Siddhartha Ahluwalia 0:43

Hi folks, Siddhartha Ahluwalia this side, your host at Neon Show and managing partner at Neon Fund, a fund that invests in the best of enterprise AI companies building between the US-India corridor. I have today with me Ashu Garg, General Partner at Foundation Capital. Ashu, I’m privileged to host you twice on the Neon Show.

Once five years ago, that was online almost.

Ashu Garg 1:07

Yes.

Siddhartha Ahluwalia 1:08

Yes. And today, you know, we are sitting in Bay Area in your home.

Ashu Garg 1:11

Thank you so much for having me. I’m excited and thank you for coming to my house.

Siddhartha Ahluwalia 1:14

No, it’s a real pleasure. You know, I don’t want to dive straight into, you know, what we want to discuss. You have been putting out a lot of amazing LinkedIn content and also on the Foundation Capital, you know, website.

Right. So I want my conversation today to be structured around that because, you know, what I see is your bullishness on AI. Yeah.

Where is it going? Right. It’s amazing.

Right. I learn a lot from it. And I want my audience also to learn from it.

So my first, you know, where I want to start is you have mentioned that reasoning and reinforcement learning, you know, combined is a new recipe for AI apps. Right. And the new kind of products and services that are becoming possible with shift from RAG to reasoning plus reinforcement learning, you know, is a new paradigm.

Right. Tell us more about that, you know, and tell us what these terms mean.

Ashu Garg 2:09

So if you look at the history of recent AI, you know, we can go back. AI has been around in some shape or form for many decades. But if you look at recent AI, you know, two and a half years ago, a little over two and a half years ago, when ChatGPT launched.

Siddhartha Ahluwalia 2:23

Yeah.

Ashu Garg 2:23

The idea of a conversation interface was innovative.

Siddhartha Ahluwalia 2:27

Yeah.

Ashu Garg 2:28

It was a game changing paradigm. And the first wave of software companies that you saw were essentially variations of that conversation interface. You know, you can ask a query and it gives you a response.

Siddhartha Ahluwalia 2:42

Yeah.

Ashu Garg 2:43

We saw that in customer service. We saw that in legal. We saw that a bunch of verticals.

But if you fast forward today, three things are going off. First and foremost, the underlying model providers are your biggest competitors. I think code generation is the most obvious example.

We’re seeing that in code generation. I think we will see that increasingly in other categories. Code generation that’s happened.

Content creation. If you go back in time and look at companies like Writer, Jasper, and a hundred others in the same category, they’re essentially sort of a wrapper on top of the underlying models. Now the models do everything.

You can get, you know, models and memory. The models can adjust tone. The models can write blog posts for you.

So I think we’re seeing the first coming back. The three things going in the market is we’re seeing the water level rise because the underlying models themselves are competing with the downstream application companies. The second thing is we’ve seen agents be the rage over the last 12 months.

Like everything is an agent. Agents, swarms of agents, systems of agents. And agents are simply a way for us to connect individual instances of a query and stitch them together to replace a complex workflow.

So something as simple as, you know, take a very simple agent. Ridiculously simple customer service. You ask for a refund.

That refund request you might make on a, let’s say an Amazon website, in turn is going to trigger five, six, seven to 10 actions. They want to check, did you buy the product? When did you buy the product?

Did you provide feedback? Was it delivered? They have some workflow that some human being was going through.

Each of those steps has to be stitched together. But those workflows are dynamic in the case of a human. And so those agents are then applying reasoning and the output of every step then influences what the next step is.

That’s why reasoning is like the notion. Agents is the visible part. Under the hood, you need reasoning for agents.

More broadly, we’re seeing, you know, ChatGPT itself offered deep research and a variety of more complex capabilities. Those capabilities require sequential reasoning. And so that’s where reasoning comes in.

The third, what we are seeing with reinforcement learning is the importance of post-training. These models themselves, you need to ground them in certain data. You need to sort of incorporate the fact that if you’re going to go through very complex workflows, they need more specialized knowledge and information around that workflow.

And that really comes from the post-training. And that really, the underlying technical, you know, technology is reinforcement learning. So those are the three big things going on in the space from a technology standpoint.

Siddhartha Ahluwalia 5:40

And what has, you know, you have mentioned in your post that in this third wave that we are talking about, reinforcement learning has emerged as the core ingredient for building models and applications where AI acts as an autonomous worker. We are not talking about automation. We are talking about autonomous, which is like a huge leap, you know, from just building workflows now to AI imagining workflows.

And I’m hoping that, you know, now what we are talking about is new workflows that human had never imagined AI will start doing also that. Right. And this is a shift from the co-pilot, which everybody was talking about. 12 months ago. Right.

So what is going on here?

Ashu Garg 6:20

Yeah. So I think, you know, if you step away from the technology, you know, reasoning and reinforcement learning are the technical innovations and reinforcement learning is interesting because reinforcement learning predicts transformer models. So it’s the coming back of reinforcement learning that has been fun for people who’ve been in the space.

But if you step away from the technology, ultimately what is going on is we’re seeing the shift from co-pilots, which assist a human to an autonomous agent that for a certain class of tasks can replace a human. This is not AGI. I mean, there’s all these fancy words, AGI, ASI, you know, whatever those are in the future, time will tell.

But what these autonomous agents are able to do is they’re able to do what Waymo does for driving for some other tasks. So think of Waymo. I think Waymo is a great example of an autonomous agent.

And under the hood, there are a series of agents and Waymo actually does a lot of reinforcement learning under the hood.

Siddhartha Ahluwalia 7:22

And, you know, the other thing that I want to touch upon is we touched upon the third wave. Right. And all these waves happen very quickly.

I’ve never seen or nobody has ever seen the speed of change, the speed of change. Right. And then I want to talk about agentic web where new product opportunities in a world where AI agents are, you know, as you mentioned, automating complex business workflows, but they are also like platforms or operating systems now getting developed, you know, for these agents, you know, and there are new security, privacy, governance written for the entire agentic web.

So now we were talking about earlier web, which was basically dominated by Chrome. Right. And now we are stepping into, you know, with Perplexity launching their own browser, agentic browser and agentic web.

What does it mean to you?

Ashu Garg 8:11

Look, ultimately, there are multiple things going on here. First and foremost, the world we’re headed to is each one of us will have one or multiple personal agents. So today you’re asking me the questions I could imagine a world five years ago, five years from now, where your agent is actually conducting the interview.

And who knows, maybe my agent is answering the questions. But jokes aside, you know, agents are increasingly representing individual humans in certain contexts. Today, those contexts are narrow.

It’s not general purpose. Over time, those contexts will get broader. In a world where you have agents acting on behalf of humans, the underlying infrastructure of the web has to change.

Take a browser and take websites. Websites have been optimized for the last 25 to 40 years. Some websites go back 40 years, but most websites are 25, 30 years old.

These websites have been optimized to make it easy for human beings to find things. But the way a human being finds things is very different from the way AI finds things. And so if you are optimizing, if you’re starting from scratch, you took any website, take amazon.com and say, hey, I’m going to completely reimagine this website to make it easy for an agent to find something that would look very different. So I think we’re going to see changes in the US. Now, websites are manifested through a browser. And so everyone who’s trying perplexity and others are saying, OK, how can I come up with a new browser that then gives me an unfair advantage in this flow?

But the same is true for payment systems. If you look at current payment systems, if you look at Stripe and all of Stripe’s fraud analytics and fraud systems, they’re designed to test for a human being. And is there a human fraud happening?

But imagine if now there’s no human being, you know, Ashu Garg is not doing a transaction. Ashu has 25 different agents. The agents don’t have their own identity to it.

They don’t have their own credit card. They’re using my credit card, my identity, but they’re behaving in ways that are similar to me, but not the same as me. So that creates its own challenge, that notion of.

So now payment systems, fraud analytics will have to change. Think of how you pay for transactions. You might say, OK, I’m going to tell my agent that it’s OK to buy a plane ticket whenever the price drops 5%.

I can set that, but 5% from what? Am I willing to let the agent make a $1,000 decision, a $5,000 decision, a $50,000 decision? Agents, when they start making decisions on behalf of a human, ultimately someone’s wallet is involved.

So you now have to put some guardrails around the agent behavior. Those guardrails require some system. Then now this agent is in turn, let’s say they are starting to do microtransactions.

So let’s say you have an agent that is using small amounts of compute. Are you going to pay for those microtransactions with the credit card? Or should you be paying for those microtransactions with the cryptocurrency?

The challenge is credit cards have a minimum. There’s a typically 25 to 30 cent minimum transaction fee. If you buy something for five bucks, 25 cents is not that much.

But if you buy something for a dollar, the transaction fee becomes. So in a world of micropayments, which is more common with agents, like you and I don’t want to spend time doing, I took a $1. Let’s say the transaction is only $0.01. Now you have a 25x markup on your $0.01 transaction. So in a world where the cost of the activity comes to close to zero, because an agent is doing the activity, the infrastructure to enable the activity will all have to change. So I think that’s what’s underway today. And that is going to cause us to reinvent the internet as we know it.

I mean, there are so many aspects of the internet that are optimized for human activity. And if you started to optimize them for an agent, you would think about the world differently.

Siddhartha Ahluwalia 12:27

Are you using any agents which are autonomous right now for yourself?

Ashu Garg 12:32

You know, I experiment with a lot of things. I have agents that, you know, look at my email, I have agents that read my email. I have agents that help me, you know, prepare for meetings.

So I’m trying a lot of things. Autonomy is a very, you know, broad term. Am I comfortable today letting an agent respond to an email on my behalf?

The answer is no. But that’s also a function of the fact that, you know, my work emails have value. You know, if I told a founder by, you know, the agent by mistake said, yeah, we’re ready to sign a contract.

That would be problematic. But you know, the value you place on emails is different for I think people will start using agents to respond to them sooner than I would.

Siddhartha Ahluwalia 13:16

And have you been ever so excited about what’s happening in the ecosystem as of today?

Ashu Garg 13:21

You know, the last time I was this excited was in the late 90s. Like I have very vivid memories of 97, 98. I saw the world, you know, in the fall of 96, I was working with McKinsey in Hong Kong.

And I remember dialing up, dialing in into private servers to be able to download data. You know, data pipelines existed. I would need to do work.

I was working with a client in China and I would go, I would log into some system in the US and, you know, we’d be in a dial-up connection for four hours. I’d get a $2,000 dial-up bill and I would download five analyst reports. By 97, I could do that on the internet.

And that transformation was magical. In 96, if I was traveling, I remember going from Bombay to Beijing in December of 96. There was no weather.com that I had access to. I still, you know, I left Bombay. And Bombay, even in December, is pretty warm. I showed up in Beijing in December and it was snowing and I was in my shorts and my chappals.

So, you know, not great weather for shorts and chappals. Weather.com changed that. By the late 90s, I was starting to use, you know, Amazon.

The idea that I could literally, you know, think of something, order it on the internet, it would show up in my house three days later was magical. Today, we take all of those things for granted. So, I think the transformation today is as dramatic as the transformation of the internet.

Nothing in the last 25 years has come close.

Siddhartha Ahluwalia 14:59

And what about, you know, accountability now? Like, if Siddhartha’s agent commits something which Siddharth never gave permission to, who is accountable then?

Ashu Garg 15:14

Ultimately, Siddhartha’s agent. Siddhartha is accountable. But you’re raising a great point.

That’s part of why, you know, whole new infrastructure on how these agents are managed, how you set guardrails becomes important. Maybe you said it was okay for the agent to spend $5, but you may not want the agent to spend $50,000. If you give the agent a credit card, you have a high limit.

And, you know, if it makes a mistake, who’s responsible? If you give it a debit card, well, then it depends on how much your bank account has. So, I think all of these things will have to get worked through.

We’re very early in this journey. And I think of this as roughly 1997 in the internet. Things are moving very quickly.

I anticipate that over the next five years, we will see the same level of progress that we saw over a decade in the internet. Because we are right, each wave rides on it, on the prior wave. And I think it’s especially exciting how quickly these benefits will become global.

If you think about the internet, the internet was predominantly a US phenomenon till the 2010s. I mean, internet, if you take India as an example, internet existed. I mean, I set up an ISP in India in 99.

But as a matter of actual mass consumption, the internet is a 2010-2012 phenomenon in India, roughly 15 years after the US. Today, you’re seeing that ChatGPT consumption in India is actually simultaneously with the US. I think it is as much consumption, perhaps more.

Siddhartha Ahluwalia 16:58

More. Sam Altman and Arvind from Public City are more excited, because the data…

Ashu Garg 17:04

No one’s paying for the tokens, that’s the problem. But yes, they’re collecting a lot of data. But certainly, we’re seeing very similar usage patterns without the time shifting that we saw on the internet.

Siddhartha Ahluwalia 17:17

And who do you think, you know, in this wave has the advantage? Is a second time founder, is a first time founder, what mindset is required to take, to ride this wave?

Ashu Garg 17:29

You know, I think in the AI wave, in some ways similar to the internet, because the transformation is as disruptive. I think the founders that approach it with the naivety of a first time founder, and the curiosity of a first time founder, and the intensity of effort of a first time founder, I think are most likely to succeed. Now, I don’t think that doesn’t mean second time founders can’t be successful.

But what I tell, you know, I work with a lot of IITians and IIT Builder Foundation. And I had this conversation with them yesterday, I said, just remember that your competition is some guy that works 16 to 18 hours a day, seven days a week. If you can’t compete at that intensity level, you have a problem.

Your competition has no baggage of the past. They don’t know how marketing was done in the past. And so they’re going to do it afresh.

Your knowledge of how marketing was done 10 years ago, has some strengths, but also has a lot of baggage. So how do you strip away the baggage of the past is a challenge for repeat entrepreneurs. So I don’t think you can make generalizations, but certainly, I think the value of experience has diminished, and the value of intensity and curiosity has exploded.

Siddhartha Ahluwalia 18:52

And what about geographical location?

Ashu Garg 18:56

You know, on one hand, I think the internet, and now AI is leveling the playing field. As you know, I talked to entrepreneurs in India, and very often they’re using the same buzzwords, the same language, and have the same context on technology, for example, as an entrepreneur sitting in Silicon Valley. At the same time, I do think that there is an incredible amount of value to being in the right locations and communities.

If anything, even in a world of AI, human connection doesn’t go away. Human connection matters. And I think what Silicon Valley offers is, for one thing, if you’re trying to build a tech, there are many other things that Silicon Valley sucks at.

But if you’re trying to build a technology company, and your ambition is to build the next Google, the next Facebook, the next Microsoft, I think the human support system that is available here is the best on the planet. Like, nothing else comes close. Maybe in a decade, maybe in two.

I think Bangalore is an interesting ecosystem, I think, in a decade or two. I think Bangalore, you know, has a shot at being there. And that’s not to say great companies can’t be built in Bangalore.

Great companies will be built in Bangalore, will be built in Seattle, will be built in Austin, will be built in New York. But your odds are higher if you build it in Silicon Valley.

Siddhartha Ahluwalia 20:28

And right now, the question today is, whose AI driven ecosystem will become users default? Right? What are your views as this economy is evolving?

Ashu Garg 20:40

You know, if I were to answer that question, I would have, I will retire. I think, I don’t think there will be a default AI system. I do think that for different use cases, you will see different winners.

Definitely for the horizontal consumer or prosumer use case, think of ChatGPT as that, there is definitely, you know, a battle royal between OpenAI, Google and Entropic today. And there is no doubt in my mind that Microsoft and Facebook meta that is will be in that battle. So I think there are five players.

Siddhartha Ahluwalia 21:20

Would you count perplexity in that battle right now?

Ashu Garg 21:23

At that abstraction, I would not count perplexity, to be honest. I think in the battle for truly owning the consumer interface, ultimately, those are the five players that are investing in the underlying model technology as well.

Siddhartha Ahluwalia 21:35

Yeah.

Ashu Garg 21:35

In varying degrees, Microsoft is piggybacking on ChatGPT. I think perplexity has a shot at that. So I wouldn’t rule it out.

I think Arvind has done a remarkable job. I think I am blown away by what he has accomplished. And I think he has, you know, he has put a stake in the ground.

I think he has a shot. But to be fair, I don’t think it’s in the same zip code. I think Apple will want to play in that game, maybe with perplexity.

But really, when you think about the players today, I think the three contenders today are really Anthropic, Google and OpenAI. And I think Microsoft and Meta are sort of fast followers. And Apple is the dark horse.

Siddhartha Ahluwalia 22:22

The other question I have is, you know, you have mentioned this, that it’s no longer about the model quality. It’s about who offers the best end to end experience for consumers, for enterprises and for developers.

Ashu Garg 22:36

So, again, to be fair, I do think model quality matters. I think the rate of change in pure model performance has slowed down. And so today, the diff between one model and another is less than it was two, three years ago.

And over time, I think we will see that rate of change slow down even more. I also think that open source has demonstrated that while it can’t compete with the SOTA models, it’s only six to 12 months away. I think where value is getting captured increasingly, and you see that in the way OpenAI is, OpenAI has pivoted the company in the last 12 months.

Siddhartha Ahluwalia 23:19

What do you mean pivoted the company?

Ashu Garg 23:20

I mean, OpenAI was primarily a model provider. ChatGPT, when it was launched two and a half years ago, was meant to be, you know, a test bed for the model as a way for them to collect training data.

I mean, today, the most important thing about OpenAI is ChatGPT. They’ve gone really deep in coding, and they want to really make a better coding with their agent systems. They’re really going after applications more broadly.

So OpenAI has pivoted from being a developer focused company to being a consumer and prosumer focused applications company. I mean, you can call it a pivot, you can call it, you know, an evolution, but it’s a very dramatic change in the span of the last 12 to 18 months. And I think the reason people are going in that direction is ultimately, consumers or enterprises want to problem solve.

They’re looking for solutions to problems, and they’re willing to pay because of the nature of these solutions. They’re willing to pick real money. I mean, we’re seeing, I have a startup that’s, you know, barely six months off the ground.

They’re about to sign what might be a $30 million deal.

Siddhartha Ahluwalia 24:28

Wow.

Ashu Garg 24:29

First deal. First deal they’re going to sign.

Siddhartha Ahluwalia 24:31

How did they pass the procurement?

Ashu Garg 24:34

You know, because they were able to demonstrate to the customer that the alternative to using them was to spend $80 to $100 million with Accenture.

So $30 million seems like a lot of money, but relative to spending $80 to $100 million, that’s not that much money. And so we’re starting to see customers pay very large amounts of money for outcomes. But then in order to be able to deliver an outcome, you have to have more control into it.

Siddhartha Ahluwalia 25:07

So you are saying at one point of area, there are these five large companies, you know, that are dominating the agentic era, as we speak, as we are entering into an agentic era. I will not call it an internet anymore. Right.

So where’s the opportunity for startups? You have mentioned there are some, you know, startups where verticals have advantages, right?

Ashu Garg 25:27

Look, I think there’s infinite opportunity for startup. If you think about things that human beings do, the vast majority of things require specialized knowledge, expertise, access to specialized systems, and the ability to deeply integrate into complex high value workflows. And that’s, I think, where startups will win.

You know, we have a company in our portfolio, Tenor, that does the front office patient intake processes for specialty medical service providers, you know, think dialysis clinics, sleep apnea, a variety of, you know, psychiatry centers, so on and so forth, you get a patient, you know, that’s referred to you, you have to go through a patient referral process, you then have to contact the patient’s schedule in parallel, you have to reach out to get their insurance information, reach out to the insurance company, get a prior auth.

Once the actual first meeting happens, you have to then submit a claim, you might provide the person a prescription that prescription may require for the justification that sets a prescription for a CPAP machine insurance is just not going to say yes, they’re going to ask for a lot of documentation, they may ask for some follow up studies, that entire process is automated by a company like Tenor. I have a company called Reggie, which is completely automating the front end of prospecting, you know, you haven’t, you have SDRs, SDRs are pulling data from a zoom info, or a clay, they are then figuring out based on that data and existing data, who should I follow up with, what messaging should I send, when those people, you know, respond to those emails, and you have to follow up all the way to the point where it becomes a meeting that has to be scheduled for a sales rep, that process can have hundreds of steps, and in some ways, an iterative cycle over months, that process gets automated by Reggie. Will people pay a lot of money for it?

Yeah, absolutely. But Reggie has to demonstrate that you’re, you know, you’re delivering business value, that if you’re saving them a million dollars of cost, you can get a couple of hundred thousand dollars. When you demonstrate you’re saving a million dollars, or you’re creating a million dollars of new pipeline, or whatever that metric might be.

Siddhartha Ahluwalia 27:47

So you’re saying it’s in the specialized and complex workflows, that startups have to tap in. But then the question there is, you know, all these companies, all the startups are built on top of these model companies, right? And these models are also learning what are the workflows that these startups are doing.

And we have seen in the past, right, for example, with what happened with windsurf, right? And there are more examples that these model companies are now launching their own vertical application.

Ashu Garg 28:15

Absolutely. So look, that’s, you know, it’s a full circle the way we started off 30 minutes ago, which is every startup in the AI application space, every, you know, services, software startup has to first and foremost internalize that their vendor is also their competitor. The model provider is both vendor and competitor.

And that’s a hard place to be. And they are, they are actually educating their vendor on how to compete with them. You know, that’s, that’s the inherent challenge.

And so if you recognize that challenge up front, then you have to start thinking about what is the data that you’re not going to send to them, send to the model provider? What is the data for which you’re going to use an open source model versus a closed source model? What is the role of reinforcement learning in sort of post training of a model for a customer specific situation?

So for example, I have a company called Player Zero. And while they do use GPT-5 and other proprietary models, the core code base of the company for their customers, what Player Zero does is it helps you debug problems in your code, and it actually simulates a pull request. So you can proactively identify likely problems.

For both of those, they’re using reinforcement learning extensively. So they build a graph of your code, that graph of your code, they don’t share, because no one wants their code to go to ChatGPT. Like that’s proprietary information for a customer.

So they build a proprietary graph, customer by customer. And they may use ChatGPT or GPT-5 or GPT-4 to query the graph. But they don’t actually expose the graph itself to OpenAI. Secondly, you know, they’re doing a whole bunch of post-training of the models to get, you know, better performance for specific use cases. So, you know, one of the things, if I’ll connect this back to sort of where we started off, we’re seeing an explosion of entrepreneurs, and even though coding has gotten commoditized, technical skills have not.

So, in some ways, the founders today that I see being successful are more technical than the founders in the SaaS world, which also means that repeat entrepreneurs who grew up in a less technical world are struggling with the technical complexity. They’re struggling with the pace of change on technology.

Siddhartha Ahluwalia 30:43

Does this also mean that if you have distribution muscles in the SaaS era, they’re not valid in the agentic era?

Ashu Garg 30:51

You know, life is never black and white, so I wouldn’t say it’s not valid, but I definitely think that the power of distribution in SaaS does not apply one-to-one in the agentic era.

Siddhartha Ahluwalia 31:02

Can you share an example why?

Ashu Garg 31:05

Look at Salesforce, an incredible distribution machine. It’s the best B2B sales and distribution machine that has been created in the last 25-30 years. Look at how their agentic, you know, agent force, agent Einstein, agent whatever, agent Salesforce.

I don’t think their product is doing that well.

Siddhartha Ahluwalia 31:27

Who is eating their lunch right now?

Ashu Garg 31:30

You know, I think there’s dozens of companies. Look, today, change takes time. I think there are dozens of companies that are poised to eat Salesforce’s lunch, but those companies, despite the distribution disadvantage, are positioning themselves to do so.

Now, could Salesforce buy a few of them? Could Salesforce be in? Nothing, I won’t rule it out.

I mean, I don’t think we should ever assume that the value of distribution is zero. I just think in a world where disruption is rapid, distribution alone is not sufficient. Product innovation becomes far more important.

Siddhartha Ahluwalia 32:07

So, are you working as hard as you were working in 97 right now?

Ashu Garg 32:12

You know, I don’t think I could work as hard as I was working in 97, and that’s my challenge. I think about that a lot. I mean, 97, I was in my mid-20s, and, you know, I worked 16 hours a day, seven days a week.

Now, I have two kids, and, you know, my body is not as strong as it was in 97. And so, I have to work smarter, but I also have to work hard. I think this is a time that as an entrepreneur, you have to recognize that you’re competing on sheer intensity, on your curiosity, on your intellect, and not just on your experience.

Siddhartha Ahluwalia 32:52

But one thing my listeners and I have, right, that access to enterprises are not easy for young entrepreneurs, right? So, how much does relationship matter? Like, gray hair matter in entering?

Ashu Garg 33:07

See, look, again, things aren’t binary. Does experience matter? Absolutely.

Do relationships matter? Absolutely. If anything, relationships matter more today because people are putting more at risk.

If you’re going to a company and say, I’m going to outsource this process to you, that’s pretty scary for them too. So, if they’re going to do something that’s scary for them, they want to do it with someone they trust. So, human connection and skills that come with human connection matter.

And those skills compound over time. Relationships are compounding assets. And so, someone in their 40s or 50s has value over someone who has no relationships at all.

And, you know, I’m very lucky that in venture, if you compound correctly, you can continue to compound for a very long time. I think the same thing can be true for entrepreneurs. And absolutely experienced entrepreneurs have that benefit.

What they have to balance those relationships with is, can they dance at the pace at which the technology is changing? Ultimately, the pace of change is such that their product has to change at that pace. And many entrepreneurs are doing it, many very experienced.

I’m an investor in many second, third-time founders. I just came out of, you know, my last meeting was with Mohit Aran, who’s starting a new company, going after, you know, sales tech, Saifin. And this is his third company.

And Mohit had two Decacorns and, you know, this will be his third. And he’s, you know, he’s working at an intensity that any young person will be hard-pressed to match. But he’s choosing to do that.

And I’m just saying that all experienced founders will have to choose to execute at that intensity level.

Siddhartha Ahluwalia 34:58

You talk about, you know, for building durable AI startups today, focus on domains where you can control the most valuable data and feedback loops. What do you mean by that?

Ashu Garg 35:10

See, it comes back to your comment a few minutes ago that your competitor is the model provider, who’s your vendor. And if your competitor is not GPT-5, your competitor is GPT-7. If you cannot have a product that can compete with GPT-7, you’re dead on arrival.

Because you want to spend the next five years losing money to the point that you can have a profitable business five years from now, and you’d be competing with GPT-7. So you’ve got to really ask yourself, what do I, what am I going to do that GPT-7 will not be able to do?

Siddhartha Ahluwalia 35:47

But I don’t know the roadmap of.

Ashu Garg 35:49

So I think you have to take a guess. I think that’s what makes this hard. That’s what’s going to separate winners from losers, is you have to be able to, you have to really get into the weeds and say, what is going on at the Open AIs in the world?

What are the SOTA labs doing? What kind of post-training work are they doing? How are they incorporating RL in there?

What kind of data sets are they accessing? So what can I assume they will do so I can take advantage? But also, what do I need to do on top of that?

And that’s where getting into high value workflows, proprietary data sets, fine-tuning and post-training as sources of advantage will become important.

Siddhartha Ahluwalia 36:30

So right now, what we observe in the market, people like Mark Zuckerberg, they’re not placing advantages on companies. They’re placing advantages or placing their bets on people, acquiring the best of people for hundreds of millions of dollars. Why is that happening?

Is proprietorship or IP not of any value?

Ashu Garg 36:52

IP absolutely is of value and entrepreneurship is absolutely of value. I think there are two different trends going on simultaneously, which we often conflate with each other. First and foremost, the four or five large platform companies are all facing some level of antitrust.

Siddhartha Ahluwalia 37:09

What is that?

Ashu Garg 37:10

So antitrust in the US is when the Department of Justice, DOJ has to approve an acquisition or is investigating you for certain trade practices, which they might deem to be monopolistic or they might harm the interest of consumers. So over the last five years, there’s a lot of antitrust activity underway. Acquisitions are taking many, many months, sometimes as many as 18 months to get approved.

So in a world where you go from GPT-4 to GPT-5 in a little more than 18 months, if you acquire a company, it’s not going to get acquired for 18 months, you’re dead on arrival. How do you make acquisition decisions? So because of the antitrust issues, companies are choosing to lift out teams, including the IP.

But the bet they’re making is the IP is changing so rapidly, you take a licensed copy of the IP, you leave a copy behind, and that team, when you acquire the team along with the licensed copy, in 18 months, the existing copy will not be of any value. And so it’s just a new structure for what is effectively an acquisition. So that’s one thing that’s going on.

The second thing is going on is that the value captured by talent in large companies has exploded. And why is that? Because if you take Meta as the example, which has been seemingly throwing away hundreds of billions of dollars of people, look, Meta will spend well north of 50 billion, I think some number between 50 and 70 billion dollars on the underlying data set on GPUs, essentially.

So if you’re going to spend 50 billion dollars on GPUs, don’t you want to spend 5% of that or 10% of that making sure those GPUs are doing the right thing? So because of the scale of CapEx, the willingness to spend a portion of that on the hundred people that matter that will maximize the value of that CapEx, that numbers have exploded. And so today it’s rational, if you’re spending 50 billion dollars on GPU, that you’d be willing to spend 5 billion dollars on assembling the team.

And ultimately, the set of people that can do these things are not infinite. You know, whether some people will say there’s 100 people that matter in AI, technically, some people will argue it’s 1000. But that number is between 100 and 1000.

It’s not 10,000.

Siddhartha Ahluwalia 39:48

And why are these skills not transferable or repeatable?

Ashu Garg 39:53

The skills are transferable and repeatable, but they diffuse with time. These are not, because of the pace of change, this stuff is not being taught in a college course. You can’t go learn how to fine tune a model.

You can’t go learn how to train a model. Because frankly, open AI doesn’t know. Like if you were to talk to the bleeding edge teams at open AI and say, what are you going to do six months from now?

The answer is they don’t know. And we live, you know, as Turing, we work with all of the large model providers, all the soda labs on a variety of data and post training services to help them sort of improve the models. And their challenge and our challenge as their partner is they don’t know what they will need from us in six months.

So we’re both guessing where the future is headed. And so the only people that can actually have those skills in the near term are the people who are in the mix. If someone went to sleep for six months and woke up again, they would be like, oh my God, the world has changed around me.

The techniques are different. The tools are different. And unlike software, traditionally, these are not architectural innovations alone.

The core architectural innovation transfer model is invaded. There are other second order architectural decisions. So it’s not that there are no architectural decisions, but a lot of the innovations feel more like recipe changes.

You know, you make some dal and then, you know, how much salt you put and how much tarka you put, that’s what separates the dal, you know, dhampuk from the dal, you know, that someone else might make. But salt and tarka is not something that you can quantify necessarily. It’s a black art in some ways.

Siddhartha Ahluwalia 41:43

So you’re saying these 100 or 200 people that have been.

Ashu Garg 41:47

I would say 100 to 1000. I think the number is closer to 1000, but say 1000.

Siddhartha Ahluwalia 41:50

But let’s say only 1000 people who have been working at the cutting edge or bleeding edge of AI in the last five years, you know, know these black arts. It was very hard for them, startups, because, you know, they’ll never be able to attract either one of these folks, start a company. But what’s in these folks for start a company when they’re already making 100 million dollars?

Ashu Garg 42:14

You know, some people want to make more than 100 million, you know, 100 million is the starting point to do something more interesting. So I think there are a lot of entrepreneurs who could make 100 for choosing to do the startup. But again, most startups are not and should not be innovating at the model layer.

I think startups have to figure out how do I innovate at the application layer in a way that I am future proofing my business to innovation by the model layer. And so the people at the application layer are different. That’s not the 1000.

When I say there’s 1000 people, those are the people that are really innovating at the model layer.

Siddhartha Ahluwalia 42:51

And, you know, right now, if you have to, this is the last part of the podcast, we have just five minutes left. If you have to summarize what you’re looking for in entrepreneurs right now for the next maybe few months, because the world is changing so fast, I’ll not you ask, I’ll not ask you for the next one year. Yeah.

Ashu Garg 43:06

You know, I think for me, what I look for an entrepreneur in some ways is actually stayed relatively consistent. I’m, I’m a big deal. I invest in people that have big dreams.

So I look for ambition of scale, I’m looking to build Decacorns, and $100 billion companies, I’m looking to find my next Databricks. So I think Databricks will be a trillion dollar company one day worth, you know, order of magnitude 100 billion today. I’m looking for founders who have an aspiration to build the next Databricks.

That’s where it starts for me. Then I’m looking for founders who have the ability to build Databricks aspiration is not enough. So what does it mean to build the next Databricks?

For me, most of those founders tend to be technical, they need they tend to have a technical insight, they tend to have a, they tend to be AI native in that they understand what’s going on in the ecosystem. In terms of personal characteristics, they have a level of intensity, a level of grit, a level of resilience, and an ability to learn and relearn. There is, you know, second to none.

So those are the kinds of things I look for, I look for ambition. I look for skills, what I think of as capabilities. And then I look for a certain set of personal attributes around what kind of person are they and what, you know, how will they react in difficult situations.

Siddhartha Ahluwalia 44:36

And in your view, right, in the last five years, how many such Decacorns have been created in the ecosystem that matter?

Ashu Garg 44:45

You know, if I don’t have the numbers of the cup, if you look backwards, my guess is that there are many 10s of Decacorns that have been created. Databricks is not the only one. Obviously, open AI and Tropic are huge success.

They’re all $100 billion plus both $100 billion companies. I think there’s, you know, if I had to guess, there’s a dozen, maybe a couple of dozen Decacorns. To me, the more interesting question is not the past is the future.

I think the number of Decacorns is going to explode. I think many more Decacorns will get created over the next decade than over the last decade.

Siddhartha Ahluwalia 45:20

And why do you think so?

Ashu Garg 45:21

Because of this innovation. See, whenever there is a disruption, whenever there is an inflection point around change, it creates opportunity. And I think this opportunity is, you know, to founders out there, I really do think there are almost unlimited opportunities to create new Decacorns.

But you have to find the right market, you have to find, you know, the right use cases, and you have to have a clear sense of why you believe you can be a winner, a winner, and not a winner today, a winner against ChatGPT, a winner against GPT-7.

Siddhartha Ahluwalia 45:57

And currently, the norm is, you know, companies like Lovable are getting created going to 100 million ARR in eight months. Are you seeing the same trend in your portfolio on companies near to you as well?

Ashu Garg 46:08

You know, I wish I could say that every company I invest in gets to 100 million in eight months, you know, that is a one in a thousand shot. We definitely have some incredible companies in our portfolio. But you don’t see that kind of growth often.

What we are seeing, and I talked about it earlier, is we are seeing customers willing to pay large amounts of money for, you know, output. And therefore, we’re seeing deal sizes are much larger than we saw before. We’re seeing consumers willing to pay for technology and willing to pay for AI, which historically consumers were less willing to pay for technology.

Most technology was monetized through ads. We’re seeing much subscription models be much more successful today. So I definitely think the rate of revenue acceleration is exhilarating.

And, you know, we’re seeing some pretty exciting stuff. But still, the zero to 108 months, I wish that happened more often.

Siddhartha Ahluwalia 47:09

But are you seeing in a large span of time also, like two to three years?

Ashu Garg 47:12

We are seeing much more rapidly. Absolutely. Yes.

Siddhartha Ahluwalia 47:17

My observation has been, and correct me if I’m wrong, that the companies winning today or seem to be winning like 100 million ARR today are more attacking the green field right now, which is building new use cases for prosumers or consumers like Lovable, OpenAI carved out new budgets for people as well as companies. So should founders attack the existing industries and try to replace ServiceNow, Salesforce, or should they focus more on the green field side?

Ashu Garg 47:48

You know, I think the beauty of technology is that both are opportunities. If you look at code generation tools, where there has been the most zero to 100, you know, the fastest zero to 100 ramp have been code generation tools. You know, I think you could argue it both ways, you could argue that code generation is a new category, because no one could generate code before it, there was no Lovable before Lovable.

Or you could argue that Lovable is, you know, eating into categories like Wix, and, you know, other site creation tools. Because we’ve had traditional configuration type of tools that allowed you to do self service website creation self service mobile app creation. So what is greenfield versus brownfield itself is unclear in today’s world where they are.

But ultimately, I would say the fastest growing companies find a wedge in the market that is high value. They find a wedge where in addition to being high value, the pain is acute enough that people are willing to take some risk and, you know, sign a check with you. Now, sometimes those checks in the case of Lovable are relatively small, in it, you need many, many, many 10s of 1000s, even hundreds of 1000s of customers to get to 100 million.

And in other cases, you know, the checks are large. And when you have more of a traditional enterprise sale, and you can get to $100 million with 100 customers. I think both both answers are possible, but they have different trade-offs.

Siddhartha Ahluwalia 49:19

And my last question today, is this still possible to create companies like Scale AI, that you have one during in your portfolio, but because the data sets today required are infinite.

Ashu Garg 49:32

You know, ultimately, we are living in a world where data is more important than code. And so the value of data will continue to grow. And once you internalize that, I think companies that help build manage scale data sets will have value.

So absolutely, I think, you know, data companies is going to be an important category, whether they have proprietary data or generate data, whether the data is customer specific or cross customer, there are many good flavors. That said, I think providing data to the SOTA labs is a very hard business. It’s a great business and scale proved to be very successful.

Turing has been very successful. There are a couple of other players. But it is not one for the faint of heart.

You know, you’re going after an incredibly demanding customer, and you have to ask yourself, what am I going to do that none of these other players will be able to do. And so if you have something unique to bring to the party, absolutely.

Siddhartha Ahluwalia 50:44

Thank you so much, Ashu.

Ashu Garg 50:45

Thank you so much. Thanks for having me on the pod.

Siddhartha Ahluwalia 50:47

Yeah, amazing to have you again.

Ashu Garg 50:49

Thanks again. Thanks so much.