Mohit Aron, a proven tech founder of two multi-billion dollar technology companies, Nutanix and Cohesity. As you are starting to build now, how did you think about technology moats in your current company?
What is an LLM? It's basically intelligence. For the LLM to be effective, it needs tools, and the tools in the age of LLMs is basically APIs. Moats are the depth of what an API can do. People who say that SaaS is dead are really confusing the UI part of the SaaS with the API that the SaaS provides. I think it's a mistake to spend time tuning LLMs. I think if you truly want to penetrate the enterprise, you don't.
Thanks Mohit for doing this. I'm really excited to talk to you about this topic of moats in the age of AI as the guy who has seen almost 20 years of technology shifts and you are coding every day. You're deep in technology. There's a lot happening in the AI-first world where startup founders are told, 'Oh, your technology is a wrapper on top of some AI,' or 'You are a thin UI layer,' or 'You don't have proprietary data,' or 'You're just fine-tuning models.' So, first principles, the topic is how do we think about moats and technology, what will persist in the age of AI. And so, as you are starting to build now, how did you think about technology moats in your current company?
Let me start with drawing an analogy. Let's go back to 3,000 years when we had cavemen. Do we agree that the caveman also had intelligence? The same intelligence we have, that person also had it. So what did the caveman not have? He or she did not have the right tools, right? And so what they could actually do was kind of limited. Now with the LLMs, what is an LLM? It's basically intelligence. So for the LLM to be effective, it needs tools. And the tools in the age of LLMs is basically APIs. The moat in the AI age is actually these APIs. So think about maybe you want to build an AI agent that can book vacations for you. Unless you have access to a site like Expedia, which becomes the API, well, your agent is not going to be able to book a vacation, right? So I think the moats are the depth of what an API can do. That's, I think, the biggest moat in the age of AI. That moat is not going away. And so I've also heard people talking about SaaS is dead and what not. They're really confusing the UI that SaaS offers, right? In the age of AI, when AI agents are going to reach out to a site like Expedia, the UI that Expedia offers may have less value. I mean, it still has value because humans are also going there. But since AI agents are going there, the beauty of the UI is less important, but the service the API provides is still important, right? And therefore, I don't believe that SaaS is going away. People who say that SaaS is dead are really confusing the UI part of the SaaS with the API that the SaaS provides service.
Yeah, that's right. That's such a clarifying sort of analogy and the caveman and the tools. So, and then if I extend that as you're saying, so the startups need to evaluate the value of their underlying service. Yes. SaaS user interface can be a voice interface, can be a chat interface, which is nice moving to, but the value of the underlying service is where value accrues to eventually. That is correct. And you also talk about ecosystems and which sort of connect to these tools. Talk about that a little bit. So the next thing I've heard is that, well, what if I vibe code everything? The problem is that the existing service may live in an ecosystem. So there are other people building tools that integrate with you, right? And they will not integrate with this vibe-coded thing. So another moat comes from the fact what your ecosystem is, right? Who else interoperates with you? And if there are tons of things that interoperate with you, even if someone were to vibe code something in like a day, that thing, even if it provides the same functionality, very likely it's not going to be able to integrate well with that ecosystem. And so that becomes a moat by itself. The third, I would say, is let's say eBay is an example. All the buyers and sellers are there. If you vibe code a clone of eBay, suddenly, you know, the moat from eBay is not going to go away, right? Same thing for WhatsApp. You can vibe code a clone of WhatsApp, but given that all of our friends are on WhatsApp, they're not going to start using this vibe-coded service. So this is where some of these moats come from.
Makes sense. And maybe even going back to your tool analogy, one thing which we see a lot is that there are workflows which the enterprise user is accustomed to in their day-to-day. So it could be a hiring workflow or a quality assurance workflow. Those workflows, in my view, have fundamental value. Now, the user interface of those workflows can change. I can engage with voice or chat or something. But when you talk about the value of the underlying tooling, right, are you talking about workflows as well? This goes more into that interoperation. So let's say you're using an ATS, something for recruiting, and that has some workflows. Well, whatever works with it is used to those workflows, and especially if it's a third-party tool that works with it, that knows that workflow and works with it. So, if you try to replace that with a vibe-coded thing which may not have the same workflow, then these external things will not be able to work well. I mean, if they somehow use the same API, right? And so that itself becomes a moat. So, that goes into that interoperability that I was talking about, that ecosystem.
Makes sense. And then moving to the next point, data, right? So now recently I read an announcement from one of the large SaaS companies that they won't let these AI agents leech on their data to build stuff on top, and it talks to your interoperability point as well. Does sort of having proprietary data become a moat? If yes, then will companies build walls around it? Will Salesforce say tomorrow that nobody else can take Slack data or Salesforce data? Or how do you think about data? And then there's the model training part of data as well. We'll come to that next, but how do you think of the borders of data?
Data is the new oil. Data is always important, but what protects the data? Data sits behind APIs, basically. If the APIs are closed, well, that becomes a moat. You can't access the data. Like I said, unless you can interoperate with that, you have a moat. But eventually, do you think, especially given how there are a lot of AI agent companies now, these systems of records will have to, because in one way the customer owns the data, right? Do you think eventually the customer, like these big companies, will have to give in to the customer and make it interoperable? Or will Salesforce and Workday and ServiceNow make it a point to stop this? So, let us say hypothetically, I think you mentioned Salesforce, some new company comes and you're able to migrate the Salesforce data to that, and that company also provides that ecosystem, that interoperability, then, you know, that will be a threat to Salesforce. But unless moving the data is easier, providing that ecosystem is harder, right? And that's where the moat comes from.
And then data from a model training perspective, right? A lot of enterprises, at least we see in our business, they want their own proprietary LLMs or something fine-tuned to them. In the age of AI and context memory for the AI agents themselves, how important do you think startups and early-stage companies should focus on that part of it? Or how big is the importance of data and model?
Anecdotally, it's becoming less and less important. Okay. So I've heard so many people just say that, look, just use an off-the-shelf LLM, a vanilla LLM without any fine-tuning, and just use basic RAG to make it work. Now, that said, you can improve latency some by fine-tuning and baking in some of the knowledge. But that said, there is always going to be some RAG anyway, because after all, not everything can be taught. There's just new data, right? And so that data needs to be sent as context to the LLM anyway. But you can definitely train it some to improve latencies. But I think we started talking about step one startups. I think it's a mistake to spend time tuning LLMs. I mean, first, just use the vanilla LLMs, send them the context. I mean, these days context windows are getting larger and larger. So send them whatever you want to send them in the context. And once your product is mature, somewhat stable, then you can actually work on reducing the latencies by fine-tuning those LLMs. But just don't try to do that too early.
Got it. And so if you are advising somebody who is like an early technical team of a step one startup or early-stage company, or looking to join, and investors say this often, 'Oh, you're just an AI wrapper on top of an AI service.' Your mantra is look at the tooling, look at the underlying value of the service being provided, look at the ecosystem, and look at some of the data moat, right? If the, you know, again going back to the service you provide which sits behind an API, if that is deep and phenomenal, that's what matters. Right? Look, eventually AI is just you're tapping into these things and providing some goodness, but the depth comes from what sits behind an API, and if that is good, that's your real moat. So one related question to that, and this is not for a step one company, but it's from the perspective of like these days a lot of companies are getting funded where they're disrupting a legacy ERP system or a legacy CRM system with an AI-first ERP, an AI-first CRM, an AI-first accounting system. You know, every top fund has funded a company in these categories. And in that way, in your definition, they actually don't have the ecosystem. They're starting off, their service is very nebulous compared to a NetSuite or something, right? How do these companies compete with NetSuite? Or who will stay two years from now, three years from now?
Yeah, so it's a very, very astute observation. Let's take an example. Let's say, I think you mentioned CRM, let's start with that. Let's say somebody says that I vibe-coded a CRM, use my CRM, and you go to a big customer, and the customer says, 'Well, maybe your CRM can do everything that my legacy CRM can do, but I have this big ecosystem that works with my CRM, and that ecosystem doesn't work with you.' So that big customer is not going to buy. So now to sell, what you're going to do is you're going to go to a customer which doesn't have that ecosystem, and that customer more likely than not is like a small company. So you end up selling to SMB companies. And one thing about go-to-market is that the selling motion to sell down-market is a very different selling motion compared to the selling motion up-market, right? Up-market you need sales teams, you need a sales guy. Down-market you need leverage, you sell through the channel, you sell using inside sales people, you sell through a lot of marketing. And once most of your revenue starts coming down-market, your selling motion basically becomes attuned to that. So you become an SMB company, and that's the problem. I think if you truly want to penetrate the enterprise, you don't go that route. You basically, again, it's a time-tested thing, you go the route of solving a pain that nobody else can solve and insert yourself, get a foothold with that, and then over time you can build into an ERP system or a CRM system and surround yourself with that ecosystem. But you need an insertion point. You can't just come and one day say that I'm going to replace the CRM. That's madness.
