In the spring of 2025, Nvidia wrote a check for roughly US$700 million to buy a roughly 20-person startup founded by a man who, more than a decade earlier, had written a deep-learning framework during his PhD at Berkeley. The framework was called Caffe, and almost no one outside AI research remembers the name. Yet most of the models you use today were trained on software it helped invent.
Yangqing Jia is that man. He is the least famous of the three figures profiled here, and arguably the one whose work touches your life most directly — not through a product with his name on it, but through the plumbing every other AI product depends on.
The framework almost nobody remembers
Jia grew up in Shaoxing, Zhejiang, earned bachelor's and master's degrees at Tsinghua University in automation, then pursued a PhD in computer science at the University of California, Berkeley. Around 2013, as a graduate student, he released Caffe — short for "Convolutional Architecture for Fast Feature Embedding." It was the first deep-learning framework adopted at scale inside real companies, used by Microsoft, Yahoo, Nvidia, and Adobe.
Its quiet innovation was a "model-as-configuration" design: you described a network in a text file, and the engine did the rest. That made deep learning reproducible and shareable in a way earlier tooling was not. The industry saying that later circulated — "studying physics means you can't avoid Newton; doing deep learning means you can't avoid Jia" — overstates it, but points at a real fact: Caffe was the on-ramp that taught a generation of engineers to train neural networks.
From Caffe to the tools everyone uses
Jia did not stop at one framework. He joined Google Brain in 2013, where he became a core author of TensorFlow and worked on GoogLeNet, the network that, in 2014, beat human-level accuracy on the ImageNet image-classification benchmark. In 2016 he moved to Facebook (now Meta) as an AI architecture director.
There his impact compounded. He shipped Caffe2 in 2017, co-created the ONNX open model format with Microsoft, and helped drive the 2018 merger of Caffe2 and PyTorch into PyTorch 1.0. Today, PyTorch is the de facto standard on which the majority of the world's AI research — and an increasing share of production systems — is built. Jia's fingerprint is on three of the four foundational toolchains (Caffe, ONNX, PyTorch) that moved deep learning out of the lab.
A detour through China's cloud
In March 2019, Jia returned to China to join Alibaba as a vice president, leading its big-data computing platform and later the Alibaba Cloud Intelligence Computing Platform. He also steered ModelScope (魔搭), an open-source model community designed to give Chinese developers a home-grown hub for sharing large models (大模型) — a direct, if quieter, contribution to China's AI infrastructure stack.
That experience mattered for what came next. Running GPU clusters at production scale taught him the operational pain of AI infrastructure: not the math, but the plumbing — scheduling, deployment, and cost.
Lepton, and the Nvidia chapter
Jia left Alibaba in March 2023 and, with Junjie Bai, a fellow ex-Meta engineer, founded Lepton AI in Silicon Valley. The idea was narrow and technical: a Python-native GPU cloud where a few lines of code deploy a model onto accelerators, abstracting away the operational mess. The team was tiny — about 20 people — and raised a US$11 million seed in May 2023.
In April 2025, Nvidia acquired Lepton AI for a reported ~US$700 million, among the larger tuck-in deals in Nvidia's push beyond chips into cloud software. Jia joined Nvidia as a vice president of systems software and led the resulting product, DGX Cloud Lepton — a marketplace connecting developers to GPU capacity across cloud providers. Combined with its ~US$1.2 billion purchase of Run:ai, Nvidia had by then spent close to US$2 billion assembling a GPU-cloud software layer.
What the departure really means
In June 2026, the semiconductor research firm SemiAnalysis reported that Jia had left Nvidia, roughly a year after the acquisition. Reporting indicated DGX Cloud Lepton underperformed expectations and that a promised 2026 open-sourcing of Lepton's core was ultimately not delivered — a tension between his lifelong open-source stance and Nvidia's hardware-centric business model. Around the same time, Jia took an advisory role at the decentralized GPU-compute firm Hyperbolic Labs and founded a new startup, Intent Lab, focused on multi-agent software engineering — what he has framed as moving "from an agent as a tool to an agent as a team."
The through-line is consistent: Jia keeps returning to the layer everyone else treats as solved. Frameworks first, then model formats, then GPU clouds, then agent orchestration. Each time, he bets that the boring infrastructure decides who can actually ship AI.
Honest limitations
This profile draws on Baidu Baike, SEC-style IPO and acquisition reporting, and coverage from TechCrunch, Bloomberg, Reuters, Forbes, and SemiAnalysis (as relayed by Chinese outlets). I have not interviewed Jia, and I have excluded his private life and any political stance; neither is relevant to his engineering record.
Several points deserve skepticism. The ~US$700 million Lepton price was never officially disclosed by Nvidia; it is reported as "high hundreds of millions" by multiple outlets, and I present it as a reported figure, not a confirmed one. The account of Jia's Nvidia departure and the reasons behind it rest on a single research firm's reporting and subsequent media summaries; specifics about internal performance are not independently verified here. Finally, "most models run on PyTorch" is a strong industry consensus, not a precise measurement, and varies by sector.
What readers can do now
- If you train or deploy models, learn PyTorch and ONNX directly. They are Jia's most durable legacy and the default lingua franca of modern AI — understanding them is understanding the substrate.
- When evaluating AI infrastructure claims, separate the chip from the software layer. Jia's career shows the scheduler and deployment tooling, not just the GPU, determine real-world cost and speed.
- Watch Intent Lab and the multi-agent "team" idea. If a startup (创业公司) can turn AI assistants from solo tools into coordinated teams that ship working software, that is the next infrastructure layer worth tracking.
