Palo Alto agrees $4.5 billion Applied Compute deal to secure enterprise AI
Edited and reviewed by AI Ping Editorial Team
3-line summary
- The agreement acquires Applied Compute in an all-stock transaction.
- Existing customers will continue receiving service.
- Closing is expected in fiscal Q3 2027, subject to regulatory clearance and other conditions.
Palo Alto Networks announced a $4.5 billion all-stock agreement to acquire enterprise AI platform Applied Compute on November 18, subject to adjustments.
Palo Alto says the combination will let enterprises train, adapt and deploy open-weight models with security built in, in public-cloud or sovereign environments. The announcement describes an acquisition agreement and integration plans. Applied Compute is to continue serving its existing customers.
The SEC report covers stock, warrants and other equity rights. Palo Alto will calculate issued shares at $337.16, its volume-weighted average closing price for the 15 trading days through November 17. The final share count remained undetermined.
Open-weight models expose trained parameters for deployment and adaptation. Applied Compute’s existing AC2 platform spans model adaptation, production inference and continued improvement using customer data. Its product description includes rewards tied to customers’ evaluations, visibility into individual runs and extraction of training signals from production traces. In its managed arrangement, customers design experiments while Applied Compute handles training, serving and infrastructure. It advertises serverless and VPC deployment choices, role-based access and audit logs, with customers controlling what trains and deploys.
Public technical documentation describes connecting a customer’s existing agent environment. The customer’s software keeps responsibility for tools, files and task state. AC2 records model calls and uses a grader to score results; the scores feed training that updates model weights. Customers implement the grader that defines what counts as a successful result. Their harness server accepts tasks from AC2 and returns status and outputs. The documentation says that server need not have a public endpoint; the customer remains responsible for operating it.
A separate serving workflow deploys either a catalog model or a saved training checkpoint. Checkpoints first undergo conversion to a serving format. Once the deployment is ready, applications can send requests using an OpenAI-compatible API format. The documentation says scheduling and loading weights can take several minutes. These are capabilities of the current platform.
Customer content has separate contractual terms
The September 17 privacy policy places business-customer datasets, prompts, outputs, traces and model weights under customer agreements and a data-processing agreement. Website, account and marketing information has a separate scope.
Applied Compute says it uses customer content to develop, train, evaluate and serve models for that customer, at the customer’s direction and as allowed by contract. It excludes identifiable customer content from training models for others. Aggregated or de-identified derived information may be used where the customer agreement permits it.
The separately published terms say customers retain ownership of their data. Individual models and tools may carry their own conditions, which control when they conflict with the general terms. That document is marked October 28, 2025. The acquisition announcement does not present a revised customer contract.
Applied Compute’s researchers are to join Palo Alto. Co-founder and CEO Yash Patil will become co-chief AI officer after closing. Regulatory clearance and customary conditions remain; the company expects closing in its fiscal third quarter of 2027.
Sources
- Acquisition agreement announcement — November 18 — Palo Alto Networks
- Form 8-K: consideration, share-price reference and closing conditions — Palo Alto Networks · SEC
- AC2 product capabilities, managed deployment and model control — Applied Compute
- Bring your own harness: evaluation and model training — Applied Compute Docs
- Serving: catalog models and training checkpoints — Applied Compute Docs
- Privacy Policy — September 17, 2026 — Applied Compute
- Terms of Service — October 28, 2025 — Applied Compute