The open vs closed AI debate is no longer a theoretical skirmish among researchers: it is a live business decision being made inside startups today, one that can determine cost structure, product defensibility, and fundraising narrative before a company has found its first hundred customers.
That decision is the subject of a session titled ‘The Open vs. Closed AI Debate Is Just Getting Started,’ coming to the Builders Stage at TechCrunch Disrupt 2026, scheduled for 13–15 October in San Francisco. Nvidia’s Nader Khalil, Director of Developer Tech, and Sydney Sykes, Global Head of VC Partnerships, will lead the discussion.
The Open vs Closed AI Decision Is a Business Call, Not a Philosophy
The gap between open and proprietary models has narrowed fast. At ICML 2026, 145 papers cited Nvidia’s Nemotron open models and datasets. Nvidia itself had 74 papers accepted at the conference, and approximately 2,000 accepted papers cited Nvidia GPUs across research areas including robotics, autonomous vehicles, and biomedical science.
Proprietary frontier labs have not stood still either. The result is a market where the question is no longer whether open models can be useful, but where each approach makes commercial sense for a specific company at a specific stage.
Nvidia’s own position on the question is deliberately hybrid. At GTC earlier this year, chief executive Jensen Huang argued that the future is not proprietary versus open, but proprietary and open together. That framing is sensible as a statement of direction. It becomes harder to act on when you are a founder deciding which infrastructure to build your product around.
If two models deliver comparable results, does the cheaper one win? What if one grants tighter control over customer data? Does owning more of the stack create genuine defensibility, or does it create infrastructure you now have to maintain indefinitely? And if the leading model changes every few months, how tightly should a product be coupled to any single one of them? These are the questions Khalil and Sykes will address on the Builders Stage.
What Nemotron 3 Super Reveals About the Hybrid Reality
Nvidia’s own model releases illustrate where the open vs closed AI debate is heading in practice. Nemotron 3 Super, released on 11 March 2026 under the NVIDIA Nemotron Open Model License, is a 120-billion-parameter open model built for agentic workloads. Companies are already pairing it with proprietary models rather than choosing one or the other outright.
The architecture behind it is worth understanding, because it shows how efficiency trade-offs work at scale. According to Nvidia Research, Nemotron 3 Super uses a hybrid Mamba-Transformer Mixture-of-Experts architecture. Mamba layers deliver 4x higher memory and compute efficiency relative to standard transformer layers; transformer layers handle advanced reasoning; and a Latent MoE technique activates four expert specialists for the cost of one at inference, producing 5x higher throughput for agentic workloads.
A Multi-Token Prediction capability allows the model to predict multiple future words simultaneously, yielding 3x faster inference, while a context window of up to 1 million tokens supports extended reasoning tasks. On the RULER long-context benchmark at 1 million token context length, Nemotron 3 Super outperforms GPT-OSS-120B and Qwen3.5-122B. In an 8k token input and 64k token output setting, Nvidia’s NIM model card records up to 2.2x higher inference throughput than GPT-OSS-120B and up to 7.5x higher throughput than Qwen3.5-122B. Nvidia is releasing pre-trained, post-trained, and quantised checkpoints alongside the training datasets.
For founders evaluating open models, those numbers matter in a concrete way: they describe the cost of running inference at scale, which feeds directly into gross margin calculations.
Two Builders, Two Angles on the Same Stack
Khalil brings infrastructure credibility to the session. Before joining Nvidia as Director of Developer Tech, he co-founded Brev.dev, a San Francisco-based platform that streamlined the development, training, and deployment of AI models within CPU- and GPU-based cloud instances. IT Pro reported that Nvidia acquired Brev.dev in July 2024 with the aim of simplifying access to GPU resources across multiple cloud providers.
The integration has since deepened. Nvidia’s developer documentation describes how DGX Spark users can register their device on Nvidia Brev to extend development capabilities beyond the local device, effectively linking local AI work to broader cloud infrastructure.
Sykes arrives from the other direction. As Nvidia’s Global Head of VC Partnerships, her perspective covers what venture investors look for when assessing AI companies, including where differentiation actually sits when competitors can access the same proprietary API.
That framing gets to the deeper question behind the open vs closed AI choice: your model is not your moat until it is, and the conditions under which it becomes one depend on proprietary data, workflow depth, distribution, and whether you can maintain an advantage as the underlying models keep improving. Neither an open nor a closed approach delivers that automatically.
TechCrunch Disrupt 2026 runs 13–15 October in San Francisco. Early registration saves up to $200 before prices rise on 25 September.
