NVIDIA's Model & GPU Push

NVIDIA is widening its artificial intelligence stack with free hosted application programming interfaces for roughly 80 models and a new RTX PRO 4500 Blackwell server chip for virtualized enterprise work. (build.nvidia.com) The hosted model catalog on NVIDIA’s Build platform now includes free endpoints for third-party models such as MiniMax M2.7 and DeepSeek-V3.2, alongside NVIDIA’s own Nemotron family. NVIDIA says NIM exposes those models through standard APIs for app integration and self-hosted deployment options. (build.nvidia.com, developer.nvidia.com) MiniMax M2.7 is listed by NVIDIA as a 230 billion-parameter model for coding, reasoning and office tasks, while DeepSeek-V3.2 is listed as a 685 billion-parameter reasoning model with long context and built-in agent tools. Both are available as free endpoints on NVIDIA’s model platform. (build.nvidia.com, build.nvidia.com) NVIDIA also published new details this week on the RTX PRO 4500 Blackwell Server Edition, a 165-watt, single-slot data center graphics processing unit aimed at virtual desktop infrastructure, graphics workloads and lighter artificial intelligence development. NVIDIA says the chip pairs with vGPU 20 software and delivers nearly 1.9 times the graphics acceleration of previous NVIDIA architectures in virtualized environments. (developer.nvidia.com, nvidia.com) Virtualization software slices one physical GPU into many remote workstations, so companies can run designers, analysts and developers from shared servers instead of individual high-end PCs. NVIDIA’s vGPU sizing guide names the RTX PRO 4500 Blackwell Server Edition as its primary recommendation for NVIDIA Virtual PC deployments and says it supports Multi-Instance GPU partitioning. (docs.nvidia.com) The software push reaches beyond model hosting. Adobe and NVIDIA said on March 16, 2026, that they would work together on next-generation Adobe Firefly models and creative, marketing and agentic workflows, and NVIDIA said this week that Adobe’s CX Enterprise Coworker uses NVIDIA Agent Toolkit components and Nemotron models in a live demo. (news.adobe.com, blogs.nvidia.com) That adds another layer to NVIDIA’s recent effort to sell more than chips. NIM packages models as inference microservices, while tools such as NeMo Agent Toolkit and Dynamo are aimed at the orchestration, monitoring and scheduling problems that appear when companies move from a single chatbot to multi-step agent systems. (developer.nvidia.com, developer.nvidia.com) NVIDIA has been building that ladder for more than a year. In March 2025, it introduced the broader RTX PRO Blackwell workstation and server line for designers, developers and data scientists, and in 2024 it opened free NIM access for developer program members as a way to seed usage before enterprise deployments. (nvidianews.nvidia.com, developer.nvidia.com) The immediate effect is that developers can test leading open and third-party models on NVIDIA-hosted endpoints, then move those workloads onto NVIDIA software and hardware inside corporate infrastructure. The same pitch now spans model discovery, agent tooling and the server GPUs meant to run those jobs at scale. (build.nvidia.com, developer.nvidia.com, developer.nvidia.com)