NVIDIA CEO Jensen Huang: Open source models will become the largest models in the world, and AI agents will reshape all industries.

Written by: Techub News

Introduction

At a special fireside chat during the NVIDIA GTC 2026 conference, founder and CEO Jensen Huang stepped down from the role of keynote speaker to become a host, engaging in a high-bandwidth, unscripted deep discussion with the founders and CEOs of nine cutting-edge AI companies, including LangChain, Perplexity, Cursor, Reflection, Mistral, Open Evidence, AI2, Black Forest Labs, and Amp. This conversation is significant because it transcends conventional discussions surrounding major closed-source model companies like OpenAI and Anthropic, shining a spotlight on the burgeoning "AI ecosystem"—the "creators" who build tools, platforms, vertical applications, and open-source models. Huang set the tone from the start: the future of AI is not a "binary" choice, but a prosperous ecology of "open-source and closed-source coexisting," where aggregated open-source models will ultimately surpass any single closed-source model in scale.

Summary

The rise of open-source models: Huang pointed out that aggregated open-source models are already the second-largest models in the world and will eventually become the largest. This does not replace closed-source models but forms a diverse, complementary foundation for AI technology.
AI as a "system," not a single model: True AI capabilities come from a "system" or "computer" composed of multiple models, tools, file systems, and connectors. The models themselves are merely "instruments" in this orchestra.
The era of agents: Marked by phenomenal open-source projects like OpenClaw, AI is moving from "answering questions" to "taking action." Agents capable of executing complex, multi-step tasks will become "digital colleagues" for individuals and enterprises.
Revolution in vertical industries: Agent systems will trigger transformations across nearly all industries, including healthcare (e.g., handling insurance authorizations), law, logistics, and engineering, with their core value being the handling of repeatable, multi-step deterministic workflows.
The value of openness: control, customization, and trust: For enterprise-level and mission-critical applications, open-source models offer unmatched control, deep customization capabilities, and auditability of system behavior—these are the foundations for building trust.

Beyond the Model Competition: AI is a "System"

At the beginning of the conversation, Jensen Huang attempted to correct a common misconception: AI is not equivalent to large language models. He emphasized that AI is a "system" made up of multiple models and a vast array of other technologies. Creating solutions for different domains requires a massive amount of peripheral technology. Therefore, today's focus is not on discussing the opposition of "open-source vs. closed-source"—which he regards as a false proposition—but on celebrating an ecology where both coexist and thrive.

Arvind Srinivasan, CEO of Perplexity, illustrated this point with a vivid analogy: "Models are just instruments, sub-agents are the musicians, and the work AI does for you is the symphony they perform." The concept of "Perplexity Computer" he proposed aims to be an "operating system" that orchestrates all AI capabilities (coding, writing, multimodal content generation). In this system, users simply delegate tasks, and the system decides which model to call and what tools to use, thus freeing them from vendor lock-in.

Harrison Chase, CEO of LangChain, added that there is emerging a "third type of company" that occupies the space between foundational model companies and pure application companies. They leverage the best API models available on the market while also developing their own models, combining both to create optimal products for specific verticals. He predicts that in the next year or two, a new type of "composite agent" will emerge capable of handling complex tasks that can take hours or even days by distributing workloads among models with different specialties, becoming smarter than any single model.

Misha Laskin, CEO of Reflection, introduced the concept of "Harness Engineering," which encompasses everything built around models—how to connect tools, when to compress, which sub-agents to use, etc. He noted that even closed-source labs depend heavily on meticulously designed "harnesses" for their success.

Open-source Models: Not Just "Catch-up," But a Foundation for Innovation

A common misconception about open-source models is that they will always lag behind cutting-edge closed-source models by several months. Both Mistral CEO Arthur Mensch and AI2 research head Hanna Hajishirzi strongly opposed this view.

Arthur Mensch pointed out that there is no fundamental difference between open-source and closed-source models. AI models serve as the foundational knowledge infrastructure, which inherently "craves openness," just as books transitioned from closed to printed formats and science evolved from alchemy to academic journals. He predicts that in the coming years, an ecosystem of open-source models will emerge that is on par with strong closed-source models in capabilities.

Hanna Hajishirzi emphasized the value of openness from the perspective of research acceleration. AI is advancing at an incredibly fast, exponential rate, with an enormous amount of knowledge to learn that cannot be solely accomplished by large labs. Open models, infrastructure, data, and research outputs allow many talented researchers worldwide to contribute, advancing AI science from diverse angles—this is a very "positive-sum game." Her team has already fostered much new research by utilizing open models throughout their entire development cycle (including data, weights, and infrastructure), such as proving that hybrid models are theoretically and empirically more efficient than pure transformer architectures.

Huang concluded that pre-training (gaining foundational knowledge and generalization capabilities) is just the beginning; the bulk of future computational costs will be in post-training (learning skills). While closed-source models may be the best "generalists," the vast majority of value comes from "specialists," and open-source models are the fertile ground for cultivating specialists in various fields.

Agent Revolution: From OpenClaw to Industrial Redefinition

The climax of the conversation focused on OpenClaw, which has recently ignited excitement in the community. Huang referred to it as the "perfect embodiment of modern computers," being the first complete, open-source agent system. It possesses capabilities such as working memory, file system access, task scheduling, and I/O interaction, effectively acting as a "new type of computer."

Participants analyzed the significance of OpenClaw and the future of agents from various perspectives:

Technological turning point: Misha Laskin believes that OpenClaw represents a matching of model capabilities with "harness" calibration, providing the "body" needed for the "brain" to operate within computers. Harrison Chase pointed out breakthroughs in models for code generation and their ability to execute command-line operations (CLI) enable previously coding-trained models to handle nearly all knowledge work.
Industrial application blueprint: Daniel Nadler, CEO of Open Evidence, depicted revolutionary scenarios for agents in the healthcare sector: while doctors sleep, agents could automatically handle insurance denial appeals, retrieving necessary information from patient records to advocate for critical treatment for patients. He emphasized that this multi-step, repeatable, template-based workflow paradigm exists in every industry in the U.S., including law and healthcare.
Enterprise-level challenges: Arthur Mensch calmly pointed out that expanding agents from personal tools to enterprise organizations faces complex data, governance, and compliance bottlenecks. Enterprises need primitives that can manage everything within a unified control plane to ensure security and observability. Huang humorously interjected, using "CEO privilege" as an example to illustrate safety principles: an agent should typically only be allowed to perform two out of the three actions of "accessing sensitive information," "executing code," or "communicating externally," unless it is the CEO.
New frontiers: visual and physical worlds: Robin Rombach, CEO of Black Forest Labs, reminded everyone not to focus solely on code. Visual intelligence (understanding, simulating, and generating visual content) represents a critical new frontier in interacting with the physical world and robotics. Agents need to learn how to collaborate with these visual models to unlock entirely new fields in manufacturing and content creation.

The Core Values of Openness: Control, Customization, and Trust

When discussing the importance of open-source models for their respective industries, "control," "customization," and "trust" repeatedly emerged as key themes.

Arthur Mensch outlined two core values of open-source models for enterprise software: 1. Control and resilience: When agents are at the execution layer, enterprises need full control over their deployment, having a "shutdown button" to avoid dependence on external APIs that may be turned off or malfunction. 2. Deep customization: In fields involving the physical world, proprietary IP, or specific engineering knowledge, enterprises need to inject proprietary data into models. Open-source models allow for such deep customization, enabling the construction of dedicated agents that understand the physical world.

Daniel Nadler likened closed-source large models to "800-year-old parents": they are very smart in their reinforced directions but difficult to change their worldview. The shape of society is specialized, and future AI should reflect this necessity, needing digital "specialists." Open-source models are the necessary foundation for cultivating these "specialists."

Amp CEO Anjney Midha elevated the discussion to the level of "trust." As AI is applied in critical missions such as healthcare and national defense, the need for extremely high reliability demands great trust. A closed-source system that cannot introspect, self-host, and relies on third parties is "delegating trust." In contrast, open systems are more likely to establish trust through scrutiny and risk management, which is one of their greatest advantages in entering critical fields. He also called for "open infrastructure" (such as AI computing grids) to avoid power hoarding and infrastructure monopolization, allowing open-source models to continue developing at the forefront.

Conclusion: A New Era of Computing

At the end of the conversation, Huang summarized the multiple turning points currently occurring: the arrival of the era of agent systems, and the integration of powerful models with agent systems giving rise to highly valuable enterprise solution platforms. He predicted that this year we will begin to see real commercial returns on investment (ROI) take off, starting with coding but rapidly spreading to all industries.

Arvind Srinivasan poetically concluded the entire discussion: “Computers have become cool again. We are building a completely new computer on GPU runtime.” This may very well be the future that Huang and all the CEOs present are collaboratively crafting: a next-generation computing platform that fundamentally reshapes all work and lifestyles, grounded in both open-source and closed-source models, with agents as the interface and powerful computing as the engine.

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。