Transcription
It has never been easier to launch apps. People who couldn't write a single line of code six months ago are now launching AI apps and raising VC rounds. But when you look under the hood at the actual economics of those apps, the picture changes dramatically. So, is AI software business worth it? If you strip away investor pitches and VC rounds, let's talk about the economics of AI software apps and why the vast majority of them are bad businesses. Episode six of my series AI hype versus reality. Let's dive in.
Before we talk about the numbers, here is what you need to know about margins. Look at this picture. This pretty much sums it up. Now, traditional B2B margins range between 70 and 90%. With best-in-class SAS applications achieving around 80% or higher. This is because software businesses are highly scalable and the marginal cost to serve each additional customer once a platform is built is near zero. Remember the video about tech sales where I was telling you that tech sales is a fantastic job if you can't do sales because there is no inventory or no physical product. Now that's exactly the reason why software is so scalable. Now the cost of acquisition of each new client can be different depending on the product and depending on the business. On average, if we take for example Shopify and you come to them as a customer and you want to do business with them and use their platform, when you go and sign up, it costs almost nothing for Shopify to have you as a client. Any additional cost to serve you is well covered by the amount of money that you're going to pay Shopify.
Now, let's talk about AI native SAS. Here's the data. 30 to 60% margin at best. Genai native SAS and especially LLM wrappers which the vast majority of them are have 30 to 60% margins with 60 being best-in-class products. Now even the most mature GI products like Anthropics Claude are at 55%. Listen to this one more time. Best-in-class is 55%. Now there's also evidence that the products sitting in the upper range of margins tend to have a lot of dead subs. Now what are dead subscriptions? A dead sub is when a user is paying but not using. And this specifically applies to those cases when you guys purchase access in bundles. I don't know if you've come across those videos on Instagram. I see them all the time where they sell you a bundle of AI tools. AI tools for content creators, for product managers, for data folks, you get the idea. Genai products are still riding that wave of hype, allowing companies to get some revenue through dead subs. But this is not going to last long.
One more thing. Remember this old video when I was sharing my thoughts on how to pick a career in tech that lasts. I want to bring you back to the concept of the hype cycle one more time. Now this is where Genai is at the moment. It is very common among tech trends that are reaching the peak of the hype cycle. But sitting at the peak of the cycle has absolutely zero correlation with long-term benefits. For instance, let's look at the blockchain trend by my favorite example. Remember how revolutionary it all sounded back when cryptocurrencies became a thing? But when you really dig into use cases around blockchain, there aren't that many things that need to be solved with blockchain and the ones that cannot be solved in any other way. Same thing with augmented reality, for example. I don't know if it's an unpopular opinion, but really there aren't that many use cases outside of entertainment that are truly monetizable when built on AR.
Can software companies stay afloat for many years despite being unprofitable? Yes, and there are numerous examples. Asana sustained years of losses while continuing to operate and grow. Monday.com has been on a continued unprofitable streak. Marquetto lost 36 out of 60 million in cash when filing for an IPO and numerous losses while still expanding. And this strategy works and it works more or less okay when it's a B2B and not AI native product. But with pure genai, the unit economics and the profitability path is entirely different. So we can't and shouldn't compare it to the traditional business models of SAS.
Now why is there such a stark margin difference? In traditional SAS after initial R&D and platform investment serving more customers adds very little extra cost. In Gai native or GPT wrapper products there are major ongoing costs per user API calls compute time licensing sometimes per output moderation. Once again in traditional SAS when you acquire a new customer and in B2B's case the customer is a team of people that you on board onto your product the biggest cost for the business is almost always service related a dedicated CSM support specialist sometimes implementation or delivery manager point is those are services and those services are not being used consistently in Genai native products costs are not only rising they can even become exponential so the companies have to cap usage to at least somehow make those ends meet for example Chad GPT cost OpenAI $700,000 a day in 2023. Now, the cost did go down and this year the sources site a range starting $100,000 to several hundred,000 per day depending on usage and models. But multiply this across multiple users or particularly prolific AI super users and users can individually cost OpenAI more than $200 a month which is one of the most expensive plans. OpenAI had to meter usage even at premium tiers because the cost per user could easily surpass revenue. Another example, GitHub Copilot launched at $10 a month, but it's been widely reported that the cost of a user for Microsoft was almost $30 a month to serve, meaning they were losing money per each active user. And shortly after they estimated that the cost for power users were actually around $80 per developer per month. And lastly, midjourney, they did offer a low cost plan, but they had a very strict limit to how many images you could generate. And that was because each image consumes significant GPU resources. And the funny thing is that if you think about it, we as users get used to this all you can eat usage behavior that when we hit those limits and we are on a paid plan, we go, "Okay, I'm not paying for this." I personally had a majour subscription a couple of years ago and I canceled it within a few months. But this payw walling is necessary to prevent runaway costs from a small set of very heavy users. So on the surface it's all software but in Genai costs for the company who run the business become exponential especially when they offer the all you can eat or unmetered fixed rate plans. If a GPT rapper startup puts a low price on AI usage or offers a beefy premium and doesn't limit expensive features, a minority of users can generate costs that will scale exponentially. But the paradox is that when they do put in usage controls, the users don't like the experience.
So how do AI companies make money and do they make money if you strip away VC funds? Let's play product management 101, shall we? As of August 2025, it is reported that Chad GPT has approximately 700 million total user base. Now, different reports show different numbers, but all of them agree that by the end of the year, they're expecting to reach 1 billion users, which leads me to believe that they're at approximately 7 to 800 million users right now. Now, of those 7 to 800 million users, how many of them are paid? 10 million on Chad GPT plus plan. Now, I did the math for you and that's less than a 2% conversion. If you're a PM or a founder listening to this, I would love to hear your thoughts on this conversion ratio. But let me tell you, when I hear 2% conversion to paid, my eyebrows get raised, especially when we're talking about a worldwide disruption. This is alarmingly low. In product management terms, your PMF goes out the window or you never had it in the first place. So, how is it possible that one of the most visited websites in the world, a major widely recognized disruptor, has less than 2% conversion rate? To be clear, I have found sources that show five or six or 7% conversion, but honestly, if you do the math, it should be less than that. I truly don't know where they got that number from. But if you want to correct me and it's truly five or six or 7%, let me know. But even with that number, a 7% conversion for a major disruptor is alarmingly low.
Moving on, in December last year, OpenAI published the following numbers. 300 million weekly active users and then the number went up to 400 a month later. Does anyone notice anything unusual? Weekly. Why weekly? Why not monthly? Monthly is a much more common SAS metric than weekly. Nevertheless, they chose to go with weekly. One more thing, weekly active users went up 100 million, but the site traffic hasn't changed. How is that possible? Personally, I think that's getting a little bit into the territory of vanity metrics. It's a nice big round number that you can show to the whole world, but does it really translate to revenue? And last, but not least, the API revenue data. This is not official data published by OpenAI, but if you believe the research that this company did, it becomes pretty obvious that AI is not going to replace us all. At least not right now. And now to put this in perspective, let's look at OpenAI's biggest competitors. monthly active users, Chad GBT 3 to 400 million. Its main competitor Claude just 3 million. Gemini 47, but that is heavily boosted by Google's reach. And Copilot 33 million, which is again boosted by Microsoft's reach. Let those numbers sink in. The Chad GPT, one of the most visited websites on the planet, has more than 93% of users on a free plan. Depending whose conversion numbers you believe. And this means that every free user generates losses for OpenAI because the computing costs are still spent. And let me remind you that Chad GBT is the absolute leader of Genai products, the top of the pyramid. The difference in market share between Chad GBT and Claude, its biggest competitor, is a factor of 100. If you assess Chad GBT as a product purely by looking at the numbers, forget that it's OpenAI, it's actually doing worse than an average SAS product. Do you see what I'm getting at? This is why I'm saying that traditional SAS metrics don't work with Genai native products. This is why I'm trying to prove that as disruptive and as mind-blowing as AI seems, we're so early in the real adoption cycle. And when you start seeing those metrics, it becomes pretty evident that LLMs have reached their peak. Sure, models get better. They code better. The error rate is lower. They meet higher benchmarks. But it's almost like the invention of a car. Sure, we have Teslas. We have Bugattis. We have Hondas and Fords. Better or worse, faster or slow, electric or gas, it's a car. It's not becoming a teleportation machine. All of these products essentially deliver the same value. It's clear that Chad GBT's reach is driven largely by media hype and much less by actual product value. The economics of foundational AI companies like OpenAI or Anthropic are royally unprofitable and the way they exist today. Their business models are extremely unsustainable. And now that we're moving into the age where LLM becomes a commodity, we will be seeing drastic changes in their pricing, bundling, and growth strategies as a whole because they have to maintain the market share. So, right now, we're witnessing a rat race on the way to profits.
Now, don't get me wrong. Those companies that are building foundational models are not typical AI businesses. And I'm not bashing them by any means. The revolution that they're making, the groundbreaking progress that they're bringing to this world is something that requires very heavy investment even though it's not pumping out money right away. So, yes, I'm using their metrics because their metrics are first of all available. Secondly, those are the biggest AI products on the planet. So, it's easy to use them as benchmarks. But I want to be very clear. I am aware that companies like OpenAI should not be judged based on their VC rounds. Companies like them should be racing because their technology is what everything else is built upon. The point I'm making in this video is that there are four, five or 10 companies like OpenAI. Everybody else who keeps raising rounds are not OpenAI.
Let's talk about the AI price wars. The AI SAS space is experiencing rapid price wars, further squeezing margins as much as possible, especially for wrappers and tools layered on top of foundational models. I suppose AI companies will be expanding their product lines to try and upsell users more aggressively because they have to make those ends meet. And as the hype and that initial fascination keeps wearing off, companies and investors start paying much closer attention to the economics of AI apps. How to price AI software so it becomes profitable is a very difficult question. Claude is doing a really smart thing in my opinion. I love the sonnet family of models and when I was in the free plan they gave me just enough to test it out but then they put in those query limits and they renew every few days. So the experience is good enough for you to get the value and annoying enough to make you upgrade if you're a savvy AI user. Now B2B AI apps often price depending on the impact that it creates. So in a nutshell, everybody's trying to experiment and come up with a working model, but on average, the profitability of a genai native software is very far from traditional B2B or even B2C SAS.
So after hearing all of this, let's think about how we can tell which AI SAS business can actually become sustainable and profitable as a business model. Here's what I think is a good litmus test. A traditional SAS startup with AI features for tasks that can be automated. Instead of building or investing or using a Genai native startup, you have to be looking for a SAS startup with AI features for tasks that can be automated. And this company or startup has to be working off of traditional SAS benchmarks, not bubbled up AI metrics. The point where you know that your product is valuable is when it is valuable without AI and when it could solve the real problem without AI. If there is AI component in it and it makes something better or faster, fantastic. But AI should not be the defining factor of the product.
Now are there companies that are still profitable by being purely Gaii native and building their positioning around Genaii? Yes, and some of them work. But when it happens, it usually follows the same scenario. The vast majority of genai startups that work are the apps that work with large amount of text-based data and documents in various industries be accounting, HR, sales or legal. And what they do is for example they stitch together contract data with invoices, automating contract customizations, connecting to CRM tools and basically automating the back and forth between various parties. Now is it an example of a billion-dollar problem or a billion-dollar business? No. But it is a working business, a working AI business. And you can get pretty creative with how you sell it. It can be a standalone product that you sell to businesses. It can be a plug-in for other products. It can also be sold as an API. Now, it's not going to make billions in revenue, but it is a sustainable and viable business model. The most difficult problems that I personally haven't seen any AI startup solve yet are those lengthy, effort consuming, old school problems in very boring industries. US enterprise world is full of boring business problems that need to be solved. It's also one of the reasons I told you in my earlier videos that if your job gets automated, the world of enterprise and especially legacy enterprise tech will always be there for you. Under one of my most recent videos, somebody commented that legal tech is really easy to automate because legal industry is text-based. I beg to differ. I work in legal tech, a very classic traditional industry. And let me tell you that if you're able to solve a problem that is prevalent in industries like accounting, legal, or pharma, you actually do have a chance at building a unicorn. And that's a pretty big if. I'm not talking about a GBT wrapper that sends contracts back and forth. I'm talking about automating truly difficult analytical or consulting work done by, for example, a criminal lawyer or corporate lawyer. Pretty much anyone who consults on cases or handles rounds of negotiations. A friend of mine, a very experienced lawyer, once told me that if you can build a piece of software that can negotiate, really negotiate with a fellow lawyer or in court, you'll be the next Elon Musk. Because that is the kind of problem that is incredibly difficult to crack and the one that can bring you a ton of money whether your business is geni native or not. It's the kind of problem that will outlive blowing up VC rounds or hype cycles. There is nothing sexy about it, but that is where the gold mine is.
Conclusion. It's easy to get a user to try product with big promises and even more so with FOMO and scare tactics like AI will take over your job. Don't be left behind by progress. And it's even easy to get people to pay. What's really difficult is to deliver value. Value that a few can replicate. And how do you know you're delivering value? The same metrics that have always mattered. Retention and conversion rates. Those two things will outlive blown up VC rounds or hype waves. So before you start thinking, "Oh my god, this program or this app is going to replace my job." Ask yourself, would this app solve a real problem without AI? If the answer is no, you're fine. The hype won't last forever, but boring, profitable businesses will. And that's it for today. As always, we hope this was helpful. We'll see you next time. Bye.