Palo Alto Networks agreed to buy Applied Compute because it wants a role in two decisions its customers face as they adopt AI. One is how to protect an AI system. The other is which model to run, how to teach it a particular job and where to run it. Applied Compute brings the technology and researchers to work on that second decision. Palo Alto Networks also wants the team to build better models for its own cybersecurity products. Acquisition announcement
The companies signed a definitive agreement on November 18, 2026. The all-stock transaction is valued at US$4.5 billion, subject to adjustments, and remains pending. Palo Alto Networks expects it to close in its fiscal third quarter of 2027, subject to regulatory clearance and other customary conditions. For customers, that distinction matters. The agreement establishes what Palo Alto Networks intends to buy. Product availability will follow its own timetable. SEC filing
What does $4.5 billion buy?

Under the merger agreement, Palo Alto Networks will acquire Applied Compute’s outstanding stock, warrants and other equity rights. It will pay in its own common shares. The SEC filing gives a reference price of $337.16 per share, based on a volume-weighted average over the 15 trading days ending November 17. The company said the final number of shares to be issued was still unknown. One outstanding US regulatory condition is the expiration or termination of the applicable Hart-Scott-Rodino antitrust waiting period.
The price is particularly striking because Applied Compute had just raised $350 million at a reported $3.25 billion valuation. It had also announced a $100 million revenue run rate. On those headline figures, the agreed acquisition value is about 38% above the financing valuation. That gap gives a sense of how quickly the company’s fortunes changed, although a private funding round and a purchase of the entire business give investors different rights. A run rate also projects a recent pace of business forward. It tells readers little about how much revenue the company actually earned over the preceding year.
What Applied Compute actually does
Applied Compute’s platform, called AC2, helps companies adapt an existing AI model, run it for users and improve it when results fall short. A customer can bring examples of good work, define tests and see whether a revised model performs better on the tasks its employees or customers actually do. Applied Compute offers infrastructure for these steps and researchers who can work alongside the customer’s engineers.
Three terms explain why this capability caught a cybersecurity company’s attention.
| Term | What it means for an enterprise |
|---|---|
| Open-weight model | The model’s learned numerical settings, or weights, are available to others under a licence. An enterprise may be able to run and adapt its own copy. It still needs to check the permitted uses and what other materials the developer releases. |
| Post-training | Engineers take an already trained model and teach it a more specific behaviour. A security team might use examples reviewed by experienced analysts to teach the model which incidents need urgent attention. |
| Inference | The trained model processes a new request and produces a result. Every time the security team submits another incident for review, it consumes computing capacity, even if the model’s training finished months ago. |
One example comes from outside cybersecurity. DoorDash co-founder and CTO Andy Fang said his company incorporated its quality standards into training to improve menu accuracy. Within cybersecurity, Semgrep co-founder and CTO Drew Dennison said its team used years of labelled security findings to adapt models to its internal tools. Both accounts appear on Applied Compute’s website. They illustrate the common ingredient across two very different businesses. Each already had people who could recognise a correct answer and records that could help teach a model to produce one.
Why Palo Alto Networks wants the research team
Palo Alto Networks first assessed Applied Compute as a prospective customer. CEO Nikesh Arora said its Unit 42 security research group reviewed the startup’s security foundations. That history helps explain management’s confidence in the technology. It also identifies whose assessment this was. Customers evaluating the combined offering will still need evidence specific to their deployment. Palo Alto Networks announcement
The planned leadership arrangement gives a clue to the team’s importance. Applied Compute was founded by three former OpenAI employees. Its co-founder and CEO, Yash Patil, is expected to become Palo Alto Networks’ co-chief AI officer after closing. The researchers would help train both large and small proprietary models across the buyer’s offerings. On the earnings call, Arora said Palo Alto Networks was already training models for more than 35 specialised use cases. He described work in progress, leaving open which capabilities customers will receive first.
Consider a security operations team piloting a model to prioritise alerts. It might correctly identify nearly every routine incident and still send one serious incident to the back of the queue. Experienced researchers can help decide which past cases to reserve for testing, how heavily to penalise a missed incident and when the model should hand a decision to a person. Those choices explain why acquiring the people who design evaluations could matter as much as acquiring the software that runs them.
The competition to run enterprise AI
Arora expects enterprises to use a mixture of open-weight models and closed models accessed through providers. There is an economic reason for that mix. A business might value a hosted model’s breadth when employees ask unpredictable questions, while a smaller model could handle a high volume of similar requests. Applied Compute already offers a model router, software that selects a model for each request. Clay’s head of AI, Jeff Barg, said a router built with Applied Compute cut inference costs for its Claygent product by up to 47% while maintaining quality. That figure describes one customer’s workload.
Palo Alto Networks will enter a market where customers already have options. Amazon SageMaker JumpStart, for example, lets AWS users deploy, fine-tune and evaluate pretrained models. An enterprise already running workloads on AWS has a reason to start there. Palo Alto Networks’ opportunity is to persuade the same buyer that its security expertise, research support and choice of deployment environment justify bringing in another provider. TD Cowen analyst Shaul Eyal characterised the proposed purchase as an expansion beyond traditional cybersecurity into enterprise AI.
What public and sovereign cloud customers should expect
Palo Alto Networks intends to support public cloud and sovereign environments and keep serving Applied Compute’s existing customer groups. A public cloud lets an enterprise use infrastructure operated by a cloud provider. A sovereign environment adds requirements around jurisdiction, location or operational control. The two can overlap. A provider might operate a regional service under specific local controls. Buyers need the details of that arrangement to decide whether it meets their requirements.
Take a multinational insurer that wants to summarise claims in a particular country. It could run the model in an approved region and still have an unresolved question about the diagnostic log created when a request fails. Does that log contain the claimant’s text? Who can open it during a support call? Does a backup cross the border? Asking a vendor to trace one failed request can reveal more about control of data than a list of data-centre locations. Teche Time covers related vendor questions in its guide to GDPR compliance software for SMEs in France.
Applied Compute currently describes serverless and virtual private cloud deployment options, role-based access and audit logs. An existing customer planning a renewal should ask which of those controls its contract covers, who handles support through the transition and whether it can export its trained model and evaluation data. A prospective customer should ask to see the service running in the intended region. For background on the distinction between rented infrastructure and the responsibilities a customer retains, see our virtual private server guide. Applied Compute’s current deployment options
The financial test comes after closing
Management set a measurable ambition on the earnings call. Arora expects Applied Compute to contribute more than $1 billion to next-generation security annual recurring revenue by fiscal 2029. He also outlined plans for more than 100 megawatts of global inference capacity over three to five years, supported by less than $3 billion in capital spending through fiscal 2030. The capital plan is separate from the acquisition price. Both the revenue figure and the infrastructure plan are management expectations.
Analysis. The financial risk lies in the timing. Computing capacity costs money before enough customers are using it to spread that cost across a large volume of requests. Sovereign deployments add another constraint. Spare capacity in one country may be of little use to a customer whose workload must stay in another. Investors will learn more from evidence of repeat usage, regional demand and the cost of serving customers than from a count of pilots alone.
There is also a practical test of integration. The same researchers could be asked to improve Palo Alto Networks’ products, support existing Applied Compute customers and help win new enterprise accounts. A customer running a production model will notice quickly if that stretches the team too thin. Support response times, the ability to roll back a poor update and access to an engineer familiar with the model are tangible measures of how well the integration is working.
FAQs
Has Palo Alto Networks completed the Applied Compute acquisition?
As of November 19, 2026, the companies have signed the agreement. Palo Alto Networks expects closing in its fiscal third quarter of 2027, roughly February through April 2027. Regulatory clearance and customary closing conditions are still outstanding.
Will customers have to switch away from closed AI models?
Palo Alto Networks has described a strategy that accommodates both open-weight and closed models. A company could keep its existing hosted assistant and evaluate a specialised model for one data-sensitive or high-volume task. The announcement gives no migration deadline or requirement to switch.
Does access to the weights make an AI model cheaper to run?
That depends on demand and the quality required. For an occasional task, a hosted service may spare a business the cost of maintaining idle capacity. For a frequent, predictable task, a smaller model may earn back the cost of adapting and operating it. A useful pilot measures cost per accepted result, including retries and human corrections.
Final Words
The logic behind the deal is clearest at the point where an enterprise puts a model into daily use. Someone has to decide whether its answers are good enough, where sensitive requests can go and who intervenes when it fails. Applied Compute gives Palo Alto Networks a way to participate in all three decisions. The $4.5 billion question is whether customers will trust one provider with that much responsibility and keep paying it once the pilots end.

