$700 million.

That's the price Jensen Huang paid to acquire LeptonAI in April 2025. A startup with just 20 people -- $35 million per head -- hailed at the time as a legendary AI Infra acquisition. LeptonAI founder Jia Yangqing, creator of the Caffe framework, core author of TensorFlow, and former VP at Alibaba Cloud, entered Nvidia with his team and his halo to serve as VP of Systems Software.

One year and two months later, the myth shattered.

On June 29, a tweet from semiconductor research firm SemiAnalysis exploded across AI circles: Jia Yangqing had left Nvidia. On the day the news broke, traffic to LeptonAI's open-source community pages surged 400% as developers flocked to GitHub to check on the promised open-source model weights -- there was nothing. That 20-billion-parameter open-source model, which Nvidia had officially touted as something that would "reshape the AI inference landscape," never saw the light of day. The flagship product DGX Lepton was virtually invisible within Nvidia's product matrix, failing to become the AI Infra benchmark it was expected to be.

According to multiple sources, Jensen Huang had long been dissatisfied with LeptonAI's operational results. The promises made at acquisition -- "maintain independent operations + open-source roadmap" -- were gradually eroded within Nvidia's vast commercial machine. The contradiction between Jia Yangqing's open-source ideals and Nvidia's closed-source commercial logic ultimately became irreconcilable.

This isn't one person's departure story. It's a microcosm of the past two years of AI Infra entrepreneurship -- when everyone is building "AI infrastructure," the only ones truly making money are the ones selling shovels.

The $700M Gamble: Why LeptonAI?

To understand why this acquisition caused such a stir at the time, you first need to understand Jia Yangqing's standing in the AI community.

Jia Yangqing is one of the pioneers of the framework era. In 2013, during his PhD at UC Berkeley, he wrote Caffe -- the open-source project that kicked off the deep learning framework era. After graduating he joined Google Brain, becoming one of TensorFlow's core authors. He then moved to Facebook, leading development of Caffe2 and PyTorch 1.0. In 2019 he returned to China to join Alibaba as VP of Alibaba Cloud Intelligence, overseeing AI platforms and big data computing.

In 2023 he left Alibaba to found LeptonAI, positioned as "AI-native infrastructure" -- building LLM inference and deployment the cloud-native way. In the AI Infra赛道 of the time, LeptonAI wasn't the best-funded or the buzziest, but its team configuration was lavish: a core team of around 20 almost entirely composed of senior infrastructure engineers from Google, Meta, and Alibaba.

Why was Nvidia willing to spend $700 million on a 20-person company? The surface reason was LeptonAI's technical accumulation in inference optimization and model serving, but the deeper reason was that Nvidia needed to fortify its moat at the software layer.

In early 2025, Nvidia was facing a delicate moment: on the hardware side, AMD's MI400 series was starting to win training contracts, Google TPU v6 was running at full capacity in its own data centers, Amazon's Trainium2 deployment exceeded 100,000 chips; on the software side, while PyTorch remained mainstream, inference engines (vLLM, TensorRT-LLM, SGLang) were proliferating, diluting Nvidia's control over the software stack. Integrating LeptonAI's model serving capabilities into the DGX product line could tighten Nvidia's "hardware + software" bundle.

And Jia Yangqing carried a label Nvidia valued most: open-source community credibility. Nvidia needed someone with sufficient prestige in the open-source community to bridge the rift between its commercial closed-source strategy and the open-source community. At the time of acquisition, Nvidia officially stated that LeptonAI's open-source projects would continue to be maintained, and the promised 20B-parameter open-source model would be released as planned.

That promise became the biggest point of contention.

The One-Year Itch: How Open-Source Promises Got Devoured by the Commercial Machine

Post-acquisition LeptonAI went through a classic case of big-company integration syndrome.

First was strategic drift. Jia Yangqing's original plan: build on LeptonAI's inference platform to create a cross-hardware, neutral model serving layer -- the Nginx of the AI era, not tied to any chip vendor. But inside Nvidia, this positioning fundamentally conflicted with the company's core objective of "selling more DGX." You can't sell DGX while simultaneously building a neutral platform that lets customers run equally well on AMD/Google chips.

DGX Lepton's productization was an uphill battle. According to former employees, the team spent enormous time adapting to Nvidia's latest hardware and CUDA versions, while support for non-Nvidia hardware was repeatedly delayed. That promised open-source 20B model kept getting deprioritized internally -- because a truly open-source model that runs efficiently on multiple chips might not serve Nvidia's hardware lock-in strategy well.

Second was talent and culture clash. LeptonAI's team was quintessential open-source culture: rapid iteration, public discussion, community-driven. Nvidia is a sales-oriented hardware company with long decision chains, extreme secrecy culture, and layers of legal approval required before any open-source release. A former LeptonAI engineer who joined Nvidia privately complained: "At Lepton we shipped weekly; here a single release note takes six weeks of approvals."

But the core contradiction was the fundamental divergence between open source and closed source.

Nvidia's business model is built on CUDA ecosystem lock-in -- developers writing code in CUDA can't leave Nvidia GPUs. The more open its software stack, the stronger the hardware lock-in (because more people use it), but Nvidia will never open-source the truly low-level performance optimizations and hardware details. Jia Yangqing's open-source ideal -- making AI infrastructure more accessible and neutral -- had no place in this commercial logic.

This isn't Nvidia's fault, nor is it Jia Yangqing's fault. It's a clash between two fundamentally different business philosophies: one that believes open ecosystems grow the pie, and one that believes vertical integration maximizes profits. In the supply-constrained AI compute market of 2025-2026, the latter clearly holds more sway.

It's Not Just Jia Yangqing: The Collective Dilemma of AI Infra Startups

Zooming out from Jia Yangqing to the entire AI Infra赛道, you'll find his predicament isn't unique.

Over the past two years, AI Infra was one of the hottest funding categories. Inference optimization, model deployment, vector databases, Agent orchestration, feature platforms, LLM gateways... dozens of startups flooded into each subcategory. But reaching mid-2026, a brutal reality emerges: the vast majority of AI Infra companies haven't found sustainable business models.

The reason is simple: AI Infra's paying customers are primarily LLM companies and cloud providers, and both categories are building infrastructure themselves.

OpenAI is building its own chip, Jalapeno -- taped out in 9 months in partnership with Broadcom, planned for mass deployment by year-end, directly challenging Nvidia's inference market. Google, Microsoft, Amazon, Meta, and even SpaceX are all pursuing custom silicon. From OpenAI to SpaceX, big tech is collectively "de-Nvidia-ing," and it's no secret. On the LLM company side, Anthropic has built its own inference engine and deployment stack; DeepSeek has done its own inference optimization since day one; ByteDance, Alibaba, and Baidu all have their own Infra teams.

When your potential customers are all building what you're building, who do you sell to?

Vector databases are a textbook example. In 2023-2024, dozens of vector database companies raised funding, but by 2026 PostgreSQL's pgvector extension already satisfies 80% of use cases, and major cloud providers (AWS, GCP, Alibaba Cloud) all include vector search as a standard database feature. Independent vector database companies are seeing their survival space rapidly compressed.

Agent orchestration frameworks face the same dynamic. LangChain was once the star of this赛道, but by 2026 developers increasingly prefer using models' native function calling and built-in Agent tools from major platforms. AWS launched the open-source Blocks framework; GitLab 19.0 embedded Agentic AI into merge requests and supply chain security; Apple launched the Core AI framework optimizing on-device AI for its custom chips -- when platforms bake Agent capabilities into infrastructure, middleware startups get disintermediated.

Beyond Acquisition: Do AI Infra Founders Have Another Way Out?

Jia Yangqing's ending can easily create an illusion: that the endpoint of AI Infra entrepreneurship is getting acquired by a big company and then getting assimilated. But reality tells a different story.

The AI Infra companies breaking out in 2026 are taking entirely different paths.

One category is "deep vertical." They don't build general-purpose platforms; they dive into AI infrastructure for specific industries -- like inference compliance platforms for medical AI, real-time decision engines for financial AI, edge deployment solutions for manufacturing AI. The barriers in these domains aren't about how advanced the technology is, but about industry know-how and compliance certification accumulation -- things big companies don't want to do and can't do well.

Another is "open-source community + commercial services." They don't make money selling software licenses; they use open-source projects to build community and standards, then monetize through enterprise editions, managed services, and technical support. Hugging Face is the poster child for this model, though its challenge is maintaining independence surrounded by giants. More niche examples include AI observability and AI security startups that build developer mindshare through open-source tools, then sell compliance and governance solutions to enterprises.

The third is "betting the opposite direction." While everyone builds infrastructure to "make LLMs run faster," some founders are building tools to "make LLMs cheaper" -- model compression, edge inference, domestic chip adaptation, mixed-precision scheduling. Against a backdrop of rising compute costs (Kimi's B2B lead Huang Zhenxin admitted frankly that "all model vendors are raising prices, the core reason being rising compute costs"), tools that save customers money have more staying power than tools that help them spend it.

The Framework Generation Exits, the Agent Generation Arrives

Jia Yangqing's departure also marks the end of an era.

He's a representative of the "framework generation" -- Caffe, TensorFlow, PyTorch, the deep learning frameworks that laid the foundation for AI today. That generation of engineers believed in the power of open source, believed good technical standards could unify ecosystems, and believed that "making great tools accessible to more people" was a technologist's mission.

But AI competition has entered a new phase. Models are increasingly capable; framework abstraction layers keep rising; most developers no longer need to understand low-level CUDA kernels or distributed training details. Developers in the Agent era don't care about "how to make the model run 5% faster"; they care about "how to make Agents reliably complete tasks." Infrastructure is shifting from "model-developer-facing" to "Agent-builder-facing."

This shift is clear from recent product releases: AWS's Blocks is a backend framework for Agents; Herdr is an Agent multiplexer in the terminal; BrowserBC lets humans teach all Agents to operate web pages with one click; Lore lets coding agents remember a team's decision history. These tools' target users aren't engineers training models; they're developers building Agent applications.

The framework generation completed their historic mission -- taking deep learning from the lab to industry. But the AI infrastructure story doesn't end here. Jia Yangqing's next move hasn't been announced, but it's certain that infrastructure geniuses like him won't stay away from the field for long.

The $700 million acquisition wasn't a failure, but it genuinely didn't meet anyone's expectations. Nvidia didn't get the open-source community influence or software ecosystem consolidation it hoped for; Jia Yangqing didn't realize his vision of open, accessible AI; LeptonAI's products didn't shine within the Nvidia machine. It's a three-way loss, but also a valuable lesson: in AI's winner-takes-all market, if an independent infrastructure company can't build irreplaceable depth, acquisition isn't the finish line -- it's the beginning of dissolution.

For founders currently on the AI Infra track, Jia Yangqing's story is a mirror. It reflects a brutal truth: in this compute-is-king era, technical idealism alone isn't enough. You either drill into a niche that big companies don't want and can't do well, or build a community and ecosystem powerful enough that you can't be easily acquired and assimilated -- or you accept the fate of being integrated and then assimilated.

This isn't pessimism. It's reality.