Open Models: Hype, Reality, and What’s Next

The panel brought together an almost absurdly high-caliber group of founders and researchers. Jensen Huang, Founder and CEO of NVIDIA, was joined by Harrison Chase, Co-Founder and CEO of LangChain; Michael Truell, CEO and Co-Founder of Cursor; Misha Laskin, Co-Founder and CEO of Reflection AI; Aravind Srinivas, CEO and Co-Founder of Perplexity; Mira Murati, Founder and CEO of Thinking Machines Lab; Arthur Mensch, Co-Founder and CEO of Mistral; Daniel Nadler, CEO of OpenEvidence; Hanna Hajishirzi, Professor and Senior Director of NLP at AI2; Anjney Midha, Founder of AMP PBC; and Robin Rombach, Co-Founder and CEO of Black Forest Labs. Between them, they represented every layer of the emerging AI stack — from foundational research and model development to orchestration frameworks, developer tooling, and domain-specific applications. What follows is my attempt to share the most important ideas from that discussion and why they matter.

The Third Category

For the past few years, the conventional wisdom has sorted the AI software landscape into two clean buckets: foundation model companies that sell capability via API, and application companies that build products on top of those APIs. It is a tidy mental model and, as this panel made clear, it is already becoming obsolete.

A third category is emerging, and it may ultimately be the most interesting one: companies that simultaneously consume the best API-level foundation models and build their own specialized models, weaving both into a single product experience. Cursor is perhaps the clearest current example. Rather than simply wrapping a frontier model, Truell described building proprietary models tuned specifically for the mechanics of code editing, speed, context precision, format adherence, while still routing certain tasks to frontier APIs for their superior general reasoning. The insight is not that one approach beats the other; it is that the most competitive products are beginning to treat model selection as a dynamic routing decision, not a static architectural commitment.

Srinivas extended this logic with striking clarity when describing Perplexity Computer. The framing he offered deserves to be quoted directly in spirit if not in word: AI is not the model. AI is the system. Perplexity Computer, as he described it, is an orchestration layer that coordinates every AI capability, coding, writing, multimodal generation, across multiple models and multiple clouds. Sub-agents, in this architecture, are like musicians. Models are the instruments. The resulting work is the symphony. No single instrument plays everything; the conductor's job is knowing when to bring each one in.

Harness Engineering: The Discipline Nobody Named (Until Now)

One of the most practically important ideas to emerge from the panel is what Chase called “harness engineering”, and the naming matters because it gives a discipline a shape it lacked before. The harness is everything that surrounds the model: the connectors, the tools, the memory compaction logic, the routing rules, the schema design that makes CLI-based orchestration possible. It is not the model. It is not the application. It is the nervous system that makes a collection of models behave like a coherent system.

This distinction has deep implications for developers and technical architects. Closed-model labs have, for some time, been quietly building very sophisticated harnesses, the infrastructure around their models that largely accounts for why their products feel qualitatively different from a raw API call. As Chase noted, the developer community is only now reckoning seriously with the need to build comparable harness infrastructure for their own environments. For enterprise practitioners especially, this is a call to action: the competitive advantage of the next few years will be built not just in model selection, but in the quality and sophistication of the harness built around it.

Three Transitions, and Why the Third Is Different

Several panelists converged on a shared periodization of the AI era. The first transition was generative AI, the ability to produce coherent, useful text by predicting tokens, which turned out to unlock emergent reasoning in ways nobody fully anticipated. The second was the rise of reasoning-oriented models: systems that could slow down, decompose problems, and work through multi-step logic chains before committing to an answer. The third and the one that dominated much of the panel's discussion is the agentic transition: models that don't just answer questions but take actions, hold state across time, coordinate with other models, and operate inside enterprise environments with real consequences.

Laskin offered an analogy that crystallized why this third transition is categorically different from the first two. When AlphaGo was developed, a system of roughly 60 million parameters, it demonstrated something that had not been true of AI systems before: it never stopped learning. Its only constraint was the economics of compute. Now that reinforcement learning works in language as it does in games, that logic applies to the full breadth of human knowledge. The question of whether we will spend ten billion or a hundred billion dollars to cure a disease is no longer a philosophical one. It is becoming an economic one. Frontier science, Laskin argued, is entering an era where its limits are defined not by what is knowable but by what is fundable.

The Agent as Computer

Perhaps the most structurally important idea in the entire session was the reconceptualization of what an agentic system actually is. The panel, in different ways, kept returning to the same conclusion: the agent is not a feature. It is a computer.

OpenClaw, an open-source agentic system discussed by multiple panelists, was cited as the first complete open-source agentic computer. Within weeks of its release it became the most popular open-source project in history, which is itself a signal about the latent demand for this type of system. What makes it a computer rather than a chatbot is a specific set of capabilities: persistent memory, file access, task scheduling, input/output management, networking, and a growing library of skills. These are the primitives of computation. What changes when they are orchestrated by a language model is that the programming interface shifts from explicit code to natural language intent.

Murati drew out an important architectural nuance here. Universal agents, systems capable of doing anything, are real, but they are slow and expensive. The more practical and near-term architecture for enterprise is what she described as systems of prescriptive sub-agents: specialized agents that operate within defined parameters, run sequentially or in parallel, and are assembled into workflows calibrated to specific business functions. The coding agent, she argued, is a particularly powerful template for this because coding itself encodes business process. An agent that can write and execute code can, in principle, do almost anything a knowledge worker does, and the harness built to orchestrate coding agents can be ported with relatively modest effort into legal, sales, finance, and operations workflows.

Enterprise Governance: The Hard Part Nobody Wants to Talk About

For all the excitement about what agents can do, the panel spent serious time on what they must not do — and on the governance infrastructure required to enforce those constraints at industrial scale. Nadler, drawing on his experience at OpenEvidence where agents operate in healthcare environments, framed the core tension with precision: agents can access sensitive data, execute code, or communicate externally. Enterprise security best practice, he argued, is that an agent may do any two of those things simultaneously, but never all three at once. The exception, offered with a dry humor that landed well, is the CEO.

This is not a trivial constraint. It represents a fundamental design principle for enterprise agentic architecture: the boundary of an agent's capability must be proportional to the sensitivity of its operational context. Building systems that enforce this, with actual governance primitives, hosting controls, and audit trails, is the organizational bottleneck that will determine which enterprises succeed in the agentic transition and which ones will spend years recovering from an incident. The panel was clear that this infrastructure is not yet commoditized. Building it is among the most important near-term challenges in applied AI.

The Open Model Argument, Made Properly

The open versus closed model debate has generated more heat than light in most public forums. This panel managed to advance it substantively. Mensch and Hajishirzi were the most direct advocates for open models, but notably, even the builders of closed systems acknowledged that open models will play an essential role in the agentic ecosystem.

Hajishirzi offered the clearest framework for why openness matters structurally. Knowledge infrastructure, she argued, has historically yearned toward openness, the printing press, scientific journals, open encryption standards. Each time knowledge was locked behind access controls, progress slowed for everyone, including the entities doing the locking. The same dynamic is playing out in AI. When post-training APIs were released openly, research accelerated dramatically. The implication is that even if proprietary frontier models remain the best general reasoners, the specialists, the models fine-tuned on physical-world data, domain-specific corpora, or enterprise IP, will deliver the majority of real-world value. And specialists, by their nature, require open, customizable foundations.

Mensch added a dimension that carries particular weight in enterprise conversations: trust. Open models allow introspection. Enterprises can examine weights, understand training processes, impose guardrails, and audit behavior in ways that closed APIs simply do not permit. For mission-critical domains, healthcare, legal, financial, this introspectability is not a preference. It is a requirement. Even companies whose core product is a closed model, Mensch observed, will increasingly use open models inside their agentic systems precisely because they need components they can fully control and verify.

Midha and Rombach brought the infrastructure dimension into focus. Beyond model openness, the panel argued for what might be called open infrastructure: an AI grid with secure baseload and elastic compute capacity that allows small organizations to access the compute required for frontier work. The historical warning embedded in this argument is sharp: societies that hoarded compute infrastructure in earlier technology transitions created structural inequalities that took generations to correct. Getting the infrastructure layer right in AI, ensuring that access to compute is not purely a function of which companies can afford a data center, is as important as getting the models right.

What This Means Going Forward

Coding was the first domain where this became undeniable, where AI moved from a curiosity to a measurable accelerant of business output. The panel was unanimous that this is no longer a coding story. The patterns established by agentic coding systems, the harness design, the orchestration logic, the governance primitives, are migrating into every domain where work can be expressed as a sequence of computable steps. Healthcare authorization workflows. Legal document review. Financial modeling. Sales pipeline management. The list is not theoretical. It is already being built.

What the GTC panel illuminated, above all else, is that we are not in an era of AI experimentation anymore. We are in the early innings of an era of AI deployment, and the competitive dynamics of that era are fundamentally different from what came before. The winners will not simply be the organizations that chose the best model. They will be the organizations that built the best system: the harness, the governance, the orchestration, the specialized agents layered on top of general ones, the compute infrastructure that can scale without sacrificing agility.

The open model is not a consolation prize for organizations that cannot afford a frontier API. It is an architectural necessity for any system that needs to be trusted, inspected, customized, and controlled. That argument, made by some of the most credible builders in the industry, in one of the most important rooms in technology this year, deserves to be taken seriously, by anyone building, deploying, or depending on AI systems that actually matter.