AI Trends

Palo Alto Networks Acquires Applied Compute for $4.5 Billion; Three Gen Z Founders Exit After 18 Months

36Kr · AIUpdated Nov 19 16:31

AI summary

Palo Alto Networks has entered into a definitive agreement to acquire AI infrastructure company Applied Compute in an all-stock transaction valued at $4.5 billion. Its team of approximately 70 people will join Palo Alto Networks, and 23-year-old co-founder and CEO Yash Patil will become co-chief AI officer. Through its AC2 platform, Applied Compute provides enterprise post-training and runs custom models. It has used Qwen and GLM to reduce customers’ inference costs, and its annualized revenue has reached $100 million. Palo Alto Networks plans to make its services available to more than 70,000 customers worldwide and has set a target of adding $1 billion in ARR by fiscal 2029.

Why read this

SelectedAI score 82

The acquisition demonstrates the commercial value of enterprise post-training services while also highlighting the trade-offs among on-site delivery, recurring subscriptions, and the security and regulatory compliance of open-source models.

Topics

Tags

Body · AI translation

As Anthropic heads toward an IPO, Silicon Valley is witnessing the birth of a new generation of wealth legends. Meanwhile, in another corner of the frenzy, three post-2000 founders who left OpenAI have already secured their own “big outcome”—less than 18 months after starting their business, they sold the company for $4.5 billion (about RMB 32 billion).

On November 18 U.S. Eastern Time, American cybersecurity giant Palo Alto Networks announced that it had entered into a definitive agreement to acquire AI infrastructure startup Applied Compute for $4.5 billion in an all-stock transaction. After the deal closes, the company’s roughly 70-person team will join Palo Alto Networks in its entirety, while 23-year-old co-founder and CEO Yash Patil will become the company’s co-chief AI officer.

The company has grown rapidly in recent months. Just over two weeks earlier, on November 2, Applied Compute had announced a $350 million funding round at a post-money valuation of $3.25 billion. Palo Alto Networks’ acquisition price was $1.25 billion above that valuation, representing a premium of about 38% after barely more than two weeks.

Ordinarily, building this kind of enterprise business requires both an understanding of models and familiarity with customers’ workflows, so founders are typically seasoned veterans with years of industry experience. But Applied Compute’s three founders—Yash Patil, Rhythm Garg, and Linden Li—are only 23, 24, and 25 years old.

From meeting at Stanford to leaving OpenAI together, they took less than a year and a half after founding the company to prove out this “big business” in a Silicon Valley crowded with giants.

From prospective customer to buyer—all because of a security assessment

When Palo Alto Networks first approached Applied Compute, it was not looking to make an acquisition. It simply wanted to see whether it could purchase the startup’s services.

Palo Alto Networks chairman and CEO Nikesh Arora revealed in the acquisition announcement that the company had initially contacted Applied Compute merely as a “prospective customer” seeking to evaluate its post-training and inference platform.

Following the professional instincts of a cybersecurity company, Palo Alto Networks specifically dispatched Unit 42, its elite threat intelligence team, to conduct a “frontier AI security assessment” of Applied Compute.

According to Arora, the assessment convinced them that Applied Compute had built the best post-training and inference platform they had ever seen and that its research team was exceptional. Palo Alto Networks therefore pivoted from procurement discussions to acquisition negotiations and moved to buy the entire company outright.

As one of the world’s largest cybersecurity companies, Palo Alto Networks currently has a market capitalization of nearly $280 billion. Its latest quarterly revenue reached $3.45 billion, up 39% year over year. Over the past year, the giant has been highly active in the M&A market, completing two acquisitions valued at $21 billion and $3.35 billion, respectively. Even within Palo Alto Networks’ acquisition portfolio, the $4.5 billion purchase of Applied Compute is no small deal.

According to documents filed with the U.S. Securities and Exchange Commission (SEC), the $4.5 billion transaction will be paid for in Palo Alto Networks stock, using the average share price over the preceding 15 trading days—$337.16 per share—as the reference price.

For Palo Alto Networks, the economics of the deal are not unfavorable. Applied Compute had just raised $350 million about two weeks earlier, and the transaction also includes approximately $500 million in cash and prepaid compute commitments held by the startup, while Palo Alto Networks will fund the entire purchase price by issuing shares. After the acquisition was announced alongside better-than-expected financial results, Palo Alto Networks’ stock surged about 10% in after-hours trading.

That a Silicon Valley giant would move so quickly to acquire a startup founded less than a year and a half ago, immediately after completing its assessment, was largely due to the technical foundations and entrepreneurial experience of its three young founders.

The three founders are Stanford alumni and former colleagues at OpenAI. Patil was the first to join OpenAI. While in college, he proactively emailed Altman seeking an opportunity to participate in the development of ChatGPT and later joined through the research residency program. He subsequently referred Garg, who in turn referred Li.

During their time at OpenAI, Patil worked on post-training infrastructure and Codex-related projects, Garg researched reasoning models, and Li was responsible for machine-learning systems. Together, the three covered multiple stages of model training and operation, forming the technical core of their future startup.

The rise of Chinese open-source models provided an important opportunity for them to strike out on their own.

DeepSeek-R1 was released in January 2025. At the time, the three young researchers were skiing at Lake Tahoe. As they later recalled, they spent the breaks between ski runs discussing the newly released Chinese model and the future of open models.

Drawing on their experience at OpenAI, they realized that because open models already possessed strong foundational capabilities, companies could continue training them to make the AI more familiar with their own business processes. A few months later, the three left OpenAI and founded Applied Compute.

Their business progress quickly won recognition from investors. In June 2025, before the company had even emerged from stealth, it secured $20 million in funding led by Benchmark. After raising $80 million in April this year, its valuation reached $1.3 billion. In August, it launched its core AC2 platform. By the time of its early-November funding round, customers had completed more than 1,700 training runs on AC2 since July, while its inference service was processing more than one trillion tokens per day.

Discussing the acquisition by Palo Alto Networks, 23-year-old CEO Patil said in a statement that they had founded the company to make it easy for any enterprise to train, run, and improve its own specialized models, and that joining Palo Alto Networks would put security at the heart of every product while accelerating access to some of the world’s largest enterprise customers.

Using Chinese models to save major US companies money and hassle

The key reason Palo Alto Networks was willing to pay the hefty price of $4.5 billion was that Applied Compute had zeroed in on the pain points large enterprises encounter when using AI.

On its website, Applied Compute sums up its business as “Specific Intelligence”: rather than having enterprises use general-purpose large models directly, it uses their own business data to train proprietary models tailored specifically to them.

For many large companies, frequently calling leading closed-source models from OpenAI or Anthropic is not only slow and expensive, but general-purpose models also lack an understanding of their specific internal standards. Banks, after all, have their own risk-control requirements, while food-delivery platforms have their own content-review rules.

More and more enterprises are now paying for model “post-training.” A recent Menlo Ventures survey found that the share of AI builders surveyed who had conducted post-training rose from 6% to 10% within a year, with 91% of them using external partners.

Applied Compute positioned itself as a “middleman” between large models and enterprises: off-the-shelf open models on one side and enterprises with specific needs on the other. It selects models, continues training them for customers’ tasks, and then helps customers deploy them.

Although it occupies the position of a middleman, the way its people work is more like an embedded team. Silicon Valley calls engineers who work deeply inside customer organizations FDEs, or forward-deployed engineers—a model first popularized by Palantir that has clearly gained traction in AI circles this year. Applied Compute’s engineers must go into customers’ companies and understand how they work before they know how to tune the models; the trained models ultimately run on its proprietary AC2 platform.

Chinese open-source models, meanwhile, gave Applied Compute an even greater edge as it helped major US companies drive costs down further. According to Menlo Ventures’ estimates, Chinese models accounted for 89% of the open-model-related spending it tracked.

On the one hand, enterprise customers urgently need affordable, capable open-source models to control the cost of handling massive volumes of tasks, and Chinese models such as Qwen, Kimi, and GLM have open weights and generally cost far less to use than leading US closed-source models. On the other hand, open-source models are merely foundational platforms when first adopted, while enterprises lack the talent and systems needed to train them—creating exactly the need for Applied Compute to help turn Chinese open-source models into tools tailored to their needs.

Judging from the case studies disclosed on Applied Compute’s website, Chinese open-source models have shouldered much of the workload across multiple projects.

Its partnership with US marketing software company Clay, for instance, targeted the cost of using AI. Clay’s AI products must gather large amounts of corporate information for users every day, and the questions involved vary widely in difficulty. Calling the most expensive model for every question would be far too costly.

Using Alibaba’s Qwen, Applied Compute trained an AI dispatcher responsible for “assigning tasks.” It first assesses the difficulty of a question and then selects the most cost-effective model to handle it. A/B testing in production showed that this routing approach reduced inference costs by as much as 47% while keeping key quality metrics stable.

At legal AI unicorn Harvey, Zhipu AI’s GLM was put to use instead.

Legal documents leave no room for error, and when AI searches large collections of documents for answers, it must also accurately identify the supporting provisions for lawyers. The two companies trained and evaluated GLM-5.2 using public legal materials and synthetic tasks. Evaluations showed that the trained model delivered better answers on specific legal-document tasks, while reducing the average inference cost per cell by 54.8% compared with Claude Sonnet 5.

With the cost-effective foundation provided by Chinese open-source models and the ability of its embedded engineers to fine-tune them, Applied Compute grew its annualized revenue to $100 million in just a year and a half, bringing it to Palo Alto Networks’ attention as an acquisition target.

Selling to a cybersecurity giant could resolve outsourcing controversy and security risks together

Its annualized revenue had already surged to $100 million, and it had raised $350 million just two weeks earlier, with its valuation still soaring.

Yet at this point, the three young founders chose to sell the company to Palo Alto Networks at a 38% premium. Will it prove to be a good home?

The first issue concerns controversy over its business model and the challenge of expanding its roster of major customers. Until the acquisition, outsiders continued to debate whether Applied Compute was truly an AI platform or an outsourcing services company that relied on engineers to take on individual projects.

To investors, these are two entirely different prospects. If it is the latter, every new customer requires a substantial number of engineers to learn the business and build the training workflow from scratch. As the customer base grows, staffing requirements swell with it, making the business decidedly unappealing.

On the other hand, it also remains uncertain whether the company can consistently sell profitable inference services after training customers’ models. The company previously disclosed that only 40% of its customers used both its training and inference services.

Joining Palo Alto Networks will at least broaden its sales channels. Under the plan, Applied Compute will be plugged directly into Palo Alto Networks’ global network of more than 70,000 customers. As for recurring subscriptions, Palo Alto Networks has set a clear target: by fiscal 2029, the business must generate $1 billion in additional annual recurring revenue (ARR), ten times its current annualized revenue of $100 million.

Second, a major company will now foot the bill for its cash-intensive computing investments. Palo Alto Networks plans to build more than 100 megawatts of global inference computing infrastructure over the next three to five years and spend less than $3 billion in capital expenditures (CapEx) on it between fiscal 2027 and fiscal 2030.

In return, besides continuing to sell the platform externally, the young team will also serve internally as Palo Alto Networks’ “AI R&D engine.” Although Palo Alto Networks does not develop general-purpose foundation models, its firewalls and security software process enormous volumes of attacks every day. After 23-year-old Patil takes over as co-chief AI officer, the team—which is slated to expand to 100 researchers by the end of fiscal 2027—will train proprietary models of various sizes for all of Palo Alto Networks’ security products.

Finally, having a cybersecurity giant behind them could also help ease enterprise customers’ security concerns about using open-source models, particularly Chinese ones.

Since the second half of this year, relevant U.S. authorities have repeatedly expressed concern about American companies using Chinese AI models and ordered federal civilian agencies to stop using DeepSeek products. Because the origins of open-source foundation models are considered sensitive, many traditional multinational companies and financial institutions are attracted by their low cost, but their chief information security officers (CISOs) are often unwilling to sign off, fearing that using core data to train open-source models could lead to data leaks or compliance risks.

This barrier, which is extremely difficult for a startup to overcome on its own, happens to be Palo Alto Networks’ core area of expertise.

In Palo Alto Networks CEO Arora’s view, every company will eventually run a mix of open- and closed-source models. Buying the training platform outright will allow security protections to be embedded in the very first step of model customization, giving enterprises the confidence to train open-source models in their own environments.

Arora even said during an earnings call that the business opportunity created by the acquisition could ultimately be “bigger than the business we have today.”

While Silicon Valley is still debating whether “AI outsourcing” is a good business, these three founders born after 2000, who built on Chinese open-source models, have already secured their exit with a $4.5 billion acquisition agreement.

This article originally appeared on the WeChat public account “Alphabet AI” and was written by Sibo. It is republished by 36Kr with permission.

Source

36Kr · AI