Palo Alto Networks agreed to acquire Applied Compute, a startup that helps companies train, customize and run open-weight AI models, in an all-stock deal valued at $4.5 billion — the cybersecurity giant's second-largest acquisition ever, trailing only its $25 billion purchase of CyberArk earlier this year. The deal, announced alongside Palo Alto's fiscal first-quarter results, is expected to close in the company's third fiscal quarter, which runs through April 2027, pending regulatory approval.
Applied Compute was founded in 2025 by a team of former OpenAI researchers and has grown fast: the company reached a $100 million annualized revenue run rate and, earlier this month, closed a $350 million funding round led by investor Elad Gil at a $3.25 billion valuation. Palo Alto's offer values the roughly 50-person company at nearly 13 times that figure. Co-founder and chief executive Yash Patil will become Palo Alto's co-chief AI officer once the deal closes, and Applied Compute's research team is expected to help build AI models across Palo Alto's security products.
Palo Alto chief executive Nikesh Arora framed the acquisition as a bet that enterprises increasingly want to run open-weight models — rather than only closed, proprietary ones — inside environments the companies themselves control. He said open-weight models have grown from under 5% to more than 60% of AI token traffic over the past year, and argued that conventional security tools can't keep up with how freely enterprises can now customize and deploy their own models. "This could be bigger than what we do today," Arora said of the new training-and-inference business. Palo Alto said it plans to build out more than 100 megawatts of its own computing capacity over the next three to five years while continuing to rent capacity from cloud partners; chief financial officer Dipak Golechha said most of that spending would go toward GPUs and servers rather than facilities.
Not every outside observer was convinced the deal solves what it promises. Sanchit Vir Gogia of Greyhound Research noted that most AI application failures happen at runtime, an area Palo Alto has already addressed since 2025 through its model-agnostic Prisma AIRS product, and questioned how much genuine data sovereignty a US-headquartered company can offer given that American firms face the same jurisdictional exposure as the big cloud providers. Gogia advised enterprise customers to insist on a costed exit plan, clear ownership of trained model weights, guarantees that their data won't be used to train other customers' models, and change-of-control rights before signing on.