NVIDIA Details Infrastructure Behind Latest OpenAI Models and Benchmarks

OpenAI's latest model releases represent a continuation of dependence on specialized GPU infrastructure. GPT-5.2, described by OpenAI as their most capable model for professional knowledge work, was trained and deployed using NVIDIA Hopper and GB200 NVL72 systems. The more recent GPT-5.3 Codex, an agentic coding model released in February, ran entirely on GB200 hardware. Both models have published benchmark results: GPT-5.2 achieved reported top scores on GPQA-Diamond, AIME 2025, and Tau2 Telecom, while GPT-5.3 Codex set new highs on SWE-Bench Pro and Terminal-Bench evaluations.

NVIDIA's comparison metrics show measurable generational improvements in its own hardware stack. According to the company's MLPerf Training submission, GB200 NVL72 systems demonstrated 3x faster training performance than Hopper on the largest tested models, alongside claims of nearly 2x better performance per dollar. NVIDIA's upcoming GB300 platform shows further gains, claiming more than 4x speedup versus Hopper. These claims rest on standardized industry benchmarks, though they reflect NVIDIA's own testing environment.

The infrastructure landscape extends beyond language models. NVIDIA highlighted support for models across multiple modalities: genetic sequence analysis (Evo 2), protein structure prediction (OpenFold3), drug interaction simulation (Boltz-2), and medical imaging synthesis through its Clara platform. Video generation models from Runway, including Gen-4.5 (positioned as the top-rated model on Artificial Analysis leaderboards), were trained entirely on NVIDIA GPUs. A separate Runway announcement describes GWM-1, a general world model also trained on Blackwell hardware.

NVIDIA Blackwell availability spans major cloud infrastructure providers. The company lists AWS, Google Cloud, Microsoft Azure, CoreWeave, Lambda, Oracle Cloud Infrastructure, and Together AI as offering Blackwell-powered instances. This distribution reflects both competitive pressure in cloud AI infrastructure and the centrality of NVIDIA's hardware to current model training workflows.