NVIDIA and DigitalOcean Collaborate on Always-On Agentic Technology Stack

NVIDIA and DigitalOcean Collaborate to Enhance Open-Source AI Development

NVIDIA and DigitalOcean are making strides in the open-source AI landscape, emphasizing the need for production-ready generative AI models. During a recent DigitalOcean Deploy session titled “Open by Design,” NVIDIA’s VP of Generative AI, Kari Briski, and DigitalOcean’s SVP of AI, Salman Paracha, discussed the growing demand for flexibility and openness in AI development. Their collaboration aims to empower developers with tools and infrastructure necessary for creating robust agentic AI applications.

The Challenge of Open Models

The rise of open-source AI models has not been without its challenges. While many models are available to developers, their long-term viability often remains uncertain. Briski pointed out that many open-source models are launched but not regularly updated, leaving developers without essential improvements or support. This gap prompted NVIDIA to create the Nemotron family of multi-modal models, released in March 2026, designed specifically for agentic AI applications.

Nemotron models allow developers to build applications requiring advanced reasoning and high computational efficiency. By leveraging these open models alongside NVIDIA’s software libraries, developers can ensure their projects remain up-to-date and evolve over time. The availability of these resources is crucial as it enables developers to experiment with agentic applications that can operate continuously and adapt to changing needs.

Local Deployment and Control

Running open-weight large language models (LLMs) locally offers developers greater control over performance, privacy, and customization. A recent tutorial on deploying NVIDIA’s Nemotron 3 Nano on DigitalOcean’s GPU Droplets illustrates how developers can utilize dedicated GPU infrastructure without relying solely on hosted AI APIs. This approach enhances flexibility while allowing for experimentation with efficient open models.

Briski emphasized the importance of treating these models as a library that can be continually improved upon. This philosophy extends beyond just model development; it encompasses orchestration frameworks—tools that manage the lifecycle of agents, memory usage, tool calling, and scaling—which are vital for building effective agentic systems.

Addressing Developer Challenges

Despite the availability of tools and resources, many developers still face significant hurdles when attempting to create durable AI-native applications. Paracha noted that understanding whether a project can compete with well-funded AI companies remains a daunting challenge. The lack of standardized evaluations makes it difficult for developers to assess their ideas’ viability accurately.

To address this issue, Briski highlighted the need for more comprehensive test cases and benchmarks across various applications. Current academic evaluations often fail to reflect real-world scenarios effectively. NVIDIA is collaborating with industry leaders like Synopsys and Cadence to develop relevant benchmarks that will facilitate broader development in agent creation.

Furthermore, sub-agent workflows—where complex problems are divided into smaller tasks managed by individual agents—are gaining traction among developers. This method allows for better management of tasks while maintaining oversight over the overall system architecture. As development practices evolve, understanding how these sub-agent systems function will become increasingly important.

The Future of Token Economics

As AI systems continue to scale, token usage presents new challenges for developers aiming to build sustainable business models around their applications. Briski pointed out that counting tokens will evolve as model architectures change; thus, focusing on the value delivered by products becomes paramount.

NVIDIA is actively working on improving token efficiency through innovations like hybrid-state-space transformers in their latest Nemotron Model. This approach contrasts with traditional dense models combined with mixture-of-experts (MoEs), allowing for more efficient computation while maintaining performance standards.

The shift from dense megamodels towards sparser architectures signifies a broader trend within the industry. By adopting solid-state models (SSMs), which reduce computational demands by streamlining processing layers, organizations can enhance their operational efficiency while managing costs effectively.

Collaboration Between NVIDIA and DigitalOcean

NVIDIA’s applied research team is committed to exploring new architectures and collaborating with the open-source community to drive innovation forward. Briski noted that sharing ideas within this community is vital for progress in AI development.

DigitalOcean is also focused on expanding its ecosystem around open-source projects by developing technologies like Plano (a data-plane technology) and researching small action models (SAMs). These advancements aim to enable efficient task completion without requiring extensive reasoning tokens or lengthy context windows.

The concept of an open harness architecture allows developers to integrate various agents seamlessly while managing their lifecycle events effectively. Paracha emphasized this approach’s importance in fostering choice within the ecosystem rather than imposing proprietary solutions.

What This Means

The collaboration between NVIDIA and DigitalOcean marks a significant step toward democratizing access to powerful generative AI tools for developers worldwide. As organizations transition from proprietary solutions to open-source frameworks, they can expect increased flexibility and innovation in building agentic applications across various industries. The emphasis on continuous improvement, developer support, and effective resource management will empower individuals and teams alike to harness the full potential of generative AI technology.

For more information, read the original report here.