Confirmed 18 Nov 2026 · Labs A one-minute read

A firewall company is paying $4.5 billion in stock to train its own models.

Palo Alto Networks agreed to buy Applied Compute, which post-trains and serves open-weight models, in an all-stock deal. Caveat: it closes in the fiscal third quarter of 2027 and needs regulatory clearance.

Until now

Security vendors sell tools that watch the models other companies run.

Now

One of them will own the training stack and run models itself.

Why it matters

Applied Compute is 13 months old and was valued at $3.25 billion two weeks ago. Palo Alto is paying 38% more in its own stock and making the founder a Co-Chief AI Officer. The company that sells enterprises their perimeter has decided the perimeter now includes the weights.

Established

  • A definitive agreement values Applied Compute at $4.5 billion, all in stock.
  • Yash Patil becomes Co-Chief AI Officer at closing; the research team transfers.
  • Palo Alto says it will train its own large and small models.

Not established

  • The share count. The release gives a value, not a number of shares.
  • That it closes. Regulatory clearance is a condition, not a formality.
  • Applied Compute's revenue. The $100 million run rate is reported, not filed.
The five-minute readThe record, the timeline, the weight and 4 sources

The record

Palo Alto Networks furnished the announcement as Exhibit 99.2 to a Form 8-K read at sec.gov, accession 0001327567-26-000030, accepted November 18, 2026 at 16:37:04, period of report November 17, 2026. The same filing reports Item 2.02 results of operations, Item 3.02 unregistered sales of equity securities, Item 7.01 Regulation FD, Item 8.01 other events and Item 9.01. The release, dated November 18, 2026, says Palo Alto Networks, Nasdaq ticker PANW, announced "a definitive agreement to acquire Applied Compute", described as a platform for post-training and inference of open-weight AI models. Consideration is all stock, valued at $4.5 billion, subject to adjustments; no share count is given. Closing is expected in Palo Alto Networks' fiscal third quarter of 2027, subject to customary conditions including regulatory clearance. The Applied Compute research team joins Palo Alto Networks, and co-founder and chief executive Yash Patil becomes Co-Chief AI Officer after closing. Chairman and chief executive Nikesh Arora is quoted: "Every enterprise is going to run a mix of open and closed models." Arora says the company's Unit 42 Frontier AI Assessment confirmed Applied Compute's security foundations. The release says Palo Alto Networks will use the team to train its proprietary large and small models, and will offer secure training, adaptation and deployment of open-weight models in public clouds or sovereign environments. The fiscal first quarter 2027 earnings call was set for November 18, 2026 at 1:30pm Pacific Time. Reported and not in the filing: Applied Compute was founded in 2025 by Yash Patil, Rhythm Garg and Linden Li, all former OpenAI researchers; it has raised $510 million; it closed $350 million at a $3.25 billion valuation in early November 2026, led by Elad Gil with Nvidia, Kleiner Perkins, Lux Capital, Blackstone, Shopify Ventures and HubSpot Ventures; annualised revenue passed $100 million inside 18 months; and customers include Microsoft and DoorDash. The $4.5 billion price is 38% above the $3.25 billion mark. Not established: the share count, the adjustment mechanics, retention terms, and whether an antitrust review follows.

How it got here

  1. Oct 2025

    Applied Compute launches with $80 million, founded by 3 former OpenAI researchers.

  2. Nov 3

    Applied Compute closes $350 million at a $3.25 billion valuation, led by Elad Gil, with Nvidia in the round. Reported annualised revenue above $100 million.

  3. Nov 12

    Palo Alto Networks' Unit 42 says attackers used Chinese AI models to breach 10 organizations. Edition 089's security card.

  4. Nov 17

    OpenAI ships the first model it rates High for improving AI itself, and Microsoft ships MAI Code 2. Edition 100's labs card.

  5. Nov 18

    Palo Alto Networks announces the $4.5 billion all-stock agreement after the American close, alongside fiscal first quarter results. Patil becomes Co-Chief AI Officer at closing.

Who has to react

CrowdStrike, Fortinet, Zscaler and Microsoft Security, now competing with a model trainer; Nvidia, Kleiner Perkins, Lux, Blackstone, Shopify and HubSpot, holding a two-week-old mark repriced 38% higher in PANW paper; Applied Compute's customers Microsoft and DoorDash; antitrust reviewers; and every post-training startup whose exit now has a comparable.

What to watch

  • Would strengthen itPalo Alto Networks ships a product that trains or serves customer-owned open-weight models, or a second security vendor above $1 billion of revenue buys a model-training company, by September 30, 2027.
  • Would weaken itThe agreement is terminated, or Applied Compute is run as a standalone unit with no Palo Alto product integration announced, by June 30, 2028.
  • Would strengthen itPalo Alto Networks discloses Applied Compute revenue at or above $100 million for a reported period without naming a security product built on it, by June 30, 2028.

Why it weighs this much

Signal strength
CostMedium
CommitmentHigh
ReversibilityModerate
PersistenceStructural

Cost is medium, not high: the price is paid in PANW stock, so no cash leaves, and the real cost is dilution plus whatever the adjustments allow. Commitment is high. A definitive agreement furnished to the SEC, with a named officer title and a transferring research team, is not walked back quietly. Reversibility is moderate because regulatory clearance is a condition, definitive agreements do collapse, and an all-stock price moves with the buyer's shares. Persistence is structural if it closes: a security incumbent that trains models does not go back to only selling tools. We hold cost to medium because the currency is paper.

The words

Post-training
Everything done to a model after the expensive base training: fine-tuning, reinforcement learning and alignment on a customer's own data. It is where most enterprise value now sits.
Open-weight model
A model whose parameters you can download and run yourself, rather than calling someone's service. The operator, not the lab, controls where it runs.
All-stock deal
The buyer pays in its own shares rather than cash. No money leaves today; existing holders are diluted instead.
Inference
Running a trained model to produce answers. It is the recurring cost of AI, as against the one-off cost of training.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.

8 contributions. Thank you.

You have built
a record here.

Keep it: create an account to verify your contributions and put your name to the answers you choose. Your record stays yours.

Keep my record ↗
Optional: help show whose perspectives are missing
Kept with your answers under a random id, never with your name. Used for one thing: showing you, on your record, where you stand next to people like you, once at least 20 of them have answered.

Add your signature

Stand behind your judgment.

Create an account to attach your name to this answer. Your other answers show no name unless you sign them too.

Sign up to add your name ↗

Already have an account?