Jensen Huang's Open Model Manifesto: What NVIDIA's GTC Panel Reveals About AI's Real Architecture

An analysis of NVIDIA GTC 2025's open model ecosystem panel where Jensen Huang and AI leaders laid out the blueprint for how AI actually works - compound agents, harness engineering, and why open beats closed for everything that matters.

2026-03-28 By Sean Weldon

At NVIDIA GTC 2025, Jensen Huang gathered the leaders building AI's real infrastructure - not the chatbot layer that gets the headlines, but the model ecosystem, agent frameworks, and compute grid underneath. What emerged was less a panel discussion and more a manifesto for how AI actually works when you strip away the marketing.

The core message: AI is not the model. AI is the system. And that system is open.

The Stat That Reframes Everything

Jensen opened with a number that should end the "proprietary vs open" debate:

Open models in aggregate are already the second-largest model in the world. My sense is they will likely become the largest.

This isn't philosophical. It's measured. Across every industry, every application, every fine-tuned variant and domain adaptation, the aggregate compute devoted to open models already rivals the largest frontier labs. The trend line is clear.

But Jensen immediately killed the binary framing:

"Proprietary versus open is not a thing. It's proprietary AND open."

The real architecture isn't one or the other. It's a system where frontier closed models handle general reasoning while open models handle everything domain-specific, customizable, or trust-dependent. Every production AI deployment is already a hybrid.

Models Are Transistors, Not Products

The most clarifying analogy of the panel:

"Models is a technology just as transistors is a technology, not a product. It's a product to TSMC. To me, it's a technology that we integrate into other products."

This reframes the entire industry. Foundation model companies (OpenAI, Anthropic, Meta) are TSMC - they make the components. The value creation happens in what's built on top: the orchestration, the domain expertise, the agent systems that turn raw inference into work.

This leads directly to the panel's most important structural observation.

The Third Type of Company

The AI industry isn't two layers (model labs + app companies). It's three:

Foundation model companies - build the base technology
Specialized agent companies - use best-of-market APIs while building their own domain models
Application companies - deliver end-user products

That middle layer is where the action is. These companies build compound agents - systems that mix multiple models together, routing tasks to whichever model handles them best. Token efficiency from open models. Reasoning from closed models. Domain expertise from fine-tuned specialists.

The orchestration layer around these models has its own name: harness engineering. It encompasses sub-agent coordination, tool selection, context assembly, and the entire machinery that turns a model call into useful work.

"These sub-agents are musicians and the models are just instruments. The work that AI gets done for you is the symphony or the music that they play."

Three Inflection Points

The panel identified a clear progression in AI capability:

Inflection 1 - Generative AI: Without generation, there's no reasoning and no tool use. This was the prerequisite for everything.

Inflection 2 - Reasoning (O1): Models learned to think through problems. Chain of thought, extended inference, structured problem-solving. This made agents possible.

Inflection 3 - Agentic Systems: Applications now have the capability for industrial-scale deployment. Agents evolved from simple tool-calling to co-workers handling multi-hour or multi-day complex workloads.

The key insight about coding agents:

"Almost all work could be specified as code. That's the reason why it's so important."

Coding agents aren't just for software engineering. They're a general-purpose interface to everything a computer can do. Business processes, compliance rules, operational procedures - all of it can be codified, which means all of it can be delegated to agents via CLI interfaces.

The Two-of-Three Security Framework

The most immediately actionable idea from the panel was the enterprise governance model. Every agent has three potential capabilities:

Access to sensitive information
Code execution authority
External communication

The rule:

"You should allow someone, including an AI, two of those three things at one time, but not all at one time. Except unless it's the CEO."

This is elegant and practical:

Info + Code, no Comms: Agent can process sensitive data and run computations but can't send it anywhere
Info + Comms, no Code: Agent can read data and report on it but can't modify systems
Code + Comms, no Info: Agent can execute tasks and communicate but has no access to sensitive data

Any combination of all three without proper authorization creates an exfiltration risk. Most enterprises have no framework for this today. This one is worth adopting immediately.

Why Open Models Win for Everything That Matters

The panel laid out the case for open models not as ideology but as engineering requirements:

Control: Organizations own and modify the entire stack - models, orchestration, execution code. No API dependency, no surprise deprecations, no vendor lock-in.

Customization: Physical systems, domain-specific applications, and any context where the company has proprietary IP to embed requires a model you can modify. Closed models can't incorporate your proprietary knowledge at the weight level.

Trust: Mission-critical applications in healthcare, defense, and national security require introspection. You can't trust what you can't inspect.

"For mission-critical applications, you're going to have to find a way to trust them. And as far as I know, open models are one of the fastest ways to trust a system."

The metaphor that landed hardest:

"Closed models are like 800-year-old parents set in their ways. Open models enable new perspectives and specialization."

And the strongest argument for specialization:

"The shape of AI is going to reflect the shape of society. Society is specialized and it's going to continue to be specialized."

We don't send everyone to a general practitioner. We have cardiologists, neurologists, oncologists. AI should mirror that. Digital twins of specialists, not a single averaged intelligence.

The Bitter Lesson Holds

The panel repeatedly returned to a fundamental economic relationship:

"Revenue scales linearly with compute. The more you buy, the more you make."

This isn't a hope. It's observed. And it validates continued massive investment in compute infrastructure. Pre-training currently consumes 90% of compute, but this proportion will shrink dramatically as post-training techniques (RLHF, reasoning training, domain adaptation) dominate.

The historical proof point: AlphaGo was a 60-million parameter network that beat the world's best Go player. Not because it was large, but because reinforcement learning at scale enabled strategic compute allocation for breakthrough capability. Scale matters, but how you scale matters more.

The Infrastructure Warning

The most forward-looking concern from the panel: open infrastructure is as critical as open models, and it's consolidating too fast.

If only a handful of companies can afford the compute to train and serve models, openness at the model layer becomes meaningless. You can download the weights, but you can't run them.

The solution discussed: AI grid computing - a shared infrastructure model with secure base load capacity and the ability to spike, preventing hoarding and waste. An AI foundry model so companies don't need to be gigantic to participate.

The historical parallel is the industrial revolution. Electricity was transformative, but only because the grid was open. If one company owned all the power plants, the revolution would have stalled.

What This Means for Practitioners

If you're building with AI today, this panel lays out the playbook:

Stop thinking about models, start thinking about systems. Your competitive advantage is harness engineering - the orchestration, tools, and domain knowledge around the model, not the model itself.
Build compound agents. Mix models. Use closed models for reasoning-heavy tasks, open models for domain-specific work, small models for classification and routing. No single model wins everything.
Adopt the two-of-three security framework immediately. Before your agents have access to production data, decide which two capabilities each one gets.
Invest in open models for anything domain-specific or mission-critical. The customization and trust advantages compound over time. Closed model APIs are fine for prototyping; open models are required for production.
Watch the infrastructure layer. The model layer is democratizing. The compute layer isn't. Your long-term AI strategy depends on access to affordable, scalable inference.

"Computers are cool again. The runtime is GPUs. Token king."