Palo Alto Networks Buys Applied Compute to Secure Open-Weight AI Models

| 2 Min Read
Palo Alto Networks is buying Applied Compute for $4.5B in stock to train, host and secure open-weight models, adding compute and a supervisory layer.

Palo Alto Networks wants to own more of the AI stack, not just the security layer that watches it. Its $4.5 billion all-stock purchase of Applied Compute, a platform for post-training and inference of open-weight models, is the clearest sign yet. The company says the deal will let it train, deploy and secure these models for enterprises, whether they run in the public cloud or in their own sovereign environments.

The deal is expected to close in Palo Alto's fiscal third quarter of 2027, which runs from February to April. Applied Compute co-founder and CEO Yash Patil becomes co-chief AI officer, and his research team will help build models across Palo Alto's product lines. The startup will keep serving AI-native companies, digital natives and large enterprises.

The pitch is that enterprises can adopt AI securely whether they pick open-weight models, closed frontier models, or both. The company's statement, released with its fiscal first quarter 2027 results, describes Applied Compute as a leading platform for post-training and inference of open-weight models.

From security vendor to model host

Chairman and CEO Nikesh Arora said on the earnings call that beyond building AI into its own products, "we will help train and serve secure sovereign implementations of open-weight models to the enterprise." He also said open-weight models rose from under 5% of token traffic to more than 60% over the past year. He didn't say where that figure came from, so treat it as a company claim until someone verifies it.

Asked by JP Morgan analyst Brian Essex about customer demand, Arora argued that deterministic security controls "cannot anticipate every scenario that an agent is going to conduct." Enterprises building AI applications "will have to train a supervisory AI model at the same time." His core point: Palo Alto can't secure models it had no part in building. "I cannot go in and intercept and understand somebody else's application," he said.

Sanchit Vir Gogia, chief analyst at Greyhound Research, thinks Arora has a point. A generic filter can't know how a particular application is supposed to behave. Still, most failures in AI applications happen at runtime, through prompts, tools and permissions, regardless of who trained the model. Palo Alto has already sold model-agnostic protection through Prisma AIRS since 2025, which makes this acquisition more about owning the model side than filling a gap.

Gogia's sharpest line: "A supervisor trained by the same hands as the system it supervises is a colleague, not a check." His practical advice for CIOs is to make sure someone other than the trainer tests the result.

Compute and the sovereignty question

Training and hosting these models means buying hardware. CFO Dipak Golechha said on the call that the spending "will mostly go to GPUs and servers," with a smaller share going to the facilities that house them. Palo Alto will keep sourcing from its cloud partners while adding dedicated capacity over time. Arora put the target at more than 100 megawatts of compute within three to five years.

Gogia notes that Palo Alto already has sovereignty credentials for its security data, including customer-held keys and a European Cortex service governed by Deutsche Telekom. "None of that yet extends to GPUs," he said. A US-headquartered operator also faces the same jurisdictional questions as the hyperscalers it rents capacity from, which matters if buyers care where their weights and inference actually run.

"A sovereign hosting offer with no named region, operator or date is a promise looking for an address," Gogia said. His checklist for compliance teams is simple: ask for a region list, the operator, and a dated certification plan. Those three answers will tell you more than the marketing will.

Arora went further on the call, saying the training and inference business "could be bigger than what we do today." That's a bold claim for a company whose core market is network and security tooling. Here's the thing, though: if the GPU buildout lands, the hardware is the easy part. The harder question is whether customers trust the same vendor to host models and guard them.

One vendor, more of the stack

The deal follows Palo Alto's $25 billion purchase of CyberArk, which added identity security to its platform. Gogia says Applied Compute would now add the models themselves to a supplier already sitting at the center of many customers' firewalls, security operations centers and identity controls.

"Palo Alto is asking enterprises to let the guard build the vault it guards," he said. One team watching both the model and the telemetry can spot trouble sooner. But a single fault, breach or pricing dispute would land "on the protection and the protected together."

His advice for CIOs is blunt. Get a costed exit in writing before any model work moves across. Applied Compute customers should use the months before closing to lock in ownership of trained weights in an exportable format, no training on their data without consent, a right to exit on change of control, and price protection that survives bundling. "Terms are cheapest to win while the suitor is still courting," Gogia said.

Watch the next two quarters. If Palo Alto names regions, operators and dates for its sovereign hosting, the bundled model-and-security pitch gets real. If it doesn't, enterprises will be buying a promise.

Source: Christopher Davis · www.csoonline.com

Comments

Please sign in to comment.