On Tuesday, November 18, Palo Alto Networks announced it will acquire Applied Compute, an 18-month-old startup that builds a platform for post-training and inference of open-weight AI models. The price is $4.5 billion, all in stock. The target’s reported revenue is a $100 million annualized run rate — company-reported, not audited. That is roughly 45 times the run rate, for a company founded in May 2025.
The numbers are the hook, but they are not the story. The story is what Palo Alto Networks is buying its way into: not another security product, but the layer underneath the AI that enterprises are about to run — the training, the serving, and the improving of models tuned to a company’s own data. For a CISO, that matters twice over. Your security vendor is about to also be your AI platform vendor. This post is the deal, the product, the team, the money, and the questions that decide whether it works.
The deal
Palo Alto Networks entered into a definitive agreement on November 18, 2026 to acquire all of Applied Compute’s outstanding stock, warrants, and other equity rights for $4.5 billion, subject to adjustments, paid entirely in Palo Alto Networks common stock.
The share count is set by a formula, not a fixed number. Shares issued at closing will be priced at $337.16 — the volume-weighted average closing price over the 15 consecutive trading days ending November 17. At that price, $4.5 billion works out to roughly 13.35 million shares, about 1.6% of Palo Alto Networks’ shares outstanding. The shares go only to accredited investors, as a private placement, with no registration statement.
Close is expected in Palo Alto Networks’ fiscal third quarter of 2027 — the quarter ending April 30, 2027 — subject to customary conditions, including expiry of the waiting period under the Hart-Scott-Rodino Antitrust Improvements Act. Two things worth noting about what is not public: the merger agreement itself was not attached to the 8-K, so there is no disclosed breakup fee, outside date, or adjustment mechanics.
For scale: $4.5 billion is about 38% above the $3.25 billion valuation Applied Compute set just two weeks earlier, more than three times its April 2026 valuation, and roughly 45 times its reported revenue run rate. By any standard multiple, the price is rich. Palo Alto Networks is not paying for what Applied Compute earns today. It is paying for what it believes the open-weight layer will be worth.
What Applied Compute actually does
Applied Compute was founded in May 2025 in San Francisco by three former OpenAI researchers who had been friends at Stanford before working together at OpenAI. The product is a platform that takes strong open-weight models and adapts them to a customer’s own data, then serves them in production — and keeps improving them on live production feedback.
The core product is AC2 (Applied Compute Agent Cloud), introduced in August 2026. It bundles training, inference, and continual improvement in one system: customers fine-tune custom models on their own data, serve them on a dedicated autoscaling endpoint, and keep improving them as production data comes in. The technical pieces that matter are asynchronous reinforcement learning (data generation and weight adjustment run simultaneously), self-distillation (no separate teacher model), decision-model training, evaluation tooling, and enterprise controls — VPC deployments, role-based access, audit logs. The company calls the result “specific intelligence” — retraining open-weight models on a company’s data and ways of working, rather than renting a general model by the token.
The second product is Shift-1, launched October 21, 2026: a frontier model router for coding and knowledge work, trained on AC2. It routes each request — not once per task — between a stronger and a cheaper model, using a custom 4-billion-parameter decision model trained on roughly 150,000 coding-agent states. The claimed savings are 55% lower Codex spend versus GPT-6 Astra while preserving SWE-bench Verified accuracy, and 24% lower Claude Code spend versus Opus 5.5. One honest flag: the company’s own pages disagree on the Claude number — the launch post and research overview say 24%, the routers page says 35%. I am using 24% because it appears in two places, and the disagreement itself is worth knowing.
The models at the center of the stack are open-weight families: Qwen, Kimi, GLM, DeepSeek, Nemotron, Beam. Nvidia has used Applied Compute to post-train its own open-weight Nemotron models — and Nvidia is both a customer and an investor. That detail matters, because it means the platform’s credibility was established by the company whose hardware the whole stack runs on.
The team
The founders are the asset Palo Alto Networks keeps naming. Yash Patil, co-founder and CEO, worked at OpenAI on post-training infrastructure and the Codex programming assistant. Rhythm Garg, co-founder and CTO, helped build the o1 reasoning model. Linden Li, co-founder and Chief Architect, designed large-scale training infrastructure for reinforcement learning. About 70 employees will join Palo Alto Networks, per the earnings presentation.
After the close, Patil becomes Palo Alto Networks’ Co-Chief AI Officer, and the research team will help embed AI across every Palo Alto Networks offering and train its proprietary large and small AI models — the company says it currently has more than 35 AI models in training. Two things have not been announced: who the other Co-Chief AI Officer is, and what roles Garg and Li take. On the first, the only public evidence is that CEO Nikesh Arora has described himself as the company’s Chief AI Officer — which makes him the likely counterpart, but it is not confirmed. On the second, the press release says only that the Applied Compute team will join. For a deal priced on a research team, those two silences are worth watching.
The money
The funding path is the steepest part of the story. Seed: $20 million in June 2025, led by Benchmark, at a roughly $100 million valuation — press-reported, not a company disclosure. Then $80 million in October 2025 as the company exited stealth. Then a $80 million Series B on April 8, 2026, at a $1.3 billion post-money valuation, led by Kleiner Perkins. Then $350 million on November 2, 2026, at a $3.25 billion valuation, led by Elad Gil, with Kleiner Perkins, Lux Capital, Hanabi, Nvidia, Blackstone, Shopify Ventures, and HubSpot Ventures among the participants. By public tallies, roughly $530 million raised in 18 months.
The valuation path runs about $100 million, to $1.3 billion, to $3.25 billion, to a $4.5 billion acquisition — a 2.5x step-up in the seven months before the deal, on the back of a revenue figure that needs careful handling. The company said on November 2 that it had grown “from zero to $100 million in annualized revenue run rate” in under 18 months. A run rate annualizes a recent period; it is not revenue recognized over a completed year. No audited revenue, profitability, gross margin, or retention figure has been disclosed anywhere. Do not read $100 million as annual revenue, because it is not that.
What is disclosed about the business: named customers include Microsoft, Shopify, Mercor, Harvey, DoorDash, Cognition, Clay, and Decagon, across life sciences, cybersecurity, legal, financial services, and logistics. Two case studies carry the commercial argument. Harvey, the legal AI company, had a model trained for its Review Table product that improved answer and citation quality while costing less than half what a frontier alternative charged at inference. Clay, the sales-data firm, trained and deployed a router that cut production inference costs by up to 47% without a quality loss. The scale signals: enterprise count more than doubled since the last fundraise, 1,700-plus training runs since July, more than 1 trillion tokens a day, and about 40% of customers using the platform for both training and inference.
Why Palo Alto Networks is buying
The stated rationale is a bet on a hybrid future. From the press release: “Palo Alto Networks firmly believes the future of enterprise AI is hybrid, and that a large share of enterprises will run on open-source and open-weight models tuned to their own data, cost, and control needs. Today, no large provider is delivering those models securely. With Applied Compute, Palo Alto Networks intends to be that partner.” Arora’s version, in the same release: “Every enterprise is going to run a mix of open and closed models, and every one of those models needs to be secured. Applied Compute has built the best post-training and inference platform we’ve seen, and its research team is exceptional. We saw that firsthand as a prospective customer, and our Unit 42 Frontier AI Assessment confirmed their strong security foundations.”
And on the earnings call, Arora went further — beyond building AI into its own security products, the company will “help train and serve secure sovereign implementations of open-weight models to the enterprise,” and told CNBC that if it works, the opportunity could “dwarf the ability that we have in cybersecurity.”
The buyer can afford the bet. Fiscal 2026 revenue was $11.48 billion, up 24.5%, though net income fell to $307 million, down 72.9% — the cost of the acquisition pace. First-quarter fiscal 2027 revenue was $3.445 billion, up 39% year over year, ahead of expectations, with next-generation security ARR at $9.73 billion, up 66% — though the headline growth includes CyberArk, acquired in February 2026. Guidance was raised the same day: fiscal 2027 revenue of $14.52–$14.57 billion. The AI stack it is bolting Applied Compute onto is already substantial: Precision AI across the platforms, Prisma AIRS at nearly $170 million in ARR five quarters after general availability, Cortex XSIAM at roughly $800 million in ARR, and Idira (the rebranded CyberArk identity business). This is the third big AI-adjacent acquisition in eighteen months, after Protect AI in July 2025 and Chronosphere for $3.35 billion in January 2026.
The potential
Applied Compute will continue to serve AI-native companies, digital natives, and large enterprises, and Palo Alto Networks will bring the platform to its global customer base of more than 70,000 customers. That distribution is the growth case: a platform that sold to a handful of sophisticated AI-native buyers now gets sold to the largest enterprise security customer base in the industry.
The targets attached to the deal, from the earnings presentation, are specific. Palo Alto Networks expects the acquisition to add $1 billion in annual recurring revenue from its next-generation security business by fiscal 2029. It aims to establish more than 100 megawatts of global inference capacity within three to five years, using cloud partners and its own infrastructure, with projected capital spending of less than $3 billion between fiscal 2027 and 2030. Management estimates the deal could reduce operating margin by up to 100 basis points in fiscal 2027 and 200 basis points in fiscal 2028.
The market’s first verdict was positive. Shares rose about 10% in the regular session on Wednesday, November 18, with rivals rising in sympathy — CrowdStrike up about 5%, Zscaler about 3%, Fortinet about 2%. TD Cowen’s Shaul Eyal called it an extension of “Palo Alto’s push beyond traditional cybersecurity into enterprise AI.” IDC’s Frank Dickson put the tension plainly: “Let’s face it, the price is rich,” but “this is a smart bet” — and framed the deal as the mirror image of the big-IT-buys-security trend: “Google bought Wiz. ServiceNow bought Armis. Cisco bought Splunk. Now Palo Alto Networks is buying its way into their AI stack.”
The questions
A $4.5 billion price for a $100 million run rate is the first question, and it does not go away by repeating the growth story. The multiple is roughly 45 times a company-reported, unaudited run rate — a figure that annualizes a recent period and says nothing about profitability, margin, or retention. Palo Alto Networks is paying for the team and the trajectory, and it is paying the top of the market for both.
The second question is the stack itself. The open-weight models at the core of Applied Compute’s platform include Chinese-origin models — Qwen and Kimi among them — which are drawing security scrutiny and export-control attention in Washington. A platform whose value is “run any open model, tuned to your data” inherits the provenance question for every model it touches. Palo Alto Networks’ answer is that security is the point of the acquisition — Unit 42 assessed the foundations, and the platform ships with VPC deployments, access controls, and audit logs.
The third question is execution. Palo Alto Networks is simultaneously integrating Chronosphere and CyberArk, both acquired in the first two months of 2026, while closing this deal in fiscal Q3 2027 after an HSR waiting period that is still pending. The 8-K’s own risk factors name “the ongoing integration of other recent acquisitions.” The margin drag is the visible cost; the invisible one is management attention spread across three integrations and a 100-megawatt build-out. And because the merger agreement has not been filed, there is no public breakup fee or outside date if the deal stalls.
For a CISO, the question that matters most is the governance one. If your security vendor becomes your AI platform vendor, the model layer and the security layer stop being separate purchases and start being one relationship. That is the consolidation argument Palo Alto Networks is making, and it is a real one — fewer vendors, one throat to choke, security built into the training and serving layer rather than bolted on after. But it also means the questions you used to ask two vendors — where does my data go, which models touch it, who governs the router that decides which model answers — now get asked of one vendor, whose answer is its own platform. The diligence that matters is not whether the platform works. It is whether the governance around it is as good as the engineering inside it. Palo Alto Networks evaluated Applied Compute as a customer and had Unit 42 assess it. Ask for the same privilege: evaluate the combined offering as a customer, and assess the governance the way you would assess any vendor that holds your data and your models.
The close is set for the quarter ending April 30, 2027; the $1 billion ARR target for fiscal 2029. Between now and then, the thing to watch is not the share price or the press releases. It is whether the platform that won Harvey and Clay and Microsoft on its engineering keeps winning them inside a 70,000-customer security company — and whether the security that justified the price shows up in the product, not just in the announcement.





Leave a Reply