NVIDIA's architecture runs every AI model in every field of science and every modality, including American and international models.

Judge as of 2026-09-04. The claim is universal: every model, in every field of science and every modality (including American and international models), runs on NVIDIA's platform. True: major model families across modalities (text, vision, audio, video, biology/chemistry, weather and climate, physics, and so on) and from US and non-US developers (for example, Chinese open-weight models) are documented as running on NVIDIA GPUs, and there is no documented notable model that cannot run on NVIDIA hardware. Mostly true: near-universal support, with only minor or obscure exceptions. Mixed: broad support, but a meaningful category of models or modalities is documented as unsupported or impractical on NVIDIA. Mostly false: significant, well-known models are documented as running only on non-NVIDIA hardware (for example, models tied to proprietary TPU or other accelerator stacks with no NVIDIA deployment path), so the universal claim overstates real broad support. False: NVIDIA support is limited to a minority of models or modalities.

Set before the research began.

1. E1 Support Matrix for NIMs NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “While not explicitly validated, the model-free NIM can be used with any model supported by the underlying backend (vLLM) version.”

2. E2 Support Matrix for NIMs NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “The following table lists the supported profile configurations for moonshotai/kimi-k2.6:”

3. E3 Support Matrix for NIMs NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “stepfun-ai/Step-5-Preview. This model is served”

4. E4 NVIDIA Earth-2 Powers Regional AI Weather Forecasting in the United Arab Emirates | NVIDIA Technical Blog NVIDIA Technical Blog · 19 March 2025 Own record

   “The local forecasting model for the UAE is trained and deployed on G42’s cluster of NVIDIA H100 GPUs, consisting of nodes of eight H100s each.”

5. E5 Support Matrix for NIMs# NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “This page lists supported models, deployment profiles, and verified hardware SKUs for NIM LLM and VLM.”

6. E6 Support Matrix for NIMs# NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “The following table lists the supported profile configurations for moonshotai/kimi-k2.6:”

7. E7 Support Matrix for NIMs# NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “While not explicitly validated, the model-free NIM can be used with any model supported by the underlying backend (vLLM) version.”

8. E8 Support Matrix for NIMs# NVIDIA (docs.nvidia.com) · 19 October 2026 Own record

   “runtime allocations and cannot start on smaller GPUs.”

9. E9 Huawei Open Sources 505B openPangu AI, Drops Weights And Code - Open Source For You Open Source For You · 4 August 2026

   “it says completed its entire pretraining run on Ascend 910B NPUs without using Nvidia GPUs.”

10. E10 Support Matrix for NIMs# NVIDIA (NIM documentation) · 29 October 2026 Own record

    “GLM-5.3 is served by the SGLang Model-Free NIM 2.1.4 container,”

11. E11 Support Matrix for NIMs# NVIDIA (NIM documentation) · 29 October 2026 Own record

    “Image and video input are not supported. Refer to the linked guide for the multi-node Helm configuration and deployment requirements.”

12. E12 bionemo-evo2 NVIDIA (BioNeMo Framework documentation) · 6 January 2025 Own record

    “The original 40b-1m context trained model only supports hpper fp8”

Almost all of the evidence comes from NVIDIA itself. Its NIM support matrix, dated October 2026 (after the judging date), lists models such as kimi-k2.6, GLM-5.3 and Step-5-Preview as supported or served through NIM (E2, E3, E10). Other NVIDIA sources show a UAE weather model (E4) and the Evo2 biology model, which has hardware restrictions, running on NVIDIA hardware (E12). The page's own scope is LLM/VLM only, and the wider "any model" statement is described as not explicitly validated (E1, E5, E11). No evidence covers audio, video or physics models. The one independent source shows only that Huawei trained openPangu without NVIDIA GPUs, not that the model cannot run on them (E9). That leaves the universal claim neither confirmed nor refuted as of 2026-09-04.