Introduction
The top 10 supercomputers powering AI research in 2026 have shifted dramatically after China’s LineShine took the TOP500 crown in June 2026 with 2.1 exaflops of sustained Linpack performance. Three regions now host exascale AI machines, with Europe joining the United States and China after JUPITER went online at Forschungszentrum Jülich. These systems are not only ranked by raw speed but also by mixed-precision throughput that trains trillion-parameter models. Research teams use them to design proteins, simulate fusion plasmas, and train sovereign large language models for public benefit. Access programs like INCITE in the United States and EuroHPC in the European Union have opened hours to industrial partners and startups. Energy use has become the dominant design constraint, with each top system consuming between 20 and 38 megawatts at full tilt. This guide ranks and explains the top 10 machines shaping the direction of global AI research in 2026.
Quick Answers on the Top AI Supercomputers in 2026
What is the fastest AI supercomputer in the world right now?
China’s LineShine system holds the top rank on the June 2026 TOP500 list with 2.1 exaflops Linpack and roughly 3.0 exaflops of mixed-precision AI performance, operated at a national supercomputing center in eastern China.
Which US system is the most powerful for AI research?
El Capitan at Lawrence Livermore delivers 1.74 exaflops Rmax and over 43 mixed-precision exaflops, making it the most capable American machine for both classified and open AI research workloads at scale.
Which processors are used for AI-driven supercomputing in 2026?
Leading AI supercomputers run AMD Instinct MI300A, Nvidia H100 and Blackwell GB200, Intel Max GPUs, and custom Chinese accelerators connected by Slingshot-11, InfiniBand NDR, or NVLink switch fabrics inside each rack.
Key Takeaways
- Three regions now host exascale AI supercomputers: the United States, China, and the European Union, which joined after JUPITER came online.
- El Capitan, Frontier, Aurora, and Fugaku anchor open science AI research in the United States and Japan.
- AMD Instinct and Nvidia Blackwell dominate AI accelerator share, while interconnect fabric choice decides training scalability.
- Access to the Amazon’s $110 million for AI research style programs now runs through EuroHPC, INCITE, and ACCESS on the top 10 systems.
Table of contents
- Introduction
- Quick Answers on the Top AI Supercomputers in 2026
- Key Takeaways
- Understanding the Top 10 Supercomputers Powering AI Research
- How the TOP500 and MLPerf Benchmarks Rank AI Capability
- LineShine (China) Leads the Top 10 Supercomputers Powering AI Research
- El Capitan at Lawrence Livermore Pushes Exascale AI for National Security
- Frontier at Oak Ridge Keeps Expanding Open AI Science
- Aurora at Argonne and the Rise of AuroraGPT
- Fugaku at RIKEN Powers Japan’s AI and Climate Research
- JUPITER at Jülich Makes Europe an Exascale AI Region
- LUMI in Finland Scales Pan-European AI for Startups
- Leonardo at CINECA Blends HPC With Industrial AI
- MareNostrum 5 Brings Sovereign AI to Southern Europe
- Alps at CSCS Pioneers Grace-Hopper AI Workflows
- The GPU and Interconnect Architectures Behind These Systems
- How to Implement AI Research Projects on These Supercomputers
- Real-World Examples of AI Supercomputers in Practice
- Ethics, Risks, Energy Costs, and Export Controls Shaping AI Supercomputing
- The Future of AI Supercomputing After the Exascale Era
- Key Insights on the Top 10 Supercomputers Powering AI Research in 2026
- Real-World Applications Running Across the Top AI Supercomputers
- In-Depth Case Studies from the Top 10 Supercomputers Powering AI Research
- Frequently Asked Questions on Top AI Supercomputers in 2026
Understanding the Top 10 Supercomputers Powering AI Research
The top 10 supercomputers powering AI research in 2026 combine dense accelerator pools with high-bandwidth memory and wide interconnect fabrics. These parts keep mixed-precision throughput near theoretical peak across thousands of nodes during a full training run.
AI Supercomputer Training Time Explorer
Pick a system and set your model size and token budget to estimate wall-clock training time at the system’s mixed-precision throughput. Numbers are approximate and assume typical scaling efficiency.
El Capitan (43.5 AI EF)
500B parameters
10 trillion tokens
Mixed-precision figures sourced from TOP500.org, system vendor briefings, and the MLPerf HPC benchmark suite. Model-training FLOP estimate uses the 6 N * D heuristic.
How the TOP500 and MLPerf Benchmarks Rank AI Capability
Building on that definition, the TOP500 list has ranked the world’s fastest supercomputers since 1993. The list uses a double-precision Linpack score called Rmax to compare them. Linpack remains a useful proxy for general HPC capability but tells only part of the story for modern AI research teams. A machine can hit a strong Linpack number yet fall behind on model training, where lower-precision math and memory bandwidth matter more. That gap pushed vendors to publish mixed-precision AI exaflop figures reflecting FP16, BF16, and FP8 math performance. These numbers now sit alongside Linpack in press releases from every leadership lab around the world.
Beyond Linpack, the MLCommons group runs the MLPerf HPC benchmark suite that scores real AI workloads like OpenCatalyst, DeepCAM, and CosmoFlow at scale. These benchmarks show scaling to tens of thousands of GPUs and expose weaknesses in storage and network fabrics that Linpack hides. Centers also publish HPCG scores, a sparse solver that captures how memory latency slows real applications. Readers who care about which system will train their model fastest should weigh MLPerf HPC results over Linpack rank. HPCG and MLPerf results are available for every top 10 system on the public rankings portal.
Beyond raw speed, energy per training run and total cost per epoch have also entered the benchmark mix. The Green500 list, a sibling of TOP500, ranks systems by gigaflops per watt and crowns efficiency winners rather than raw speed. Top machines trade raw peak for efficiency through liquid cooling, low-voltage silicon, and rack-level power sharing techniques. Procurement teams now look at Green500 alongside MLPerf when they decide which system will run billion-parameter training jobs. For AI buyers, this is similar to how fleet managers look at energy-efficient AI training techniques rather than peak horsepower alone.
LineShine (China) Leads the Top 10 Supercomputers Powering AI Research
Moving from benchmarks to systems, LineShine displaced US machines from the TOP500 summit in June 2026. The move ended an eight-year American run and marked China’s first return to the top since Sunway TaihuLight. The system sits at the National Supercomputing Center and delivers 2.1 exaflops Rmax on the public TOP500 page. It uses homegrown accelerators connected by a custom high-bandwidth fabric, a pivot forced by continuing US export controls on advanced Nvidia parts. The June 2026 TOP500 announcement ranked LineShine first on sustained performance. The system shows that the global AI compute landscape now includes a non-US flagship for the first time in nearly a decade.
Beyond raw speed, LineShine targets large-model training and scientific AI that China once ran on imported hardware. The system has been used for GLM-5 and DeepSeek-style models that rival Western frontier models on reasoning tasks reported in academic conferences. Climate simulation, drug design, and satellite image analysis share the queue, giving the facility a mixed open-science and policy mandate. Analysts note that LineShine’s gains underline a growing split in the AI compute supply chain between Western and Chinese ecosystems. The system also supports sovereign AI workloads that must stay on domestic hardware for regulatory reasons.
Even with those gains, operational details remain tightly held, a pattern consistent with earlier Chinese top systems that did not file MLPerf numbers. The lack of published MLPerf HPC results makes direct comparison with Frontier or El Capitan harder, and some Western analysts treat the AI exaflop claims as aspirational marketing. Even so, Linpack is publicly verified by the TOP500 committee, and the 2.1 exaflops figure has been independently reviewed by external examiners. The political backdrop frames LineShine as part of the broader US vs China AI frontier. Chinese vendors are now briefing allied labs on their accelerator roadmap in hopes of building a parallel global AI supply chain.
Looking ahead, LineShine also matters because it influences the next round of US and European procurement decisions. US exascale buyers have already briefed Congress on the need for a response system at 3 exaflops or higher soon. European partners are debating whether to specify a second exascale node after JUPITER to keep pace with the new reality. The next 24 months will likely see both sides race to redefine what counts as a leading AI supercomputer. That race will keep this list fluid through 2027 and beyond, with yearly reshuffles likely.
El Capitan at Lawrence Livermore Pushes Exascale AI for National Security
Shifting to the US flagship, El Capitan at Lawrence Livermore National Laboratory recorded a Rmax of 1.742 exaflops on the November 2024 TOP500 list. That lead held for 18 months before LineShine overtook it. The machine runs 11 million AMD cores across 44,544 Instinct MI300A APUs connected by HPE Slingshot-11 Ethernet. Together these parts produce an estimated 43.5 mixed-precision exaflops for AI, enough to train a trillion-parameter model in roughly five weeks of wall time. El Capitan’s APU design integrates CPU and GPU on one package, cutting the cost of moving tensors during training. The combined compute budget makes it the leading open-plus-classified AI machine for the US national labs.
Beyond the specs, Los Alamos, Lawrence Livermore, and Sandia share El Capitan for stockpile stewardship simulations that replaced live nuclear testing after 1992. AI workloads now ride that same procurement because surrogate models trained on prior physics runs cut compute cost dramatically compared with full physics replays. Open science time is limited to a small percentage of hours, but awarded projects include fusion, materials discovery, and biothreat modeling. LLNL has published papers using El Capitan to train foundation models for inertial confinement fusion diagnostics on the system. The mixed mission keeps the machine busy and makes it one of the most demanded platforms in the world.
Beyond compute, power draw sits near 30 megawatts at sustained full-system use, delivered through warm-water liquid cooling that recycles heat for adjacent buildings. The Green500 team reports El Capitan at roughly 58 gigaflops per watt, placing it among the most efficient exascale systems on the planet. Operational cost per year is reported in the hundreds of millions of dollars, with most funding flowing from the National Nuclear Security Administration. The LLNL El Capitan platform page lists the current node counts and uptime schedule for applicants. El Capitan shows that efficiency improvements at this scale directly translate into more training throughput per dollar.
Frontier at Oak Ridge Keeps Expanding Open AI Science
Turning to open science, Frontier at Oak Ridge National Laboratory was the first machine to break the exascale barrier and still delivers 1.35 exaflops Rmax on the current TOP500 list. The HPE Cray EX system runs 37,888 AMD Instinct MI250X GPUs with Slingshot-11 interconnect across 9,408 compute nodes. Oak Ridge allocates Frontier through the INCITE and ALCC programs, which open the machine to US and international teams based on competitive proposals. A recent Nature paper credited Frontier with a protein folding run that explored more conformations than any prior academic effort on the system. Frontier remains the easiest leadership-class machine for university teams to reach because of its open peer-reviewed allocations.
Beyond peer-reviewed access, Frontier has become a key open science venue for AI models that cannot run on classified machines. Its mixed-precision throughput reaches around 11 AI exaflops, enough for large language model pretraining and multi-trillion-parameter sparse mixtures of experts. Researchers using Frontier have trained climate emulators, weather forecasting foundation models, and quantum chemistry surrogates with public code releases. The OLCF Frontier page publishes current allocation calls and system status updates for applicants worldwide. Frontier hosts roughly a dozen INCITE AI projects per year and nearly twice that number through the shorter ALCC call window. Oak Ridge has also invested in teaching tools so graduate students can scale onto Frontier without waiting years for pre-exascale expertise.
Aurora at Argonne and the Rise of AuroraGPT
Moving to Intel’s flagship, Aurora at Argonne Leadership Computing Facility reached 1.012 exaflops Rmax on the November 2024 TOP500 after years of commissioning delays. The HPE Cray EX system uses Intel Xeon CPU Max and Intel Data Center GPU Max (codename Ponte Vecchio) across 10,624 nodes. Aurora is the only exascale machine built on Intel accelerators, giving it a different software stack from the AMD and Nvidia leaders. It targets open science, with the AuroraGPT project using the whole machine to train a trillion-parameter model tuned for scientific reasoning. Aurora has become the main alternative to AMD and Nvidia platforms for groups that want hardware diversity in their training stacks.
Beyond the hardware, AuroraGPT started as a 7-billion parameter prototype and has been scaled up by the Argonne, Intel, and HPE teams in partnership with roughly a dozen universities. The project aims to train a science-specific foundation model that can reason over biology, chemistry, materials, and physics text and data. Argonne has made the Aurora system description and AuroraGPT project page openly available to researchers worldwide. The team has released checkpoints and data mixes, something no other national lab has done at this scale. That openness has attracted outside contributors from the evolution of generative AI models ecosystem.
In addition to language modeling, Aurora also runs large-scale drug discovery pipelines that pair diffusion models with molecular dynamics simulations. Pharma partners such as AbbVie have used ALCC time to screen billions of candidate molecules against disease-relevant proteins on the platform. These runs combine classical HPC with transformer models in a hybrid loop that proved hard to run on earlier systems. Argonne reports that Aurora completed a 40-million conformation docking study in under 72 hours during 2024. That case has been cited by funding agencies as proof that AI supercomputing accelerates drug design in meaningful ways.
Even so, Aurora has faced criticism for power draw and for the delays that pushed first-light from 2021 to 2024. Full-system runs push 60 megawatts, among the highest in the list, and energy cost per epoch remains a visible line item in facility budgets. Intel’s Max GPU roadmap pivoted after Aurora’s deployment, so follow-on systems will likely use different accelerators. Even with those tradeoffs, Aurora remains a key proving ground for Intel’s software stack and for AuroraGPT-style scientific foundation models. The facility plans to run Aurora through at least 2029 with incremental refresh upgrades.
Fugaku at RIKEN Powers Japan’s AI and Climate Research
Turning to Japan, Fugaku at RIKEN remains the country’s flagship supercomputer. It still ranks in the top 10 with a Rmax of 442 petaflops recorded on the TOP500 Fugaku entry. Built by Fujitsu, the system uses 158,976 A64FX ARM-based CPUs with high-bandwidth memory on-package and the Tofu-D six-dimensional torus interconnect. It is the only top system with no attached GPU accelerators, a design choice that favors flexible AI inference and classical HPC over dedicated training alone. Fugaku has become a template for sovereign AI compute programs in Asia. Its CPU-only design has survived longer than skeptics expected when the system first launched in 2020.
Beyond the hardware, Japanese researchers have used Fugaku to train the Fugaku-LLM 13B model and more recently a 70B science-tuned model that supports Japanese and English. The machine also runs the country’s civil weather models, the Earth Simulator ensemble, and large-scale COVID aerosol studies published during the pandemic. RIKEN allocates time through a national program open to academic, industrial, and startup teams for a range of projects. Japan’s Ministry of Education has already pledged funding for a Fugaku successor targeted for 2029 that will likely add GPU accelerators. The successor plans show how Fugaku’s lessons have shaped Japan’s broader AI compute strategy for the next decade.
Even on older silicon, Fugaku’s ARM CPU design has influenced follow-on AI systems in Europe and Japan that choose CPU-dense topologies for memory-bound workloads. The A64FX supports Scalable Vector Extensions, helping it run certain transformer inference loads surprisingly well without GPUs. RIKEN publishes annual energy and utilization statistics that remain among the most transparent in the world. The facility has been cited as a model for sovereign AI compute in India’s Param-1 GenAI model planning documents. Fugaku continues to attract pilot projects from Southeast Asian research consortia looking for an alternative to US cloud providers.
JUPITER at Jülich Makes Europe an Exascale AI Region
Shifting to Europe, JUPITER at Forschungszentrum Jülich entered the TOP500 in November 2025 as the first European exascale machine. The system reached Europe’s first exascale Linpack score in the November 2025 TOP500. The modular system pairs around 24,000 Nvidia Grace-Hopper GH200 superchips in its Booster module with roughly 1,300 SiPearl Rhea CPUs in its Cluster module. JUPITER was built by Eviden and ParTec for the EuroHPC Joint Undertaking at a cost of about 500 million euros. The modular design lets Jülich upgrade individual modules without rebuilding the entire facility. The system also showcases European processor sovereignty through the SiPearl Rhea CPU partition.
Beyond raw performance, JUPITER was commissioned specifically to train sovereign foundation models and to run climate projections for EU policy decisions. The system allocates 50 percent of its time to EU industrial and public-sector users through EuroHPC calls that opened throughout 2026. That is a sharp break from older European systems that privileged pure academic research and signals a political commitment to an EU AI industrial base. JUPITER’s design also reflects lessons learned from the earlier pre-exascale systems that struggled with AI workloads. The system has become the reference point for European AI compute sovereignty.
LUMI in Finland Scales Pan-European AI for Startups
Moving to Finland, LUMI at CSC in Kajaani delivers 379 petaflops Rmax and about 530 petaflops of mixed-precision AI performance, placing it among the top 10 systems worldwide. The HPE Cray EX machine runs 11,912 AMD Instinct MI250X GPUs that are the same family powering Frontier. LUMI runs on 100 percent hydro-generated electricity drawn from a nearby northern grid, with waste heat piped into the Kajaani district heating network. The facility operates year-round at near full utilization with strong user support teams. LUMI’s energy transparency has become a model for other European AI data centers.
Beyond power, LUMI is famous for its open access to industrial users, including European AI startups that have applied for free compute through the EuroHPC AI access scheme. Nokia, Silo AI, and dozens of smaller companies have run pretraining jobs on the system during the last two years. CSC publishes monthly utilization numbers and has reported more than 90 percent average occupancy across 2024 and 2025 reporting cycles. The EuroHPC LUMI system description lists current access calls and application deadlines. CSC also offers dedicated onboarding for teams new to AMD ROCm software.
Even so, LUMI’s open access has made it a magnet for health sciences AI, including AstraZeneca’s work on structural prediction and Thermo Fisher’s imaging pipelines. Yet queue times for the largest multi-node reservations have stretched past 10 days during peak periods, a bottleneck CSC is addressing through smaller dedicated partitions. The system also hosts the open-weights Poro and Viking models trained for Nordic languages. LUMI remains the clearest proof that sovereign AI compute can be built on open hardware with transparent energy accounting. European policymakers now cite LUMI when they argue for continued EuroHPC funding of new AI systems.
Leonardo at CINECA Blends HPC With Industrial AI
Turning to Italy, Leonardo at CINECA in Bologna hit a Rmax of 241 petaflops when first ranked and now sits near 266 petaflops after upgrades. The Atos BullSequana XH2000 system pairs Intel Xeon Platinum CPUs with 13,824 Nvidia A100 GPUs in its Booster module and over 1,500 CPU-only nodes for data-heavy workloads. Leonardo is funded by the EuroHPC JU and operated by CINECA, Italy’s largest supercomputing consortium with deep industry ties. The system’s blended design reflects Italy’s strong industrial research base in automotive and energy. Leonardo was among the first European systems to dedicate a large allocation to commercial AI training in 2024.
Beyond its hardware, Leonardo serves a mix of academic, industrial, and public-sector AI research including Italy’s national digital twin program and ENI’s energy models. The system was among the first European platforms to receive dedicated AI access schemes in 2024 and still runs commercial training jobs through a partnership tier. Leonardo’s modular design has let CINECA add a smaller data-centric partition for inference-heavy workloads without rebuilding the Booster. The facility publishes its allocation rules on the Leonardo user access page at CINECA. Italian pharma and manufacturing teams use Leonardo for model training that would otherwise need cloud GPUs.
MareNostrum 5 Brings Sovereign AI to Southern Europe
Shifting to Spain, MareNostrum 5 at the Barcelona Supercomputing Center began production in December 2023 and currently delivers 314 petaflops Rmax and about 230 AI petaflops in its accelerated partition. The Lenovo-built machine combines Intel Sapphire Rapids and AMD EPYC CPUs with 3,456 Nvidia H100 GPUs in its accelerated partition, plus a next-generation partition that tests emerging accelerators. It sits inside the historic Torre Girona chapel, a striking visual signature that appears in every launch photograph and press tour. The chapel setting has drawn global media attention and made the system a cultural icon. MareNostrum 5 anchors the Southern European branch of the EuroHPC network.
Beyond compute, MareNostrum 5 serves as Spain’s sovereign AI platform and has hosted training of the ALIA family of Spanish language models that cover peninsular and Latin American variants. Barcelona Supercomputing Center partners with local startups through a shared pre-commercial procurement arm that spans both Spanish and EU funding. The system has also absorbed workloads from pharmaceutical teams in Catalonia and Madrid that previously bought dedicated cloud GPUs. Its mixed partition model gives researchers unusual flexibility compared with monolithic machines. Spanish authorities now point to MareNostrum 5 as a template for sovereign AI access across multilingual regions.
In addition to language modeling, MareNostrum 5 is a political signal that AI compute does not have to concentrate in northern Europe. The system aligns with EU strategies to build regional AI hubs. The next-generation partition is being used to evaluate RISC-V accelerators, quantum-classical hybrids, and photonic chips before any production procurement decision. Barcelona Supercomputing Center publishes a monthly technical report with access times, utilization and energy data. The MareNostrum 5 technical page remains the authoritative reference for the system. The partition model is being studied by other EuroHPC centers that are planning new systems.
Alps at CSCS Pioneers Grace-Hopper AI Workflows
Turning to Switzerland, Alps at the Swiss National Supercomputing Centre in Lugano achieved 270 petaflops Rmax. The machine uses 2,688 Nvidia Grace Hopper GH200 superchips and is the first production system to run a large Grace Hopper fleet. The HPE Cray EX system supports a modular platform that lets different scientific communities carve out isolated software stacks per project. Alps was commissioned in 2024 and continues to grow through staged upgrades planned for 2026 and 2027. The modular design means each research community can pin its own CUDA, PyTorch, and compiler versions on its slice. Alps has also become a reference for AI-plus-HPC hybrid workflows across Europe.
Beyond architecture, Alps is a key proving ground for AI research in Europe because it pioneers the unified CPU-GPU memory model Nvidia introduced with Grace Hopper. That architecture reduces the cost of moving data between host and accelerator, a bottleneck that slows large transformer training on older systems. ETH Zurich and EPFL teams have used Alps for climate emulation, drug discovery, and the Apertus Swiss language model. The facility is also a reference point for the next wave of Nvidia’s new LLM models deployments on exascale fabrics. The Swiss government has already committed funding to double Alps capacity by 2028, keeping it in the top European AI compute tier.
The GPU and Interconnect Architectures Behind These Systems
Pulling back to the shared architecture layer, the top AI supercomputers in 2026 run one of three accelerator families. Those families are AMD Instinct, Nvidia H100 and Grace Hopper GH200, or Intel Data Center GPU Max. AMD powers Frontier, El Capitan, and LUMI, taking four of the top 10 slots by share. Nvidia drives JUPITER, MareNostrum 5, Alps, and Leonardo, while Intel holds Aurora as its only exascale flagship worldwide. China’s LineShine uses a domestic accelerator whose ISA is partially disclosed and whose on-die HBM sources remain under US scrutiny. These vendor choices shape the software stack every research team must adopt.
Beyond accelerators, the interconnect fabric determines whether a system can train a model that fits across its whole accelerator pool. HPE Slingshot-11 links Frontier, El Capitan, and LUMI with 200 gigabits per second per port and adaptive routing. Nvidia InfiniBand NDR and Quantum-2 switches run JUPITER, MareNostrum 5, and Alps at 400 Gb/s with NVLink inside the node. Fujitsu’s Tofu-D fabric wraps Fugaku in a six-dimensional torus that favors nearest-neighbor exchange over all-to-all patterns. These choices are as consequential as the GPU choice for frontier training.
Beyond fabric, memory bandwidth is the other defining architectural choice. AMD MI300A packages HBM3 directly next to 24 Zen 4 cores inside one APU, putting roughly 5.3 terabytes per second of bandwidth within a single socket. Nvidia Grace Hopper pairs 72-core Grace CPUs with a Hopper GPU over an NVLink-C2C fabric at 900 gigabytes per second. Intel Max GPU and Chinese accelerators rely on separate HBM stacks linked through on-package silicon bridges. These memory topologies decide whether training stays memory-bound or compute-bound across modern AI supercomputers.
In addition to compute, storage and parallel file systems complete the picture, with Lustre, DAOS, and AI and cloud computing inspired VAST and WEKA deployments. Frontier and El Capitan use Lustre on NVMe arrays delivering tens of terabytes per second of aggregate read throughput. Aurora runs DAOS for object-based storage that matches AI data loader patterns at scale. These storage choices affect everything from checkpointing cadence to the cost of the data lake that feeds each training run. Procurement teams increasingly list storage as a line item equal to the accelerator budget in new proposals.
How to Implement AI Research Projects on These Supercomputers
Moving from hardware to access, researchers in the United States apply for time on Frontier, Aurora, and other DOE systems. The INCITE and ALCC programs, both overseen by the Office of Science, handle allocations. INCITE awards the largest production allocations annually through a peer-reviewed process with proposals due each spring. ALCC runs a second channel aimed at projects aligned with specific DOE mission areas across research groups. The DOE Leadership Computing INCITE call page publishes the current call schedule. Teams need to run benchmark scaling before they submit so reviewers can judge readiness for the system.
Beyond the US, European researchers apply to EuroHPC calls that run several times per year on JUPITER, LUMI, Leonardo, MareNostrum 5, and Alps across the continent. The EuroHPC AI access scheme has dedicated calls for AI startups, academic teams, and SMEs with lightweight proposals and fast turnaround. Industrial partners can also buy time through national cost-recovery programs, a route that growing AI startups use to avoid cloud GPU scarcity in tight markets. Non-EU researchers can participate through institutional partnerships with EuroHPC hosting sites. Call deadlines and expected award timelines are posted on the EuroHPC Joint Undertaking portal.
In addition to Europe, Japan’s HPCI consortium opens Fugaku to foreign researchers through a shorter annual call. China’s Tianhe, LineShine, and Sunway machines run national-level access that is largely domestic. In every case, allocations come with requirements to publish results and to share derived code and data with the broader community. That transparency distinguishes public supercomputers from private cloud contracts and is one reason the Qualcomm AI-focused data chips conversation often stops short of covering open national systems. Teams planning to implement AI research on these supercomputers should plan a six-month lead time from proposal to first compute. That lead time covers proposal writing, review, resource allocation, and software environment setup.
Real-World Examples of AI Supercomputers in Practice
Moving to concrete deployments, real-world AI research on the top 10 supercomputers powering AI research in 2026 covers protein folding on Frontier and trillion-parameter foundation model training on Aurora. Climate emulation on Alps and drug discovery on LUMI complete the main European and US workload set. Each system has published outcome papers tied to a specific allocation, including a 405-million conformation protein search that ran in under 72 hours on Frontier. US and European pharma teams cite exascale time as the main reason model-based drug screening moved from pilot to production in 2024 and 2025. Those outcomes show these leadership systems now affect commercial pipelines, not just academic publications. Open allocation also means these breakthroughs are documented in the public literature.
Beyond individual projects, these runs would simply not fit on the biggest cloud GPU reservations available in 2024. For instance, a pretraining run equivalent to GPT-4 class takes around 90 days on 10,000 commercial H100s. Public leadership systems can slot that run in a single INCITE award cycle instead. That substitution effect has reshaped the economic case for buying private compute, particularly for public-interest research teams. AI access schemes at Nvidia AI factory cybersecurity labs now point users to EuroHPC first. The cost comparison alone justifies the proposal writing effort for many projects.
Ethics, Risks, Energy Costs, and Export Controls Shaping AI Supercomputing
Turning to the hard questions, AI supercomputing in 2026 faces three big risks: energy use, geopolitical export controls, and concentration of compute in a small number of operators. El Capitan, Frontier, and Aurora each draw between 20 and 60 megawatts at peak, with annual electricity bills that stretch into the hundreds of millions of dollars each year. Those numbers push every operator to invest in direct liquid cooling, warm-water heat recapture, and renewable procurement for their power. Energy growth is also drawing political attention, especially where new data center proposals strain local grids. The ethics of concentrating so much energy use in a handful of sites has become a mainstream policy debate.
Beyond energy, export controls have reshaped the global AI compute map. US restrictions on HBM, advanced lithography, and A100-class GPUs pushed China to build domestic accelerators like the one inside LineShine. The OpenAI’s move into AI chips and the broader US chip policy debate increasingly frame supercomputers as strategic infrastructure for national competitiveness. Allied governments are coordinating licenses and watch lists, and China is responding with multibillion-dollar funding for sovereign fabs. The result is a slower, costlier, and more fragmented global supply chain for AI compute. The ethics of restricting scientific computing to allied blocs remains contested in the research community.
Beyond geopolitics, concentration of compute is the third risk and perhaps the most important for ethics and equity. A small number of national labs and three or four cloud providers now hold the bulk of training capacity for frontier models. That raises questions about democratic oversight and research equity across the broader AI community. The EU is pushing EuroHPC and the US is funding ACCESS and NAIRR to broaden access across many more institutions. Even so, the ratio of frontier compute to graduate student groups has never been more skewed, and many argue that the gap will widen before it narrows. Policy on access, pricing, and sovereign procurement will shape the ethics of AI research for the next decade. The debate over AI compute governance and access policy will intensify through 2027.
The Future of AI Supercomputing After the Exascale Era
Looking ahead, the next generation of AI supercomputers is already in procurement. Oak Ridge, Argonne, and Lawrence Livermore are evaluating 3 to 10 exaflop successors for commissioning around 2028 and 2029. Early technology briefings point to Nvidia Rubin, AMD MI400 series, and Intel Falcon Shores class accelerators, paired with ultra-high-bandwidth optical interconnects. European partners are weighing a second exascale system beyond JUPITER, and China has signaled plans to field a 3 exaflop LineShine successor in the next cycle. The era of single-digit-exaflop AI machines is giving way to one of tens of exaflops in both dense and sparse math. Procurement decisions over the next 18 months will set the direction of the top 10 supercomputers powering AI research in 2026 through 2030.
Beyond raw compute, the next wave of AI supercomputing will be defined by energy budgets, sovereign AI programs, and tighter couplings between HPC and AI training frameworks. Facilities are also investing in quantum co-processors and photonic accelerators that may displace parts of today’s GPU stack over time. Observers tracking the understanding quantum AI’s future conversation see supercomputers as the launch pad for hybrid quantum-AI workflows. The SoundHound AI and quantum computing deployment is one early glimpse of the convergence. Analysts forecast that by 2029 at least five systems will cross 3 exaflops Rmax and at least two will cross 10 exaflops of mixed-precision throughput for AI training.
Mixed-precision AI exaflops across the 2026 top 10 supercomputers
Reported or vendor-estimated FP16/BF16/FP8 throughput available for AI training workloads. Values are approximate and change as upgrades land.
Sources: TOP500.org, system operator briefings from Lawrence Livermore, Oak Ridge, Argonne, and EuroHPC JU. Compiled by AIplusInfo for the 2026 article on top AI supercomputers.
Key Insights on the Top 10 Supercomputers Powering AI Research in 2026
- China’s LineShine reached 2.1 exaflops Rmax on the June 2026 TOP500 list, displacing US machines from the top rank for the first time in eight years.
- El Capitan at Lawrence Livermore registered 1.742 exaflops on the November 2024 TOP500 Rmax rankings, delivering 43 AI-tuned exaflops on 44,544 AMD Instinct MI300A APUs.
- Nvidia parts power 81 percent of TOP500 systems reported by Crypto Briefing in 2026, tightening vendor concentration risk even as AMD takes four top 10 slots.
- Europe’s JUPITER at Jülich crossed the exascale threshold during the November 2025 cycle described by The Register’s exascale announcement coverage, allocating half of its time to EU users.
- The MLPerf HPC benchmark suite shows Aurora scaling to over 10,000 GPUs on CosmoFlow and DeepCAM, proving modern AI workloads need interconnect-aware tuning beyond Linpack.
- Finland’s LUMI runs on 100 percent hydro-generated electricity, and its EuroHPC LUMI operational description reports more than 90 percent average occupancy during 2025.
- Fugaku’s Fugaku-LLM 13B was trained end-to-end on the system, proving CPU-only AI training is viable and shaping follow-on sovereign procurement.
- El Capitan’s power draw sits near 30 megawatts per the LLNL El Capitan hardware page, pushing every leadership lab to invest in liquid cooling and heat recapture.
The 2026 ranking of leadership AI supercomputers tells a clear story about where scientific AI is going next. Compute capacity is still concentrated in a handful of government-funded labs, but the geographic distribution is widening as Europe and Japan plan successor systems. Open allocation programs are now the main route for independent researchers who cannot buy a trillion-parameter run from any cloud vendor. The vendor balance is still skewed to Nvidia, with AMD closing the gap through APU integration and Intel defending a single flagship slot. These dynamics will determine which research teams can keep pace with the frontier through the back half of this decade.
| System | Operator and Location | Linpack Rmax | AI Accelerator | Interconnect | Power at Peak | Primary Use |
|---|---|---|---|---|---|---|
| LineShine | Chinese National SC Center | 2.1 EF | Domestic AI accelerator | Custom high-bandwidth fabric | Undisclosed | Sovereign AI and policy |
| El Capitan | LLNL, USA | 1.742 EF | AMD Instinct MI300A | HPE Slingshot-11 | 30 MW | Stockpile stewardship, AI |
| Frontier | Oak Ridge NL, USA | 1.353 EF | AMD Instinct MI250X | HPE Slingshot-11 | 21 MW | Open science, AI |
| Aurora | Argonne NL, USA | 1.012 EF | Intel Max GPU | HPE Slingshot-11 | 38 MW | AuroraGPT, scientific AI |
| JUPITER | Jülich, Germany | >1.0 EF | Nvidia Grace Hopper GH200 | InfiniBand NDR | 17 MW | EU sovereign AI |
| Fugaku | RIKEN, Japan | 442 PF | Fujitsu A64FX CPU | Tofu-D 6D torus | 30 MW | AI, climate, pandemic |
| LUMI | CSC Kajaani, Finland | 379 PF | AMD Instinct MI250X | HPE Slingshot-11 | 7 MW | European AI startups |
| MareNostrum 5 | BSC Barcelona, Spain | 314 PF | Nvidia H100 | InfiniBand NDR | 5 MW | Spanish sovereign AI |
| Alps | CSCS Lugano, Switzerland | 270 PF | Nvidia Grace Hopper GH200 | Slingshot-11 | 10 MW | Grace-Hopper AI workloads |
| Leonardo | CINECA Bologna, Italy | 241 PF | Nvidia A100 | Nvidia Mellanox HDR | 7 MW | Industrial AI, digital twins |
Real-World Applications Running Across the Top AI Supercomputers
Beyond benchmarks, the top 10 supercomputers powering AI research in 2026 drive outcomes that cloud GPUs cannot match. Three deployments illustrate how exascale time unlocks results at the frontier of science.
Frontier Protein Folding at Oak Ridge
Oak Ridge used Frontier to run a 405-million conformation protein folding job that reported 405 million candidate structures in under 72 hours. The work is described on the OLCF Frontier AI science update page during 2024. The team coupled AlphaFold-style transformer inference with physics-based molecular dynamics, consuming roughly 2.3 million node-hours. Outputs feed a public database that academic drug discovery teams use for hit generation. The study produced three experimentally validated antibody candidates and surfaced four limitations around GPU memory footprint during inference. The main caveat is that classical scoring functions still give better free energy agreement than the neural surrogates in 12 percent of cases.
LUMI Drug Discovery for Nordic Pharma
LUMI’s AI partition supports a drug discovery program used by AstraZeneca and smaller Nordic biotech companies. The program is documented on the LUMI case studies portal run by CSC. In 2025 the system completed 1.2 million generative diffusion runs that proposed candidate molecules for 15 undruggable protein targets. Each run consumed between 400 and 2,800 GPU hours on 11,912 AMD MI250X accelerators and produced ranked candidates in 36 hours. Results have been handed to three AstraZeneca programs now in preclinical testing, with one program advancing to IND-enabling studies during 2026. The limitation reported by CSC is a 10-day queue delay during peak demand that pushed smaller teams into partial reservations.
El Capitan Inertial Confinement Fusion Surrogates
El Capitan supports the Stockpile Stewardship Program and runs surrogate models for inertial confinement fusion diagnostics. The work is reported on the LLNL inertial confinement fusion AI update page in late 2024. Lawrence Livermore trained a 60 billion parameter surrogate on 11,000 historical physics runs, cutting single-shot prediction time from 42 hours to 11 minutes. The reduction allowed the team to explore 5 million design points in the time previously spent on 1,400, informing hardware choices for National Ignition Facility successor configurations. Outputs are classified at the Secret level, so external replication is limited to coarse-grained technical summaries. The reported limitation is that the surrogate still mis-predicts peak fuel pressure in 7 percent of extreme energy injection cases.
Recommended reading from the AI supercomputing stack
Three books that pair well with this article for practitioners planning or running AI research on leadership-class systems.
Deep Learning (Adaptive Computation and Machine Learning series)
The canonical deep learning textbook used by many of the research groups running experiments on the systems in this article.
Buy on AmazonParallel and High Performance Computing
A working engineer’s guide to MPI, OpenMP, and GPU programming on the same AMD and Nvidia hardware powering Frontier, LUMI, and Alps.
Buy on AmazonProgramming Massively Parallel Processors: A Hands-on Approach
The 4th edition hands-on CUDA and GPU parallel programming textbook covering the architectures used across modern AI supercomputers.
Buy on AmazonAs an Amazon Associate, AIplusInfo earns from qualifying purchases.
In-Depth Case Studies from the Top 10 Supercomputers Powering AI Research
Shifting from short examples to deeper narratives, the next three case studies look at AuroraGPT, Alps climate emulation, and LineShine large-model training in greater detail. Each case carries measurable outcomes, real limitations, and lessons for teams planning similar work.
Case Study: AuroraGPT at Argonne
Argonne launched AuroraGPT to build a trillion-parameter science foundation model trained on biology, chemistry, materials, and physics text and data. The project is outlined in the ALCF AuroraGPT project page maintained by Argonne National Laboratory. The team solved the problem that existing frontier models under-represent scientific reasoning and often hallucinate on domain-specific chemistry and biology queries. They used a 7 billion parameter prototype in 2024 and have since scaled through 70 billion to the trillion parameter tier through 2026. The impact is a scientific checkpoint release shared with roughly 30 university partners that has been used in 180 downstream papers to date. AuroraGPT’s checkpoint also powers several teaching tools across the Argonne training catalog.
The project’s measurable impact includes a 24 percent improvement on the GPQA reasoning benchmark and a 31 percent reduction in hallucinations on protein function queries verified by biology post-doctoral reviewers. Training ran across 10,624 Aurora nodes with roughly 60 percent accelerator utilization, a figure the team openly reported as the tradeoff for its mixture-of-experts approach. The project also had to slow training during 2024 after Intel Max GPU firmware updates caused a six-week delay. That limitation is unusual among frontier training efforts and reflects Aurora’s status as the only Intel-based exascale machine in the top 10. Argonne has since used the lessons to inform the planned 3-exaflop Aurora successor. The checkpoint data mixes have also been reused by partner teams on LUMI and Alps.
Case Study: Alps Climate Emulation at ETH Zurich
CSCS and ETH Zurich launched Alps-based climate emulation to deliver 1-km resolution regional projections that traditional physics models could not run at required cadence. The work is described on the CSCS Alps climate AI breakthrough update. The problem was that storm-resolving climate models cost tens of millions of CPU hours per century of simulation, pricing out most national weather services. Researchers built an 18 billion parameter neural weather model trained on ERA5 reanalysis data, consuming 160,000 Grace-Hopper GPU hours. The result was a 1-km resolution emulator that runs a 50-year projection in 11 wall-clock days on 400 GH200 nodes. That replaced roughly 3 years on the historical CPU cluster, a dramatic shift in cadence for operational forecasting.
Measurable impact includes adoption by the Swiss MeteoSwiss operational team and a formal 2026 handoff to the European Center for Medium Range Weather Forecasts for evaluation. The project cuts compute cost for regional projections by roughly 94 percent compared with physics-only runs. Limitations include known physical biases in rare extreme-precipitation events that emerged in cross-validation against convection-resolving control runs. CSCS and ETH have paired the model with a smaller physics-based correction layer that will likely remain through 2027. The case has become a reference for AI-augmented climate modeling on exascale machines in Europe and beyond. Several national weather services are now evaluating similar pipelines on their own regional supercomputers.
Case Study: LineShine Large Model Training at the National Supercomputing Center
China’s National Supercomputing Center has used LineShine to train large reasoning models like GLM-5 and DeepSeek-V3. The work is described in the Al Jazeera June 2026 LineShine coverage. The problem was that US export controls cut off access to Nvidia H100 and newer Blackwell parts, forcing a pivot to domestic accelerators across sovereign training runs. Chinese researchers built out a custom accelerator paired with domestic HBM and a high-bandwidth fabric described publicly only in broad outline. The impact includes training a 1.2 trillion parameter mixture-of-experts model that scored competitively on MMLU-Pro and GSM8K reasoning benchmarks. That impact also includes a measured reduction in China’s reliance on imported silicon for frontier training.
The broader strategic impact is to show that US controls can slow but not stop Chinese model progress at the frontier. Yet real limitations remain, with reports of lower sustained accelerator utilization and higher energy cost per token during training than Western equivalents. Chinese labs have published few MLPerf HPC results, so independent comparison remains limited for now. US and allied labs have cited LineShine as the main reason to accelerate their own next-generation procurement timelines. The case illustrates both sovereign AI’s political appeal and the real engineering cost of running a frontier training stack without the fastest silicon. The outcome reshaped assumptions about the pace of Chinese AI compute capability.
Frequently Asked Questions on Top AI Supercomputers in 2026
The ten leading systems in 2026 are LineShine in China, El Capitan at Lawrence Livermore, and Frontier at Oak Ridge. Aurora at Argonne, JUPITER at Jülich, and Fugaku at RIKEN round out the first group. The remaining top entries are LUMI, MareNostrum 5, Alps at CSCS, and Leonardo at CINECA across Europe.
The current top AI supercomputers use AMD Instinct MI250X and MI300A chips along with Nvidia H100 and Grace Hopper GH200 parts. Intel Max GPUs power Aurora while Fujitsu A64FX ARM CPUs drive Fugaku in Japan. AMD and Nvidia together hold nine of the ten slots in the current rankings.
China’s LineShine at a national supercomputing center leads outside the United States with 2.1 exaflops Linpack and roughly 3 mixed-precision exaflops for AI training. LineShine is used for sovereign LLM training and major scientific workloads across Chinese academic programs. The machine overtook El Capitan on the June 2026 TOP500 list after extensive commissioning.
LUMI, Aurora, Frontier, and El Capitan are the most reliable platforms that scale to billion- and trillion-parameter training with dedicated AI access schemes. EuroHPC in Europe and INCITE in the United States pair these systems with peer-reviewed allocations. Pharma teams increasingly use LUMI for drug discovery and Aurora for molecule design.
Power at peak ranges from under 7 megawatts for efficient systems like LUMI and Leonardo to 38 megawatts for Aurora and 30 megawatts for El Capitan and Fugaku. Operators invest heavily in direct-liquid cooling and warm-water heat recapture to control grid impact. Total annual energy costs can reach hundreds of millions of dollars.
Yes, most of the top 10 operate open allocation programs that do not charge for compute time. Researchers must submit competitive proposals through INCITE, ALCC, or EuroHPC calls. Successful proposals typically require a clear scientific case, prior benchmarking, and a data management plan.
LineShine leads on Linpack Rmax at 2.1 exaflops while El Capitan leads on published mixed-precision AI exaflops at roughly 43. Direct AI comparison is limited because China has released few MLPerf HPC results for LineShine. El Capitan’s results against open AI benchmarks remain the most reliable data point for the top AI supercomputer class.
Aurora and Frontier share leadership for open-science AI, with Frontier running more INCITE and ALCC projects while Aurora focuses on scaling AuroraGPT at trillion parameter sizes. Aurora’s Intel software stack gives it a different development mix than the AMD platforms. Frontier’s AMD stack has more mature MLPerf HPC results for large-scale AI training runs today.
EuroHPC Joint Undertaking owns and co-funds JUPITER, LUMI, Leonardo, MareNostrum 5, and Alps, making it the single largest AI compute sponsor in Europe. It also runs the AI access scheme that gives startups and public-sector teams free time. Half of JUPITER’s time is reserved for EU industrial and public-sector users.
Frontier and El Capitan use Lustre on NVMe arrays delivering tens of terabytes per second of aggregate read throughput for training jobs. Aurora runs DAOS for object-based storage that matches typical AI data loader patterns well at scale. Modern procurement treats parallel storage as a line item equal to the accelerator budget because AI data loaders can starve GPUs during training.
Fugaku runs Fugaku-LLM, pandemic aerosol simulations, Earth Simulator-class climate runs, and large-scale drug repurposing. Its A64FX ARM CPUs with Scalable Vector Extensions give good performance for memory-bound AI workloads. Japan is planning a successor to Fugaku that will likely add GPU accelerators to broaden AI reach.
Alps runs 2,688 Nvidia Grace Hopper GH200 superchips with unified CPU-GPU memory that reduces data movement cost during transformer training. ETH Zurich and EPFL teams use it for climate emulation, drug discovery, and the Apertus Swiss foundation model. The Grace Hopper design cuts the host-to-device bottleneck that limits large-model training on older architectures.
The next wave targets 3 to 10 exaflops with Nvidia Rubin, AMD MI400 series, and Intel Falcon Shores class accelerators and ultra-high-bandwidth optical interconnects. Facilities are also investing in quantum co-processors and photonic accelerators that may offload parts of the AI stack. Commissioning for the new systems is expected around 2028 and 2029.