初创公司如何在 OpenAI 和 Anthropic 之外胜出 | HubSpot 的 Dharmesh Shah

ChatGPT and cloud services have effectively become the new OS.

There is a lot of room for software creation outside the path that AI labs are rushing down. HubSpot has been working tirelessly ever since.

The most valuable thing is the work itself that’s happening. The interface is just a means to an end.

Building an ecosystem is very similar to building a product. You have to think about it like this. We had a very broad base of marketing applications. We chose this path because it’s the only way to solve the problem: customers are scared of this fragmentation.

Everything was fine. We had to do it all in one decision. The architectural solutions we laid out for HubSpot originally relied on a person managing all this information. Now we have incredibly powerful models that see things very differently.

What structure do we need for HubSpot for the next 20 years, knowing about this world of AI?

A few points. First…

1. The Best Time Ever to Be a Builder

Welcome to the A16Z podcast. I’m Joe Schmidt, and today we have Dharmesh Shah, co-founder and CTO of HubSpot. We’re continuing the conversation that began with an exchange of thoughts on X about my essay, “The Yellow Brick Road.” The idea is that there’s a lot of room for software creation outside the path that AI labs are rushing down. They launched HubSpot for business from a single person and are rebuilding the platform around agents.

Today we’ll talk about what’s worth building, what a sustainable software company needs to change, and what happens when AI becomes the primary way your product is used. Dharmesh, thank you for joining us.

Thank you for inviting me. Yeah, I’m very happy to be here.

Dharmesh, like LeBron, you have a one-word name. It’s pretty exciting. This is an interesting conversation for me because I wrote that essay a few months ago in response to a sense of nihilism that was prevalent in Silicon Valley, and it seems to still be going on. So it must be strange to be running a large, successful public company during this period. How does it feel? I’d like to start with a market overview.

On the one hand, it’s like the classics: the best of times, the worst of times. These are the best times because we have that energy and excitement that comes with such a large-scale transformation. HubSpot is 20 years old, and when the company was starting, the idea was that there was a new thing—the internet—that small and medium-sized businesses could benefit from for growth. But they couldn’t yet, because everything was too complicated and the tools were disparate. That’s where HubSpot started.

Now, 20 years later, there’s this new thing called AI that we think small and medium-sized businesses should benefit from, too, but it still seems scary or something. So it feels like a new, refreshed mission for the same customer base, just with new technology. On the other hand, it creates a lot of uncertainty and “fog of war,” so we, as a team, try to focus mostly on customers and let the markets do what they want. We have our own theory about how things are going and what we’re betting on, but I’m focusing on the positive aspects of it.

I’ve been building commercial software for over 30 years, literally every day. It’s the most interesting development experience I’ve ever had.

What a great time to be a developer. I completely agree, and it’s so strange that the perception is negative when literally everything works. It’s like falling out of a boat—you immediately find yourself in the water.

Maybe if I go back, I’d like to talk a little bit about that formative experience, of course, with Pyramid, and then with HubSpot, and how you got there. I’m sure people back then were saying things like, “Hey, it’s gotten a lot easier to build software, so what are you even developing now?” I’d love to hear your comments on the similarities and differences in how you build companies today and what you observe in new startups.

I’ll give a slightly geeky answer, because I’m like that. Having been in the software industry since prehistoric times—150 years ago, you know—the history of software development is always about increasing levels of abstraction, right? When I started, we still had machine language. We worked on mainframes and things like that. Then came the 3GLs, C and C++, and later 4GLs.

We asked ourselves: Can we make it easier to get from your idea of solving a problem to its implementation? It used to be very difficult, but then it became easier. We had certain jumps, a nonlinear increase in the level of abstraction. At its core, that’s what excites me about AI: it gives us new tools to work at even higher levels of abstraction. That’s what really changed, isn’t it? In my opinion, the problems facing humanity and business remain the same. But now we have a better set of tools, and I have a certain programmatic approach. This is my element.

In a way, it’s the same, and we’re going to digress a bit, because recently there was a Jev model released by developers at Typesafe. This brings us back to the fact that it’s all just software, and we have another dimension of development that isn’t transformative, but opens up new possibilities. That’s what it gives us. We’re just adding more tools to the developer arsenal and expanding the circle of people who can become developers, which is great. In many ways, things remain the same, and it’s nice to be able to work at a higher level of abstraction. That’s one of the reasons why this is a great time.

It’s always good to be involved in a startup, but especially now, because thanks to high levels of abstraction, AI allows you to iterate much faster. So you can learn lessons that would previously have taken 6 months, a year, or 2. It sounds like this: just try it, build something, release it, test it with real users and customers, and see what happens. Previously, this was simply impossible.

2. Building an Ecosystem Developers Trust

Oh my God, I totally agree with you. I wish you could see all the things that we create here at our venture firm. Being on the front lines is much more interesting. This actually brings me to the point I wanted to discuss, which is this: historically, if you look at the business you’ve built, you started with a very narrow product for a very narrow set of customers. How do you see these changes with a rapid iterative process and the ability to build much more, to create much more software much faster? I’d like to hear what you think about this.

3. The 60% That Weren't Using a CRM

I think I’ve been in the startup world for a long time, directly or indirectly. I still consider myself a startup person, and choosing ideas has always been one of the most difficult tasks, especially for novice founders who haven’t had that experience yet. Even now, 30 years later, I think I’m doing it a little better than before.

One of the common mistakes entrepreneurs make when choosing ideas is thinking that it’s a matter of careful analysis, as if some 18-column table will give all the coefficients that yield the best possible idea.

Yeah, how many times have I seen files like the U.S. Census? Exactly. “Oh, there are 18,000 law firms.”

In my experience—and this isn’t just my experience, not just one data point, but conversations with other founders—it’s almost impossible to know a priori the “quality” of an idea, its likelihood of success, or what the outcome will look like. The only real way to test an idea, or find out if it’s a good one, is to actually try it out in practice.

In the end, even if you succeed, what you come up with will often not be anything like the original idea you started with, right? The things that people now perceive as, “Oh, that was a successful business,” sometimes aren’t similar at all. They may have some similarities, maybe the gene pool of the original concepts is still there, but it’s something else.

The side effect of this kind of thinking is that, since you hardly know whether a particular idea or set of ideas you’re exploring will be successful, the smart move is to say, “Let’s release the idea as soon as possible and find out if it’s good or if we should look for something else.” That’s relevant today, going back to AI: we can iterate faster.

The only thing that remains difficult, and I would say maybe even more difficult than before, is distribution, right? It’s much easier to create software now, and significantly more people will be doing it. Therefore, there will be many more options, and it will be difficult to break through this noise and establish contact with the client.

I’m a big believer in starting with—and I’m not the only person who says this—narrowing your focus early on, because that’s what increases your chances of breaking through and learning. The second mistake entrepreneurs make is feeling the need to start with a huge market, like, “Oh, this could be a $100 billion business.” It’s great that you have ambition, but tell me what you’re going to do in the first 30 days. What exactly are you going to focus on? First, income, and then we’ll reach $100 billion.

Even now, when I explore new ideas within HubSpot—both for HubSpot Next, our innovation arm, and for new ideas within our core platform—it feels like a narrowing of the field of vision. Can we somehow… I think whoever has the shortest iteration time and the lowest learning cost wins. If you can shorten this cycle and learn more cheaply, then over time you’ll ultimately win on average, no matter what the lessons are.

4. From LegalSpot to HubSpot: The SMB Bet

Yeah. No, I still believe that this moment is the best for people with high initiative. The reality is that building software has become a lot easier, but not many people are sitting alone at 10:45 p.m. on a Saturday night, exchanging emails, working on a computer, and just trying to build something cool.

I want to go back to where I think HubSpot started. You started, correct me if I’m wrong, with a legal product. Maybe there was a legal product somewhere in there, and somehow it turned into a CRM for small and medium-sized businesses. It’s incredible how far you’ve expanded that initial tool. I’d be interested to know how you came to what became a marketing tool for SMBs, and how you expanded from there to an entire platform.

When we first started, we were in business school. That’s where my co-founder and I met, here on the East Coast. We came to understand the attractiveness of SMBs as a market. It was under-covered, and I’ll tell you a short story because it’s interesting.

We’d both worked in the software industry for a long time. He was in marketing and sales; I was in product development and engineering. We’d both worked in enterprise software, so we knew the field well. The problem with enterprise software was that it was no fun at all for a startup entrepreneur. Sales cycles were long, and revenue was concentrated. You had a small group of customers who influenced your product roadmap, vision, and other things.

On the other end of the spectrum, you have consumer markets, right? Consumer software companies often have very bimodal outcomes: if you succeed, it’s a colossal success. Often, you have to use advertising subsidies or find other ways to make money, because you can’t set a price for the product without looking stupid.

The beauty of the SMB segment—small and medium-sized businesses—was clear from the market analysis. The two of us were in business school, thinking, “SMBs have an advantage over the corporate sector because you can actually charge money for solving a problem. There’s nothing wrong with taking money, and the market has the scale of a consumer market, where you won’t have that concentration of revenue. You can have tens of thousands, hundreds of thousands, or millions of customers.” So you get the best of both worlds, right?

That was our initial reasoning. If we were launching a startup, our first idea, which we took to the MIT $50K competition, was an ERP system—a kind of Oracle for small businesses.

Yes.

Going back to narrowing the focus, we thought this looked interesting because we were just going through the transition from client-server architectures to the web. Someone should create an Oracle for small businesses, where everything is integrated, works together, and is built specifically for SMBs.

When we went to the business plan competition, we were told it was too big. We narrowed the focus and called it LegalSpot. The company name HubSpot already existed, but the business plan we wrote, purely for early market testing, focused on ERP for law firms as a narrow niche.

Going back to your original point, I was thinking about the U.S. Census Bureau data, too. There were a lot of them. We could multiply X by Y, and that makes for great spreadsheets and PowerPoint presentations in business school. So we did. We also have a tendency to test the waters.

One thing we realized early on, when we were in business school and thinking about the idea for HubSpot, was that this wasn’t an easy market to sell to. When we talked to law firms about ERP and what the solution could be, marketing and the website turned out to be the aspects they were most interested in.

They were like, “Wait, wait. Can you help me find more clients?”

That’s right. Yes.

We were like, “We have time tracking and everything.” It didn’t really matter. They wanted to know, “Can you help me find more clients who are a perfect fit?” That was the kind of thing they were asking.

We were hesitating between 2 ideas: Should we stick with ERP for small and medium-sized businesses, or should we do marketing for small and medium-sized businesses? We could almost have flipped a coin. We didn’t; we were a little more measured about it. We drew conclusions based on a small amount of data, but it seemed like life was short.

As it turns out, lawyers are the second most difficult market to sell to, and software developers are the most difficult. So we focused on a marketing product for small and medium-sized businesses. Of course, that turned into a platform and a full-fledged CRM system.

Was this thought out from day one? I ask because I think a lot of people are using new AI tools now, whether it’s voice products, as was the case in the early stages. I still think it’s a huge opportunity for many. I’m sure you’re researching this at HubSpot. Or maybe it’s some kind of chat support interface. Everyone always wants to become HubSpot for their use case, right?

Could you talk a little bit about how you went through that product journey and created the right cycles to go from the initial solution to the platform?

We started as a set of marketing applications. To answer your first question: at the initial stage, we didn’t think we would ever become a CRM platform. That wasn’t in any of our presentations. It wasn’t something we had planned.

But as the company grows, you might need a second product. This is a bit specific to HubSpot, although I don’t think it is. I believe this pattern exists even outside of CRM. You need a second product, your second act, or whatever it’s called. We’ve come to understand that.

Several things happened at the same time. We were in the marketing software business, and we realized that marketing software is very close to sales and CRM. It was a related category, so it seemed logical.

Another thing we found out when we already had a bunch of customers was that we could survey our users. This was probably a couple of years before the IPO, about 5 or 6 years after the company was founded. We wanted to see what CRMs they were using, and we found that 20% were using Salesforce. Another 20% were using some other CRM besides Salesforce.

Yes.

But 60% weren’t using CRM at all.

Yes.

For us, it was crazy. We were like, “Wait. You’re using HubSpot software. You pay us money.” These weren’t random companies; they were people paying us to generate leads. What happens to those leads?

So we concluded that these companies should use CRM. Michael Fenn and I just think very strangely. We were like, “Okay, why aren’t they using CRM?” These categories already existed. It wasn’t that there weren’t great companies with great products. We realized it was just too complicated. They were afraid of the very idea of using CRM. They thought it was too expensive and that implementing CRM would cost too much.

So we decided this was a great related niche. We needed a second product anyway. There was strategic value in recognizing that existing CRM players would probably enter the marketing space over time. Even if we wanted to stay in the marketing software category, we couldn’t do that forever, right?

All of these things came together around the same time. When we launched CRM, we solved both problems: we built CRM to appeal to that 60%. Why hadn’t people chosen a CRM yet? Was it too difficult? Okay, let’s build the simplest CRM imaginable. We focused on small and medium-sized businesses, which was our main advantage anyway. And we would eliminate the “too expensive” problem by making it free.

Mm.

Not freemium, but free. That was a certain philosophy at HubSpot: create value for the customer before trying to get something in return. If you can come to market with something useful, we could certainly subsidize it with the marketing software business we already had. That was the idea: remove the barriers and go to market.

Then, in the process of improving the product, we realized that to win in the CRM space, we needed to become a platform. It couldn’t just be one thing, because CRM, by definition, was a platform, even then. So that was the journey: from the first product to the second, expanding into a strategically related area.

Here’s another thing I think will be useful for entrepreneurs: do everything you can to solve the customer’s problem completely. How do you define it? Whatever circle we draw, what’s the real problem we’re trying to solve?

Even when HubSpot started, we had a very broad base of marketing apps, which goes against all the advice you get—and that I’ve given myself. The advice is, “Do one thing very, very well.” We did everything: content management, SEO, landing pages, forms, and so on. In each of those categories, there were great companies with great products.

The reason we decided to do everything was that it was the only way to solve the real problem customers had: they were scared of the fragmentation. We had to combine everything into one offering. That’s an idea we’ve carried with us for 20 years. We continue to do the same thing now. The question is: what’s the real problem we’re trying to solve?

It’s very interesting because, looking back, everyone always says, “From a go-to-market strategy perspective, there’s the brownfield approach, where you look for people who are already using the core systems and just replace them, hoping you’re in the middle of the cycle when everyone is ready for change. Or you go greenfield and try to bring in new companies that have never used the software.”

I think the standard thinking about HubSpot is, “It was very well thought out from the start. Let’s focus on greenfield.” Actually, it wasn’t like that. Our clients just didn’t have those systems, and we built it. If you do everything in that order, you can still become a huge company, if you can manage it.

Very interesting. What do you think about the other part of this process: building the right ecosystem around the initial offering? I think HubSpot has done that incredibly well. What are the key takeaways from the early days of building the ecosystem, and what does it look like now that the world is completely different and you can implement a lot of things differently? I’d like to hear your thoughts, given the advent of AI and how it affects the ecosystem.

I’ve been a user of platforms and ecosystems for most of my professional career, so I know what it looks like from both sides. This isn’t a criticism of any existing platforms, then or now, but what often happens is that vendors bring in developers to build apps or companies that add value to the core platform. Over time, though, something happens: intentionally or not, they seem to suppress these developers, close off opportunities, and stop working with them. We’ve seen this play out over more than 15 years.

That’s how Twitter was in its early stages, right? The original Twitter was a very open ecosystem, a platform where literally thousands of possibilities blossomed. A lot of Twitter clients were built, and a lot of people built products and companies on that initial set of APIs. It’s not just Twitter; it’s just the first thing that comes to mind. But then they kind of shut it down and became unfriendly to developers, almost hostile.

One thing that I think helped us build the ecosystem is that we took those lessons to heart. We thought, “Okay, let’s be the platform we dreamed of when we were doing things ourselves.” It’s about carrying the HubSpot brand, being friendly and being nice. We’re genuine capitalists. We like to think of ourselves as kind-hearted, sincere capitalists. Everyone is here to make money. This is what we do.

We try to be as open as possible with our partners: here are the things that work. We can’t promise we won’t do this, but here’s what we do. If we were you, here’s what we would do: here are the opportunities we see, and where we’re unlikely to go. And it worked. We built trust among a wide range of people.

Part of it is that building an ecosystem is a lot like building a product. As a developer at heart, I understand that just as we think about time to value for core customers, we need to think about time to satisfaction for the developer. How quickly can someone go from, “Oh, I got something working with the HubSpot API,” to, “Oh, I made my first dollar working in the HubSpot ecosystem”? Whatever it is, can we shorten it? Can we reduce the time to satisfaction for this category of customers?

5. Staying Off the Yellow Brick Road

Yeah, no, absolutely, and congratulations on building such a successful ecosystem. Let’s move on a little bit to your essay and the conversation we had around it. I’d be interested to hear what you think about the workflows that labs are striving for. There are a lot of things within the HubSpot experience that these horizontal products are trying to capture. How do you assess the areas you want to protect and where you can create the most sustainable value? We’d love to hear your comments.

A few thoughts. First, let me say right away that I have a very biased perspective, considering that I’m the co-founder and CTO of HubSpot.

I suppose so.

That doesn’t necessarily make me wrong. We can talk more broadly, and then focus on some specific things. To those observing from the sidelines, it looks like your “Yellow Brick Road,” and I really enjoyed your essay. The idea and analogy were very apt: there’s this well-paved, shining, bright road that leads to the Emerald City or something like that. How can you not go with it? It seems so empty at that moment, doesn’t it?

But you don’t know there’s a whole tank unit moving along this road. They’re moving faster and faster. They’re not even slow, because they’ve invented a new technology that allows them to develop great speed. That was the premise of your essay, which I wholeheartedly agree with: yes, they will race down this road. It’s brilliant and beautiful, and that’s why we have these trillion-dollar valuations and a market cap that’s going to grow soon.

But there are so many opportunities beyond this path. I’m a fan of clear concepts, so here’s a mistake I think entrepreneurs on both sides make. One mistake is to say, “Oh, well, ChatGPT doesn’t do that,” or “Claude doesn’t do that. We’ll do it. We’ll fill this gap.” Ideas like these make me nervous because they often don’t work. There are 2 things I think are somewhat wrong with this kind of thinking.

First, if you can just add the word “yet” and it still makes sense, that’s a problem. For example, “Oh, ChatGPT doesn’t know how to do that yet.” Does that ring true? If so, that’s a problem. That’s the number 1 thing. The number 2 thing is what I call evaluating the product versus the company. For example, “Oh, ChatGPT can’t do that.” Okay, but can OpenAI do it? Maybe it’s a new product, a new product line, or something else.

So you should think about how wide this road is. Maybe the road itself is getting wider. Good. Maybe it’s worth staying off the road, or at least a little off it. All else being equal, it’s better to be a little further away. Then, even when they expand, it will be somewhat predictable.

That’s the first challenge: making sure you’re choosing areas where companies developing advanced models have either structural constraints—regulatory or otherwise—or what I call “incentive constraints.” They could do it, but the likelihood of that happening is very low. I’ll give you the example of HubSpot. Let’s say any company that builds advanced models—I’ll take Sam Altman, because I’m an investor in OpenAI and I know Sam—could wake up tomorrow and say, “This CRM industry looks really attractive. There’s over $3 billion in HubSpot revenue; we’re going to make it our mission, or the mission of someone in the company, to take that money.”

6. Why Models Are Becoming Infrastructure

Are they smart enough for this? Yes. Do they have ambition? Yes. But it’s precisely these ambitions that make it extremely unlikely that the OpenAI team, Sam, or anyone at Anthropic or other advanced model development companies will go for it. My theory boils down to finding niches of opportunity off the main paths.

We know one thing for sure: advanced model companies have invested hundreds of billions of dollars in pre-training, as you all know. Now their whole business is to take the capital expenditure they’ve made, and the money they continue to spend, and decide how to sell the inference. How to sell tokens? How do we monetize what we’ve created?

Yes, perfectly.

This is a story as old as the world. They have 2 ways. One is to say, “We’re going to go directly to everyone on the planet and sell them tokens,” and do it through consumer products. They can also say, “Okay, we’ll go straight to Fortune 500 companies like JPMorgan and sell tokens directly to them. There’s no need for a middleman, and we’ll have an army of specialists to help us implement this.” That’s certainly one path, and they’re following it to some extent.

Another way is to recognize that there are SaaS companies, products, and startups that know their industry and their market, and have sales channels or something that allows them to sell tokens and act as intermediaries in providing inference. It’s not an either-or choice. They can do both. They might say, “For these kinds of things, for consumers, we want to control everything ourselves. We don’t want others to create ChatGPT analogs. We will own this.” But there will be plenty of opportunities to say, “Okay, think of AI as infrastructure, just like we thought of AWS and cloud computing 15 years ago.” It’s the same.

As entrepreneurs, I repeat, I may be biased, but I’m not wrong. I think building a business around AI is less scary now than it was in 2023. I have very few rules for business. First: never own a restaurant. Second: never be a landlord and collect rent. Third: never compete with Sam Altman, although it happened a couple of times by accident.

But really, if we go back to 2023, if I were a betting person, I would say they were just way ahead. It was as if they had invented some kind of magic elixir or something. I had no idea if anyone would ever be able to truly solve this issue and provide everything needed, including capital expenditures. Fortunately, I was wrong about that—fortunately for entrepreneurs and the whole ecosystem.

Now, I wouldn’t say the models have become a commodity, but part of the benefit of their availability is that they’ve become infrastructure and will evolve in terms of price and interchangeability. There are things like OpenRouter, because it’s easy to swap models. Even HubSpot and most SaaS companies use use-case decomposition and model routing to determine which one is best for a particular task.

A side effect of this is that you can easily replace what you need. There’s less fear because you don’t have this dependency on a supply chain where one company controls the single critical ingredient needed to create your product. We don’t have that now. We have open source code.

We have many great commercial products. As long as you stay away from the yellow brick road, it’s possible that you won’t lose out on profits because you have a choice. There is no single company that provides the one key ingredient for building an AI-powered company.

7. ChatGPT Is the New OS

Yes, I completely agree. Another aspect you touched on is that, with enough focus, you can create value as a startup—at HubSpot or anywhere else—even if you don’t develop a foundation model. You can still build high-quality software. That’s the level you’re trying to master to deliver value to the customer, and there are other great infrastructure pieces that can now be packaged and delivered to customers.

The way I think about it, and what I argue in this essay, is that there are loops, tests, and other things you can do to try to identify these niches. Hopefully, you then build some form of increasing returns to scale or an advantage that compounds at that level. I’d love to hear what you think about this at HubSpot or in any other projects you’re working on, in terms of the loops you’re trying to build into these products or the advantages you’re trying to create.

Yes, Joe, there are compounding loops everywhere here. That’s the main goal, both before and after the advent of AI. But I will say a few things. I’ll go back—I put aside one idea to get it out now and discuss it. One of the questions we get is from entrepreneurs evaluating ideas.

Today was OpenAI’s DevDay, by the way. It’s September 30, so right before this call I was taking notes: What’s there? What’s happening?

I said we’re marking the recording date for our audience, so now you know the timestamp.

Exactly, you know exactly when it’s being recorded. HubSpot launched our connector to ChatGPT over a year ago. We were the first CRM company to do this, and we have another one for Anthropic.

One of the questions we sometimes hear is, “Doesn’t this reduce HubSpot’s long-term value? If everyone comes through ChatGPT, Cloud Co-worker, or something like that, doesn’t that destroy long-term value?” I understand these concerns. My counterargument is that the interface does matter, of course, but to varying degrees. There’s a whole spectrum.

Let’s say, on one end of the spectrum—I’m making this up as I go—we take something like Uber Eats or DoorDash. They have very convenient applications, but the application itself is incidental. The real value is their network, their partnerships with restaurants, and so on, so that the driver can come, pick up the food, and deliver it. It’s the logistics and all that. That’s where the real value lies, and it’s pretty hard to displace.

But there are things where the interface is the entire product. That’s all there is. If it weren’t for the interface, there would be no value at all. I think most business apps are closer to the Uber Eats model—not in the sense that they have network effects like Uber Eats or DoorDash, but in the sense that the value lies in the work that gets done. The interface is just a means to an end.

The only reason the interface is important to HubSpot is because marketers and salespeople are trying to do something or get certain results, and the interface is how they interact with the computer. Before AI, HubSpot’s main advantage was that we understood the user’s problem and created a simple interface to help solve it. But that wasn’t all the value. You can argue about exactly where HubSpot falls on this spectrum, but I can guarantee it’s not a case where the interface accounts for 99% of the value.

My argument is that it’s all happening again. When we watched the birth of SaaS and the transition to web applications, we saw exactly the same thing. At one time, we had client-server applications that ran on our computers. Then the browser came along—Internet Explorer and later Mozilla, which you all know well—and it turned out my software could run inside the browser.

Google didn’t own the entire app stack, did it? A huge number of software companies appeared, and the browser became the interface because it was a convenient way to access the product you wanted. A browser was a better way than installing something on millions of computers.

We have a similar situation now, although it’s different, right? AI performs cognitive work. It does really smart things. It can do more things. If you look at ChatGPT, Cloud or something like that, they’re actually a new operating system.

Yes.

They’re new browsers. This is the way people get value. There will be companies like HubSpot that join in, and that’s what we saw at DevDay last year and continue to see: they’ll expand ChatGPT’s capabilities. They launched this GPT-3, added a marketplace so people could connect, and created a plugin architecture. It all makes perfect sense to me.

My fundamental philosophy, which I believe is true for most of the B2B sector, is not to overdo it with the interface. Find the best way to achieve the results your clients expect. Another kind of value you can bring is doing that as cost-effectively as possible. Which method has the least delay? Which method has the fewest obstacles to achieving these results?

A large language model is very, very good at many things. It’s the most brilliant technology in the history of mankind. But I’ve already said that anything that can be solved deterministically with a structured query should be solved that way. Don’t use inference where the problem can be solved with a query. That doesn’t make any sense.

If your question is how many customers or how much revenue we get from Ireland, you shouldn’t have to look at 8 million Slack messages, emails, or whatever to draw a conclusion.

Yes, the data is probably out there somewhere.

Yes, but it must be a database query. This shouldn’t be decided by inference or by applying a model with an IQ of 300.

8. Rethinking HubSpot's Primitives for AI

Of course, I completely agree. A lot of companies are going through this right now, trying to figure out their own ideas about pricing and consumption, which could be a topic for a separate podcast. Maybe the last question on this topic, because I think it’s very interesting how you talk about it.

With the architectural decisions you made when you built HubSpot, I’m curious how you look at them now, with the benefit of experience. I’m very interested in hearing your thoughts, because you originally designed everything so that a person could navigate the information, and you stored it in a way that was understandable to a person. Now we have these incredibly powerful models that can look at things completely differently. So I don’t know how you think about what you’d like to start doing differently, or what you’re doing to address that now as the head of HubSpot.

Yes, there are several points. First of all, I’m a classic computer scientist and software developer. I’ve been one my whole life. One of the things I believe in when it comes to large-scale software applications is the concept of frameworks and working with a set of primitives that you develop over time, because that’s what gives you an advantage and the possibility of a compounding effect.

One of the things I’ve learned at every company I’ve worked for is that if you can get these basic primitives right, then eventually you can recombine and reassemble them in new ways as the market and the product evolve. We got this part right at HubSpot from the start. We created a kind of basic operating system or framework on which we built marketing automation tools and the like.

So when we moved to products two, three, four, and so on, most of the value we needed to bring to market already existed in the form of basic elements. For example, we started with marketing automation: it’s a deterministic workflow, like sending an email, waiting 3 days, and waiting for a click. Then, when we moved into sales and CRM, it turned out that our basic marketing automation element was actually universal. We could reuse all of this technology, and on day 1 our sales automation app or product had more capabilities than anyone who’d been developing it for years. This part, in my opinion, remains relevant today.

So now we ask the question: What should the framework for the next 20 years of HubSpot be in an AI-dominated world? You mentioned one of the problems: data is no longer just structured information in a relational database, which was appropriate for certain deterministic queries. Now it’s both structured and unstructured data. It’s not that this data didn’t exist in companies before, even in our clients’ organizations. It’s just that 10 years ago you couldn’t do anything meaningful with it at a reasonable price. But now you can.

So we look at the basics and say, “Okay, we still need a structured data repository.” I’d be happy to discuss this with anyone who’s willing. There’s a need for a persistent mechanism for storing known facts and data, a kind of system of record, alongside what we consider a system of context—that is, unstructured data. And this is exactly what we’re building now.

I think this will evolve into the concept of the “second brain,” a term that I and others use. Can we collect everything that happens into a structure that probably looks like a graph? Can we store this in a kind of context graph, another term that’s already in use? That’s the idea. Can we take this, combine it with a structured data store, draw conclusions, run agents, get results, and do the work?

9. Replaying 20 Years of Code With AI

One interesting question about AI came to my mind during the conversation. What’s really great about agentic programming right now—and this is especially true for large companies—is…

This allows us to build software faster. We can test ideas faster, as we’ve already talked about, and create products faster. That’s great; we benefit from that. But you can go deeper and take on larger projects you might not have thought about before.

Now this is possible: we have a version control system, GitHub, which almost everyone uses. Imagine you say: I have something like Fable 5.1 or 6 Astra, or 6.1, which is out now. But imagine, you say: if I were younger, I would have made a different choice. If I could go back in time, this is how I would have made these primitives or implemented them as a reusable service, because we didn’t know it then, but we do now.

Theoretically, you can now say, “I want to recreate this. I want to take the codebase we used, this data structure, this language, this framework, this architecture. What if we spent $200,000 on tokens to remake this, literally recreating a variant from a multitude of possible universes? Let’s jump to the branch of the universe where we made a different choice and see what happens. Where will we go?”

One advantage of a company like HubSpot is that we have what used to be called unit tests and are now called evaluations, or evals. Can we measure the viability of what we just did? Is this still working? Does it still deliver the desired results? Does it still perform all the necessary functions?

We do this test every time new models come out. We can run our set of evals for the new models and say, “Okay, here’s the tradeoff. We lose about 3% of quality, but gain the same amount in latency or economic performance.” We can do the same thing by testing theories like, “Okay, if we rewrote this all in Rust, what would it look like? What’s the actual result?” You can test this on small amounts of work. So, in short, it’s exciting.

Yeah. I think this is an extremely interesting topic. I wrote another essay around the same time I wrote “The Yellow Brick Road,” maybe a little earlier. I even cited Workday as an example of having to start from scratch to become AI-native: here are all your advantages, but here are the disadvantages.

10. HubSpot Next and the Birth of uSpot

The advantage of HubSpot, in my opinion, is that you have a great founder who’s still actively involved in the life of the company, running the business and implementing these changes. I like the example of imagining alternative universes you could create. Are there other processes you’re trying to create at HubSpot, or teams you’re trying to form? Maybe you’re struggling with bureaucracy. I’m curious what you, as a founder, are doing now.

We have a group at HubSpot called HubSpot Next, and the idea is that it’s essentially an innovation lab—a startup within a startup. At any given time, we may have several of these projects running. The motivation is that we really treat this like a startup, and we’ve done this before.

When we were creating CRM, we said, “Okay, we’re going to create a CRM system.” And they answered, “Wait, what is this? We deal with marketing software.” We separated that team: they weren’t subject to the general rules, they didn’t have to use our stack or our brand; they just had to solve the problem. Now we continue to develop this idea.

I’ve supported and led several of these projects over the years. The one I’m working on right now, literally every night, is called Uspot: a CRM for one person, for solo entrepreneurs. This is our market, and it’s a return to narrowing the focus. HubSpot Next projects are either pursuing a new product category or idea, entering a new market, or testing a new technology.

Going back to whether you can learn from mistakes cheaply, it’s much easier for me to implement my idea of a “second brain.” Maybe a graph model works better. Maybe it’s worth connecting all social networks so that everything flows into one place. So I’m like, “Okay, let’s test it in Uspot,” because it’s a much easier and faster way to check.

The product is now exactly 30 days old, and it’s been an exciting learning experience. So, the answer to the original question—could something like this work?—is yes. It works. I think what other entrepreneurs should take away from this is that the concept of a “second brain” makes sense.

11. How to Actually Pick a Startup Idea

Going back to the conversation about “The Yellow Brick Road” and how to avoid getting hit, you shouldn’t say, “Oh, ChatGPT or Claude don’t work well with memory,” or anything like that. Yes, it will get better. Don’t try to solve that problem or close off a general use case. Instead, say, “Hey, this category of customers has problems that can be solved with the second brain concept.” What will the data entry points be? How do you put it all together? What real agents will you create, or what work will you be able to do? I think this basic principle is very powerful and applicable.

12. A Startup Within a Startup

Yeah. I think Uspot is very interesting and certainly very valuable for this type of user. This is exactly what they need. How do you use this knowledge and turn it into a product in the parent company, if that’s possible? What processes need to be set up to make this happen?

After years of trial and error, we’ve come to hold on to the idea of a startup within a startup. We do a few things. First, if the project scales—and this has happened to us many times—then we decide what to do with it.

A few years ago, for example, I launched a product called Agent.ai. It was a marketplace for AI agents, a long time ago. It was one of those stories where I accidentally became Sam Altman’s competitor. It was market-oriented, but too horizontal. It had 2 components: an AI agent marketplace and an agent builder that allowed ordinary people to create a mix of deterministic and non-deterministic agents.

We realized that half of the idea was right, but it needed the support of the entire HubSpot team. So we took it over and turned the agent builder into a product within HubSpot. That’s where it exists now.

In the case of HubSpot, one option is to create a CRM for one-person companies. We did some things right and some things wrong; it doesn’t matter. Maybe we’ll integrate it into HubSpot and it’ll become a HubSpot product. Maybe the product itself should be separate. Or maybe the market won’t work, but the idea of a “second brain” might be viable.

These are the moments when we decide what’s right for small and medium-sized businesses, for teamwork, or for something else. We just let the core team choose what they want. It’s a bit like acqui-hiring, literally. It’s like, “Here’s what we can buy at a low price,” meaning it’s already included, right?

Right.

It’s free.

That’s right. It’s free. We try to look at it objectively. What’s great about this approach—and what makes you ask, “Why not just do it within the core business?”—is that these degrees of freedom are incredibly important for experimenting at the speed you want.

One of the things I’ve been saying for a long time as a startup guy is that the problem market leaders have—it’s hard to use the term “leader” when I think of HubSpot; I still think of it as a startup—

Don’t worry.

—isn’t that they lack something. It’s not that startups have something big companies don’t. The problem is a kind of baggage, and the weight the organization carries. If they can find a way to get rid of it, they get all the benefits of a big company: distribution, customer relationships, evaluations, and 20 years of codebase they can use as data to build the next product. All of that is an asset.

So if you can drop the commitments, HubSpot Next is a way to say, “Okay, we’ll just pretend we don’t have these commitments,” like a startup. As a startup, we can do our thing, which we did. But when it comes time to leverage an asset, you can think, “Oh, it would be helpful to have the HubSpot brand here, or distribution advice.” We can do that.

13. The Rise of the One-Person Company

It’s easy right now, especially with the debate in society, to think negatively about AI. But all the data I’ve seen is so exciting. The growth is simply incredible, and I think HubSpot is a great example of that potential growth. When in human history would it have been a good idea to create such a super-individualized CRM for one person? Historically, that was almost impossible. But now we can do it.

What will this allow companies with one employee to do? We’re going to see growth like I don’t think we’ve ever seen before, and it’s so cool to dream about. I’d love to hear a little about what you’re seeing: maybe some early use cases at HubSpot, or what you’re seeing in terms of growth. It’s truly an amazing time to be alive.

The interesting thing is that I knew this would happen, but I wasn’t sure to what extent. The only difference I made with HubSpot Next, compared to previous HubSpot Next projects, is that I started monetization very early. It wasn’t to cover the cost or anything like that. One of the lessons I wanted to learn was about pricing and monetization, and what value people place on it. That was part of it.

What really surprised me was people’s willingness to use a credit card on a pay-as-you-go model. It’s unlike anything I’ve seen. I’ve launched many projects over the years at HubSpot, and in this respect, this has been the most successful.

So, there’s a signal there. Part of it is driven by the numbers themselves. If you look at U.S. Census data, in 2025 we had the highest number of new companies the United States has ever seen. It was around 5.6 million, getting closer to 6 million, which is great. This year, as of August, we already had over 4 million, so this will be another record year.

This isn’t just a one-time spike after which everything returns to normal. That’s the number one factor that’s extremely interesting. Factor number two, which is even more interesting, is that the percentage of those new companies expected to remain one-employee companies is at an all-time high and continues to rise. The U.S. Census Bureau actually tracks this data.

They divide each new company into 2 categories. The first is a company projected to hire at least 1 employee within the next 2 years. They divide them into these 2 groups, and the share of companies unlikely to hire an employee in the next 2 years is now over 70%. That’s exciting. A completely new market for entrepreneurs has emerged, like SMB was back in the day, when everyone resisted HubSpot. “You can’t make money from small and medium-sized businesses. It’s so hard. No other software company besides Intuit has ever built a multibillion-dollar business in the SMB space.” But you can. It’s pretty massive.

There will be painful moments in markets like this, but either way, I’m excited about it. I’m excited about the economy. I’m excited about all the entrepreneurs who are going to start these million-dollar companies or something like that, and about the entrepreneurs creating products for this market. It’s fascinating from all sides. So, yeah.

14. AI Is Still Just Software

No, I completely agree with that. We often talk about certain aspects of this, and I think one of the things we’ve mentioned in the past is vertical software. How did you manage to make it work? You made vertical software efficient with low annual or monthly costs, but made up for that with payment revenue, interchange fees. I don’t know if people fully realize how much the opportunity is growing, because there’s so much more you can do and so many new ways to create value for your small customers. That’s my argument against the idea of a “SaaS apocalypse”: AI, by the way, for the record—I know you know this—is also software.

Yes, yes, yes. It’s not that it’s not software, right? This isn’t some alien life form. This is software. And what actually happened is, if you look at the class of problems software was designed to solve… Okay, I’ll take a minute, because this is actually a good story. I only have a few good stories.

When we first started, in the first year of HubSpot, we raised venture capital and said, “Hey, we’re building this marketing software company. Here’s what we’re going to do.” Almost all the venture capitalists we talked to said, “Well, there hasn’t been a billion-dollar marketing software company or anything like that yet.” So we had 2 strikes: the SMB strike and the fact that we were in marketing software.

Our argument at the time was that if you look at the income statement of any business, next to every expense item you’ll see the name of a software company worth over $100 billion, excluding marketing. It’s not that people don’t spend money; it’s that marketing has historically been something of a creative craft. You have a certain budget. You spend a little bit creating the creative, and then you pour it all into advertising, billboards, and things like that.

Our thesis was that, through the internet, marketing would become a problem that software had to solve. This is why a software company didn’t make sense 5 years before HubSpot, but now someone was bound to create a huge marketing software business. Now they make the same argument: There used to be a class of problems that software couldn’t effectively solve. Now a huge set of new problems has emerged. The problems existed before, but new opportunities have emerged that we can realize thanks to AI, opportunities that didn’t exist before.

The set of opportunities is much larger now. Some software companies will disappear, no doubt, just like some companies failed to transition to SaaS, cloud technology, or something similar. It’s like the cycle of life. But overall, the opportunities are greater now than ever. If you’re into software, have talent, and are entrepreneurial, now is the best time to start a software company. Act now.

15. Lightning Round

Yes, I completely agree with you. This was incredibly interesting. I’m very grateful for your time. Finally, I’ll do a quick lightning round with you, and then everyone can go about their business. I’m not very quick to react, so…

Yeah, yeah.

Well, there are only a few questions, and they’ll be easy. I think I’ll start with something very simple: Dunkin’ or Starbucks?

Dunkin’. Although, by the way, I don’t drink coffee.

Oh, wow. Of course. What do you drink instead? Only water? Are you a cyborg?

I gave up caffeine years ago.

So, in short… This is even more impressive, Dharmesh. Okay. Instacart or Manus?

Excuse me, what?

Instacart or Manus? You know, the new agent solutions for shopping?

I have to say, Instinct.

Of course. Interesting. If you had to become an animal, what animal would you be?

I’d still choose a mammal. Maybe an elephant.

Elephant. I like it. Okay, great. What’s the most Bostonian thing about HubSpot?

Almost everything.

That’s a good, valid answer.

Yes. We’re very pragmatic; we work for the benefit of the client. I’d say we’re like a hybrid: we took the best of the West Coast—their creative thinking and ambition—but added the pragmatism of the East Coast. We just work hard.

I like it. And finally, what was the best day you had at HubSpot?

A week ago. It feels like everything is falling into place, both in my head and in my business. I can almost see everything unfolding. It’s hard to explain. I can’t really articulate it, but, yeah…

Oh, it’s so cool. So after all these years, the best day was a week ago?

Yes. Although I’ll say it more broadly: the best time I’ve had, even as a startup owner in HubSpot’s entire 20-year history, is right now. People say, “Oh, that’s impossible!” but I don’t have any direct reports at HubSpot. I’ve never had any direct reports. The reason is that I have all the advantages of scale. You’re sitting at the adults’ table. You can make long-term bets and do certain things, but without the downsides of leadership.

I can work on HubSpot, on future projects like YouSpot, or something else. So I have the best of both worlds. And now I’m having so much fun.

Yes. This is amazing. Well, Dharmesh, you’re an outstanding entrepreneur, an exceptionally kind person, and a thinker of your generation. I really appreciate your time today and your involvement in these topics. I look forward to the next meeting. Thank you.

See you later, friend.