Written by: Techub News Compilation

During a high-profile roundtable discussion at the NVIDIA GTC 2026 conference, NVIDIA founder and CEO Jensen Huang, along with founders and executives from leading AI companies such as Perplexity, Mistral AI, Cognition, AMP, and the Allen Institute for AI (AI2), discussed the future landscape of the AI model ecosystem. This dialogue transcended the simple "open source vs. proprietary" debate, deeply analyzing the forms of next-generation AI agents, the core value of open models in critical tasks, and the urgency of building open infrastructure, providing key insights for the evolution direction of the AI industry.

Beyond Opposition: The Era of Symbiosis Between Proprietary and Open Source Models

Jensen Huang set the tone from the outset: the future of AI is not a choice between "proprietary and open source," but rather a symbiosis of both. He believes that the market needs both forms to co-exist. On one hand, excellent proprietary models, as mature products, provide services through APIs to meet widespread demand; on the other hand, numerous companies across various industries and fields need to treat models as a foundational technology, transforming them into their exclusive products. This has led to the emergence of a "third type of company" that exists between foundational model companies and pure application companies. These companies leverage the best API services available while deeply developing at the model level, combining external APIs with self-developed models to create the best products for specific verticals.

This viewpoint received broad endorsement from participants. Aravind Srinivas, CEO of Perplexity, elaborated further, stating that AI is not just the model itself, but the entire system, akin to a "computer." He proposed the concept of a "Perplexity computer," which means building a system capable of orchestrating all capabilities of AI—whether it's coding, writing, or generating multimodal content. In this system, various models function like musical instruments, sub-agents resemble musicians, while the orchestration system acts as the conductor, ultimately performing a symphony together. Users simply issue tasks without worrying about which model excels at what; everything is intelligently scheduled by the orchestration system. In this architecture, open and closed models can each play their roles: open models typically excel in token efficiency and cost control, while closed models have advantages in complex reasoning and task orchestration. The models themselves are gradually becoming foundational tools like file systems and connectors.

Arthur Mensch, co-founder and CEO of Mistral AI, added from a corporate practice perspective that open source models should become the cornerstone of global AI software for two main reasons: control and customization. AI agents operate at the execution layer, and companies must have complete control over their deployment, including the control of "switches." If completely reliant on external APIs, an interruption in service or any issues could pose significant risks. Having a complete system from the model to orchestration layer provides crucial resilience. Additionally, when agents need to interact with the physical world (like operating machines, modeling physical systems), companies generate a considerable amount of proprietary knowledge that they do not wish to disclose to generalized models. Open source models allow companies to modify deeply, access time-sensitive data streams, thus building agents that understand the physical world and truly create value for teams like engineers. The product "Forged" launched by Mistral AI aims to connect models with various data sources from the physical world.

The Evolution of Agents: From Tool Utilization to "Composite Colleagues"

One focal point of the discussion was the evolution path of AI agents. Participants believed that agents are evolving from simple model calls to assistants able to use tools, and are on the verge of becoming "composite colleagues" that can undertake complex tasks lasting hours or even days.

Scott Wu, CEO of Cognition, shared their experiment: allowing an agent to build a prototype browser from scratch based entirely on intent over several weeks. When tasks become this complex, the underlying system needs to allocate workloads to different models because each model has its own specialties. Sometimes, the capabilities of top foundational model APIs need to be called upon, while other times, the company’s accumulated industry knowledge in specific areas is needed to fine-tune its own models. In the future, we will witness the rise of these "composite agents" that become smarter than any single model by integrating the strengths of multiple models.

Jensen Huang pointed out from a macro perspective on technological evolution that there is a common misunderstanding in the industry that pre-training is the entirety of model development. In fact, pre-training is merely the beginning; it provides memory, generalization, and foundational knowledge, much like needing to grasp the fundamentals of math and science to learn engineering. The model development that various labs invest huge computing resources into is more concentrated in the "post-training" phase. He predicts that the allocation ratio of training computing resources will undergo a fundamental shift in the future: a few years ago, pre-training might have accounted for 90% of computing volume, whereas in the future, this proportion will become minimal, with the vast majority of computation devoted to post-training. This means that the refinement and specialization of model abilities will become unprecedentedly important.

Scott Wu cited AlphaGo as an example to highlight the immense potential of combining reinforcement learning (RL) with language models. AlphaGo, a "small" network with only 60 million parameters, managed to defeat the world's top Go players, and its core lies in being a system that never ceases to learn; its capabilities were more of an economic question at the time (how much computing resource one is willing to invest to make it stronger). Now, RL is beginning to play a role in language models, and we will face similar economic decisions: whether to invest billions or tens of billions to have AI systems tackle a specific disease? Once RL is widely applied, solving many issues will become a matter of pure resource allocation decisions. Currently, we have seen the first generation of results in coding and enterprise-level agent applications, which will extend to breakthroughs in basic scientific problems in the future.

The Value of Openness: Trust, Innovation, and Infrastructure Grid

When discussing why open source models are crucial, several guests pointed towards "trust" as the core issue. Brandon Kumar, CEO of AMP, sharply noted that as AI applications (especially agents) become increasingly critical, they begin to enter high-risk, low-tolerance "critical task" areas such as healthcare and national defense. In such cases, trust becomes paramount. If a model is not open-source, you cannot control where it is deployed, cannot perform introspective checks, cannot self-host, and are entirely dependent on third parties, you are essentially "delegating trust." For critical task systems, we must find ways to trust them. In his view, open source models are one of the fastest ways to establish systemic trust because you can examine their components, clarify which parts are closed source (yet performing well), and based on this, deploy guardrails and manage risks.

Brandon Kumar further directed the conversation toward deeper infrastructure issues. He believes that for open source models to progress to the cutting edge, open infrastructure is equally necessary. He warned that the current AI infrastructure field is rapidly consolidating, reminiscent of the early industrial revolution when factories hoarded steam engines and coal. If each organization over-provisions and hoards computing power for peak demand, it will lead to significant resource waste and inefficiency. What we need is a mechanism similar to the "electric grid"—with sufficient secure infrastructure to meet basic loads while being able to flexibly respond to peak fluctuations. AMP is committed to building such an "AI grid." Only by combining open models with open infrastructure can we achieve efficient resource sharing and widespread innovation.

Dirk Groeneveld, a senior manager at the Allen Institute for AI (AI2), emphasized the value of openness from the perspective of research democratization. He pointed out that current AI advancements seem increasingly concentrated in a few closed labs, but access to and participation in cutting-edge research is equally vital for the broader research community, including academia and non-profit organizations. His team has built the Olmo model and the "model flow" framework designed to open the entire lifecycle of model development, including data, model weights, infrastructure, etc., allowing researchers and developers to experiment, integrate discoveries, and even build new applications. He used hybrid model architectures (like Nemotron) as an example, illustrating that it is through open research environments that communities can jointly analyze and validate why hybrid models are more efficient than pure Transformers, thereby promoting the advancement of next-generation model architectures and AI systems.

Dirk Groeneveld concluded that, strictly speaking, open source models are superior to closed source models because they promote the diffusion of innovation, research, and competition. Even within a closed-source model company, open source models will inevitably be used as part of the agent system, where closed source models are the "crown jewel," surrounded by a range of open source models. He predicts that even in the currently highly proprietary field of visual intelligence (future robotic visual intelligence systems), it will ultimately be surrounded by a large number of open source models.

The Industry Inflection Point Has Arrived: From "Is it Useful?" to "Return on Investment"

At the end of the roundtable, the discussion returned to the commercial reality of AI. Jensen Huang optimistically pointed out that the industry is undergoing an inflection point. Three to four years ago, people were questioning whether AI could only be used for chatbots; today, due to the maturation of reinforcement learning and agent infrastructure, AI is becoming genuinely useful in critical task applications. This year, the core questions in the industry will shift from "Does AI have a return on investment?" to specifically exploring "What is the return on investment of AI in various fields?"

He particularly noted that coding will be the first area to show clear commercial value, and this "coding" goes far beyond software engineering; it represents the description and solidification of business processes, commercial rules, and almost all work. With the enhancement of agent capabilities and the maturity of composite systems, the true commercial economic effects will take off this year. Participants unanimously agreed that a new AI ecosystem driven by both proprietary and open source models, with agents as the core execution unit, and built on open, grid-based infrastructure, is accelerating in formation. This represents not only an evolution of technology but also a profound transformation of the entire industry collaboration model and value creation approach.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。