Palo Alto Networks has agreed to buy Applied Compute, a San Francisco startup that helps companies build their own AI models on top of open-weight ones. The price is $4.5 billion, paid entirely in PANW stock and subject to adjustments. The deal still needs regulatory clearance. Palo Alto expects it to close in its fiscal third quarter of 2027, so sometime in the first half of next calendar year.
The timing is what jumps out. On November 2, Applied Compute closed a $350 million round at a $3.25 billion valuation, led by Elad Gil. About two weeks later a buyer is paying about 38% more than that. Early backers got a very fast markup. So did everyone who came in this month.
What Applied Compute does
The company was founded in 2025 by three Stanford friends who had worked at OpenAI: Yash Patil (CEO), Rhythm Garg (CTO) and Linden Li. Patil worked on Codex. Garg was a core contributor to o1, OpenAI’s first reasoning model trained with reinforcement learning. Li worked on the systems and infrastructure behind that kind of training.
They left with a simple bet. Open-weight models, the ones anyone can download and run, would get good enough that companies would rather own a model than rent one. Applied Compute takes a base model such as Alibaba’s Qwen or Moonshot’s Kimi, post-trains it with reinforcement learning on a customer’s own data and rules, then runs it on its own cloud. The model keeps learning from real use after it ships. The company calls the loop train, serve, improve.
Customers named publicly include Nvidia, which used the startup to post-train its Nemotron open-weight models, DoorDash, and the legal AI company Harvey, which built a document review agent with it. Revenue was reported at around a $100 million run rate this year. At $4.5 billion that’s roughly 45 times run rate. Steep, but this is the AI market in late 2026.
Why a security company wants it
On paper, Palo Alto is buying an AI training shop. Patil framed it differently in the announcement: joining Palo Alto puts security at the center of everything Applied Compute builds, and opens doors at the largest companies in the world.
Read it from the buyer’s side. Big enterprises are moving workloads off closed frontier models to cut costs and keep their data in house. Every company that does that ends up with its own model, trained on sensitive data, wired into agents that call tools and touch production systems. Somebody has to secure that model, its training data and everything the agents do with it. Palo Alto would like that somebody to be Palo Alto, and owning the place where the model gets built is the most direct way in.
It also fits a pattern. In the past twelve months or so Palo Alto has bought CyberArk for about $25 billion (identity), Chronosphere for $3.35 billion (observability) and Protect AI for around $700 million (AI runtime security). Its Prisma AIRS product passed $100 million in ARR within four quarters of launch, the fastest ramp in company history. Applied Compute pushes the AI pillar further upstream, from guarding models other people built to building them.
The catch
All-stock deals are cheap in cash and expensive in shares. Palo Alto’s buying run already diluted holders by around 14% and dragged the stock down to $139 in February before the core business sped up and the shares re-rated. Another $4.5 billion of paper adds to that.
There’s also a culture question. Applied Compute’s staff is about 85% researchers and engineers, and the team tripled this year. Palo Alto is a 16,000-person sales machine. Keeping a young frontier-lab team inside a security giant is harder than closing the deal. And Applied Compute’s customers, Nvidia included, now have a security vendor between them and their model partner. Some will be fine with that. Some won’t.
Palo Alto said it would give more detail on the deal on its fiscal Q1 2027 earnings call, held November 18.