NVIDIA Dominates June 2025 Supercomputing and AI Hardware

NVIDIA this week reported that its technology powers 81% of the TOP500 supercomputers — more than 400 of the world’s 500 fastest systems — while simultaneously unveiling new scientific AI software, expanded AWS cloud instances, and telecom AI agent infrastructure at conferences in Hamburg and Copenhagen. The announcements span hardware, software, and cloud partnerships, and come as banned NVIDIA chips are reportedly fetching double their original price on China’s black market.

NVIDIA Controls 81% of the TOP500 List in June 2025

According to NVIDIA’s announcement tied to the ISC High Performance conference in Hamburg, NVIDIA GPUs now accelerate a record 238 systems on the TOP500, while NVIDIA networking connects a record 376 systems — the majority running NVIDIA Quantum InfiniBand. The company gained 17 net systems from the previous list, with 90% of newly added machines built on NVIDIA technology.

Energy efficiency rankings also favor NVIDIA hardware. The top eight systems on the Green500 — which ranks supercomputers by performance per watt — run on NVIDIA GPUs, and nine of the top ten use NVIDIA technologies. The number-one Green500 system, KAIROS, runs on a single NVIDIA Grace Hopper Superchip. The report also noted that NVIDIA’s collective TOP500 presence now delivers more than 2x the AI training throughput and nearly 3x the AI inference throughput of every other platform combined, according to NVIDIA’s own figures.

The Grace CPU specifically gained traction: 26 TOP500 systems now use it, up eight from the prior list.

New Scientific AI Software Targets Chemistry, Astronomy, and Materials Research

At ISC this week, NVIDIA introduced three software tools under its CUDA-X umbrella aimed at accelerating scientific workflows that previously ran on CPUs over hours or days. The new offerings are the NVIDIA DAQIRI library, NVIDIA ALCHEMI NIM microservices, and NVIDIA cuPhoton (in early access), according to the NVIDIA AI Blog.

cuPhoton is designed for astronomical data processing. Running on NVIDIA GB200 NVL72 systems, it accelerated the loading and reading of FITS image files from the Rubin Observatory’s Legacy Survey of Space and Time (LSST) by 14,900x in early testing. Signal processing and analysis using 32 NVIDIA Grace Blackwell superchips ran up to 8,400x faster than CPU-based baselines. The LSST camera — described as the world’s largest digital camera at 3,200 megapixels — is expected to image roughly 20 billion galaxies over its operational life, making high-throughput data pipelines essential.

ALCHEMI NIM microservices target chemistry and materials discovery, while DAQIRI is aimed at real-time data acquisition and processing pipelines across experimental science domains.

AWS Deploys RTX PRO 4500 Blackwell GPUs in New EC2 G7 Instances

Amazon Web Services launched new EC2 G7 instances powered by NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs, delivering up to 4.6x AI inference performance and up to 2.1x graphics performance compared to the prior G6 instances, according to the NVIDIA AI Blog. The G7 instances support up to eight GPUs, 256GB of total GPU memory, and 700 Gbps of EFA-enabled networking, with up to 7.6TB of local NVMe storage.

Separately, AWS has achieved NVIDIA Exemplar Cloud status for NVIDIA GB300, a certification NVIDIA uses to indicate peak-optimized performance for training workloads on a given cloud provider. NVIDIA’s cuVS library is also now the default for GPU-powered vector indexing in Amazon OpenSearch Serverless, accelerating the retrieval layer for enterprise AI applications.

The collaboration reflects a broader push by NVIDIA to embed its hardware and software stack into managed cloud services, reducing the operational burden on enterprise teams deploying AI at scale.

Telecom AI Agents Debut at DTW Ignite 2026 in Copenhagen

NVIDIA and its partners are demonstrating autonomous network management tools at TM Forum’s DTW Ignite 2026 in Copenhagen this week. The system combines synthetic data generation, telecom-domain AI models, secure agent runtimes, and network simulation, according to the NVIDIA AI Blog. The goal is to move telecom operators from task-based automation toward AI agents that can proactively detect and resolve network issues across IT, network, and business systems without continuous human direction.

NVIDIA’s telecom autonomy platform is designed to keep humans in control of policy while agents handle execution — a design choice driven by regulatory and operational risk considerations in critical infrastructure. The company cited a prior report showing operators have already seen measurable returns from generative AI in network management and customer care, though specific figures from that report were not available in the source material reviewed.

Banned NVIDIA Chips Double in Price on China’s Black Market

The Financial Times reported that NVIDIA AI chips subject to U.S. export restrictions have doubled in price on China’s black market, though the full article is behind a paywall and specific chip models and price figures were not available in the accessible portion of the source. The headline finding points to sustained demand from Chinese buyers unable to access NVIDIA hardware through official channels following successive rounds of U.S. export controls targeting advanced AI accelerators.

This dynamic has persisted since the Biden administration’s October 2022 and October 2023 export control expansions, which restricted sales of chips including the A100 and H100 to China. The price premium on gray-market units reflects both demand intensity and the difficulty of sourcing alternatives at equivalent performance levels.

What This Means

NVIDIA’s position across supercomputing, cloud, scientific research, and telecom AI is not a single product story — it is a platform lock-in story. With 81% of the TOP500, default status in Amazon OpenSearch, and new scientific libraries tied to GB200 and Grace Blackwell hardware, the company is embedding itself into the infrastructure layer of AI workloads across sectors. Competitors seeking to displace NVIDIA face not just a hardware gap but a deep software and ecosystem dependency that takes years to replicate.

The China black market data point, while limited in detail, signals that export controls are creating scarcity without eliminating demand — a condition that sustains NVIDIA’s pricing power globally while creating a parallel supply chain problem that U.S. policymakers will need to address separately.

FAQ

How many of the TOP500 supercomputers use NVIDIA technology?

According to NVIDIA’s announcement at the ISC High Performance conference in June 2025, NVIDIA technology powers more than 400 of the world’s 500 fastest supercomputers, representing 81% of the TOP500 list. NVIDIA GPUs accelerate a record 238 systems, and NVIDIA networking connects a record 376.

What is NVIDIA cuPhoton and what does it do?

NVIDIA cuPhoton is a GPU-accelerated reference library for processing astronomical data in the FITS file format. In early access testing on GB200 NVL72 systems, it accelerated FITS image loading from the Rubin Observatory’s LSST survey by 14,900x and enabled up to 8,400x faster signal processing using 32 Grace Blackwell superchips.

Why are NVIDIA chips expensive on China’s black market?

U.S. export controls restrict the sale of advanced NVIDIA AI accelerators — including the H100 and related chips — to Chinese buyers through official channels. The Financial Times reported that prices for these banned chips have doubled on China’s black market, reflecting high demand from Chinese AI developers and a lack of comparable domestic alternatives.