NVIDIA acquires a company founded by a former Alibaba vice president.

April 8 news, NVIDIA has completed its acquisition of GPU leasing company Lepton AI.

Founded in 2023 by former Alibaba Vice President Jia Yangqing and his team of approximately 20, Lepton AI reportedly sold for hundreds of millions of dollars. Both Jia Yangqing and co-founder Bai Junjie have joined NVIDIA.

As a leading player in the global AI computing market, NVIDIA aims beyond being just a chip supplier. Amidst competition from AWS, Azure, and GCP, Huang Renxun is playing a strategic game.

By acquiring companies like Lepton AI, NVIDIA not only controls the entire industry chain from chip development to cloud service implementation but also directly targets small and medium-sized enterprises that need flexible computing power but cannot afford long-term cloud contracts, especially amidst the surge of generative AI applications.

This seemingly low-key acquisition unveils NVIDIA's ambition to build an AI empire. While OpenAI and others grapple with high computing costs, NVIDIA has quietly expanded its reach into the server rooms of many AI startups.

- 01 -

Caffe's Creator Ventures Out

Jia Yangqing, the founder of Lepton AI, hails from Shaoxing, Zhejiang, and is a legend in the AI architecture space.

With degrees in Automation from Tsinghua University, he developed the popular deep learning framework Caffe during his Ph.D. in Computer Science at the University of California, Berkeley. After graduation, he worked at the National University of Singapore, NEC America Laboratories, and Google Brain.

In March 2019, Jia Yangqing joined Alibaba as Vice President of Technology, later holding key positions such as Vice President of Alibaba Group in September of the same year.

Four months after leaving Alibaba, Jia Yangqing, along with Bai Junjie (creator of the ONNX neural network exchange standard) and Li Xiang (founder of the Kubernetes core component etcd), founded Lepton AI in July 2023.

Named after Leptons in physics, the company symbolizes changing the underlying logic of the computing world with agility.

Their conversational search engine, Lepton Search, built with just 500 lines of code, breaks down traditional AI development barriers and popularizes the cloud experience of giants like Google and Meta.

(Science Popularization: Lepton Search is an open-source conversational AI search engine developed by Jia Yangqing's team. Based on the Transformer model and knowledge graph, it aims to understand users' true intentions through natural language interaction and provide an intelligent search experience.)

- 02 -

Enhancing GPUs, Not Producing Them

Lepton AI provides enterprises with efficient and scalable AI application platforms using a cloud-native multi-cloud solution that empowers any GPU provider to quickly upgrade.

In simple terms, Lepton AI serves as the "computing power mage" in the AI world. Without owning GPU hardware, it constructs a global computing network through the cloud.

By renting NVIDIA chips from cloud providers like AWS and overlaying their self-developed intelligent scheduling system, Lepton AI helps game company Latitude.io's hundreds of thousands of players experience smooth AI character interactions and aids research institution SciSpace in accurately capturing data from vast papers.

Its unique "multi-cloud puzzle" mode integrates resources from different cloud platforms with a unified API, enabling enterprises to access high-performance professional computing power at an incremental cost of "just a few cents more per GPU hour," directly addressing developers' pain points of "not being able to afford or manage well."

In just two years, Lepton AI was hailed by SemiAnalysis as the "only player not burning money or stockpiling goods" in the global GPU cloud service golden tier.

In May 2023, Lepton AI completed an angel round of financing of $11 million (approximately RMB 79 million), with investors including renowned Silicon Valley venture capital firm CRV, Sequoia China, and Fusion Fund. These three institutions achieved investment exits within less than two years, realizing significant financial returns.

- 03 -

NVIDIA's Strategic Move

In essence, Lepton AI's main business involves renting servers powered by NVIDIA AI chips.

In the fiercely competitive AI computing market, NVIDIA's acquisition of Lepton AI for hundreds of millions of dollars marks a crucial step in its strategic transformation from a "hardware overlord" to a "full-stack service provider".

Today, cloud service giants like Amazon AWS and Google Cloud are developing their AI chips (such as AWS Trainium and Google TPU) and implementing low-cost leasing strategies, posing a threat to NVIDIA's GPU monopoly with the trend of "de-NVIDIA-ization".

Lepton AI's "cloud-native + multi-cloud integration" technology schedules global GPU resources at minimal costs. Its lightweight software toolchains, like the search engine built with only 500 lines of code, are deeply integrated with NVIDIA's CUDA ecosystem, effectively compensating for NVIDIA's shortcomings in the cloud service and enterprise software markets.

By integrating the Lepton AI technical team, NVIDIA can establish an end-to-end "chips + cloud platforms" solution, not only resisting the impact of cloud giants' ecological closed loops but also penetrating the small and medium-sized enterprise market, driving its software business revenue towards the long-term goal of $150 billion.

- 04 -

NVIDIA's Recent M&A Strategy

Strategically, NVIDIA has frequently made acquisitions in the AI infrastructure field in recent years.

In 2024, NVIDIA acquired the cluster management platform Run.ai for $700 million, the model optimization company Deci for $300 million, and successively incorporated the inference acceleration tool OctoAI and the synthetic data company Gretel.

These acquisitions span the entire AI development chain, aiming to reduce AI development costs and solidify dominance in computing power infrastructure. Run.ai optimizes GPU resource scheduling, Deci improves model energy efficiency ratios, and Gretel addresses training data bottlenecks, collectively forming a closed-loop ecosystem from chips to applications.

Currently, these enterprises are deeply integrated into the NVIDIA AI Enterprise suite. For example, Run.ai's technology boosts GPU utilization by 40%, and Deci's automated compression algorithms help customers cut inference costs by 30%.

This "puzzle-style M&A" strategy has proven highly effective, with NVIDIA's cloud and software business revenue exceeding $1.5 billion in 2024, nearly a five-fold increase from three years ago.

- 05 -

Global AI Cloud Service Market Landscape and Challenges

The global AI cloud service market is growing rapidly at a compound annual growth rate of 38%, projected to surpass $200 billion by 2025.

Currently, the market exhibits a "three-legged stool" pattern, with NVIDIA leading with hardware advantages and ecological integration, AWS and Google Cloud competing through self-developed chips and low-cost strategies, and Alibaba Cloud and Huawei Cloud accelerating domestic chip (like Ascend 910B) and open-source ecosystem development.

However, the intensifying Sino-US tariff issue poses a significant challenge. If it escalates further, NVIDIA may face a double blow to its "hardware and software integration" strategy. Specifically:

Chip export restrictions could weaken its hardware foundation, and data sovereignty competition may lead to regional cloud service fragmentation.

For instance, the US plans to impose a 25% tariff on data center equipment, increasing server procurement costs by 3%-5%, prompting enterprises to shift production to Mexico or Southeast Asia.

Simultaneously, geopolitical factors may accelerate technological substitution. China is promoting the coordinated development of Ascend chips and the MindSpore framework, while the European Union is advancing the construction of the Gaia-X sovereign cloud. The global AI infrastructure may split into three major technological camps: the US, China, and Europe. This fragmentation not only raises research and development costs but may also diminish global AI innovation efficiency.

For China, if the US tightens chip export restrictions, it may exacerbate computing power cost pressures on Chinese enterprises in the short term but foster long-term breakthroughs in local technologies.

Tianzhi Intelligent Chip has swiftly adapted domestic GPUs to the DeepSeek model, and companies like DeepSeek are exploring low-computing-power-dependent models to overcome hardware barriers.

Coupled with policy-supported industrial chain collaboration (like the "New Generation Artificial Intelligence Development Plan"), it may accelerate the formation of a parallel ecosystem "de-NVIDIA-ized".

This global game will not only reshape the global computing power supply chain but may also lead to the splitting of technical standards, transforming the AI infrastructure market from a "performance race" to a new competitive stage of "ecological fragmentation".

The content of this article is for reference only and does not constitute investment advice. This article refers to relevant content from Z Finance, AI Time Journal, etc., during the creation process, and thanks are hereby given. The image is from WeChat images.