¶1Engineers, small business owners, and enterprise leaders, it's time we sit down, look at the elephant in the room, and answer the question head-on. Is Anthropic stealing your data while you pay for it? This video and this channel, it's not hype or fear-based. If we can't answer the question without supporting ground truth data, we don't answer at all. But, there is sufficient evidence to support the fact that Anthropic is competing with its customers, and I'm not the only one asking this question.
¶2Satya Nadella, CEO of Microsoft, says this, "You essentially pay for intelligence twice. Once with money and again with something even more valuable, the proprietary knowledge you must reveal to make that intelligence useful. The better you want the model to perform, the more of that knowledge you have to feed it." This is a brutally true take on AI and on agents. And that's not all. We have Alex Karp, CEO of Palantir, saying this, "What the technical customers want is control over their compute, their models, their data stack, their alpha." Keyword, alpha.
¶3"They want to know that the means of production is not being transferred to someone else. Who owns the data? Are the prompts secure? Is this being transferred to you?" Microsoft even went as far to ban Fable 5 over the 30-day ZDR data retention policy. I don't agree with both of their takes 100%, but their views and points are valid.
¶4I've been engineering for over 15 years now, [music] and I've become hyper aware of companies that use my data and use your data to fuel their business. Language models and agents are far worse [music] than Google Search. It's not just the query we're handing over to these companies. It's, as Satya and Alex Karp mentioned directly, it's the intelligence [music] of our business and therefore our livelihood. In this video, we answer the question directly, "Is Anthropic [music] stealing your data while you pay for it.
¶5Let's visualize the problem properly. This is what it looks like. Your business hands over your IP through prompts, agents, traces, workflows, and we pay twice here. As Satya said, you're paying in cash and you're paying the IP you hand over. And you can imagine if the model labs wanted to, they could in secret learn all about your business, about your domain, because what are we handing over?
¶6Not just money, right? The money is in comparison not the important part. It's the know-how. It's the business intelligence. This is the problem.
¶7If this is happening, it's only a matter of time until every model lab takes over every industry. That's the extreme version of this. Now we have to figure it out, is this true? So here's what we're going to cover in this video. This is what you're going to get.
¶8We're going to answer the question, is Anthropic stealing your data? Talk about the aggregate bucket problem. We're going to look at the ground truth facts. I'm not going to make anything up here, okay? I'm looking at the ground truth data and reporting it right back to you so you can make the decision you need to make for your work, for your business, for your livelihood.
¶9We'll talk about commodity agents versus IP agents. Very important distinction to make. And then we're going to break down the solution, of course, with open models being a key, key piece of this. So this is the road map. If this interests you, stick around.
¶10Let's jump right into this, okay? So is Anthropic stealing your data? No, they are not. But they use it anonymized in aggregate. Let's break this down.
¶11What do I mean by that? Think about a data tumbler. You, me, all the other companies in the world using intelligence, all of our prompts go into this bucket, they anonymize it, and now they have aggregate trends. And by they, I'm talking about Anthropic, I'm talking about OpenAI, I'm talking about Gemini, I'm talking about the owners, the true owners of the models. Because remember, when you prompt in Cloud Code, in Codex, you are renting the model.
¶12Let's look at another example. The coin tumbler idea is so interesting, right? You can put in your Bitcoin, you can put in your prompts, you can put in any types of information, you spin it all around, and then what do you still get out of that? It's anonymized, it's randomized. Look at what comes out.
¶13You can still identify the value proposition that each prompt, that each agent, that each trace delivered. So it's not your data anymore. This is fair play. It's anonymized. They say that they do this, and they do.
¶14But, as you can see here, it's still a map of the market. This is a market map. Coding, design, legal. I chose these on purpose, and you know why already. There is a chain of evidence of Anthropic competing with its customers.
¶15Remember Cursor? And then all of a sudden we had Claude Code. Figma MCP usage went crazy. What do you know? Claude Design.
¶16Security usage, Mythos. Claude Security. Life Science, Claude Life Science. What do you know? You know, [laughter] like the the evidence is almost insurmountable at this point.
¶17If it was just one of these, they want to compete in one domain, totally fine, makes sense. They're a platform, they have the access to the information, but this is a bad bad trend. One is a coincidence, four is a pattern. Three is a pattern, four is a confirmation of the pattern. So the evidence is there, and although I'm pointing my fingers directly at Anthropic here, I'm talking about platforms in general, specifically AI platforms.
¶18So, you know, Google, OpenAI, Anthropic are the biggest offenders here. This is how it works. This is what it really looks like. How are they able to do this? It's a simple four-step process.
¶19They see the usage, they learn the trends, they enter the vertical, and then they cut access. Now, they're not cutting access so much, we have to be super super fair and balanced here. They're not doing this much anymore. Some of you might remember, they cut off some access to Cursor, they cut off some access to Wind Surf. So this is a concern, they're not really doing this much anymore, but it just re-emphasizes this simple fact.
¶20You and I do not own the model. We do not own the thing that is creating the real value. As much as we want to own the harness, the trace, the prompts, the system prompts, all that is good, right? The outputs, fantastic. But at the end of the day, the intelligence that fuels everything, we are renting it.
¶21And out of my 15-plus years engineering, this is something I've learned to look out for because this means this is IP that we depend on for our business to operate that we don't own. Okay, this is a dependency risk. This is keyman risk. This is key technology risk. Anyway, this isn't happening so much anymore, but the principle is the idea that matters that you and I as business owners must pay attention to.
¶22Fable was recently banned by the government. So, it's not just the companies we need to watch out for, okay? We need to defend our business from every possible angle. And if your IP can be snatched away, is it your IP? You know the answer to that question.
¶23This is not me speculating. I just want to make that super clear here. Nothing I'm saying here is like new information. This is in the terms of service, okay? You and I agree to this.
¶24Maybe a good place to stop here and really understand what this looks like. You've probably heard of Cleo. This is Anthropic's aggregated privacy-preserving analysis of data to gain insights into the real-world impacts of AI systems and the usage patterns of their products while maintaining user privacy. Very, very good, right? Love to hear that.
¶25If this wasn't here, many users, customers, enterprises would not use this technology, okay? You You just can't if your privacy is not being preserved. But, as I mentioned, you're probably aware of the Anthropic economic indexes they put out. It's a really great breakdown of how AI is being used across different classes, across different socio-economic statuses across the world. Of course, they break it down by domain.
¶26How do you think they got this data? They make no secret of it. It's Cleo. It's their data handling system. Aggregated privacy-preserving analysis of data to gain insights.
¶27This is in their terms of service. And this might be a good place to stop and mention how can you get out of this to some degree right here in Claude.ai. I'm going to hit command-shift-comma. It's going to open the settings. I'm going to go to privacy, and you can see this right in their privacy statement.
¶28Anthropic may conduct, and by may, it's they do, aggregated anonymized analysis of data to understand how people use Claude. They don't go on to say what they're doing with that data, but they say that they're using it to understand how people use Claude. So, very important, if you scroll down, help improve our AI models, uh you probably want to check this off unless you don't care. And if you're not doing any interesting IP defending work, keep it on, whatever. I highly recommend you click this button.
¶29Props to them though, they're not selling our data to any third parties, and they will delete our data promptly if we ask for it. But there are some catch-alls to this. As mentioned, there is a 30-day required retention policy for the Fable model because of the cybersecurity harms. I just want to directly call this out. This is not speculation, it's in the terms of service.
¶30We all agree to this stuff. This goes even further, right? Not all customer classes [music] are equal. And and if you're a business owner, you know this. There's nothing wrong with this, but as the end user here, we have to be careful, okay?
¶31There's consumers, and then there's commercial. I'll let you guess which one's privacy is better preserved. Of course, it's commercial, teams, enterprises, API users. If you're on the free plan, which probably no one is on my channel, very few, you're cooked. They're using all of your data, they're training models on it, you have nothing here.
¶32If it's free, you are the product. Never forget that. And then Pro and Max, this is one of those cases where they're pulling that data for Cleo, for these nice reports that they create, which again, I'm super glad that they do this, right? This is very valuable important information to share, these economic indices. I love that they share these, but underneath the hood, as you can see very, very clearly from the pattern that they're clearly proven to be executing on, they're a platform, and they're competing once they see the trend data, and they're going vertical.
¶33They're competing with their customers. Let's be super clear about that. So, not all customers are created equally, you really need to understand this because that whenever you boot up Claude Code, whenever you go through the API, your enterprise, your team plan, whatever you have, the controls, the data protections are not the same. Just really important to call that out. Again, this is all pulled from raw documentation.
¶34But as you can see here very clearly, commercial is much more defended, okay? which gets us into our solution space in just a moment here. Cloud code inherits your account, cloud co-work, anything you use, it just trickles down to what type of account you have. Are you a consumer or are you commercial? If you're commercial, you're going to have better privacy protections.
¶35There's only one thing keeping Anthropic, keeping all these businesses from maintaining their privacy policies. If they violate a term of service even one time, that enterprise is going to discover there's going to be a mass enterprise exodus and then the company is cooked. They're going to die. So, it's incentives that protect your and I's data, not goodwill, all right? Let's just be super clear about that.
¶36And again, just to say this, this is not me attacking Anthropic, this is not me going after them, flaming them. I'm just addressing the data in front of me. I'm just calling out what is directly in the limelight already. What they're doing in the dark, no one knows, okay? And I mean that for every one of these big AI labs with the extraordinary power and levers that they're gaining every single day.
¶37It really is like, understand the incentive and understand if you will will not be protected. Incentives make up every business decision. This is what keeps us floating, okay? [laughter] So, there are four claims here, only two of these claims survive. Try to think through which two are true.
¶38It is very clear they see the aggregate patterns, okay? This is true. Do they train on your code, on your prompts, on your incoming information? That is not true. Their terms of service says no.
¶39This is fantastic. Do they own your outputs? No. Again, their terms of service says you own them. Do they compete with your product?
¶40Absolutely. If you're in a domain that is scaling, that is making money, they see those trends and at the drop of a single decision by some PM somewhere inside the org, they can compete with you. It's that simple. It's aggregate data, it's not your data, it's everyone's data. See the kind of like legal loophole there?
¶41And it's not even like a legal loophole, this is in the terms of service, okay? So, the verdict is they are not stealing your data, but they are competing with their customers, which could be you given the right numbers, given success, given growth. They want to enter in profitable areas that they can properly compete in with intelligence. And there really is nothing limiting them from going after all the top domains and making Claude or GPT or whatever the go-to place for every type of application. Based on manpower, this is hard to do technically, but it is not impossible.
¶42It is not impossible. What a single engineer can accomplish now thanks to compute, thanks to agentic engineering, it is exponentially higher than it was a year ago, two years ago, three years ago, God forbid five years ago. It's just exponentially higher. And that means what a team can do is more exponential, and what an entire org can do that is agent first is absolutely absurd. So, that's the verdict.
¶43And the test is simple. It's like, are you in a growing domain? How big? And if you're in a big growing domain, you should expect that one of these AI labs is going to go after your domain. This is just the ground truth of it.
¶44So, enough of the problem. You understand this now. If you didn't know it before, you know it now, and it's very clear. I've been thinking about this ever since I started using this technology. I am paying and I am sending in my intelligence to get these models to act on that intelligence.
¶45Never has a technology existed like this before. Like Google search is the best closest example. You have interesting questions, trends show up in the question. This is a whole 'nother level. As Satya said, you pay twice, once with money, once with your technical know-how, your business intelligence, your IP.
¶46So, this is an unprecedented problem that us engineers must [music] face now. What can you and I, the engineer, do about it? What is the solution? The solution starts now. So, let's talk about the solution.
¶47On the channel, I don't just want to call out the problem. I don't want to fear-monger on the stuff. I don't want to point fingers. This channel is and will always be about helping you, the engineer, progress in the age of agents. We're here to learn to build software that works while we sleep.
¶48I'm here to do accomplish that myself. You know, when I can, I try to share Alpha with you. So, what can we do about the fact that AI labs are using our data in aggregate to compete directly against us. And you already have solutions going on in your mind. You understand this problem.
¶49I know a lot of engineers watching this channel, watching this video, you have thought this, too. When I hit enter, all that engineering work, right? All that agentic engineering work that I just did setting up this new fusion harness, filling out this AI developer workflow, this new software developer life cycle I've enhanced with agents, I just sent all of that to this lab. >> [laughter] >> And I know what the terms of service says, but I just sent all of that away. And now it's stuck in some database somewhere that anyone can pull out at a moment's notice.
¶50Now, as mentioned, ZDR, terms of service, blah blah blah, but what can we really do about this? Let's break it down. We first need to talk about two types of agentic coding. You know, the senior engineer mindset is always this. If you don't have to solve the problem, don't waste your time.
¶51Why would we be wasting our time trying to solve this problem? Of course, commodity agentic coding. If you're building prototypes, CRUD, boilerplate, glue, replaceable work that anyone can prompt in a weekend, the labs seeing this, other businesses seeing this, it doesn't matter at all. These are commodity agents. But then we have IP agents.
¶52This is what you want to defend as if your life depends on it, okay? Because if your business goes down, your life will depend on it. I don't mean to be drastic there, but like just connect the dots. This is your business know-how, your prompts, your traces, your domain logic, your farmed evals, your hard-earned evals you built by hand. And then of course, the most valuable of all, your user data insights that you can use based on your privacy policy and your terms of service.
¶53Okay, this is scarce. This is highly asymmetric, and this is compounding IP that only you have or very few that can compete where you're competing. So, be very, very careful with this type of work when you're prompting off to these AI labs because you could be giving them the next signal. First off, we need to distinguish between these two types of work because the ultimate test is this. If you remember nothing from this video, remember this.
¶54The test is if a competitor could read my full agent trace, would it matter? Let's just sit on that for a moment cuz this is the core of the solution. If a competitor could read your full agent trace, would it matter? This is the 80/20 of the solution here. Let's be honest though, most engineering work is commodity work.
¶55It's full stack garbage, right? Maybe some devops glue code. It's stuff that's been done by other engineers a million times. That's part of why these agents are so valuable because they're still in the normal distribution of results, of probabilities. So, you have to be really careful here because as soon as you exit that normal distribution and you enter that 20, that 10, that five, that one, that sub fraction of a single percent distribution that the models have never seen in their training data, this is IP and this is the most IP dense work that you can do.
¶56The key here is separate and understand every prompt you write. Is this commodity work or is this IP work? And the moment you have IP work inside your agent's context window, you need to be aware you are sending scarce asymmetric compounding work to the AI lab. Privacy is a stack. There's multiple elements to it.
¶57Your account, the contract, the feature, the model, the routing, and then all of your agentic engineering that surrounds this core. So, we want to think about this holistically. This isn't just about one piece. There are many ways to operate within the contracts of the AI labs and still be protecting your IP. On this channel, I have a decent segment of cracked engineers all the way from individuals, small medium size businesses to enterprise customers.
¶58Let's walk through the real solution for every class of engineer that watches this channel, for every class of engineer in general. And so, let me preface this first by saying most engineers, you're not doing IP work. I'm not saying that to be mean or anything. It's just true and there's nothing wrong with that. Okay, that's still valuable work.
¶59It's just not IP defensible work. And you know what type of work this is. It's full work, it's some DevOps work, it's a little bit of infrastructure work, but the real IP-defensible work here is like 20%, it's 10%, it's 5%, and you really want to understand, again, are you doing commodity work or are you doing IP work? For most engineers, what I'm going to say next doesn't matter at all. Keep using the APIs, keep sending over your prompts to these labs.
¶60Uh just forget about this. Don't stress yourself with more problems you don't need to have. But, for everyone else, it's time to lock in and let's walk up the AI sovereignty ladder. At each level, you gain more control. At each level, your work is truly your own in the age of agents.
¶61So, for most engineers, individuals, small businesses, solopreneurs, and enterprises, um I highly recommend you get up to this tier three level, and if and when you can, the tier four level. This is hybrid private. You use an open-weights model, you get on some GPU stack. You don't have to own the GPU metal, that's very, very expensive, cost-prohibitive, but you basically create your own endpoints on top of some service that lets you access the GPUs, and then you mount the open-weight models. I'm going to make this super clear, this is the target solution for anyone building something valuable, building something unique, building something agent-first, okay?
¶62Agent-first products. But, if you can't get there, you want to at least own the control plane. You set up a simple, small VM, a light LLM gateway, or whatever tool or service so that you can own and switch between every single model at a dime. Diversity here, model diversity is absolutely a solution. Don't depend on one model.
¶63Also, there's a really great solution here, model cloud. You get a lot of protection just by being on a third-party vendor. There aren't a ton of these, right? It's AWS, it's GCP Vertex, Google Cloud, and then there's Foundry, which is Microsoft product, okay? That's Claude's product inside of their cloud.
¶64And then there's, going down a level, there's the commercial API. You're going right through the API, you're not using a Claude or ChatGPT subscription, because this is the lowest level of protection. Your data is being used in aggregate at this lowest level. And again, if you're doing commodity work, this doesn't matter, okay? Look away, close the video, move on, okay?
¶65But, if you ever hope to build true, defensible, unique IP in the age of agents, you will have to move up this ladder to truly defend your IP. Imagine having a part of your business you don't own that you need to survive to win. That's what we're defending against here. I want to help you defend your products, defend your IP, defend your business, defend your career, okay? That's what this is really about.
¶66So, let's address each level, okay? So, if you are an individual engineer, I recommend you sit at one of these two levels, right? Stay at the subscription level and move to the commercial API level. Why is that? The commercial API has more protections.
¶67And specifically, I'm referring to Anthropic here. In the terms of service, they explicitly state they only sample aggregate data from the API. So, a little more protection, assuming you aren't, you know, tinfoil hat and you trust their terms of service, okay? Briefly talk about that in a second. If you're a big AI doomer and you don't trust big tech, big cloud, big whatever, uh we'll talk about that in a moment.
¶68But, for individuals, level one is going to be the best return on investment for both protecting your IP and for spending. A lot of people commented on last week's video, the Fusion Harness. All you could talk about was me spending API credits, me spending tokens, you know, fable and soul tokens inside the Fusion Harness. Like, yes, yes, I do. Spend to win.
¶69Compute is still relatively expensive. The price per intelligence is always going down. Price per intelligent agent task is constantly going down. But, there are other benefits to this. I hope you realize this a little bit.
¶70Your subscription is not just for you, okay? It's for the AI labs, too. And quick aside here, if you're an individual and you're above that 1 million uh revenue mark and you're doing real agentic work, non-commodity work, you should move up this tier stack, too. So, let's move on to that because that really puts you in the small, medium-size business level, 2 to 100 people. I highly recommend you move up to tier two or three here.
¶71Why is that? When you're using a model cloud, for instance, AWS, GCP, Azure, your data is not being sampled at all. Anthropic has deals with these companies to just set up mount their technology on their software hardware and then Anthropic steps away. And then you're just going right through your cloud provider. This is going to be the solution for most small and medium-sized businesses and even some enterprise customers.
¶72Just set it up here, use the cloud provider you trust and get away from these AI labs, okay? >> [laughter] >> But if you can, there's another solution here, which is to own the control plane. And really, there's a good argument to flip-flop these, you know, tier two, tier three, because the LLM gateway and setting up your VM to control the model traces isn't giving you that much. So I actually probably made a mistake here. I need to swap these tiers.
¶73But this is also another good solution. Why? Because you can toggle in and out to many, many model providers while you own all the traces. Every single one of the traces. But of course, for the enterprises, you guys just got to move up here.
¶74You have to own this stuff. I will say tier two is a piece of this. You can certainly own this, but like you have to own your control plane and you have to be setting up an open weight model on top of a service that gives you access to heavy GPUs. So I'm still saying rent the GPUs here, but here you own the model. I'll speak personally here for my business work and for any work that I do with clients, I am always pushing here.
¶75If you're doing real agentic engineering work that's scaling, you got to own the model. You just have to and you have to own the model, you have to own the traces, the evals, the routing. This is just like part of the new stack. And so over time, we're slowly going to see this. If you're serious, if you have defensible IP, right?
¶76Not commodity agents. If you have IP agents operating your business, satisfying your customers, you got to own it. You just have to because at any point, it could be snapped away from you. Really, if you're down here at the tier one or tier zero level. Um, it's possible at the model cloud level if the government says you can't serve that model, tier two is gone as well.
¶77This tier's gone, this tier's gone, this tier's gone, but as soon as you get to tier four, it doesn't matter what anyone says, they cannot take down your businesses, you know, one of the core pieces of every business moving forward, your AI, your intelligence, your LLMs, your agents, and more importantly, your agentic system starts to scale up to a real useful system. Anyone can prompt a model, very few can build software factories, and developer workflows. This is where the true IP of the future is getting built. Just to call that out, and then, you know, this last level here is just really hard. It's basically impossible to own GPUs on prem, you have to have a ton of cash, a lot of technical know-how, and a lot of capability.
¶78I just wanted to throw that out on there cuz obviously, it's the true optimal solution, everything. But this is much harder, right? So, I hope this makes sense. This is the sovereignty ladder. You want to climb this.
¶79The bigger your business gets, the more successful you are. Do not stay at these bottom levels where you are not in control. You know, open weights is obviously taking center stage right now. Kimmy K3, GLM 5.2, Minimax, the Quinn models, open weights in your infrastructure that tier 345 level is the real solution here. Like, let's just be brutally honest about this.
¶80You have to own the model. I want to be really, really clear with this. Don't confuse the overseas APIs with you owning the model. >> [laughter] >> Model diversification is only going to get you so far. I am very, very skeptical.
¶81Maybe it's just cuz I'm team America, maybe it's just cuz I'm a US guy, but like, I'm very, very skeptical of signing up for any one of these deep seek Kimmy plans, any one of these across shore plans because the terms of service there cannot reliably be reinforced. Maybe I need to like look at home and look inward, maybe our model labs are corrupt as well, and they're actually using our data to train the models and violating terms of service. I just don't believe that. But I wouldn't put it past uh some of these Chinese labs. We are competitors.
¶82You know, at the end of the day, I like to think that we can all hold hands and Kumbaya, but we have to address the facts on the ground. There is competition between US and China, and AI is the battleground for this conflict right now. So, just to really call this out, okay? Self-hosting wins on secrecy and trace ownership, but not automatically on price, okay? So, this is where things get a little harder for businesses and again, I think you want to move up the sovereignty ladder as your business grows.
¶83If you don't have a business and you're just paranoid, you don't want anyone to see what you're doing, sorry, but like don't do this stuff. It's going to be cost-prohibitive. It's too expensive. You need the means to be able to move up the ladder and the means is success. It's cash.
¶84It's product market fit. So, open weights is a really really important part of this journey, but by open weights, I do not mean overseas APIs. I think that is a mistake. We cannot reliably enforce as much as we can or can't enforce the traces against our own US model labs. I would just wouldn't for a second trust overseas APIs.
¶85I like open router, but I am very very skeptical of really depending my business IP and other engineer's business IP on these overseas discounted APIs. Download the model, use the model, own it, get the weights in your infrastructure or on some infrastructure where you can at least rent the GPUs while you own the model. If we want to look at the best, mid, and worst case scenario for what happens moving forward, I think it looks like this. The model labs are platforms. They're going to keep being platforms.
¶86They're going to keep using their data as they should, as you or I would if we owned one of these labs. It's just the right business thing to do. It's the strategic move to make. The mid case is, and I'm betting on this case, this just continues to happen. There's a great case here where they just say, "You know what?
¶87I'm just going to be a model lab. I'm going to serve the best models at the best prices. We're going to focus all on the data, certainly the GPU, the chip, the energy, and we're going to like piss off from the application layer." Wouldn't that be great? I would love that. The only one who I truly believe is doing that is Nvidia, but they also would have incentives to continue moving up the stack.
¶88But then the worst case here is for all my super paranoid engineers, shout out to you, is that the anonymization is a huge catch-all and anything we prompt into these AI labs is getting absorbed into the IP of these companies. So, what does that look like? This is speculation. This is not based on any grounded fact, but the legal alchemy could look like this. As soon as they anonymize the data, it's no longer ours, so they can train on it.
¶89Or they can use it to build products against. Again, I just want to make this super clear, this is speculation. I don't believe this is true, but probably Alex Karp fits into this scenario here, CEO of Palantir, he probably believes something like this is true or is happening. I don't buy this, but I just want to mention it. This is the worst-case scenario.
¶90Regardless of any scenario here, the defense is the same for all three. Own the traces, own the evals, keep a second model passing for whatever IP essential work you're doing. And the next logical step here is like rent the GPUs, own the models. That's like a relatively reasonable place to end up. The final answer here is Anthropic is not stealing your data.
¶91The public record says that is not true. Their incentive structure says that is false. But is your usage shaping what they build next? Yes. It's pretty easy to identify these four verticals.
¶92Likely there are more coming, you know, Claude code, Claude design, Claude life science, Claude security, Claude animation, Claude legal, Claude insert another word after Claude. They could and have all the wherewithal to build into any vertical they want, and the usage data is the map of the market, and we're giving it to them. Same verdict for OpenAI. This is what model labs are. This is not a story about Anthropic, I'm using them as an example.
¶93Anthropic has become a monster as I predicted last year. They're just the biggest, they're the ones to point at first, but these model labs, they're platforms now. The answer is to own your own GPU, set up a home lab, do that thing, but that would be premature. I think the the sweet spot is rent GPUs from some cloud service, and then put an open-weight model on top of it, own all the traces, and you know exactly what and where things are happening. Then the only data you're sending is GPU usage, which is whatever, who cares?
¶94That's the real solution here. I wouldn't prematurely try to own GPUs, that is financial suicide unless you are very, very wealthy or can write some serious checks and get some serious loans. So, the whole idea here is you want to own the learning loop, and you want to have your own router so that you can run and connect to any model of your choosing. As long as they keep performing, we'll always want the very, very frontier for commodity work and some IP work. But, when you're scaling to millions, tens of millions, billions of dollars of revenue, you really don't want your IP going to these labs.
¶95That's the top line. You want this router set up, an API you created that sends you to the models you have configured, you have set up. Let me be super clear, you know, I love open router, I love Claude, I love open AI. They've done incredible work, but as these systems grow, the incentive structure changes, and what's best for them is not what's best for you or I, the engineer, at the end of the day. This is just the nature of systems, of growth.
¶96So, what do we do? We can operate our own systems. We need to control the plane, we need to own the learning loop, we need to protect our IP while having access to the state of the art, to our workhorse models we deploy on some GPUs that we rent, where we placed open weight model, and then, of course, other cloud models. We want the optionality, we want to explore, we want to still evaluate the state space. That's the idea here.
¶97Is Anthropic stealing your data? No, but they are running anonymized aggregate use, which steers what they build next. Your traces get tumbled with everyone else's. It's not your data anymore. It's a market map.
¶98Ground truth facts, we've seen these companies go after verticalized use cases. Coding is just the beginning. They see the data, they see the trends, they know where the money is. So, why wouldn't they go after it? This is just classic incentives.
¶99Commodity agents versus IP agents. This is how you should decide if what you're doing is something you should be defending, if something you should be deploying an open weights model against, if something you should be getting on to some cloud service provider that isn't the AI model labs. Of course, I'm talking about GCP, I'm talking about Azure, I'm talking about AWS, I'm talking about any other neo cloud, any other provider of compute that is not going to use your data. And the only way to really know is to go up this AI sovereignty ladder. The more important your IP agents become.
¶100The open weight models is the true only escape hatch. So here we have to give a lot of credit to these Chinese labs. They are scorching the earth. They are commoditizing the closed sourced US labs. But we have to be very careful.
¶101You want to own these models on GPUs you're renting or owning if you can pay for it. But I would be very very careful going over to the discount overseas APIs. Unless you have no other choice. I'm empathetic toward you're in the beginning of your career, you're starting out, you're trying to grow your business, you can't spend the money on Fable, no problem, I get it. But if you can, be very very careful with this because they're likely doing even worse with your data, rolling them right into the models, rolling them right into their next product, so on and so forth.
¶102At the end of the day, you need to own everything that makes up your product. And again, for the 20% of engineers, of small, medium-sized business owners, for of entrepreneurs that this applies to, you have to be very very careful here and you have to understand is the work that you and your engineering team are doing and the rest of your team as we're deploying agents all over the place, is this commodity agent work or is this IP agent work? Is this critical business know-how, business intelligence, business workflows, agentic workflows that we've been working really hard to develop? We're talking about agent evals. We're talking about your users usage data patterns for your domain.
¶103Um own the learning. Don't let that go anywhere else. Just want to leave you with that one final image. We want to avoid this. Do not hand off your IP to a model lab without the right protections, without the right defenses in place.
¶104At the end of the day, as Satya said, we're paying twice here, once with money and again with our business's IP. We have to be very careful. We have to defend against this. We need to own our technology. We need to own our future and this is how we do it.
¶105You know, again, big fan of Claude, big fan of all the work they're doing. Same with OpenAI. I genuinely love these companies and all the engineers working at them. I think they're doing the most incredible work of the generation, frankly. But at the same time, they have grown to a size and a scale where their incentives have changed.
¶106And that means for us, for you and I, the engineer with our boots on the ground every single day, we have to defend our IP. We have to defend our intelligence that we are embedding, that we're templating, right? We are templating our engineering into the fabric of our software now, of our agents. And so as we're building out our AI developer workflows, our loops, our software development life cycles, building up to our software factory, to our automated workflows, to our true end-to-end systems, all the way up to ZTE, we always need to be on the lookout for threats. And I don't know about you, but what is more threatening to your business than a double attack on your business?
¶107One on your capital, your cash via token spend, and the other on your business's intellectual property. I don't know if there has ever been a bigger threat to a business, to engineers, to a company. So, keep your eye on this. Is Anthropic stealing your data? No, but they are using it to make plays vertically in the domains that are making the most cash.
¶108And I don't want to just point fingers at Anthropic. This is any platform. Let's just up-level all this. This is the platform play. The data incoming into their system is valuable information they can act on to increase their revenues.
¶109[music] We would do the same thing in their position. This is just about incentives. This is just about the game of business. If you enjoyed this video, if this helped you make a connection, if this helped you take action, drop a like, subscribe, comment, stay connected to information like this. Every single Monday, I show up here for engineers like you [music] to help you progress your business, your career, to help you build the most valuable software [music] you can in the age of agents.
¶110You know where to find me every single Monday. Stay focused and keep building.