AI / Acquisitions

🔥 Palo Alto Just Paid $4.5 Billion for a Startup Worth $100 Million 18 Months Ago

Palo Alto Networks is buying Applied Compute, an 18-month-old startup from three ex-OpenAI researchers, for $4.5bn all-stock.

Palo Alto Just Paid $4.5 Billion for a Startup Worth $100 Million 18 Months Ago — Tech Arcade
Photo: Aleksandr Lyaptsev / Unsplash

Three former OpenAI researchers, barely out of Stanford, started a company a year and a half ago. This week one of the biggest names in cybersecurity agreed to buy it for more than the GDP of a small nation.

Let’s get the number on the table first: $4.5 billion, all-stock, for a company called Applied Compute (Palo Alto Networks 8-K). Palo Alto Networks announced the deal on 18 November 2026, and it expects to close in its fiscal third quarter of 2027, pending regulatory clearance.

Here is the part that makes your eyebrows climb. When Applied Compute took its first cheque in mid-2025, insiders pegged it at roughly $100 million (Upstarts, via FourWeekMBA). So the price tag has gone up about 45 times in the time it takes most startups to agree on a logo.

The thesis: this is the clearest sign yet that in AI, a small team with the right pedigree and a usable product is now worth more than almost any amount of revenue it has actually booked.

🧠 Why This Matters

Applied Compute does something unglamorous and increasingly valuable. It takes open-weight AI models (the kind you can download and run yourself, as opposed to the sealed boxes from OpenAI or Anthropic) and fine-tunes them on a company’s own data, so the result behaves like a specialist rather than a know-it-all. It helps customers build custom models and agents “in days instead of the months usually required,” per the company’s own pitch (SiliconANGLE).

That is now a strategic asset, not a side feature. Palo Alto’s chairman and CEO Nikesh Arora is betting that every large company will soon run a patchwork of models and will need someone to secure all of them.

“Every enterprise is going to run a mix of open and closed models, and every one of those models needs to be secured.” — Nikesh Arora, Chairman and CEO, Palo Alto Networks (8-K)

Arora claims open-weight models jumped from under 5% of token traffic to over 60% in a single year (Network World). He did not cite a source for that figure, so treat it as a sales pitch with a number attached. But the direction is real, and it explains why a security company suddenly wants to be in the model-training business.

📊 Deep Dive

Applied Compute was founded by three ex-OpenAI technical staff: Yash Patil, who worked on the Codex coding assistant and is now CEO; Rhythm Garg, who helped build the o1 reasoning model; and Linden Li, who worked on training infrastructure (SiliconANGLE). Its customer list reads like a tech conference speaker lineup: DoorDash, Shopify, Mercor, Harvey, Cognition.

Here is the valuation climb, as reported (the later figures are company-stated and unaudited, so hold them loosely):

  • Mid-2025: roughly $100M valuation on an early round led by Benchmark
  • 2026: an $80M raise with Sequoia, Lux and others joining
  • November 2026: the company said it raised $350M at a $3.25 billion valuation, led by investor Elad Gil
  • 18 November 2026: acquired for $4.5 billion by Palo Alto Networks

The company also says it hit $100 million in annualised revenue run rate less than 18 months after launch (company post, via FourWeekMBA). If that holds, the buyers paid roughly 45 times run-rate revenue. For a company this young, that is less a multiple and more an act of faith.

⚠️ The Catch

A $4.5 billion all-stock deal is not cash in a founder’s pocket on day one, and the strategic logic has sceptics. Sanchit Vir Gogia, chief analyst at Greyhound Research, pointed out the obvious tension in asking a security vendor to both build AI models and police them.

“Palo Alto is asking enterprises to let the guard build the vault it guards.” — Sanchit Vir Gogia, Greyhound Research (Network World)

His point lands. Most AI failures happen at runtime, through prompts, tools and permissions, regardless of who trained the model. Buying a model-training shop does not automatically fix that, and it hands one company both the lock and the key. Gogia’s advice to customers was blunt: get a costed exit in writing, insist on owning your trained model weights in an exportable format, and demand no training on your data without consent.

There is also the small matter of integration. Palo Alto is still digesting its $25 billion acquisition of identity firm CyberArk (Network World). Bolting a GPU-hungry model business onto a software company is a different kind of problem.

🎯 What Happens Next

Once the deal closes, Yash Patil becomes Palo Alto’s Co-Chief AI Officer, and his research team folds into the broader product line. Finance chief Dipak Golechha has said most of the new compute spend will go on GPUs and servers, with Palo Alto renting from cloud partners first and building dedicated capacity over time. Arora wants more than 100 megawatts of compute within three to five years, and floated that the new business “could be bigger than what we do today” (Network World).

Watch for regulators too. All-stock mega-deals clear more easily than cash ones, but a cybersecurity giant absorbing an AI-infrastructure startup will draw a look.

🧩 Bigger Picture

This is the acqui-hire taken to its logical extreme. Big incumbents have decided that the fastest way to own AI talent and technology is to buy whole companies before they have had time to prove a durable business. The ex-OpenAI diaspora is the hottest commodity in the market, and a Stanford dorm’s worth of researchers can now command the price of a mid-cap industrial firm.

For founders, the lesson is intoxicating and a little unnerving: in this cycle, pedigree plus a working product plus 18 months can equal several billion dollars. For everyone else, it is a reminder of how much of AI’s value sits in a very small number of heads, and how aggressively the giants will pay to rent them.

Applied Compute’s three founders left OpenAI to build an alternative to it. Eighteen months later, they sold the alternative to a firewall company for four and a half billion dollars. In this market, that counts as patience.


Sources

❓ Quick answers

How much is Palo Alto Networks paying for Applied Compute?

$4.5 billion in an all-stock deal, announced on 18 November 2026, expected to close in Palo Alto's fiscal Q3 2027 (source: Palo Alto Networks 8-K).

What does Applied Compute do?

It runs a platform for post-training and inference of open-weight AI models, fine-tuning them on a customer's own data; it was founded by three former OpenAI researchers including CEO Yash Patil.

Why is Palo Alto Networks buying an AI model startup?

CEO Nikesh Arora argues every enterprise will run a mix of open and closed models that all need securing; after close, Yash Patil becomes Palo Alto's Co-Chief AI Officer (source: Network World, 18 Nov 2026).