A deep-learning framework that fit on a graduate student's laptop ended up reshaping how an entire industry serves models. The person behind it, Jia Yangqing (贾扬清), is one of the few engineers who has sat at the center of both China's cloud-AI build-out and the global chip race.
His story is less about a single product than about a recurring instinct: turn research infrastructure into something thousands of teams can actually use.
From Tsinghua labs to Caffe
Jia earned his bachelor's and master's degrees in Control Science and Engineering at Tsinghua University, then a PhD in Electrical Engineering and Computer Science at UC Berkeley. During his doctoral work he wrote Caffe, an open-source deep-learning framework that became one of the first widely adopted training tools — used inside companies such as Microsoft, Yahoo, NVIDIA and Adobe.
Caffe mattered because it made convolutional networks approachable. Before it, running a serious vision model usually meant wiring together research code. Caffe gave practitioners a clean, fast, documented toolkit.
The Silicon Valley years
After Berkeley, Jia moved through the labs that defined modern AI infrastructure:
- Google Brain (2014–2016): he contributed to TensorFlow and co-created GoogLeNet (Inception).
- Facebook / Meta (2016–2019): he created Caffe2, co-created the ONNX interoperability standard, and co-led PyTorch 1.0, the merger of research and production tooling that many teams still run on.
Those projects shared a theme: stop every company from reinventing the wheel, and let models move between frameworks.
Running Alibaba Cloud's AI platform
In 2019 Jia returned to China as an Alibaba Group vice president. He led Alibaba Cloud's Computing Platform and the System AI Lab at the Damo Academy (达摩院), scaling the data-analytics and AI products that power much of Alibaba's internal and external cloud business.
That role is the part most relevant to neuroai.site readers: it put a Chinese-born researcher in charge of one of the country's largest AI training-and-serving stacks. The habits he formed there — huge usage volumes, low inference cost, platform thinking — would resurface at his next company.
Lepton AI: GPU cloud as a product
Jia left Alibaba in 2023 and founded Lepton AI with Junjie Bai, another former Meta and Alibaba engineer. The startup raised about US$11 million in seed funding from CRV and Fusion Fund.
Lepton's idea was narrow and practical: lease NVIDIA GPU servers from cloud providers, pool them, and wrap the pool in tooling that lets developers build, deploy and optimize large and multimodal models without managing the hardware themselves. It was infrastructure arbitrage with a developer-friendly interface — a control point between models and GPUs.
The NVIDIA acquisition
In April 2025, NVIDIA completed its acquisition of Lepton AI. The price was reported in the several-hundred-million-US-dollar range, but the terms were never officially disclosed. Jia and the Lepton team joined NVIDIA.
NVIDIA then relaunched the product in June 2025 as DGX Cloud Lepton, a marketplace that routes developers to GPU capacity across participating neoclouds — with NVIDIA sitting in the middle as the management layer rather than the landlord. For NVIDIA, the deal bought a working demand-side relationship with the developers choosing where to run.
Why this matters for China's AI stack
Jia's arc is a clean illustration of how China's AI talent flows through global infrastructure:
- Tsinghua supplied the fundamentals.
- Berkeley and the U.S. labs supplied the open frameworks (Caffe, PyTorch, ONNX).
- Alibaba Cloud supplied the proof that platform-scale serving works at Chinese volumes.
- Lepton supplied the MaaS (model-as-a-service) instinct that NVIDIA wanted.
None of it is a "China vs. U.S." story. It is one engineer repeatedly solving the same problem — make models cheap and easy to run — on whatever stage he happened to be on.
The pattern worth copying
The through-line in Jia's career is not loyalty to any flag — it is a repeated bet on open interfaces. Caffe lowered the barrier to train; ONNX let models travel; Lepton pooled GPUs so developers stopped caring whose silicon ran them. Each step removed a tax on the next person building. That instinct, applied to the embodied-AI and BCI workloads this site covers, is exactly what lets a lab in one country run on hardware made in another. The lesson for builders is unglamorous: invest in the interchange layer, because the applications always outrun the infrastructure.
Honest limitations
This profile relies on Jia's own public pages (UC Berkeley EECS, his professional profiles), independent reporting (seedtable, citing The Information, TechNode, Data Center Dynamics, Forbes), and Chinese tech press. The acquisition value is reported, not disclosed, so it should be treated as an estimate. His exact title and scope at NVIDIA after mid-2025 are not detailed here because public descriptions vary; the core facts (Caffe, Berkeley, Alibaba Cloud leadership, Lepton founding, NVIDIA acquisition) are well corroborated.
What readers can do now
- If you build on Chinese clouds, study Lepton's "pool GPUs, wrap with tooling" pattern — it is now part of NVIDIA's DGX Cloud Lepton.
- Read the Caffe and ONNX histories to understand why open model-interchange standards still matter for embodied-AI and BCI workloads.
- Track Alibaba Cloud and Volcano Engine (火山引擎) as the two largest proofs that platform-scale model serving was born in China's cloud wars.