – not FLOPS. By those measures, AI campuses now dwarf even the largest national lab systems.
Sovereignty Is Now a Site-Selection Factor #
Governments increasingly treat compute as strategic infrastructure and are funding national capacity rather than relying solely on hyperscalers.
Humain (Saudi Arabia). A subsidiary of the kingdom's Public Investment Fund began rolling out
the first phaseof Nvidia GB300 GPUs in 2025: an 18,000-GPU system as part of a broader build targeting up to 500 MW.EuroHPC AI Factories (European Union). At ISC 2026 in June, Nvidia announced
35 new AI supercomputersin development across Europe, calling it the continent’s largest one-year expansion of supercomputing capacity. The program includes upgrades at Barcelona Supercomputing Center and Jülich, extending the AI factory model centered on Germany’s Jupiter exascale system.
Watching for the Next Checkpoint #
Several milestones remain ahead in 2026: Meta's Prometheus site is slated to come online with its first gigawatt of AI capacity later this year; installation of Alice Recoque continues in France; and the next TOP500 list will be published at SC26 in November – the next opportunity for a new system to challenge LineShine’s record.