Will Falcon, CEO of Lightning AI, ends up arguing that AI infrastructure is consolidating into a new kind of cloud: software primitives for training, inference, development and agents bundled with owned GPU supply. He says Lightning's merger with Voltage Park makes it the first "full-stack AI neocloud", built on low debt and enterprise customers at a time when he expects debt-laden neoclouds to face a credit squeeze. The host discloses a year-long fellowship at Lightning AI.
The merger. The host gives the combined figures: 400,000 developers and companies, over $500 million in ARR gained in under two years, and 35,000 GPUs, which he calls the third-largest neocloud. Lightning had never sold compute; customers brought their own cloud. He calls AWS premium-priced and built for CPU workloads, so AI teams stitch point products together at a cost in security and overhead. Of hundreds of GPU-first neoclouds, Lightning partnered with the top seven or eight, but enterprise customers such as Cisco struggled on them until Voltage Park's responsive "white glove" support won and retained them. Lightning alone grew revenue at least 30–40x in the past year yet lost deals for lack of compute, while neoclouds offered little beyond Kubernetes. Falcon says the company was standing up its seventh data center and built Cursor's training cluster.
What an AI cloud is. Falcon regrets that "inference" stuck as a name and says a production inference server is like a web server with different batching and streaming. Against the claim that roughly 99% of GPU use is inference, which the host offers, he recalls that at Facebook in 2019, 99% of workloads ran on one GPU only because multi-GPU training was hard; training will grow once tooling exists. His analogy is S3 in the early 2000s: storage became a primitive within what was later called the cloud, and inference, vector databases, training infrastructure and agents will bundle the same way rather than each becoming a company.
Studios. The core product is a persistent cloud development environment that behaves like a laptop: Claude Code can run overnight without risk to a local machine, 20 can run at once on different CPUs and GPUs, any editor or SSH can connect, teams share environments, and enterprises can audit them and run them on-prem or in their own cloud. Every Lightning developer works on Studios. Vibe-coding agents aim to close a talent gap for data scientists used to scikit-learn and NumPy, handling the MLOps and spinning up clusters of 1,000 GPUs.
Neocloud economics. Asked about a McKinsey article calling neocloud economics fragile, Falcon says many rivals borrowed against startup customers or, once public, against market caps. With neocloud prices down about 60% or more in six months, he expects loans to be called and a mortgage-crisis-like need for a TARP-style rescue. Voltage Park has no debt and its CEO, now on Lightning's board, is a former finance executive; Lightning expects to take on controlled debt in the next few years. Enterprise contracts fund a free tier and the claimed 70% cost cut, which he ties to his own need for compute as a student.
PyTorch Lightning's origins. As a Columbia undergraduate around 2015, Falcon decoded images from retinal recordings with autoencoders and GANs, first in Theano, using a personal "research lib". Rewriting it in TensorFlow across four self-built four-GPU boxes cut an ImageNet run from 21 days to about an hour and yielded his only NeurIPS paper. He co-founded NextGen Vest, which used reinforcement learning to suggest questions for staff texting low-income students and got them about $60 million in aid; CommonBond acquired it around 2017–2018. At NYU under Kyunghyun Cho and Yann LeCun he open-sourced the library and took it into FAIR, then a group of about 100–150 people. Maintaining it slowed his research and he nearly shut it down; then about 10 term sheets arrived in a few days and he founded the company.
The library today. Falcon counts about 500,000 engineering hours in PyTorch Lightning and rejects the idea that it is not for generative AI: Nvidia's NeMo is built on it, AWS has pre-trained LLMs with it and LinkedIn trained a 100-billion-parameter LLM with it. Teams wanting more control move to Lightning Fabric, then raw PyTorch.
Grit. Falcon arrived from Caracas at 13 without English and left a two-year ESL programme in 4 or 5 months. He was one of about 20 officers selected from roughly 12,000 applicants for Navy SEAL training but, after pneumonia and Hell Week injuries, never finished. He credits the junior-officer leadership course taught by Leif Babin, later the book Extreme Ownership, which he recommends, for his leadership, and says the trait he relies on is solving whatever annoys him quickly.