Palo Alto Networks entered into a definitive agreement to acquire Applied Compute in an all-stock transaction valued at $4.5 billion. The acquisition brings Applied Compute and its post-training and inference platform into the cybersecurity provider to help organizations adopt open-weight AI securely. The transaction is expected to close during the fiscal third quarter of 2027.
Palo Alto Networks Applied Compute Acquisition Hybrid AI Vision
Palo Alto Networks chairman and chief executive officer Nikesh Arora says that the future of enterprise AI is hybrid. According to Arora, a large share of organizations will run open-source and open-weight models tuned to their own data, cost, and control needs. Prior to selecting the startup, Palo Alto Networks evaluated nine providers before deciding on the platform, Arora says.
Post-Training Inference Platform and Commercial Growth
Applied Compute operates a platform for post-training and inference of open-weight AI models. According to Palo Alto Networks, the combined operations will help organizations train, adapt, and deploy open-source models securely across public and sovereign clouds. The startup reached an annualized revenue run-rate of $100 million within roughly 18 months of inception, according to The420.in.
Its inference-serving business began processing 1,000,000,000,000 tokens per day just 12 weeks after launch, according to The420.in. The enterprise client roster includes technology firms such as Microsoft, Shopify, DoorDash, Harvey, Cognition, Clay, and Decagon, according to The420.in. Palo Alto Networks says Applied Compute will continue serving AI-native companies while the buyer brings the platform to its global customer base.
Yash Patil Leadership and Research Team Integration
Yash Patil, co-founder and chief executive officer of Applied Compute, will become co-chief AI officer at Palo Alto Networks following the close of the transaction. Patil founded the company in May 2025 alongside former OpenAI researchers Rhythm Garg and Linden Li, according to India Today. Palo Alto Networks says the acquisition brings an AI research team that will help embed AI across every offering by training proprietary large and small models alike.
Palo Alto Networks plans to build a team of 100 AI researchers by the end of the year by incorporating the startup research group, according to Calcalist Tech. Applied Compute also embeds engineers with customers to encode institutional judgment into reward functions, according to Modal.
Financial Terms and Fiscal 2029 Projections
The all-stock transaction is subject to adjustments and customary closing conditions, including regulatory clearance. The startup previously raised $350 million in a funding round led by Elad Gil that valued the company at $3.25 billion, according to The420.in. Palo Alto Networks expects Applied Compute to contribute more than $1 billion in annual recurring revenue to its Next-Generation Security business by fiscal 2029, according to Calcalist Tech.
Questions about Palo Alto Networks Applied Compute Acquisition
What is the core focus of Applied Compute technology?
Applied Compute provides a platform for post-training and inference of open-weight models, allowing companies to train custom models using proprietary data and run them for everyday business operations, according to Palo Alto Networks and India Today.
How does the reinforcement learning training loop operate?
According to Modal, the reinforcement learning training loop involves rollouts where agents attempt tasks inside replayable environments, evaluations scored against reward functions, and inference serving trained models in production.
What infrastructure does Applied Compute use for execution?
According to Modal, Applied Compute uses container infrastructure to provide rapid startup, filesystem and network isolation, and snapshot semantics for replayability while handling parallel CPU computation for grading live agent behavior across thousands of concurrent execution traces.
Why are reinforcement learning rollouts computationally demanding?
According to Modal, training reinforcement learning models requires running inference and sandboxes simultaneously to maintain accelerator utilization and prevent idle hardware during execution.


