AI supercomputers are not just about speed; they’re about pushing the boundaries of what’s possible.
These machines use thousands of powerful GPUs to train and run artificial intelligence models, making them perfect for deep learning and generative AI tasks. They’re the tools that help us unlock the potential of AI, and they’re only becoming more powerful.
Today, the most powerful AI supercomputer is Colossus, built by xAI in the United States. It initially used around 100,000 NVIDIA H100 GPUs and remains unrivalled in scale. Colossus is being expanded with even more H100 and new H200 GPUs.
These AI supercomputers are not just marvels of technology; they’re also a testament to human ingenuity and collaboration. In this article, we will take a look at the list of the most powerful AI supercomputers.
We will include their H100 GPU count, owners, and the countries in which they are located. We will compare their scale, technology, and impact, showing how these machines are shaping the future of technology on a global scale.
Which Are The Top AI Supercomputers In The World With H100 GPU count?
Based on the data provided by EpochAI, here are the top 10 most powerful AI Supercomputers in the world with a H100 GPU count:
| Rank | Compute Cluster | H100 Equivalents | Owner | Country |
| 1 | xAI Colossus Memphis Phase 2 | 200,000 | xAI | U.S. |
| 2 | Meta 100k | 100,000 | Meta AI | U.S. |
| 3 | OpenAI/Microsoft Goodyear Arizona | 100,000 | Microsoft, OpenAI | U.S. |
| 4 | xAI Colossus Memphis Phase 1 | 100,000 | xAI | U.S. |
| 5 | Oracle OCI Supercluster H200S | 65,536 | Oracle | U.S. |
| 6 | Tesla Cortex Phase 1 | 50,000 | Tesla | U.S. |
| 7 | Lawrence Livermore NL El Capitan Phase 2 | 44,143 | U.S. Department of Energy | U.S. |
| 8 | CoreWeave H200s | 42,000 | CoreWeave | U.S. |
| 9 | Lambda Labs H100/H200 | 32,000 | Lambda Labs | U.S. |
| 10 | Anonymised Chinese System | 30,000 | N/A | China |
xAI Colossus Memphis Phase 2
xAI’s Colossus Phase 2 is the world’s largest AI supercomputer. It utilises approximately 200,000 Nvidia GPUs and operates on 150 MW of grid power, supplemented by Tesla Megapacks.
The total power draw reaches 300 MW, sufficient to power approximately 300,000 homes. The cluster uses Nvidia Spectrum‑X Ethernet and BlueField‑3 SuperNICs. This enables data to move quickly with low latency, even at massive scale.
Colossus was built in just 122 days and began with 100,000 GPUs. It expanded to 200,000 within 92 days—a significant engineering feat. The cluster trains xAI’s Grok language models and integrates into X platform services. xAI plans to scale to one million GPUs in the future, Phase 3.
Meta 100k
Meta’s Meta 100k cluster utilises over 100,000 H100 GPUs to train its Llama 4 AI models.
The company is investing heavily in new data centres—some planned almost the size of Manhattan—to support AI scale and future AGI ambitions.
Meta’s AI Research SuperCluster (RSC) evolved from earlier clusters with A100 GPUs and petabytes of storage.
It now features open-source infrastructure, high-throughput storage, and custom hardware stacks for reliability and scalability. This public investment signals Meta’s strategic bet on AI leadership.
OpenAI/Microsoft Goodyear Arizona
The Goodyear cluster in Arizona, a colossal system co-owned by OpenAI and Microsoft, operates on approximately 100,000 H100‑equivalent GPUs as part of the “Stargate” Azure AI infrastructure.
The system, optimised for OpenAI’s largest models and deep learning workloads, features high-density GPU racks, custom networking, and advanced storage pipelines, showcasing its cutting-edge capabilities. The site serves as a critical backbone for real‑time AI services via Microsoft Azure.
xAI Colossus Memphis Phase 1
Phase 1 of Colossus launched with 100,000 H100 GPUs. It was completed within 122 days at a former Electrolux factory in South Memphis.
That initial build powered the first versions of Grok, xAI’s flagship language model.
Colossus Phase 1 proved xAI’s ability to rapidly build at scale using high-efficiency liquid cooling and Spectrum-X networking.
Oracle OCI Supercluster H200S
Oracle’s OCI Supercluster offers up to 65,536 Nvidia H200 GPUs in the cloud. It delivers up to 260 exaflops of FP8 performance—much higher than earlier H100-based systems.
It features super-fast RoCE v2 networking via ConnectX‑7 NICs and 400 Gbps GPU-to-GPU links, plus a 200 Gbps front-end for data movement. Oracle makes it publicly available to enterprise and government users, with H200 rates similar to H100 pricing ($10 per GPU-hour).
Conclusion
AI supercomputers, the backbone of global innovation, are equipped with tens of thousands of powerful H100 or H200 GPUs, designed to train the most advanced AI models in history. The U.S. currently leads with giants like xAI’s Colossus, Meta’s RSC, and Microsoft’s Azure clusters.
These systems demand massive power, cutting-edge cooling, and huge investments. As more companies strive to build even larger systems, the role of AI supercomputers in driving global innovation becomes more pronounced.
This list highlights the world’s top AI supercomputers—by size, ownership, and location—giving a clear view of where the future of AI is being built.
