Acquisitions

Palo Alto Networks Buys Applied Compute to Train and Secure Your Open-Weight Models

AI Security · Acquisitions

Palo Alto Networks agreed to buy Applied Compute, a platform that trains and runs open-weight AI models, and will use its research team to build models of its own.

By Shashi Bellamkonda · November 18, 2026
$4.5B
All-stock deal value (Palo Alto Networks, 2026)
FQ3 2027
Expected close, Palo Alto Networks fiscal year (Palo Alto Networks, 2026)
70,000+
Palo Alto Networks customers (Palo Alto Networks, 2026)
Key takeaway
Palo Alto Networks will sell a platform for adapting and running open-weight models with security controls built in. The same research team will train Palo Alto Networks' own large and small models. Your security vendor is about to become a model builder too.

Palo Alto Networks signed a definitive agreement on November 18 to acquire Applied Compute, a San Francisco company that takes open-weight AI models, trains them on a customer's own data and runs them in production. The all-stock deal is valued at $4.5 billion (Palo Alto Networks, 2026) and is expected to close in Palo Alto Networks' fiscal third quarter of 2027.

Open-weight models publish their parameters for anyone to download, which lets you run them where you choose and tune them on data you would never send to an outside provider. You pick them when control over weights, data location or running cost outweighs the convenience of a closed model behind someone else's interface.

Yash Patil, Rhythm Garg and Linden Li left OpenAI in 2025 to start Applied Compute. Forbes reported in September that Nvidia used the company to post-train its Nemotron models, and that DoorDash, Microsoft and Harvey work with it. Patil becomes Palo Alto Networks' Co-Chief AI Officer once the deal closes.

Nikesh Arora, chairman and CEO of Palo Alto Networks, said his team first evaluated Applied Compute's platform as a prospective customer.

Palo Alto Networks gets a product and a model lab in one purchase

The product is the training and serving platform. Palo Alto Networks plans to offer it to its customer base so you can adapt and deploy open-weight models with security controls built into each step, in public clouds or in sovereign environments where data has to stay inside a country or a private network. Applied Compute will keep serving the AI-native companies and large enterprises it already works with.

The model lab is the research team. According to the announcement, that team will train Palo Alto Networks' proprietary large and small models and help put AI into every Palo Alto Networks product. Model training becomes an in-house job at a security company, sitting next to the threat research its Unit 42 group already does.

This deal moves Palo Alto Networks from guarding AI to building it

Palo Alto Networks has spent 2026 adding security layers around AI. I wrote about its Koi acquisition and the agentic endpoint category in April, its plan for a checkpoint on every AI transaction in May, and its reported interest in Clutch for AI agent credentials earlier this month.

Those deals protect AI that another company trained. Applied Compute gives Palo Alto Networks a hand in the training itself, the stage where a flaw can be built into the weights long before a runtime filter sees a single prompt.

"Every one of those models needs to be secured."
Nikesh Arora, Palo Alto Networks
Key takeaway
A filter in front of a model sees prompts and answers. It cannot see what a fine-tuning run changed inside the weights, so Palo Alto Networks is buying a seat earlier in the model's life.

Fine-tuning puts model risk on your team

Once your developers download weights and train them on internal data, the lab that released the base model no longer controls what ships. Researchers from Princeton, Virginia Tech, IBM and Stanford showed in 2023 that a small fine-tuning job can strip safety behavior out of an aligned model, even when the people running it mean no harm. A poisoned dataset or a swapped adapter file can change behavior the same way, and a prompt filter downstream has no record of either.

Three steps for your team before the deal closes:

  • List every open-weight model you run: the base checkpoint, the training data, each adapter, and who approved its move to production. You need that inventory whichever vendor you choose.
  • Ask your Palo Alto Networks account team which security controls ship inside the Applied Compute platform at close, and which arrive later through integration with its existing AI security products.
  • If you use Applied Compute today, ask how pricing, sovereign deployment options and support terms change under new ownership before your next renewal.

A security vendor that trains its own models owes you a way to check them

Palo Alto Networks will put the models its new team trains inside the products that inspect your traffic and your agents. You will rely on AI the vendor built to judge AI you built, and that calls for the evidence you would ask of any model in production: how the vendor tested it and how often it retests.

Arora said Unit 42 ran a Frontier AI Assessment on Applied Compute before the deal, so Palo Alto Networks already has a review method for outside platforms. Asking how that same method applies to the models Palo Alto Networks trains for itself is a fair first question for your account team. The company planned to share more detail on its fiscal first quarter 2027 earnings call, held the same afternoon as the announcement.

CIO/CTO Viability Question

Your security vendor will soon train its own models and sell you a platform to train yours. Before fiscal Q3 2027, ask Palo Alto Networks to show you the fine-tuning audit trail you will get at close, from training data lineage to post-tune safety test results. If it cannot produce that trail, keep your open-weight training pipeline under your own controls until it can.

Works Cited

Nieva, Richard. "Meet the Ex-OpenAI Guys Building a Cheaper Open Source Alternative." Forbes Australia, 3 Sept. 2026, www.forbes.com.au/news/innovation/meet-the-ex-openai-guys-building-a-cheaper-open-source-alternative/.

Palo Alto Networks. Press Release on the Acquisition of Applied Compute, Exhibit 99.2 to Form 8-K. U.S. Securities and Exchange Commission, 18 Nov. 2026, www.sec.gov/Archives/edgar/data/0001327567/000132756726000030/ex992pressrelease_november.htm.

Qi, Xiangyu, et al. "Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To!" arXiv, 5 Oct. 2023, arxiv.org.

Disclaimer: This blog reflects my personal views only. Content does not represent the views of my employer, Infotech. AI tools may have been used for brevity, structure, or research support. Please independently verify any information before relying on it.