Siddhartha Ahluwalia 0:42
Hi, this is Siddhartha Ahluwalia, your host at Neon Show and Managing Partner at Neon Fund. Today, I’m glad to host Ashu Garg, General Partner at Foundation Capital, third time on Neon Show. Ashu, so glad to have you back again.

Ashu Garg 0:54
Thank you for having me again. This has definitely become a habit.

Siddhartha Ahluwalia 0:57
You and Jaya, the article that you wrote on Context Graph has, you know, cornered, I think, all the VC inflows in the last one quarter, and it’s just keeping on compounding, right? So today, what we want to discuss is about the three things, what everyone agrees on, what questions are still unanswered, and what comes next, right? Just to start with, you know, so to our audience who don’t know or who haven’t read the article, can you explain Context Graph, graph and layman?

Ashu Garg 1:24
Absolutely. You know, think of a Context Graph as the institutional memory for an organization. At its simplest level, it is an aggregation of the knowledge of why decisions were made, how they were made, who made them, and the why often has to be inferred from the how and the who.

And so with that institutional memory of all that information that exists within an organization, which is partly in human beings’ heads and is partly in a variety of digital traces, is the Context Graph, in layman terms, and I can give you a little bit more of a technical definition in a minute, if you’d like.

Siddhartha Ahluwalia 2:02
Sure. We’d love to learn that.

Ashu Garg 2:03
So at the root of a Context Graph is this notion of a decision trace. So as humans and agents collaborate to execute business processes, as you go through the 10 or 15 steps, often across multiple systems of record, across multiple communication systems, you know, if you’re making a decision, you might have a Zoom call, you might have an email trail, you might have a Slack message, you may have a phone call, and you might make a bunch of system calls to a bunch of, you know, systems of record.

When you stitch all of that together with a time and state component, you get a decision trace. And historically, organizations do not have these decision traces in a form that can be processed. They have the outcome.

So systems of record capture the output of these decision traces. So what happened at the end? If you signed a deal with XYZ company at ABC price and gave, you know, so-and-so discount, that information is captured.

But that’s the outcome of a set of decisions that you made. And that set of decisions you made is a decision trace. When you aggregate decision traces together for an organization in, you know, in the context of a specific business process or sphere, depending on the abstraction, that becomes a Context graph.

Siddhartha Ahluwalia 3:16
And when you wrote this article, why did it hit such a nerve? Why did it go viral?

Ashu Garg 3:22
So Jay and I, and you know, Jay has been my co-conspirator on this for the last year. For the last year, we’ve really been asking ourselves this question of why have agents struggled to realize the potential? The promise of agents, you know, which we were all making 12, 18 months ago, has struggled.

And while we’re seeing improvements, we’re seeing improvements in single-player mode, we’re seeing improvements in single-chat mode. What you really want is a multiplayer mode with state and memory across time and more automation. We have seen our portfolio companies work through these issues.

And as Jaya and I worked through this with companies like Maximor in finance automation, Player Zero in production engineering automation, and many others, what we saw was that there was a common thread to the answer. And that common thread is the Context Graph. And so our belief is that the most successful AI startups of the future will be those that are able to build a mode by building a unique Context Graph.

Siddhartha Ahluwalia 4:29
And today, is there a technology that can build this Context Graph?

Ashu Garg 4:34
Absolutely. The technology is not the barrier here. There are many companies.

I mean, I talked about Maximor, I talked about Player Zero, there is Tessera who’s doing it for ERP migrations and automations. There is Olive that is doing that for sales and Reggie for demand generation. So we have many examples of this in our portfolio.

And there are many examples outside our portfolio, to be fair. I think it is very hard, technically, I don’t want to minimize the technical challenge. It requires a combination of building an organizational ontology.

So you know, the good old world of knowledge management and knowledge graphs. It requires very often the use of some sort of graph based approach, hence the context graph. Now, you could have a graph in the form of a graph database, you could have a relational database with a graph layer.

There are many implementation approaches, and it requires actually capturing state. So I have a portfolio company, Conviva, that is really working on data infrastructure, where state becomes a first class citizen. Because the sequence of events matters, and the timestamp of those events matters.

And historically, that’s not been true. Most databases have not really factored that. So event streams matter.

So all of this is hard. But the best in class companies today have built context graphs, and they will be the winners of the future.

Siddhartha Ahluwalia 5:45
So what delta value are these kinds of companies and the context graph is adding, because what you’re talking about is institutional memory that gets passed on or across the institution, like Amazon for leaders, right? For example, in Amazon, the context is you should follow the leadership principles, right? And if you’re deviating from it, you get charged by the leaders on coming back.

Ashu Garg 6:05
So you know, for human beings, you can articulate principles at the level of decision, you know, Amazon’s leadership. It’s a great example. And then human beings are expected to interpret the eight or 10 Amazon leadership principles in a variety of situations.

How they interpret them, what decisions they make in real world situations, that’s a decision trace. Agents need that context, that’s the context graph. Some total of decisions that human beings made in real world situations, in interpreting those decision principles, that’s a set of decision traces in the context graph.

The reason this matters is that in the absence of a context graph, we are struggling to get into a virtuous cycle of automation. You know, you can do zero short question and answers, which are GPT and Gemini, that’s the model. And the models have gotten 100x better.

As the models are getting better, the quality of those responses is getting better. But human beings are not, you know, don’t function that way. If you’re going through the process of selling, let’s say you’re selling some product to Disney.

You’re trying to sell Disney. Let’s say you’re going to sell your podcast rights to Disney. Think about the steps in the process.

You are going to be talking to dozens of different people at Disney. That process takes 12-18 months very often. You might eventually decide you wanted to get a 40% revenue cut.

They were pushing for 20. Think of the back and forth that happens and all the different situations in which those interactions happen. If you now want an agent to automate that decision making.

So on your side, you’re like, oh, I’m only going to deal with Disney once, why the hell would I want to automate it? But think about it from Disney’s perspective. They’re talking to podcast owners like yourself, very successful podcast owners, but they’re talking to thousands of them.

So they’re looking for an agent that’s going to automate a series of those decisions. That automation requires some memory of what other human beings did. It’s a staircase approach.

Today, the interface might be a human being asks the agent, okay, tell me what was done in the past. And human beings make all the actions. Step two might be the agent starts to make recommendations.

So you’re not providing data. The agent is the manager and the human being is the employee, conceptually. We can imagine a world where the dynamics are changing, but the conversation is human because you don’t want to talk to an agent, you want to talk to a human, but the decision is being made by an agent.

The agent is the manager. So that’s the continuum. We can also imagine a world where there’s no human being involved.

The Disney agent talks to your agent and they strike a deal. To get to that, and that might not make sense for everything, but there are many situations where that makes sense. To get to that nirvana, we need context graphs.

Siddhartha Ahluwalia 8:45
And how would these agents store the context graph? Because let’s say, it’s like I’m talking to Ashu right now, right? So my agent is only talking over my either WhatsApp, email or Slack with your agent, right?

But the context that we have right now, all the relationship we have without an agent, the agent is not able to capture.

Ashu Garg 9:02
So that’s what makes this hard. And the question is, what abstraction do you store this information? We believe that the best, the companies that win in the era of context graphs will find a way to deliver enough value that they earn the right to build a context graph.

If you go, if Maximor were to go to a finance team and say, I’m going to sell you a context graph. Why? You’re like, well, I don’t know.

I didn’t wake up today by looking to buy a context graph. No customers, you know, I love my idea, but that’s not what customers want to buy. So Maximor has to go to finance teams and convince them that I will help you solve the business problems you care about.

That could be cash reconciliation, it could be, I want to have audit ready automation, I want to have a certain level of verifiability. So the business value proposition doesn’t change. As a founder, you have to identify that.

By delivering a certain level of automation, you might say, look, I’m not going to close the books for you. I’m going to identify, maybe just even something as simple as analytics, I’m going to help you automate the analytics of closing the books. And if that value proposition is sufficient for a customer, they will give you access to the data.

You then as the provider, Maximor in this example, has to stitch that data together. Now it’s entirely possible that I’ve been given access to email trails, I’ve been given access to the ERP system, next week most likely I’ve been given access to the billing system Stripe, I’ve been given access to Zora because that’s what I’m doing my subscription billing and so on. I may not have access to your WhatsApp.

So I’ll have incomplete decision basis. But if I do a certain level of automation, the next step will be for you to say, okay, you know what, I’m getting value out of this, let me integrate my WhatsApp into this as well. So that’s the ladder.

And there’s a virtuous cycle of as agents deliver more value, organizations and human beings will be more comfortable sharing more data.

Siddhartha Ahluwalia 10:44
So let’s take example of Maximor since you mentioned them a couple of times. What delta they are able to deliver to their customers in reconciliation as compared to any XYZ tool, because they’re using context graph.

Ashu Garg 10:55
So look, obviously this varies from customer to customer, but the common pattern we’re seeing is we’re seeing a 30 to 80% reduction in exceptions. So you see pretty dramatic, because eventually finance reconciliation is all a function of handling exceptions. The book should close automatically, this invoice should tie to this purchase order should tie to this payment.

And whenever it doesn’t, that’s an exception. Nirvana would be zero exceptions. It’s all automated away.

We’re not there yet. And there will be a time, but we’re seeing dramatic changes. But I’ll give you another example of a context graph in this ladder.

I have a company called Vivid, started by another Ashu garg, and Varun Kacholia. So Varun and Ashu, who are also the founders of Eightfold, had this idea that we need to capture the institutional memory of a human as a digital twin. So same thing, human beings have state.

Let’s imagine there was a digital twin of you, and a digital twin of your wife. And so when you’re sitting here in the US, your wife can query your digital twin to say, hey, have we paid this bill? Have we done this?

Have we done that? Or I need to do XYZ. So a digital twin of human in a business context, I could do a personal example there, that’s what Vivid is building.

And so where is the context graph coming from? Well, it starts off by you give it access to some data sets. You might say, most people give it access to email, they give it access to their Google Docs or SharePoint or Dropbox folders.

Most people don’t want to start off by giving access to all Zoom calls or phones. But with the set of data it has, it starts to build a context graph of you. And so when you ask the model questions about yourself, hey, did I pay the bill?

Or I talked to Ashok two years ago, can you remind me what I talked to him about? It starts to deliver value to you. If employees in your organization, let’s say you had a context, you had a Vivid twin for every single employee.

You have a finance, you have a CFO, you’re sitting in the US, in the middle of India, you’re trying to figure out, well, we were trying to close this term sheet. What happened? I mean, it’s sitting in someone’s email, it’s sitting in some, you know, the term sheet is sitting in some SharePoint or the Google Docs.

If it’s signed, the signed version is sitting. Well, the digital twin can answer that question. So depending on the role, depending on the person, maybe it’s 20%, maybe it’s 80%.

But somewhere between 20 and 80% of the questions you have are answerable by a digital twin in that case. Because that twin has built an institutional memory of you, the individual, your role, your organization, and that’s the context graph. Now, over time, you can imagine that twin is willing to do more than just answer questions.

Because this is such a broad product, like a digital twin has such broad ramifications, the current approach is not to automate decisioning, it’s to make information available. But as I said, that’s a staircase. Tomorrow, you can imagine that the CFO starts to say, hey, it’s configured that if you were to ask for something, instead of giving you a response, it can actually make a decision on the CFO’s behalf.

Let’s say an employee would ask for approval for an expense, the system could approve that expense, because it has the decision traces of what would the CFO have done in the past. And if it’s within those parameters, it will approve it.

Siddhartha Ahluwalia 14:06
So in many cases, this context is a mode of an organization, right? So why would an organization typically try to share it with a startup that they are working on? Because then the mode is out there.

Ashu Garg 14:20
Look, the most valuable asset of a company is its context graph. And the reason context graph have not been stolen was, it’s very hard to build, as you said, in companies with organizations are rightly nervous about, hey, if I let everyone build a context graph of me as a customer, what is my organizational value? So first and foremost, you know, context evolves, context graph will evolve.

And so my first response to that is, relax, human beings are not that easily replaceable. That said, on a more serious note, I do think data security, data governance, access controls, which agents can access what parts of a graph. See, the downside of stitching everything together is now it’s all available in one place.

And so I would just take, you know, the interaction we talked about in case of Vivint. Vivint spent a lot of time thinking about this issue, because as they stitch together your emails and all of your data into a digital brain of you and a digital brain of your CFO, and you can query your CFO when he or she is asleep, well, the CFO’s email has and data might have all of their professional information, but it will also have their personal social security number, it may have their credit card number, they had a fight with their spouse, it may have all of that, you know, in the same email, because the lines between a personal and professional lives are blurred. And so a large part of the access control technology for Vivint is stripping that up.

And many startups like Vivint will build that within their infrastructure, and it’s a core part of their their own IP. But for most companies, you will want to buy a third party stack. And I have a company called Skyflow.

And Skyflow, and you’re definitely someone you should you should talk to in your podcast. Skyflow is one of those data security technologies, which, while it has existed prior to the context of context graphs, is your purpose built. So Skyflow is purpose built for securing context graphs, it will allow you to sort of control the data that goes in.

So as emails, for example, are flowing into the context graph, in real time in line, it can strip out PII, personally identifiable, you know, information. So you don’t want people’s social security number on a context graph, perhaps, and perhaps you do, perhaps you want to last four digits. So it can strip out in a way that makes sense.

And by the way, it can, it can, it can mask and anonymize the data. So if we strip it out, put something else, it’s a key value pair that you have sort of, you know, you can go back to it in a more secure heart. Second, you actually have to have identity controls.

Your agent is allowed to act, you may have a shared context graph between you and your CFO. When your CFO wants to access the context, it’s okay for her agent or his agent to access personal information. When you’re accessing that context, it’s not okay for you to access personal information.

What you can access versus an employee in the organization can access, because context is a function of pairwise relationships. When I ask you a question, how things are going, your wife asks you the question, how things are going, and your CFO asks you the question, how things are going, you have very different answers to the same question. That’s context.

There’s pairwise context there, and that has to be captured in a context graph. But then when the agent is answering on your behalf, access controls become really important.

Siddhartha Ahluwalia 17:25
So have you created a new market on context graph through the work that you and Jaya have done?

Ashu Garg 17:29
We believe context graphs will be the basis for the next trillion dollar AI opportunity. Like all markets, it’s a combination of things old and new. You know, in the last six weeks since we talked about context graphs, SaaS stocks have collapsed 30 to 40%.

I don’t think I’m going to get any credit for that. I’ve definitely lost some money, a lot of money actually. But I think it’s a sign of the era of systems of record is coming to an end.

That doesn’t mean they’ll go away. You know, databases still exist. Databases will exist forever.

New types of databases will create. I think systems of record will exist. But the value they capture relative to the total pie will decline.

So in absolute terms, they might still do well. And so the correction might be an overreaction. But the relative value is clearly the market is signaling.

And we agree that existing systems of record are not the natural owners of the context graph. And so they will capture a declining share of the value that’s created by technology. And I think entirely new categories will get created.

I have a portfolio company, Tessera. They’re automating ERP maintenance and migration. They want a very large contract.

Tessera won a $25 million contract from a Fortune 50 corporation. And they went from $0 to $25 million in bookings. When does that happen for a startup?

But the more interesting question is why. And I wish I was sitting in the boardrooms in that Fortune 50 company when this decision was being made. That decision race would have been fun.

But my understanding of it is that they were willing to give Tessera $25 million because they cut $75 million out of their SI bill. Not an Indian SI in this case, but an SI. So that $75 million that was going to Tinker, Deloitte, and E&Y, and Accenture, or one of these players, instead was shared as $50 million of saving by the Fortune 50 company and $25 million by Tessera.

Is that a new market? Not really in this case. It’s a cannibalization of an existing market.

The other thing that will happen is that in the case of a Maximor, you might actually choose to, instead of scaling your finance team from five people to eight people, you may do the same work with five or more work with five. You’re capturing a larger share of the opex of a company. That’s a completely new market.

If a Vivint allows people to be twice as productive, maybe they will spend half as much time at work, or maybe they’ll spend 80% as much time at work, a little more leisure, and companies will do much more. I think of this as really it is expanding the pie for everyone. I think first and foremost, I’m excited about the fact that it will allow us to automate the mundane in our lives.

It will both enable and force. There are two sides of the coin, human beings to step up our game. As long as human beings step up our game, we will become managers of agents.

I think humans have a choice. We can be managed by agents, or we can be managers of agents. My hope is that we will all become managers of agents.

Siddhartha Ahluwalia 20:28
I’ll give another example from our portfolio. You know Samay Kohli from Budy. Yes.

Great guy. In Budy, they have created AI teammates for senior living homes. These AI teammates are not EAs.

They create tasks for the sales team. You didn’t talk to the senior for this amount of time, or this is the right time to talk, or this is the time to give them a wine, this is the time to visit their house. The AI teammate, it doesn’t have a UI, first of all, and it creates tasks for the entire sales team.

Ashu Garg 20:58
Those sales people are being managed by agents. Absolutely.

Look, that’s the future. Managed by agents, manager of agents. We all get to pick.

Siddhartha Ahluwalia 21:06
Yes.
But, you know, have you thought about enough? What is the complete infra require? Would these context graphs reside on Databricks, Snowflake, or the complete new infra required?

One is to store them, one is to access them. And what are the opportunities these are creating, right, in terms for new companies?

Ashu Garg 21:25
Look, both Databricks and Snowflake are remarkable companies. And I was very lucky to be a seed investor in Databricks specifically. And both companies have leaned in to the infrastructure opportunity for systems of agents and therefore for context graphs.

You know, Databricks has agent bricks, and Databricks has agent infrastructure, including all the backend data processing and data, you know, they acquired Neon, the database company, and so on. So you can build context graphs on Databricks, and they can be the enabling infrastructure. You can also use, you know, other databases.

There are people using Neo4j. I spoke to the CEO of Neo4j a few days ago, and they have customers building context graphs on top of that. In the case of women, they’re building on top of good old-fashioned Postgres, you know, with a lot of sophisticated, you know, code and programming there.

In the case of Olive, they have built 80-plus SLMs, small language models, because the way they capture the context is by fine-tuning SLMs on particular parameters in that business process. So I think the implementations are manifold. The common theme is that this is hard today.

And our focus is on finding exceptionally strong technical founders who deeply understand the enabling infrastructure required to build a context graph, deeply understand the middleware and application layer infrastructure to build agents on top of existing models. In many cases, not all cases, deeply understand model training and model fine-tuning, because they are doing fine-tuning and training of their own models on top of the foundation models. Those three things, we think, are the bare minimum from a technology standpoint.

And then on top of that, you have to understand the business problem, because ultimately, you have to solve a business problem versus a technical problem if you’re building an application.

Siddhartha Ahluwalia 23:09
So we talked about some of the vertical, like Maximor in Finance, right? So do you think the first large application would be vertical or horizontal, and which would eventually win?

Ashu Garg 23:21
Ultimately, I think it’s not an either-or. So I think the trillion-dollar-plus opportunity will get captured by 100-plus winners in context graphs. So I think there are many, many ways to win.

And some of those paths to success will be horizontal and some will be vertical. And we’re investing in both. For example, my partner, Joanne, has a very successful context graph company in Tenor, which is going after Hespian.

Siddhartha Ahluwalia 23:48
So what I’m learning from you is, you were already working on it, but it’s just that nobody named the nomenclature in the last year.

Ashu Garg 23:55
Yeah, you know, I think, look, new technology, it is not a new technology. You know, it’s not like Spark, which at its core was, it didn’t exist, and then one day it existed. And the basis of Databricks was Spark.

But today, when you think Databricks, most people don’t even know about Spark. So very often ideas start with a single technical breakthrough and compound from there. In the case of context graph, that’s not it.

It is an approach that has been emerging, I would say, for the last couple of years. Because, you know, if you go back in time, it was only two years ago that people started talking about agents. And we had coined the term systems of agents.

And, you know, there was conversations around agent harnesses and this and that, all of which were an attempt to figure out how do we, you know, how do we have large agent swarms? I mean, so that human beings can be managed by agents and become managers. As we have all done that, I think we started to realize that there was a gap.

And that gap and the solution to that problem, that there’s a problem, the solution to that problem is context graphs. And so it’s emerged over time for JR, Joanne and I, who’ve really been in the trenches at Foundation working on this. It has come from the interaction with our portfolio, first and foremost.

They have tried to articulate what are they doing? What is their source of defensibility? How are they driving increasing automation or pass?

How are they capturing institutional memory? And we’ve had a dozen conversations with a dozen different companies. I’ve only given you three or four examples.

And then somehow at some point it clicked that there’s a common pattern. They’re all building a context graph. And many more are outside our portfolio.

Siddhartha Ahluwalia 25:27
I would definitely like to have a digital twin of Ashu where rather than forwarding you an email, would this company be interesting? I’ll just forward it to the digital twin and no.

Ashu Garg 25:34
I have one.
I have one. It’s still in beta testing. But Vivian has helped me build a digital twin of Ashu.

And I can query it. My assistant can query it. And my assistant is happy because, you know, I’m sometimes hard to find.

Siddhartha Ahluwalia 25:44
Because sometimes when emails are not answered, I think that, oh, is Ashu busy or is the company not interesting enough?

Ashu Garg 25:50
I apologize for that. But the twin will solve that.

So, you know, we have to ask the other Ashu guy who’s the CEO of Vivian.

Siddhartha Ahluwalia 25:57
Yeah. We’d love to see it in production.

Ashu Garg 25:58
Absolutely.

Siddhartha Ahluwalia 25:59
You mentioned about a system of agents, right? To solve this problem. Can you dive deeper into this?

Ashu Garg 26:07
So, I think what we’re all realizing is that the future is agent swaps.

Siddhartha Ahluwalia 26:15
What do you mean?

Ashu Garg 26:15
So, that you have multiple agents.

And the way a complex business process will get automated is agents will be task specific. In the tasking, it’s sliced and diced in a variety of ways. And so why do you need multiple agents and why do you want to have task specific agents?

Because you need guardrails. You need to be able to bound. Agents do hallucinate because models do hallucinate.

Human beings also make mistakes. We all forget, you know. If we didn’t forget, we would have no driving accidents.

And we know that the agents driving way more are actually much safer than human beings. So, I think you can have superhuman performance for specific classes of tasks. But it requires us to break down the aggregate tasks into smaller tasks, have agents that do smaller tasks.

And the systems that build agents, you know, is what we call the system of agents. And that’s the, you know, I think of this technology, as I said, in three buckets. I think of it as the context graph, which is the enabling infrastructure.

I think of it as the model layer. And I think of it as the agent harness. And when you put this together, you get a collection of agents, a system of agents.

An agent swap, pick your favorite word.

Siddhartha Ahluwalia 27:20
So, I assume each of the companies, the 11 companies that you have mentioned about, they have built a system of agents.

Ashu Garg 27:25
Yes. So, we are either investing in two things. We’re either investing in systems of agents or we’re investing in enabling infrastructure.

The common theme across these is the context graph. And so, companies like Skyflow that I mentioned or Fortanix are enabling technologies for context graphs. Whereas companies like Maximor, Vivint, Tessera, Player Zero are actually, you know, building those context graphs in the context, no pun intended, of a particular business process or a particular vertical or function.

Siddhartha Ahluwalia 27:54
I find one of your companies, Arise, very interesting because it sits in the infra layer for this context.

Ashu Garg 28:02
Absolutely, yes. So, Arise is another great example where, you know, for your audience, Arise is really the monitoring layer. You know, if you have AI agents and you have models making decisions, you need some way to monitor what they’re doing.

They are the police force and the CIA all put together for the agent work. They monitor what’s called evals. They see what the agents did.

They see what prompts you provided as a human being, what was the response. And it’s very meta. They use models to monitor models because the model output is so large.

Human beings cannot monitor models. So, it’s a very technical team that is building one of the core components of the enabling infrastructure for context graphs.

Siddhartha Ahluwalia 28:51
In previous era or in the SaaS era, the teams were, you know, now you have mentioned technical teams a few times, right? I’m going to hop on that. It required a very strong GTM-C or a GTM co-founder.

Has it changed in the recent agentic era where you are mentioning, like, very strong technical teams?

Ashu Garg 29:07
So, what remains true is that technology always struggles to find product market fit. And ultimately, the companies that succeed find product market fit. I mean, Databricks started in 2013, but it took three, four years and a lot of going out in circles for them to get to the point where they had product market fit.

So, companies need to find product market fit. And ultimately, it is the job of the CEO to find product market fit. You can hire salespeople.

You can hire marketing people. You can hire product people. But the difference between a great and an average CEO is their ability, at the zero to one, to find product market fit.

Not the job of a CEO changes every year. So, once you have product market fit, there are other challenges. But that’s the first step.

Sometimes, deeply technical founders have a natural passion for product, have a natural passion for go-to-market, and have the learning ability to become go-to-market first founders. That’s my favorite kind of founder. Huge success for me when I look at Ashutosh at Eightfold and Vivint, Kabir at Nisera, and many other CEOs that I’ve worked with over the years.

You know, Mohit Iyer at Cohesity. They are brilliant technologists. Very few people can go toe-to-toe with Mohit at Cohesity and Nutanix and build a better storage layer.

He’s one of the top maybe two or three people in the world who can do that. And at the same time, when JP Morgan or Morgan Stanley calls and says, I want to talk to somebody, they want to talk to Mohit. So that’s rare, but that’s the founder we look for. A technical founder that has a passion and has acquired the skills in go-to-market.

Siddhartha Ahluwalia 30:48
So necessarily storytelling has taken a backseat then because naturally the CEOs had the flair for it.

Ashu Garg 30:54
Absolutely not, I disagree. I think storytelling is, see at the heart of humanity is storytelling. Religion is a story.

Money is a story. Ultimately the very human existence beyond the difference between humans and animals is stories. So at the very heart of humanity is storytelling.

The question for us is how do we find as investors, we need to find founders who have enough technical depth. And really, you know, we are pushing the limits of technology and the limits are changing by the month, by the week. So if you were technical five years ago and are no longer technical, that’s a challenge.

So how do you find the balance between deeply technical founders who are exceptional storytellers, storytellers with investors, exceptional storytellers with their employees, and most importantly with customers?

Siddhartha Ahluwalia 31:45
So enterprises are now aware of context graphs. Where do they start?

Ashu Garg 31:49
You know, ultimately the most successful enterprises of 2030, let’s zoom out, were early in this trend. I believe that four years from now, the winners and losers on the S&P or the Nifty in India, the winners and losers will be separated by their ability to embrace context graphs.

Siddhartha Ahluwalia 32:12
That’s a very powerful statement.

Ashu Garg 32:13
I think that would be true because the companies that embrace these context graphs, we are a very different productivity curve. Context graphs will drive organizational productivity. And when you sum that up, it drives societal productivity.

So where do you start? Well, that depends on the organization. If you are the CIO organization in a Fortune 100 company and you are thinking about an SAP or an Oracle migration, well, call 1-800-TESSERA.

Because if you are looking to drive a major system of record migration without exploring what the context graph first companies have to offer, I think you’re being remiss in your job. Now, whether it’s the right company for you or not, obviously every company has to make its own decision. But I think every CIO has an obligation to their organization to explore that.

If you are the CFO of an organization and you have 300, 400, 500 people, whether they’re in India or in the Philippines or in the US, doing accounts reconciliation, you have an obligation to explore context graphs. You have to ask yourself, why do I have 500 people doing this? Can I have 100 people doing accounts reconciliation and 400 people adding some other value to the organization?

If you are a VP of sales at a large company, you have to ask yourself, typically in a large organization, half the company is in go-to-market. Even in the most sophisticated technology, we have many thousands of people in go-to-market. I think we have to ask ourselves the question, which of these individuals will be managed by agents and which of them will step up to manage agents?

And both will happen. But if we don’t have agents in the mix, we will be left behind in the productivity race. And I think that’s the future.

Siddhartha Ahluwalia 33:51
So essentially, what you’re saying is, and maybe you can correct me, you’re betting the firm on context graphs.

Ashu Garg 33:57
We are betting foundation on two megatrends, AI and blockchain. We believe that AI is going to reinvent how humans live, work and play. And we’ve talked a lot about that.

Context graphs is an enabler of that. We also believe that blockchain is going to reinvent money as we know it. And the very nature of money is going to change because of blockchains.

And we’re betting the firm on those two things. And we’ve had tremendous success with blockchain. We were the first investors in Solana.

We’ve been very successful in many other blockchain technologies. Blockchain is a more nascent, and therefore we’re seeing more ups and downs in that technology. But it’s just as powerful as AI as a technology.

Within AI, there are many bets you can make. And it’s not an either-or. I mean, we believe that there is a post-LLM world.

And in a post-LLM world, you will see new model architectures. So we’re already betting on new model architectures. We believe that in a post-LLM world, we will see completely new approaches to agent harnesses and agent frameworks.

And so we’re looking long and hard. And at the same time, we believe that a critical piece of the AI stack is the context graph. And so we are making a very large investment in context graphs, whether it is in application companies, the systems of agents, or the enabling technologies, like Skyflow and Protonix, Enterprise, and many others.

Siddhartha Ahluwalia 35:22
What I’m observing is the narrative in AI is shifting towards winner-takes-all. Is that true?

Ashu Garg 35:28
I don’t believe that at all. I don’t believe that. I don’t see the evidence.

If anything, I think the evidence is the opposite. If we were having this conversation a year ago, people would have, and many people, including myself in varying degrees, were saying, we think OpenAI is going to win. A year ago, many people were saying, hey, we see Llama really, Meta was back in the game.

And we were all counting out Gemini. And Anthropic was sort of a niche. Fast forward a year, Gemini is back in business, baby.

So that started it out. And so Gemini is a legitimate alternative to ChatGPT. The Gemini model stack is a legitimate alternative to the OpenAI stack.

Anthropic has really exploded in the enterprise and has carved out a very legitimate and defensible market segment for itself. And we don’t talk much about Llama today. But I won’t rule out Facebook.

I mean, they’re down the road. I would not rule them out. I think a year from now, we could be having a very different conversation.

So this is by no means, even the model layer is not a winner-takes-all market. In fact, the question we have to ask ourselves is, given how competitive intensity is increasing in the model layer, will the model providers be able to capture returns proportionate to the investments they’re making? And I think the answer is a no.

It is unclear how much value the model providers will capture in the long term. I mean, even OpenAI has shifted from purely releasing the best model to try to win at the application layer, which is ChatGPT. So I don’t believe that’s a winner-takes-all at all.

I do think that enduring companies have to build moats. Like ultimately, if you get commoditized, you’d be the one of a thousand people in any one category. So the old adage that in any category, there are three winners, and then a long tail of maybes or have-nots, and there’s almost sort of a 60-30-10 split between one, two, and three, I think continues to hold.

And so our goal as investors is to find the one, two, or three in every category. I think Context Graphs will create 100 plus winners, and my goal is to be in as many of them as possible.

Siddhartha Ahluwalia 37:32
Also, you’re saying, because you are telling this is a trillion-dollar category of Context Graphs, so 100 plus winners are going to emerge.

Ashu Garg 37:40
Trillion dollars in revenues, by the way. Hopefully many more trillions in market cap.

Siddhartha Ahluwalia 37:45
But coming back to the infra, like LLMs have short-term memory. Context Graphs require long-term persistent memory. So do we need to solve memory first before we talk about Context Graphs?

Ashu Garg 38:00
Look, memory is inherent in Context Graphs, and memory is inherent in technology. So when you say, do we have to solve memory, I think we have solutions. A database is memory.

You have relational databases, which capture a point in time. You have event stream databases. There are databases which capture the time dimension of memory.

You have structured and unstructured databases. So I think many of the components exist. There is also a ton of innovation that is being done and has been done around model memory, around memory between how we have short-term state versus long-term state.

And yes, there are many, many fundamental questions about memory that are unsolved. So what makes this opportunity exciting and challenging, both at the same time, is the Context Graph of today may not be the Context Graph of tomorrow, because the stack is changing. And innovations in memory, whether it’s model memory or agent memory or system memory, whether it’s short-term or long-term memory, innovations in memory will change the nature of the Context Graph.

Siddhartha Ahluwalia 38:52
Have you thought about, as the technology emerges, what will Context Graph be replaced by?

Ashu Garg 38:59
You know, it’s six weeks old. So hopefully, Context Graph is the idea of 2026. And we’ll see what happens in 2027.

Siddhartha Ahluwalia 39:06
And you know, we discussed about, but just for our audience, I want to hop on again. For any organization, for Foundation Capital, for Neon or Disney, where will the Context Graph reside?

Ashu Garg 39:17
You know, where the Context Graph resides will depend on the business function and the organization. But let me step back. First and foremost, every organization will have multiple Context Graphs.

There is no one world model or one world Context Graph in any organization. That’s our bet. An organization like Disney may have 30, 40, 50, 100 Context Graphs.

You know, just like there’s no one data lake to rule them all in any organization. Now you may choose to have one Context Graph for all your customer data. You may choose to have a Context Graph for all your legal data.

And by the way, two Context Graphs, your customer Context Graph and your legal Context Graph may have intersecting data, may query each other. And agents may query both Context Graphs to arrive at an answer. Some of these Context Graphs, many of these Context Graphs, I would say, will be embedded in startup companies, hopefully not startups at that point, public companies, but today companies like Maximor and Tessera and Olive and Vivin that are, you know, built by these companies, by technology providers.

But our own, you know, the data rights to those graphs will be owned by their customers. So many of those Context Graphs will be built that way. And every Global 2000 company will build many Context Graphs in-house.

And when they build it in-house, that’s when they need the data control layer or the data access, data security layer from a Skyflow. That’s when they need the, you know, decision trace evals layer from Arise and so on and so forth. And many of those companies will build it on top of Databricks or something.

So both will be true.

Siddhartha Ahluwalia 40:56
So clearly what you believe in and the world that we are going towards is startups have an advantage over incumbents because Context Graph doesn’t come natural to incumbents. Like, for example, for Salesforce, they have to rewrite themselves. For ServiceNow, they would have to rewrite themselves.

So over a long period of time, then 10 years, the system of record should be replaced, the current system of record.

Ashu Garg 41:18
You know, systems of record companies have a choice. They will either choose to embrace Context Graphs or many of them will, you know, sort of bury their head in the sand the way Nostridge does, and the rest will be history for them. For the companies that choose to embrace Context Graphs, and I think, you know, whether it’s ServiceNow or Salesforce, at least those companies, I think, have been very, very proactive about embracing to that credit.

They still have a set of very hard decisions to make. They are build versus buy. How do I build or do I buy?

To what extent do I want to be backward compatible with the existing Salesforce architecture, which is their strength, that is their mode, but that is also their constraint? To what extent do they want to be interoperable with other systems? Context Graphs are not bound by system constraints.

You’re not going to build a Context Graph on top of your Salesforce data, you’re going to build a Context Graph on top of your sales process. And every business process touches dozens of systems in a large company. And so incumbents will have to figure out how do I become interoperable with other systems of record, other technology providers.

And some will build and some will buy. And some will perish. And some will build and buy and then perish.

So I think it’s a huge opportunity for startups. But I also think that incumbents, you know, have strategic advantages. And so I am bullish that they will also capture some share of this value.

And so I’m a buyer of Salesforce. And I’m a buyer of Salesforce stock. I’m a buyer of Snowflake stock.

I’m a buyer of ServiceNow stock.

Siddhartha Ahluwalia 42:49
So there is a debate on the biggest value capture in Context Graph. So human in the loop versus handoff execution. What are your views, you know, how in the building of this Context Graph?

Ashu Garg 43:00
Look, ultimately, both will be true. Humans have to be in the loop. And so I think human in the loop is very important.

And in some cases, humans will be in the loop to make the final call. In other cases, humans will be in the loop as an FYI. And in yet more cases, humans will, you know, just have evals to evaluate the decision traces, because so many decisions are made by agents, that even the FYI is an over, you know, is an over load of information.

So all things will be true. And, you know, I think it’s, it’s, it’s, I personally don’t spend too much time thinking about the details of that mix, five years from now, I think it’s going to change, it’s going to change what it will look like in 2030 is different from what it looked like in 2026. There are only three things I know for sure.

I know for sure that over the last 10, 20,000 years, human beings have continued to flourish despite or human beings have continued to flourish and built upon technical innovation. And so that will be true over the coming decades. So I’m very bullish in the human race.

Second, I believe it to be true that the rate of change of automation as a result of AI will continue to accelerate. And the winners and losers on the S&P and the Nifty will be decided by their willingness to embrace automation and their willingness to invest behind. I think the third most important thing context graphs, you build organizational context graph, whether you build it in house, or you build in partnership with startups, whether you build some in house and some in partnerships, but every organization needs to embrace context graphs, every organization needs to embrace automation, and every organization needs to have the confidence that like every generation before it, we will benefit from this innovation and not cow or bow down to it.

Siddhartha Ahluwalia 44:51
So till now has any enterprise called you that we are hiring a chief context graph officer?

Ashu Garg 44:55
I hope not. I hope not. I’m still gainfully employed at Foundation, not looking for that job.

But are they creating such kind of roles? I think we’re definitely seeing a lot interest from chief data officers, chief, a lot of companies have hired chief AI officers, and who knows tomorrow, chief context graph officers.

Siddhartha Ahluwalia 45:11
And what there have been various debates, you know, like from Dharmesh, from Arvind, from Glean that, you know, people are believing is just the thing that we are very far away because till now, even semi autonomous agents are not delivering.

Ashu Garg 45:24
Look, I think if you go back to the debate around context graphs, both Dharmesh and Arvind I think have been great collaborators on this discussion, as has Aaron Levie at Box or Jamin from Altimeter or, you know, animation player zero and many others. I think it’s very consistent that the idea of a context graph is shared by most pioneers of application or AI technologies. It is also very clear that it is a word that means different things to different people.

And it is also very clear that the very nature of context graphs will evolve. And so the timeline, you know, is a little bit like a mirage, it’s around the corner. And so we can build context graphs today.

And the companies that I mentioned have built context graphs today. And I think we will look back in three years and say, Oh, my God, that stuff was so primitive. The context graphs of three years, you know, from now, would be 10 to 100x better than today’s context graphs, because the enabling technology will make them better.

And the decision traces you capture will make them better. See, key to the context graph idea is the notion that decision traces, you know, compound. As you capture decision traces, you earn the right to capture more decision traces.

And that compounding of decision of capturing decision traces is what makes for a richer context graph. And as the context graph gets richer, you need more enabling technology.

Siddhartha Ahluwalia 46:50
You also talk about the orchestration layer. Can you dive deep into it? Like, where startups would win an orchestration layer, right?

And do you believe this is the right layer where, you know, value will be captured?

Ashu Garg 47:01
So ultimately, you know, as we talked about, if I go back to the beginning, the context graph is the aggregation of decision traces. Decision traces can only be captured if you’re in the orchestration or you’re in the flow of those decisions. If you ask a human being to sit down and write down a decision trace, you know, they’re going to, you’re going to get, you know, a bunch of, you know, swear words from them.

And even if they’re willing to do it, they can’t because most human beings don’t, are not a template. When we make the same decision 20 times, 30 times, you make it slightly differently. We may end up with the same outcome or we end up with different outcomes.

So capturing decision traces requires capturing the decisions as human beings interact with agents over time across hundreds, thousands, hundreds of thousands of decisions. The only system that can capture that are systems that are orchestrating those decisions, whether that is a startup like Maximor or Tessera or Vivint, or it could be an incumbent that builds an application. But if you’re not in the flow of the decision, if you’re not orchestrating those decisions, you have no ability to capture them.

A database vendor can’t capture them because they’re not in any one database. An analytics vendor cannot capture them because they capture the outcome, not the why and the how. And the only people who capture the how decisions are made are the orchestrators of those decisions.

And you can infer the why from the how.

Siddhartha Ahluwalia 48:27
So one of our companies called PitCrew is by third time founders, Rishi and Sameer. They sold their first company to Freshworks, second to LegalZoom. The third company that they’re building is the orchestration layer for wealth advisors in the US.

Ashu Garg 48:38
Look, you take wealth advisors, it’s a highly automatable opportunity. Most wealth advisors have a lot of back office, compliance processes, a lot of processes to automate the execution of what high network clients ask them to do, whether it is money management, or it is helping them buy their planes or helping them manage their real estate. So I think it’s a big opportunity.

So congratulations to them.

Siddhartha Ahluwalia 49:00
Thank you so much, Ashu. Love the discussion.

Ashu Garg 49:02
Thank you so much for having me on the show. And thank you for making Context Graphs the idea of 2026.

Siddhartha Ahluwalia 49:07
Thank you so much for building this idea.

Ashu Garg 49:09
Thank you for having me.