0:00 Palo Alto Networks is buying the artificial intelligence startup Applied Compute for $4.5 billion in an all stock deal. 0:08 Right. $4.5 billion for a company that has existed for what, roughly a year and a half? Yeah. 0:15 I mean, if you actually calculate the math out loud, that comes out to an astronomical valuation for every single month that the company has even been alive. 0:23 It really does. The velocity of that transaction completely defies traditional software metrics. 0:28 Which introduces the central mystery for you to consider today as we explore this Why does an established legacy cybersecurity giant need a young open weight artificial intelligence model builders so desperately that they would happily pay that kind of premium? 0:42 Well, I think the purchase price connects directly to the fundamental architecture of the product they shipped. 0:48 How so? To understand the valuation, you really have to examine the mechanics of what these developers actually brought to the market because they basically bypassed the entire traditional software development life cycle. Right. 1:00 So we are looking at founders Yash Patil Lingli and Rhythm Garg. These are former Open AI staff who stepped away from building closed models and they built something highly specific, and they somehow managed to take this company from zero to a $100 million annualized revenue run 1:17 rate in that incredibly short lifespan. 1:19 If you're not subscribed yet, take a second and hit follow on whatever podcast a you're using. It helps us Kee making this. 1:25 We appreciate you being here. So reaching 100 million in recurring revenue that fast in the enterprise sector, I mean that means they bypass the usual experimental pilot phases entirely, right? Exactly. 1:38 In traditional software sales, you know you spend years running small test programs, but Applied Compute prove their value immediately to clients like DoorDash and Harvey. 1:48 Which are two organizations with wildly different data sets and compliance requirements. Oh, entirely different. 1:53 You look at DoorDash and they handle massive logistical data, millions of moving parts, drivers, restaurants, real time consumer interactions. Yeah, and then Harvey deals with highly sensitive, confidential legal documents for major law firms. Right, and the fact that both 2:08 rely on the same core technology points to the founders central philosophy. They firmly believe every company should own its intelligence. 2:15 Which was a pretty unusual belief when they started out. 2:18 I mean, most people just assumed enterprises would simply rent intelligence via an application programming interface, you know, sending their data out to a massive model hosted somewhere else. But these founders recognized that training, serving and improving proprietary models 2:32 would just become a required part of running any modern corporation. Right. 2:36 So Apply Compute focuses entirely on customizable open weight models. Yeah, they give the enterprise the fundamental building blocks to host the intelligence internally, which allows businesses to run these models directly on their own proprietary data. 2:49 We should probably define what an open weight model actually is just because the distinction really matters here. 2:55 Go for it. So in a neural network, the weights are the mathematical parameters that determine how the model connects different concepts and words. 3:04 OK, when a company releases an open weight model, they publish those underlying parameters. You can actually download the model, put it on your own hardware and adjust those connections based on your specific information. 3:17 Right. And Palo Alto Networks President BJ Jenkins stated clearly that the future of enterprise computing is hybrid. Yeah, customers refuse to choose between just relying on closed models hosted by third parties and running open weight models internally. 3:31 They absolutely require both. I mean, they want the general reasoning of a closed model for drafting, say, standard public relations emails, but they demand open weight models running securely in their own environments for the most 3:44 sensitive operations. And we should expand on what it actually means to run these models securely in a proprietary environment. Yeah, let's do that. 3:52 When an organization decides to bring an open weight model inside their own walls, they are often setting up what is known as a sovereign environment. Right. 4:00 So picture of physical server room or like a heavily partitioned section of a cloud network that is completely isolated. Exactly. 4:08 In extreme cases they are air gapped, meaning there is physically no cable connecting those servers to the public Internet. Wow. 4:15 The enterprise controls the hardware, the software, the weights of the model and all the data flowing into it. 4:21 The demand for that level of control really reveals a shift in how these massive organizations view the value of artificial intelligence. We are totally moving past the novelty phase now. Yeah, Access. 4:31 To general foundational artificial intelligence is fast becoming a commodity. Think about how commodities function in the physical world. Water and electricity are essential, right? But because everyone has access to them at roughly the same price, possessing them gives the 4:45 factory no competitive advantage over the factory across the street. Right. 4:49 And the digital equivalent is happening right now. 4:52 You can get a highly capable reasoning engine from multiple massive tech companies today. They can all write functional code. They can all summarize public news articles, and they can all draft a polite corporate apology. And because your competitor has 5:05 access to that exact same baseline intelligence, possessing it offers absolutely zero competitive differentiation. If you both use the same cognitive engine, you were just running the exact same plays on the field. Exactly. 5:18 The real asset, the version of artificial intelligence that is actually worth paying a premium for, is the one trained explicitly on a company's internal knowledge. 5:27 And this is where the approach of applied compute becomes incredibly tangible for the businesses using it. 5:32 You can really visualize the exact inputs making this proprietary intelligence work. Picture your own workplace for a moment. Just think about the sheer volume of unstructured information floating around your network. Right, you have shared drives 5:46 filled with hundreds of differently formatted spreadsheets. You have endless chat logs, transcribed meeting notes, e-mail chains debating slight adjustments to pricing tiers. And a general language model knows how to read a financial statement, Sure, but it does not 6:01 know why your specific supply chain failed last quarter. 6:04 No, it doesn't. It doesn't know the unwritten rules your sales team uses to authorize discounts for legacy clients. Right, so when a company uses the tools built by Applied Compute, they pull all of that specific internal knowledge. They compile their last 50 6:20 customer emails. Exactly the specific phrases that keep repeating during successful sales pitches and their unique internal compliance policies. 6:29 They make that chaotic, highly specific internal data the grounding source that their artificial intelligence tools work from every single time a query is made. 6:38 Yeah. It completely changes the nature of the tool. You are creating an expert who has memorized every single interaction your company has ever had. Right, if you ask this internal model a question about a client, it doesn't just give you a 6:51 generic best practices answer. It tells you exactly what the client complained about 3 days ago. 6:57 Well, the account manager promised them in response and whether or not your current inventory can fulfill that promise. It becomes hyper relevant to your daily operations. But wait, if I connect a highly capable cognitive engine to my company's private 7:10 communications, haven't I just created a massive security nightmare? Oh, completely. 7:15 That hyper relevance introduces a terrifying level of sensitivity. Because a model that does useful work inherently needs access to sensitive systems. 7:24 You can't have a highly effective assistant that is locked out of the filing cabinet. 7:28 It needs the keys. And handing those keys to a digital entity introduces severe systemic threats that legacy cybersecurity architectures simply were not built to handle, right? If that model understands the unwritten rules of your pricing margins, what happens when a 7:43 competitor figures out how to query it? 7:46 This vulnerability is the exact reason Palo Alto Networks is aggressively buying their way into this specific layer of the technology stack. Yeah. 7:54 And there are three very specific risks introduced by deploying these powerful models internally. 7:59 Model tampering, data leakage and potential misuse. 8:02 Let's detail the mechanics of each to really understand the complexity of the problem being solved here. 8:08 Start with model tampering because this does not look like a traditional cyber attack where someone steals a password or crashes a server. No, Model tampering is incredibly insidious. You are subtly altering the neural pathways of the artificial intelligence itself. 8:24 Right, the technical term is often data poisoning. 8:27 Think about how a neural network learns. 8:29 It ingests massive amounts of texts and adjusts its weights to form statistical connections. So an attacker who gains even low level access to a company's network might subtly alter the training data over a long period. 8:44 Exactly. They might inject hidden text into spreadsheets that the model reads during its nightly update process. The mechanism there is fascinating. It's similar to slowly teaching a puppy the wrong trick over several months. 8:58 Yeah, the owner does not notice the behavior changing until the habit is permanently formed. The attacker manipulates the weights of the open weight model so that it functions perfectly normally 99% of the time. Right, so the security system 9:10 thinks everything is completely fine, but the logic engine has been fundamentally compromised to react to a specific hidden trigger phrase. Well, when the model sees that trigger phrase and it invoys, it deliberately approves a 9:22 fraudulent transaction without flagging it for human review. 9:25 It's the digital equivalent of hypnotizing an employee to ignore a specific person walking through the front door. 9:31 Exactly. The traditional firewalls and antivirus software are completely blind to this because the threat is baked into the logic of the allowed application itself. Then you have data leakage, which is perhaps the most immediate headache for any 9:43 corporate technology officer. Oh yeah, this happens when the model accesses sensitive data to formulate an answer, but fails to respect the complex web of human access permissions. 9:53 I might do a scenario where the artificial intelligence is tasked with reading the chief executive officers emails to help draft executive summaries. Right. 10:02 So a junior marketing employee who only has clearance to see public marketing materials logs into the internal chat interface. OK, they type a seemingly innocent prompt asking what the company's goals are for the fourth quarter. And the model functioning 10:16 exactly as it was designed to do, surges its memory for information about the fourth quarter. 10:21 Yeah, and it finds the CEO's draft emails discussing a highly classified merger and a planned series of layoffs. 10:28 The model synthesizes that information and casually presents it to the junior employee. 10:33 It does not understand the social concept of a secret, it just understands the statistical connection of the words requested. The company has inadvertently leaked its own highly classified strategy to its entire staff simply because the model synthesized information across 10:48 different permission levels. And the third threat, potential misuse, shifts the focus from the model being tricked or making a mistake to the model being utilized as an autonomous weapon against the company hosting it. 11:00 This brings in a concept that really frames the modern corporate environment. We are looking at artificial intelligence as a powerful digital workforce operating at 0 cost. If you think about traditional labor, scaling your workforce requires immense capital, 11:15 physical office space, human resources departments, complex management structures. Scaling an artificial intelligence workforce costs almost nothing. 11:24 In comparison, a company can deploy 10,000 digital agents to process invoices, analyse contracts and write code simultaneously. But if you have an army of free labor executing tasks across your network, your primary job description as a company alters completely. 11:41 You are no longer managing the execution of the work. 11:44 Your main job becomes supervising these digital entities and protecting yourself from others misusing them you. 11:50 Become a warden of a highly capable, completely tireless digital staff. Right. 11:55 And if an intruder compromises an employee's account, they can take control of those internal digital agents. 12:01 They can use the company's own intelligence against itself. 12:04 They can command the internal model to automatically generate highly personalized phishing emails aimed at the executive board. Exactly, utilizing the internal tone, the specific jargon and the inside jokes the artificial 12:17 intelligence learned by reading the company's private chat logs. 12:21 And these scenarios are happening right now. 12:23 The convergence of artificial intelligence and cybersecurity is accelerating faster than anyone predicted. 12:31 Intelligence and security agencies globally, including the Cybersecurity and Infrastructure Security Agency, the Federal Bureau of Investigation, and the Office of the Director of National Intelligence, are actively reporting a wave of 12:43 attacks driven by these models. These agencies are dealing with the reality that adversarial nations and criminal organizations are already using these models to accelerate their offensive capabilities. Attackers are using artificial 12:57 intelligence to find vulnerabilities and software code faster than human defenders can patch them. 13:02 They're writing more convincing social engineering campaigns and automating the exploitation of complex networks. 13:08 The companies building the most advanced foundation models recognize the severity of this They are doubling down heavily on security training for their own products. 13:16 Because they understand that shipping a powerful cognitive engine without a robust security containment strategy is incredibly reckless. The entire industry is coming to the realization that intelligence and security can no 13:28 longer exist in separate silos. They have to be integrated from the ground up. Hold on though. 13:32 I need to challenge the economics of this specific deal for a second. Sure, Palo Alto Networks is a Titan in the industry. They generate billions in revenue. They employ thousands of the most elite security engineers on the planet. 13:46 They have massive research and development budgets. 13:48 Yeah, they do. Why do they need to spend $4.5 billion in stock to acquire an outside team that has only existed for 18 months? Why not simply task their own engineers to build a secure framework for open weight models? And that is a completely logical 14:03 question. And to answer it, you really have to look past the technology itself and explore the go to market theory. OK. 14:09 The astronomical price tag attached to applied compute is not just paying for their code or their algorithms, it is paying for time. Time. 14:16 Yeah, time is the one asset that even the most well capitalized tech giants cannot easily manufacture in a laboratory. 14:23 To understand why time is worth billions of dollars, we need to thoroughly detail the mechanics of enterprise software sales. 14:29 It's a world governed by intense friction and extreme lock in. 14:34 Let's walk through a hypothetical scenario of how a typical massive corporation actually buys security software. 14:40 OK, picture a chief information security officer sitting in a boardroom. They decide the company needs a new security platform. They can't just swipe a credit card and download an application. 14:51 They issue a formal request for proposal. 14:53 They evaluate 5 different vendors over six months. 14:56 They run complex proof of concept tests. 14:59 They spend weeks negotiating with the procurement department over pricing and liability clauses. 15:04 The entire cycle can easily take a year or more, and once they finally signed that contract, they have to deploy the software across 100,000 employee laptops and thousands of servers. 15:14 It is an incredibly painful, disruptive process that often causes downtime in friction across the entire workforce. 15:21 The changing enterprise security vendors is frequently compared to performing open heart surgery on a patient while they are running a marathon. Because of that immense pain, corporations sign multi year contracts, usually locking in 15:33 for three to five years, and they absolutely avoid changing vendors during that period unless it is absolutely necessary. This dynamic means a software provider like Palo Alto Networks only gets one, maybe 2 genuine opportunities a year to have a strategic conversation with 15:49 their largest customers about renewing or expanding their services. You wait years for that renewal conversation to appear on the calendar. 15:56 Now imagine the absolute nightmare scenario for a vendor executive walking into that boardroom. 16:02 The customer security team sits down and says our Board of directors is reading every article about artificial intelligence efficiency and they are demanding that we deploy custom open weight models to process our internal data by the 16:15 end of the year. How does your platform secure these models against the data leakage and tampering risks we just discussed? If the vendor executive answers, we have our engineers looking into that and we plan to release 16:27 a feature sometime next year, the conversation is effectively over. The customer cannot wait a year to satisfy their board of directors. 16:34 If a competitor, even a young startup, has already shown up with a working proven solution for these specific threats, that competitor takes your seat at the table. 16:42 They do not just take the budget allocated for artificial intelligence security either. No, they don't. 16:48 Once they are inside the network, once they have endured that grueling deployment process, they eventually replace your firewalls, your endpoint protection, your entire software footprint. You lose the entire multimillion dollar account because you missed the window on a single 17:04 critical capability. When you evaluate the conclusion on that $4.5 billion price tag through that specific lens, it starts to make harsh economic sense. 17:14 Buying a team that has already shipped a working product, a team that has already secured high profile clients like DoorDash and Harvey, costs an absolute fortune, but the product arrives in time for the current sales cycle. 17:26 They are essentially paying for a guaranteed slot in a sales conversation that is already scheduled on the calendar. 17:32 They are buying the ability to walk into that boardroom tomorrow morning and say, yes, we secure open weight models. 17:38 We have the best team in the world doing it, and here's exactly how we integrate it into the tools you already use. 17:43 You are protecting billions of dollars of recurring revenue by spending billions of dollars today. 17:48 Palo Alto Networks wants to achieve something much larger than just protecting their existing revenue streams, though This acquisition points to a desire to expand their mandate entirely. They want a foundational say in 18:02 how enterprises build artificial intelligence, not just how they secure it after the fact. That represents A fundamental expansion of their corporate identity. 18:12 Traditionally, a cyber security company sits on the perimeter. 18:15 Right, the customer builds an application and the security vendor puts a wall around it to stop bad traffic from getting inside. But with artificial intelligence, that perimeter defense model completely falls apart. You cannot just put a wall 18:28 around a cognitive engine and assume it is safe. 18:31 To explain why the traditional perimeter fails, we should use a highly specific analogy regarding a confidential memo. 18:37 Yeah, if you have a highly sensitive physical document sitting on a table in an office, the absolute best time to determine who is allowed to read that confidential memo is before the artificial intelligence model ever scans it. 18:51 If you let the neural network ingest the confidential memo 1st, and then try to apply a security filter afterward to stop the bottle from talking about the memo, you have already lost control of the information. The knowledge is baked into the 19:04 model's weights. The artificial intelligence understands the concepts, the relationships, the strategic planning and the financial figures within that document. 19:14 The information is distributed across billions of parameters. 19:18 It is similar to trying to remove a single drop of blue dye after it has been stirred into a swimming pool. 19:24 You cannot easily isolate and extract the knowledge once it has been synthesized. An intelligent user will eventually find a way to prompt the model to reveal that synthesized knowledge, bypassing whatever clumsy filter you slapped on the output phase. They might ask the model to 19:38 write a fictional story about a company with financial constraints matching the exact numbers in the confidential memo. And the model will comply, leaking the data in a creative format. 19:48 Therefore, the security layer has to be embedded into the data retrieval and training process itself. 19:54 The security must act as the plumbing that dictates how data flows into the cognitive engine in the first place. 19:59 This requires massive architectural shifts within the enterprise environment. This acquisition allows Palo Alto Networks to embed enterprise grade security from the very start of deployment. They're shifting from being a reactionary shield blocking attacks at the perimeter to 20:15 being a foundational builder architecting the environment from the ground floor. They want to dictate the fundamental rules of data ingestion. 20:24 Whether a company is deploying these models in a massive public cloud infrastructure or locking them down inside heavily restricted air gapped sovereign environments, the security layer has to be the first component installed, not the last. 20:37 When you understand they're trying to position themselves as the foundational layer for all future enterprise computing, the executive vision driving this massive purchase makes a lot more sense. I want to highlight exactly what 20:51 Chief Executive Officer Nikesh Arora said about the potential of this acquisition. What'd you say? 20:56 He stated publicly that if this bet works, this will dwarf the abilities that we have in cybersecurity. 21:02 Wow, you really have to pause and react to the weight of that statement. Yeah. 21:06 This is the executive leader of one of the largest, most successful cybersecurity companies on the planet. 21:13 Their core business generates billions of dollars annually by protecting the world's largest networks. 21:18 For him to publicly state that this move into artificial intelligence infrastructure will dwarf their traditional capabilities is an astonishing claim about where he believes the global economy is heading. He is effectively broadcasting 21:31 to the market that the business of building and managing secure internal models is going to be significantly larger than the entire existing market for keeping hackers out of traditional software. And he backs up this aggressive ambition with incredibly strong financial data. 21:47 In their first financial quarter, their revenue hit $3.45 billion. Which represents a 39% jump from the previous year. Gender reading a 39% increase in revenue at that immense scale is exceptional performance. 22:02 It gives them the financial leverage required to make these massive bets. They even raise their annual forecast to a range between 14.52 billion and $14.57 billion. They are operating from a position of immense strength, using the profits from their 22:16 legacy business to fund their expansion into the new economy. 22:20 What is truly fascinating here is the market validation of the specific strategy. Typically, if a mature, publicly traded company announces they are diluting their stock to spend $4.5 billion on a startup with only a year and a half of 22:33 operating history, Wall Street punches them severely. 22:36 The stock price drops because investors worry about overpaying, integration risks and a lack of focus on the core business. But in this specific case, shares jumped over 7% following the earnings and acquisition announcement. The market actively rewarded the 22:50 aggressive expansion. Investors are looking at the exact same enterprise reality the chief executive is looking at. They understand that infrastructure for internal models is the next mandatory spending category for large corporations, and they want Palo 23:03 Alto Networks to capture that massive upcoming budget. 23:06 Zooming out to look at the macro view of this industry, the speed of value creation we are witnessing right now is almost difficult to process. Trace the timeline of Applied Compute. They go from inception to securing backing from some of the most powerful entities in 23:20 finance and technology, including Kleiner Perkins, Lux Capital, NVIDIA and Blackstone. They reach a valuation of $3.25 billion and then within weeks they finalize a $4.5 billion exit. The inclusion of NVIDIA in that investor list acts as a very strong signal. 23:39 NVIDIA provides the physical hardware, the graphics processing units powering this entire technological revolution. 23:45 But hardware is useless to an enterprise if they are too scared to run their data through it, right? 23:50 NVIDIA recognized that Applied Compute was building the necessary software layer to make that hardware usable and secure for cautious corporations. NVIDIA needs this security layer to exist so their chips continue to have a market among risk 24:03 averse banks and hospitals. This rapid exit establishes entirely new rules of acquisition for the industry. 24:09 It completely rewrites the playbook for founders. 24:12 For smaller companies building enterprise tools in this space, the ultimate buyer might not be the end customer. 24:17 The exit strategy is not necessarily launching an initial public offering where you attempt to sell software to banks and hospitals directly competing against established giants. The most logical buyer becomes the legacy software vendor who already holds the long term 24:31 relationship with those banks and hospitals and who desperately needs to protect their turf from new threats. 24:37 You build the essential component, the missing puzzle piece, and you sell it to the giant who already owns the distribution channel. It is the highly efficient, though incredibly expensive, way for the technology ecosystem to upgrade its capabilities. The giants use their highly 24:52 valued stock as currency to acquire innovation they simply cannot build fast enough internally. 24:57 Despite these massive numbers and the astronomical improvements in model capabilities we have seen recently, we should remember a specific comment from founder Linden Lai. He explicitly noted that enterprise adoption of these tools is still in the early 25:10 innings. We see the headlines about models passing complex professional exams or writing entire software programs, but the actual deployment of these models deep inside corporate networks is just beginning. 25:21 Most companies are currently just trying to figure out how to securely connect a basic chat interface to their internal wikis without accidentally exposing human resources records to the entire staff. The sophisticated use cases 25:34 where autonomous digital agents are negotiating contracts with vendors, managing global supply chains in real time, and analyzing proprietary financial data without human supervision are still years ahead of us. Which means the architectural 25:50 foundations being laid right now through acquisitions exactly like this one, are going to dictate how the global economy operates for the next decade. Security can no longer be an afterthought applied to software. 26:02 In the age of artificial intelligence, security is the foundation that dictates what the software is even allowed to learn in the 1st place. You cannot bolt a firewall onto a brain. You have to construct the brain's pathways securely from the moment it is activated. 26:16 If an artificial intelligence model needs to remain explainable, verifiable, and auditable while operating on your company's most sensitive data, how long until the distinction between a security company and an artificial intelligence company completely disappears? 26:31 If you're not subscribed yet, take a second and hit follow on whatever app you're using. It helps us keep making this. 26:37 We appreciate you being here. Also, check out our YouTube channel for more business and tech updates. 26:42 There's a link in the description.
Community discussion