Lightning AI, the startup behind the open source PyTorch Lightning framework, announced today that it has raised $40 million in a Series B round led by Coatue that includes Index Ventures, Bain, the Chainsmokers’ Mantis VC and First Minute capital. CEO William Falcon told TechCrunch that the new money will be used to expand Lightning AI’s 60-strong team while also supporting the community around PyTorch Lightning development.
Lightning AI, formerly Grid.ai, is the culmination of work that began in 2018 at the New York University Computational Intelligence, Learning, Vision, and Robotics (NYU CILVR) Lab and Facebook AI Research (now Meta AI Research). After Falcon began developing PyTorch Lightning in 2015 as a student at Columbia, he co-founded Lightning AI in 2019 with Luis Capelo, the former head of data products at Forbes. While working on his PhD at NYU and Facebook AI Research, Falcon used open source PyTorch Lightning and, he says, the project quickly gained traction.
†[W]We realized that the biggest challenge holding back AI adoption at scale is the fragmentation of the AI ecosystem,” said Falcon. “In 2019 I first noticed the impact of the fragmented AI ecosystem. Because AI adoption on a large scale was still in its infancy, every few months we discovered a ‘missing’ part of the machine learning stack. Why this matters is that every missing piece of the puzzle is slowing the overall pace of AI innovation… Just to get a model to go into production would take hundreds, if not thousands. developer hours that are spent purely on infrastructure.”
AI dev framework
With PyTorch Lightning, Falcon tried to decouple R&D workflows from engineering, making AI experiments easier to read and reproduce (in theory). PyTorch Lightning provides a high-level interface to PyTorch, Facebook’s popular deep learning framework, stripping away the code normally needed to set up and maintain AI systems.
The new Lightning AI platform includes a collection of tools relevant to machine learning, such as workflow planning for distributed computing (i.e., spreading workloads across multiple machines) and managing and provisioning infrastructure through code. A gallery of AI apps curated by the Lightning team is available to use or build upon, as well as a library of components that add capabilities to AI apps, such as extracting data from streaming video.
Apps built with Lightning AI can run in a private cloud infrastructure or in on-premises environments. Alternatively, Lightning AI provides a hosting platform for deploying and monitoring apps in the cloud.
“The vision I’ve pursued since my time at NYU has always been to build something like an AI operating system that would allow all the disparate parts of the AI ecosystem to work together,” Falcon said. “You don’t need to know anything about the combustion engine to drive to the supermarket; why should you know anything about it? [containers], cloud infrastructure, distributed file systems, and fault-tolerant training to easily bring your AI project to life? The current solutions give you the different parts of a working car and hope you can put them together into something you can use to take a ride.”
To that end, Lightning’s app gallery includes out-of-the-box AI models designed to perform tasks such as diagnosing pet cancer, running workloads in the cloud, and launching cloud AI projects. Falcon sees the Lightning app gallery as a way to launch “multi-cloud, distributed” AI systems at “enterprise scale” as well as a building block for the “next generation” applied AI startups.
“We found that many startups launched in the past two years could have been built as Lightning apps,” Falcon said. “We see ourselves as the Apple App Store, but for AI. Some of the world’s largest companies first launched as Apple Apps: we’re doing the same for AI.”
A burgeoning ecosystem
PyTorch Lightning and Lightning apps fall into the category of MLOps software, which supports the AI lifecycle by orchestrating experimentation, AI model training, model implementation in production, and model tracking. There is a lot of interest in MLOps. in a recent questionnaire, Forrester found that 73% of companies believe MLOps adoption would keep them competitive, while 24% say it would make them a market leader. Deloitte predicts that MLOps will be a $4 billion market by 2025.
There are plenty of alternatives to PyTorch Lightning, including PyTorch Ignite and Fast.ai. But Falcon points to statistics as proof that Lighting AI’s project is ahead of the pack: To date, PyTorch Lightning has been downloaded 20 million times and is used by an estimated 10,000 companies in production. In 2019, NeurIPS, one of the world’s largest AI conferences, adopted PyTorch Lightning as the standard for submitting PyTorch code.
†[With Lightning AI,] companies can understand what the systems they are building are doing, and even bring in non-engineering team members — from compliance, for example — to eliminate the business risk of a system that will bankrupt the company or become racist on Twitter,” Falcon said. “Companies building solutions with Lightning AI are not tied to a specific cloud provider or hardware vendor… Finally, Lightning AI is powered by the open source AI community, meaning companies can use the latest and greatest open source tools without taking months to spend one of them integrate.”
Despite its open source origins, Lightning AI could be seen as competing with startups like Hugging Face, who also provided AI app hosting services. Companies including: Comet† iterative† Weights and Prejudice and infuseAI offer a similar mix of paid and free MLOps solutions.
Falcon isn’t worried about the competition, though. He claims Lightning AI’s revenue trajectory is on track to be cash flow positive within the next 36 months; the company primarily monetizes its fully managed products. While he wouldn’t commit to hiring plans, Falcon noted that Lightning AI doubled his team to 60 people in the past year.
“I grew up in an inflation-ridden Venezuela during some of the darkest days of the financial markets. As a result, I am always careful and make sure we are well capitalized for possible long-term withdrawals,” Falcon said. raised enough capital to provide us with a runway of at least several years.”