Palo Alto Networks Buys Applied Compute for $4.5 Billion — and Plans to Train and Secure Open Models Itself

Cybersecurity company Palo Alto Networks is moving from securing other people's AI to training and running its own open models: the company has agreed to buy the startup Applied Compute for roughly $4.5 billion in an all-stock transaction.

Illustration: an open wire-mesh cage of server modules being sealed by a large dark enclosure, a metaphor for Palo Alto Networks training and securing its own open AI models after acquiring Applied Compute.
Illustration

Palo Alto Networks Buys Applied Compute for $4.5 Billion — and Plans to Train and Secure Open Models Itself

Cybersecurity company Palo Alto Networks is moving from securing other people's AI to training and running its own open models: the company has agreed to buy the startup Applied Compute for roughly $4.5 billion in an all-stock transaction. The acquisition was announced alongside strong fiscal first-quarter 2027 results, and the stock rose about 6 percent on Thursday. But the deal also raises an unanswered question: how independent is a security control performed by the same vendor that trained the model?

The Deal

Palo Alto Networks has, according to citybiz, agreed to acquire Applied Compute for approximately $4.5 billion in a pure stock transaction, with expected closing in the company's fiscal third quarter of 2027 — that is, February through April. Once the deal closes, Applied Compute's team will join Palo Alto, and co-founder and CEO Yash Patil will become the company's co-chief AI officer.

The purchase comes barely three weeks after Applied Compute in early November raised $350 million at a $3.25 billion valuation, in a round led by investor Elad Gil. The purchase price amounts to an uplift of roughly 38 percent from that round, but citybiz itself notes that the funding round and the acquisition are different transaction structures, so the figures cannot be compared cleanly.

Palo Alto's chairman and CEO Nikesh Arora did not quantify the valuation gap, but said according to citybiz: "Applied Compute has built the best platform for post-training and inference we have seen, and the research team is extraordinary. We saw that up close as a potential customer."

Who Is Applied Compute?

Applied Compute was founded by three former OpenAI researchers — Yash Patil, Rhythm Garg and Linden Li, all 25 or younger — and has according to BeInsure (based on Forbes' reporting) exceeded $100 million in annualized revenue after roughly 18 months of operation. Moneycontrol describes Patil, who is 23 and dropped out of Stanford, as an Indian-American entrepreneur who "has had quite a hectic month."

The company's customers are said to include Microsoft, DoorDash and Harvey. Patil claims some of the company's specialized models can be run at one-tenth the cost of frontier models from leading AI developers — but these are the company's own claims about application-specific results, not independently verified figures. The same applies to the $100 million annualized revenue figure.

The Numbers That Came With the Quarterly Report

The acquisition was layered on top of a strong quarter. Palo Alto reported fiscal first-quarter 2027 results on November 19: revenue grew 39 percent year over year to $3.44 billion, against analyst estimates of $3.31 billion, according to Yahoo Finance. Adjusted earnings per share came in at $1.05 versus $0.97 expected. The company also raised its full-year fiscal 2027 revenue guidance to $14.52–14.57 billion.

For the new generation of security services — the business the acquisition is meant to strengthen — Rosenblatt analyst Catharine Trebnick reported ARR growth of 66 percent to $9.73 billion, $170 million above the top of guidance, according to MSN/Benzinga.

The stock market responded: CNBC reported a 6 percent gain on Thursday, driven by both the quarterly results and the acquisition. Other sources (Yahoo Finance and citybiz) reported around 8 percent after market close — a minor discrepancy, likely due to different measurement points in time.

The Strategy: Secure, Sovereign Open Models

Palo Alto is now moving into model training and inference operations from a company that historically sold firewalls and security platforms. On the earnings call, Arora laid out a strategy in which the company will help train and operate secure, sovereign implementations of open models — models whose weights are publicly available and can be run locally by enterprises and states that do not want to send data to a frontier-model vendor, according to Network World.

Arora claimed that the share of token traffic on open models had grown from under 5 percent to over 60 percent in a year — but presented no source for the figure, something Network World also points out. The number should therefore be treated as the company's own, unsourced claim.

Concretely, Arora said the company is targeting more than 100 megawatts of compute capacity over the next three to five years, with spending mainly on GPUs and servers, according to CFO Dipak Golechha. DA Davidson analyst Rudy Kessinger wrote that the company "expects to build out more than 100 MW of global inference capacity over the next 3–5 years at an expected cost of under $3 billion in capital expenditures, with $1 billion in NGS ARR expected by fiscal year 2029."

CNBC reports that Palo Alto says the acquisition is expected to add more than $1 billion in annual recurring revenue to the next-generation security business by fiscal year 2029. Arora told CNBC: "Instead of being one of ten players in the industry delivering compute, we will be the only player that, together with Palo Alto, delivers secure compute and security capability." According to BeInsure, Arora presented to analysts the ambition as a business that could become larger than the company's existing cybersecurity operations — a company statement about the future, not a measured result. A company statement carried by Network World adds that Palo Alto will help enterprises adopt AI securely, whether they choose open models, closed frontier models, or both.

The Counterargument: The Trainer as Overseer

The strategy has met governance criticism. Sanchit Vir Gogia of Greyhound Research warned of the concentration risk in a vendor that both trains and secures the same models: an overseer trained by the same hands as the system it oversees is, in his formulation, "a colleague, not a check." He added that CIOs should ensure that someone other than the trainer tests the outcome, according to Network World.

For CIOs considering secure AI stacks from a single vendor, this means an in-house security layer cannot automatically be counted as an independent control — especially if the model weights themselves come from the same vendor. This tension sits at the heart of the deal: Palo Alto's sales pitch is precisely that secure compute and secure models from one vendor are a strength, while the classic security argument — separation of duties — says the opposite. The available evidence does not resolve which view enterprises will ultimately weigh most heavily.

What Remains to Be Answered

Several elements should be handled with care. The acquisition has been announced but not completed; closing is expected in the fiscal third quarter of 2027. The financial targets — $1 billion in NGS ARR by fiscal year 2029, more than 100 MW of inference capacity, capital expenditures under $3 billion — are the company's own projections. Applied Compute's revenue and cost claims are company figures or Forbes-derived reporting, not audited numbers, and Arora's statistics on open-model traffic lack any stated source. How Palo Alto will manage its dual role as model trainer and security overseer remains, so far, unanswered.

AIMag.no
AIMag.no
The AIMag.no editorial team covers artificial intelligence, tools, research, and regulation.

Get the best of AI MAG in your inbox

News, analysis, and ideas at the intersection of AI and society.