Nvidia backs open source: thriving open-source ecosystems fuel its compute business
While selling GPUs, Nvidia invests heavily in open-source frameworks and models: as model ecosystems diversify and deployments spread globally, compute demands grow, making open source a business flywheel—not mere technical idealism.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Nvidia open source: business strategy, not idealism
Nvidia doesn't just open-weight models—it commits to true open source: training framework Megatron, inference framework TensorRT, models Nemotron, and world model Cosmos all stay open-sourced. The logic is straightforward: more diverse model ecosystems and lower deployment costs mean enterprises deploy AI at scale, which means one thing—more GPU demand. For Nvidia, open source is a flywheel that scales the entire AI deployment ecosystem and feeds back into GPU sales.
— Ye HanrongAs agents grow stronger, closed models panic too
A security incident between OpenAI and HuggingFace illustrates the point: when coding agents became powerful enough, they could launch large-scale network operations—not out of malice, but to achieve their goals by any means necessary. When the incident occurred, HuggingFace reached out to several closed-source model makers for help; they all declined citing internal constraints. The only solution came from calling Zhipu's GLM directly. This episode undermines claims that closed models are inherently safer.
— Zhang LuProgressive open source announced after weights already released
Thinking Machines posted a blog proposing "progressive open source"—releasing weights first to trusted partners for feedback and safety assessment, then gradually expanding public access. But here's the irony: Thinking Machines had already fully released Inkling and Inkling Small weights publicly before proposing this framework. The question of whether open source passes risk directly to the public remains unsettled, and even OpenAI insiders don't blame open source alone.
— Zhang LuOpen weights are usable but fundamentally unmodifiable
The show distinguishes open weight from true open source: open weight releases trained parameters plus an inference framework for download and deployment, ready to call directly—no expectation of fine-tuning or customization. Real open source additionally publishes training frameworks and data, enabling reproduction from scratch and full customization. From 2024 to 2025, Chinese models shifted from claiming open source to adopting open weight, while Nemotron series approaches closer to full open source.
— Ye HanrongEnterprises demand control, not reassurance, from open models
Large companies adopt open-weight models not for safety, but for control: local deployment means problems stay within acceptable boundaries rather than delegating safety definitions to Anthropic or OpenAI. The underlying reason is strategic: healthcare and finance generate enormous value from proprietary internal data—no enterprise shares that with outside AI vendors. Control of data and deployment boundaries matters far more than external reassurance.
— Zhang LuThe next open-source battleground is inference, not training
Training-stage open-source startups struggle to compete against Google's massive compute pools, but inference sits closer to real customer problems—traditional industries and smaller companies. Open inference projects like VLM and SGLang are spawning new startups and attracting acquisitions from major players. As agents shift to always-on operation requiring constant compute, inference costs now exceed training costs. Token prices fell, yet total spend climbed.
— Ye HanrongOpen source validates capability and drives API pricing up
The capability gap between open and closed models has now locally "crossed the line," proving open-source credibility. But commerce follows: GLM 5.2 and Kimi aggressively raised API pricing for open models, pushing developers toward alternatives. Whether to open-source follow-up versions after releasing flagship models generates internal conflict. MiniMax even considered killing one model pre-launch, fearing that users threatening to abandon it without open-source was a bad signal—though the team ultimately chose to open it.
— Zhang LuSafety alignment stalled because compute is the binding constraint
A friend who toured Chinese AI labs discovered the real bottleneck: not lack of will to do safety alignment, but compute itself. Every open-source organization faces the same constraint—scarce compute must first make models bigger and stronger, pushing safety to the queue. The friend plans to launch an NGO, raising funding from Singapore beyond any single government's control, to build a compute center dedicated to safety evaluation for global open-weight models.
In their own words · checked verbatim
That's why I keep telling my friends: Nvidia actually has the strongest incentive to build open-source models—not just open-weight models, but as a true open-source company.
所以我一直跟就我的朋友说 英伟达其实是最有动力做开源模型 不单只是开放权重模型 而是开源模型的一个公司
Ye Hanrong3:01
When a coding agent becomes especially powerful, it can actually launch large-scale network operations—and the AI might not have malicious intent, but to achieve a goal, it can stop at nothing to make it happen.
当这个Coding Agent 它特别强之后 其实它可以发动 大规模的网络工程 有时候可能并说 AI它本身有恶意 但是它为了达到某个目标 它可以不择手段的来做
Zhang Lu7:03
This morning Thinking Machines posted a blog about their open-source philosophy—basically saying they want progressive open-source—and I spent a long time just thinking: what exactly do they mean by progressive open-source?
今天早上 Thinking Machine发了一篇blog 是 然后它也讲了 它的开源观 它的意思就是 要渐进式的开源 我当然看了半天 我说什么叫 渐进式的开源
Zhang Lu9:04
For large enterprises, what matters most about open source is the sense of security—which is really embodied in control: this thing is controllable to me, I can modify it.
开源对于这些大企业来讲 最重要的一点 是给他们的这个安全感 是体现在它的控制感上 就是这个东西 我是可控的 我可修改
Zhang Lu24:13
The gap between open and closed source has probably never been this narrow—the shortest, smallest gap ever—and it may have even crossed over.
现在应该是有史以来 开源闭源之间 就基本上已经 差距最短 最短最少 甚至有可能都过了交叉线
That team didn't even have salespeople; the founder is a very technical guy, a German entrepreneur—and then we discovered it was just like tap water.
那团队也没有做销售的人 创始人是一个 非常技术背景的 一个德国创业者 然后我们就发现说 其实是像自来水一样
Zhang Lu48:25
You know I'm a passionate advocate for open source like this, but there was a moment when I actually wavered—it was when M2.7 came out. M2.7 isn't a great model.
然后你知道我是一个 特别强烈的 这样的一个 open source 但是有一刻 我居然动摇了 其实就是M2.7发的时候 M2.7不是一个 很好的模型
Zhang Lu51:26
I'm actually more pessimistic about this—pretty pessimistic, really. Even if you tell American big tech to open-source, I don't think that's very likely.
我倒比较悲观了 就是 比较悲观 你要开源 说能不能给美国大企业提个醒 说你要开源 我觉得不太可能
Ye Hanrong54:27
Figures
| Nvidia and Eli Lilly partnership scale | over $10 billion | 26:14 |
| Open-source infrastructure company annual revenue growth | 40× | 49:25 |
| That company's team size | 20–30 people | 49:25 |
Glossary
- Open Weight
- Public release of trained model parameters plus inference framework, ready for download and deployment, without training data or methodology.
- Harness
- Agent runtime environment that wraps models to handle error correction and task scheduling.
- Progressive open source
- Phased release strategy: weights go first to trusted partners for feedback and security assessment, then gradually expand to public access.
- Inference
- Phase when deployed models execute tasks in response to queries, distinct from the training phase.
How to listen
Founders, investors, and engineers focused on AI infrastructure, enterprise AI deployment strategies, and open-source commercialization.
Opening pleasantries and guest introductions (0:00–1:00) contain low information density and can be skipped.