From "managing computing power" to "building systems" upward.

Article author and source: Lieyun Selection

Open-source expert Jia Yangqing starts a startup

The open-source expert Jia Yangqing has made new moves.

Today, Jia Yangqing officially announced his new startup, Intent Lab, where he is building an autonomous AI team called "fleet" to turn intentions into production software.

At the same time, he also showcased three early achievements: the fastest GLM 5.2 inference engine, a database engine generated from a single sentence, and a file system with formal verification.

Source: Jia Yangqing X account

With the release of this tweet, Jia Yangqing has taken on a new role.

His X profile bio has been updated to reflect his role as founder and CEO of Intent Lab, and NVIDIA has been added to his past employment history, fully confirming the month-long rumors of his split from NVIDIA.

This Chinese scientist, known in Silicon Valley as the most accomplished in the field of AI frameworks, left Alibaba, founded Lepton AI, sold it for $700 million to NVIDIA, and is now standing at a new starting point.

Jia Yangqing's second entrepreneurial venture had signs early on.

On June 29, semiconductor research firm SemiAnalysis posted a tweet stating that Jia Yangqing, founder and CEO of Lepton AI, had left the company just one year after its acquisition.

At the time, although this message had not been confirmed, it was enough to spark curiosity within the industry, especially among developers, and the tweet went viral.

Almost simultaneously with SemiAnalysis’s revelation, in early June 2026, GPU cloud provider Hyperbolic announced the hiring of Jia Yangqing as a company advisor.

Hyperbolic stated that Jia Yangqing will provide expert guidance in areas such as AI systems, cloud infrastructure, GPU marketplaces, multi-cloud orchestration, and improving GPU utilization—areas closely aligned with his previous work leading the DGX Lepton platform at NVIDIA.

Taken together, these two pieces of news have already set the stage for Jia Yangqing's departure.

And why he left is likely because Huang Renxun changed his mind, shattering Jia Yangqing's belief in open source.

Recalling 2023, Jia Yangqing left Alibaba in March and founded Lepton AI, a cloud-native AI platform, in Silicon Valley in April. The founding team included prominent figures such as Bai Junjie, co-founder of ONNX, and Li Xiang, founder of etcd.

At that time, Lepton had a clear positioning: a multi-cloud native AI inference computing platform, building a unified orchestration layer to help developers allocate GPU resources across major cloud providers, and lowering the barrier to deploying large models with its lightweight Tuna inference engine.

In short, an operations platform that bundles these complexities lowers the barrier for AI engineers to use and manage GPU clusters, allowing developers to focus on the models themselves.

Just two years after its founding, Lepton AI has been ranked in the second-tier "Gold" group of the SemiAnalysis ClusterMAX GPU cloud services rating system, alongside Microsoft Azure, Together AI, and Oracle.

This was followed in April 2025 by NVIDIA's acquisition of Lepton AI for $700 million.

This transaction was once regarded by the industry as a win-win.

Before the acquisition was completed, Jensen Huang stated at the GTC conference in March last year: NVIDIA is no longer just a chip company, nor just an AI company, but an algorithm and infrastructure company.

Jensen Huang’s logic is that NVIDIA cannot just be a GPU seller—it must become the cloud and software gateway of the AI era.

He hopes NVIDIA will enhance the software orchestration capabilities of DGX Cloud to build its own global compute marketplace, directly competing with public cloud providers and GPU cloud rivals like CoreWeave—and acquiring Lepton AI is one of the final pieces in Huang’s vision for an “AI factory.”

For Jia Yangqing's team, leveraging NVIDIA's hardware ecosystem could help bring the cross-cloud computing platform concept to life.

But soon, the ideal encountered the impact of reality.

Lepton AI's product philosophy, such as "native Python, out-of-the-box," is incompatible with NVIDIA's more closed product management system targeted at enterprise customers.

According to public reports, DGX Lepton, after its rebranding, had largely ceased external operations by mid-2025, with its performance falling short of expectations.

Yet a deeper, more irreconcilable contradiction lies in the open-source commitment.

NVIDIA pledged to open-source the Lepton core platform by 2026 at the time of its acquisition, but NVIDIA's core business model relies on GPU hardware sales.

Once the Lepton core scheduling code is opened, AMD and various domestic heterogeneous chip manufacturers can use this efficient tool to orchestrate computing power, effectively weakening their own hardware moat.

Ultimately, the open-source initiative was officially rejected.

This commitment was the foundational reason Ji Yangqing agreed to the acquisition, led his team to join NVIDIA as Vice President of Systems Software, and reported directly to Jensen Huang.

Keep in mind that Yangqing Jia’s entire career has been built on the open-source ethos of "everyone for me, and I for everyone."

He reshaped AI engineering infrastructure with open source, open-sourcing all career milestones—from Caffe and PyTorch to ONNX.

On one side are chip giants bound to maintain hardware barriers; on the other are believers in neutral, open infrastructure—these two sides are destined to coexist only temporarily.

Just a year after the "mutual pursuit," Jia Yangqing left again to start a new venture.

From the founding team perspective, Intent Lab remains the same team.

The co-founders include Bai Junjie and Li Xiang, longtime open-source collaborators from Lepton AI, but their current work is significantly different from what they did at Lepton AI.

Jia Yangqing is no longer building the next generation of computing cloud and inference scheduling platforms; instead, his new company is focused on autonomous AI software engineering.

The new product Fleet no longer aims to "help humans deploy AI models," but instead enables teams of AI agents to complete the entire process—from requirement understanding and architecture design to coding, testing, verification, and ongoing maintenance.

The core logic behind it is that AI infrastructure is undergoing a valuation reassessment.

Currently, agentic coding tools such as Cursor, Claude Code, and Codex are reshaping the foundational logic of software development.

"Vibe-coding" allows developers to describe requirements in natural language, while AI agents automatically generate, debug, and deploy complete infrastructure code.

For example, AMD's open-source Spur project is a compatible, open-source alternative to Slurm that can be directly invoked and configured by AI agents with nearly zero barrier to entry.

This also means that the complexity traditionally attempted to be encapsulated through productized AI infrastructure is being directly bypassed by agentic coding through code generation.

This is the objective context behind issues such as DGX Lepton failing to meet Jensen Huang’s expectations and the unfulfilled open-source commitments.

Choosing not to follow Lepton’s path reflects Jia Yangqing’s clarity as an entrepreneur; founding Intent Lab is simply aligning with the tide of the times.

Source: Jia Yangqing X account

Jia Yangqing wrote on social media about the core thinking behind his shift in focus: For the past several decades, we have been manually building distributed AI systems, one at a time. Now, he aims to build a system that can efficiently generate thousands of software systems in bulk.

In his view, current models write code very quickly, but there is still a significant gap between "code that runs" and "code that can be confidently deployed to production and maintained for years."

He does not view this gap as a limitation of the model, but rather as a missing layer.

This missing capability is a production system that autonomously transforms vague intentions into concrete designs, validates whether the results hold up, and continues evolving the system after it’s live—all on its own.

And this is precisely the level at which Intent Lab is working.

Source: Intentlab official website

Therefore, unlike code assistance tools such as Cursor and Claude Code, Fleet aims to deliver production-grade foundational systems that can run stably over the long term, filling this gap.

Currently, Jia Yangqing has already submitted his first three assignments.

First, Fleet redesigned the TRT-LLM inference engine to make it the fastest GLM 5.2 engine known to the team: on two Grace Blackwell nodes, output speed increased from 102 tokens/s to 647 tokens/s, a 6.3-fold improvement.

This is the result of Fleet’s end-to-end construction and implementation of optimizations.

Source: Intentlab official website

Second, starting from a single prompt, a database engine was created.

Fleet started from scratch with no existing documentation or code, taking on every role of a strong team: architecture design, coding, testing, review, and quality assurance, operating autonomously until all 6 million SQLite compatibility tests passed.

Third, we built a file system for a world where infrastructure is used by agents rather than humans.

In this example, Fleet used formal verification to identify and fix a transient corrupted state, catching a bug written by a coding agent.

These three tasks may appear unrelated, but they clearly point to what Jia Yangqing truly wants to emphasize: the capabilities of the underlying system, Fleet.

AI doesn't just handle programming tasks—it can take full responsibility for the final system, just like a complete software engineering team.

If successful, Jia Yangqing believes the economic model of computer engineering will be permanently changed.

For 50 years, the wise approach was to develop one piece of software and sell as much of it as possible to users with varying needs—even if those needs differed. Now, this can change. The world will see more software, more highly customized software.

In other words, once AI can reliably generate and operate production-grade systems, businesses or individuals will be able to simply describe their processes and constraints, and the AI will generate software specifically tailored to their needs.

Currently, Intent Lab is partnering externally to focus specifically on systems that are "too labor-intensive and have remained on the backlog due to lack of resources."

Looking back at Jia Yangqing's journey, the main thread has always been clear.

From creating Caffe and advancing the ONNX open standard to working at Alibaba Cloud and founding Lepton AI, he has consistently been doing different versions of the same thing: making computing power easier for others to access.

As AI becomes the builder of infrastructure, its next step moves from "managing computing power" to "building systems."