Key Summary
Jia Yangqing, a leading figure in the AI field, left his company just one year after it was acquired by NVIDIA for $700 million along with his team of 20 people. Behind this sudden departure were internal conflicts regarding values (unfulfilled open-source commitments), issues with product implementation (core technologies not being resolved), and external challenges (AI tools that can automatically write code diminishing the value of NVIDIA’s platform). This event is not only a personal career decision but also reflects NVIDIA’s difficulties in expanding from a hardware powerhouse to a software company, as well as the systemic challenges faced by the AI infrastructure industry.
Why Did NVIDIA Spend $700 Million on Jia Yangqing’s Team?
NVIDIA’s ambitions extend beyond just selling GPUs. Over the past two years, it has become a key player in the AI landscape thanks to its GPUs (such as the H100). However, Jensen Huang wants to take things one step further: to build its own software platform that would compete with cloud providers like AWS and Alibaba Cloud, allowing developers to directly purchase, rent, and deploy models on NVIDIA’s platform rather than through these cloud services. This would enable NVIDIA to gain a share of the market from them.
Jia Yangqing is the perfect candidate to realize this vision:
- He is a key contributor to open-source AI frameworks: Caffe (the first industrial-grade deep learning framework), PyTorch (used by almost all mainstream AI models today), and ONNX (a cross-framework compatibility standard). He has built up decades of trust within the global developer community.
- He has expertise in managing GPU clusters, having led large-scale AI infrastructure projects at Alibaba.
NVIDIA acquired not just the company Lepton but also Jia Yangqing’s technical skills and influence in open-source development.
The Fundamental Conflict: The Breaking of Open-Source Beliefs
Jia Yangqing is a staunch supporter of open-source software. His entire career has been tied to open-source projects (Caffe, PyTorch, ONNX). During the acquisition, NVIDIA promised to make Lepton’s core software open-source by 2026, which was one of the key reasons he agreed to join.
However, Jensen Huang later went back on this commitment. NVIDIA’s history shows that it is not always sincere about its support for open-source: previous projects like NIMs were nominally open-source but actually closed-source, and FlashInfer changed from open-source to closed-source. For Jia Yangqing, this was not just a difference in product direction but a fundamental conflict of values—his beliefs were betrayed by commercial interests, leading to his departure.
Product Implementation: Surface-Level Efforts, Core Problems Unsolved
NVIDIA’s corporate culture caused the Lepton team to focus on superficial aspects rather than solving critical technical issues. For example, the team spent time on changing colors and optimizing login interfaces, while the key problem of managing multi-tenant GPU clusters (where multiple users share resources without interference) remained unresolved. Cloud providers criticized Lepton, saying, “Lepton has solved all the problems with GPU computing except for the most difficult one.”
AI developers are reluctant to pay for non-core software from NVIDIA; they prefer open-source tools like vLLMs and SLURM. Since Lepton’s product did not address these core issues, it failed to attract buyers.
External Challenges: AI Tools That Automate Code Writing
Initially, Lepton’s value lay in simplifying GPU management for developers by eliminating the need to use complex tools like Kubernetes or write deployment scripts and manage monitoring logs. However, new AI tools like Cursor and Claude Code have made such simplifications obsolete. For example, developers can now use natural language commands to ask AI to deploy models on a GPU cluster and set up monitoring automatically. AMD’s open-source Spur project, which is compatible with SLURM, allows developers to create their own workflows without relying on Lepton.
This is not just a problem for Lepton but a crisis for the entire AI infrastructure industry: if a product only simplifies superficial tasks, AI tools that can write code automatically will render it obsolete.
Industry Signals Behind This Split
This departure sends two important signals:
1. While GPUs can be monopolized, the AI infrastructure market is different. NVIDIA faces greater resistance in expanding into software compared to its success with GPUs. Developers prefer open-source tools over closed platforms from large companies.
2. There is a crisis in the value of middleware platforms. Many AI infrastructure companies that raised funding between 2022 and 2024 promised to lower the technical barriers for developers. But now that AI can handle these tasks automatically, these companies need to reevaluate their offerings: do they solve superficial complexities or address underlying scarcity issues (such as the scarcity of GPUs)?
In summary, Jia Yangqing’s departure is not an isolated event; it reflects the inevitable clash of values, product strategies, and technological trends as the AI industry transitions from a hardware-driven to a software-dominated landscape.