Fireside Chat with Anurag Gupta and Paul Hsiao

Please welcome Paul Hushow, partner at Canvas Ventures, and Anurag Gupta, VP at Amazon Web Services. Hey guys, how's everyone doing today? Thanks for having us. I've always wanted to be a talk show host. I can't see you guys all, but thanks, Anurag, for giving us the opportunity. So my name is Paul Sharp, general partner at Canvas. Joined the venture business in 2003. The first two investments I worked on were Salesforce and Tableau. So, you know, really come a long way. Sit on a few SaaS boards today. So privileged and excited to have Anurag here. For those of you who don't know him, He joined AWS in 2011 and leads their analytics and transactional database today. You know, from a short span of 6 years, these guys have built a multi-billion ARR business. You know, we'll get into sort of how they did it and it just really, it's exciting to have you here.

Thanks.

How did you get to AWS?

So I was at, a startup which we sold to another company. You were on the board, of course. Yeah. And so I spent a couple of years there, you know, to get the handcuffs off. And then at that point, I couldn't do another startup for personal reasons. And so it was a question of where to go next. And so the question I asked myself at that time was, so every, you know, it's easy to look back and figure out 10 years ago, what company would you have loved to have been joined? At that time. So the question for me was 10 years from now, what's the company I wanted to join now? You know, would, and you know, would have wanted to join now. And you know, Amazon and AWS really were at the top of that list. Yep.

Bezos talks about the regret minimization. Sounds like you went through a very similar thought process there too. So what, what led you to AWS then?

So, you know, I'm a, systems guy. And so if you look at systems, there are these points where things transition like mainframe to mini computer, mini computer to PC, PC to intranet, intranet to cloud and mobile. And when you look at those transitions, basically the whole stack has an opportunity to change. And so I saw the opportunity to change the stack.

Got it. You know, and so, you know, if we think about, It's been astonishing the depth of the innovation that you guys have introduced, the pace of the product, so that you have also done. And more importantly, I'm just surprised by sort of the business model, right? You guys have turned a lot of non-consumption clients into mainstream customers, right? I guess question is about like, how do you guys approach this? Like what sort of, what was the insight there that found? In such a short period of time, you know, you could actually come into the business and build such an offering.

Yeah. So you're familiar with our working backwards process? Maybe I can fill you guys in a little bit. Yeah, please. So, you know, what we do is before we write a line of code, we work on the press release and FAQ, frequently asked questions document. And basically there are two problems when you're building something, whether what people need and whether it's possible. And so whether it's possible takes a lot of time to figure out. Um, but what people need is actually a simpler problem. And in some, particularly for engineers, it's very easy to get wrapped up in the technology and that's very different from figuring out what people want. Yeah. And so, you know, for these big services, we'd like to go into large spaces because you're, you know, you're not gonna get 90% of some small space. You're gonna, you know, if you, you wanna do well enough so even if you mess up, you're gonna do pretty well, right? And it's meaningful business. And then you know that you're gonna go into a space where you're going to be, you know, feature poor compared to what people are, at least when you launch, right? I mean, that's the general nature of minimum viable product, right? And so you have to look for a space of non-consumption. And so one of the things when we started Redshift was this notion that, hey, enterprise storage, the growth rate is about 4 times the pace of data warehousing. So why is that? And so you get into then with your customers figuring out why they're storing data that they're not analyzing. Now that's a strange thing, right? Yes. You know, and so what we figured out is that it just wasn't simple enough. It just was too expensive and was just too slow to put more data into these systems.

Yeah.

So, you know, that sort of leads us to a somewhat different model.

Yeah.

Than what they have today.

I mean, gosh, congrats on the success, right? If I track you for the last 6 years, it's more than, what, 12 products from Redshift to EMR to Elasticsearch, and now most recently Neptune. You know, I sort of think about how we sit on boards where we dream about building a $100 million ARR, $150 million ARR in 6 years. It's incredible that you guys have built multi-billion ARR in the same time frame basically from scratch, what's a 1,000x reach sort of a growth rate from—

Yeah, I mean, we don't break out anything below the AWS level, but you know, we're pretty happy with where we are.

I worship these guys if it's not, if it's not obvious to y'all. What surprised you?

So what surprises me is, is that when I first joined AWS, it was a much smaller team than it is today, obviously. And, but the people in leadership roles are much the same. And so the ability to scale without going outside has been really impressive to see.

And do you, how is that? Like how do you, you know, sort of, I think about, I sit on boards of some of these SaaS-enabled marketplaces. They are really powered by the services that you provide, right? I mean like from QuickSight and others, I mean I think about Howz, I think about Transfix. It's incredible what you have done, but yet you held the headcount sort of the same. I'm just curious about what's the methodology, what about AWS that enable you to do so?

So I think we think about scaling all the time.

Yeah.

And the key thing there for us is really working through mechanisms because, you know, we have, Jeff has this famous statement that good intentions don't work. People don't come into the office expecting to mess up, right? And you know, if they do, you know, you've got bigger problems, right? And so the question is really how do you create mechanisms so that you can make things that are scalable? And so examples of that would be, so you set a goal, maybe it's a revenue target, and then you're gonna have trackable output metrics that drive that. And then you might have input metrics that drive that output metrics. And then you maybe recurse a couple of times to say, okay, treat that input metric as an output metric. What drives that one? And you know, that just lets you watch the metrics and then you can just focus on the things that are doing really well so you can double down on them and the things that are doing really poorly so you can, you know, focus your attention, which is a fixed resource and apply into that, right? Because the things that are in the middle, that's fine, right? You don't have to pay that much attention.

Is there an example of a service that you guys have applied that methodology to?

Really all of them. You do? Yeah, it's across the board.

From Aurora to Athena, you guys have done similar things.

Yep.

Got it. So switching topic a little bit. Mm-hmm. I mean, you have an amazing vantage point on the transformation that's really happening in the stack that many of companies here build on. Just curious about what you see, you know, in the transformation right now.

So I guess I'd look at the growth of data.

Yeah.

I think that's been a really big deal. I mean, I often, you know, my estimate is, is growth data grows 10 times, 10-fold every 5 years.

Yeah.

And so that's basically 3 orders of magnitude every 15 years. So if you have a petabyte now, you're going to have an exabyte in 15 years. If you have a terabyte now, you're going to have a petabyte in 15 years. And that's a lot, right? And that actually changes how you approach things because It used to be that you built these monolithic systems and then just for the, um, because of the growth of data alone, we've had to go to distributed systems, things like Hadoop, et cetera. Um, and I think that's also in the future gonna drive a move towards things like serverless and going to multi-tenant systems and pay for what you use in very small, uh, request bursts. And, uh, you know, we could talk a little bit about that.

Yeah, totally. I mean, I think the notion that, data that 10x every 5 years, I encourage you to name it Gupta's Data Law. Because on the venture side, we've sort of followed Moore's Law for many years. That's been the defining technology curve. And what's interesting is that with the data increasing 10x, there's just a lot one could do.

And Moore's Law is basically out of bed.

It is, right? And so then how do you drive the innovation based on the data that we are collecting?

So I think, I mean, data is really like potential energy. So it doesn't do you any good if you just store it away somewhere. But if you look at, it used to be that voice recognition systems were just garbage, right? And I mean, if you ever had like a wrist that was hurt and you tried to use one of those systems, it would just, every word would have a mistake in it. And now suddenly they all work, right? Why is that? Well, you know, it's really two factors. One is, is that we have enough data to provide a meaningful corpus as a training set. And we have inexpensive compute power that we can use to sort of, you know, run the training models on and just make them better and better and better. And I think that's generally true across a lot of AI. And, you know, AI has been a pretty big change in the industry. I think we're still in the hype cycle on it, but, you know, maybe today. But, you know, it's going to be, I think, transformative. And, you know, the other thing that I've seen is, is that people's expectations change, right? Like, you, you know, you start off and you're saying to yourself, wow, it's— isn't it amazing I can talk to my phone and have it do something? And then like 2 weeks later, you're like, how ridiculous is it that I'm— look, on this voice conference and have to tap things in while I'm trying to drive my car, rather than just being able to tell it who I'm trying to talk to.

And that intelligence, how you build into that software stack makes a difference.

That's right.

As we talk about the data increasing so much, what are these data? Are these business data? Are they machine data? What are you seeing? What are we collecting here?

Yeah, so I mean, at the highest level, I guess, The internet itself is a video serving platform, right? With a small amount of data going on it. But just focusing on the data, you know, it's, there's this, you know, we used to think about data in terms of things like business transactions. So that's a tiny fraction of it now. It's all really machine to machine now. Or it's internet scale consumer or, you know, and so forth. So the real question still is, you know, what do you do with it? I can give you an example maybe and, So let's think about if I were building an expense reporting app 5 years ago, what would I do?

Yeah.

So, you know, I'd obviously make it mobile first. I'd do gestures. Maybe nowadays I'd do voice. Yeah. You know, I'd focus on productivity and so forth. So that's all the basic meat and potatoes, if you will, of building something. And it's not particularly defensive.

No.

Because anybody else can do the same thing, particularly when they have my app to go and copy, if you will. And so the question Jeff would ask—

The software development stack is so easy now, everyone could have these mobile sort of—

That's right. And so what Jeff would say is, OK, you've built a castle. Now the barbarians are coming to attack your castle. You know what you need? You need a moat around your castle. And so the question is, what's your moat? And so I think data presents a moat because it there's a flywheel for it. The more data you have, the more use you can provide into it, and then the more data you have, right? And those are positive things. Yeah. And, um, an example of that, going back to expense reporting, is, is that if you think about it, expense reporting could be viewed as a productivity tool or something like that, but it's something more than that. It's actually a measurement of transactions or interactions between producers and consumers. Right? Much like Amazon.com is. And so if I'm a consumer, wouldn't it be great if I go to Kansas City to find out what are the restaurants that people like me use? That's right. Very similar to finding out a book. Or if I'm a producer, where do I stack rank relative to other people, and what could I be doing to make it better? Or if I'm a company trying to provide this app, how do I do automated acceptance of trivial transactions to prevent people from dealing with them? Or like understand fraud where people are doing something where I can collect that. And what's great about that sort of domain in SaaS is that you don't have data for one company, you have data for a lot of companies. And there's a positive cycle there. The more data I have, the more useful it is, the more data I have, right?

And so that's, you know, I'm glad you—

you know, it's a real shift away from—

it is, right? When we think about, you know, having watched this industry for, gosh, coming up, you know, 20, 30 years. But there was a semiconductor that sort of innovation drove a lot of, and then we put the ERP CRM on the last, what, SaaS is now coming up 20 years old now, right? We've gone to the cloud transformation, we've gone to the mobile transformation. It's harder to build that moat around SaaS. I mean, we see the white spaces taking up, the pitches that come into our shops are, you know, you target a particular vertical industry like farming, like whatever the vertical industrial might be, or, you target a market segment like SMB or mid-market and start. We think about like how do you make that SaaS more intelligent, right? How do you build that something that self-learns and that train data is where something we think a lot about. Like can you put, can you actually incorporate a proprietary data into that SaaS stack?

Exactly. Yeah.

Do you, one thing if you look at the capabilities that Athena and all these have provided, it feels like everyone needs a data lake strategy, right? What is a data lake?

So a data lake is basic, basically once you have a lot of data, you know, you need to put it into a central location. So, you know, in AWS that might be S3, which is our central repository. You know, I think of S3 as sort of like Chicago O'Hare, everything can get into it, everything can get out of it, you know, so, and then around, you know, you want that data to be in some standardized open format so that, you know, the tools that you might use externally or the tools that we provide, all of it can access it. You can't afford to have your data be trapped because it's way too expensive to move it around. Got it. Or change it or transform it or whatever. And so then you need a real portfolio of different data analytics capabilities around it. You know, so in our case, we have serverless SQL in Athena, or we have a dupe in EMR, or we have, you know, data warehousing in Redshift, and, you know, so on, so on, so on. And the real goal for us is to be able to provide something so you just throw your data in and the system will go and process it and understand it and so forth. You know, we're not 100% of the way there, but it matters, right? And so, you know, you need to secure that data lake, of course, you know, because you And then around the basic core data analytics, you need some form of AI, right? So you can really gain incremental value off of it.

One of the things that we all struggle with is finding sort of, I mean, I think entrepreneurs here do as well, is the scarcity of finding data scientists, right? Including the data engineers. I'm just curious about what you've seen and suggest your sort of recommendation on that as well.

You know, I see a lot of people on LinkedIn who are suddenly data scientists. But, you know, I think in some sense the question really is what you need. You know, I don't know that it's about data science and, you know, how many PhDs you can stack from CMU or someplace like that. It's really about figuring out the value you're trying to give to your customer. It's very much back to that working backwards process. Of going and figuring out, what am I trying to do? If I'm trying to build a recommendation engine, then there are particular things I need. I need someone who knows recommendation engines, right? And the degree to which you can get, you know, the great thing is, is that most of the algorithms are well-known, published, you have open-source formats. A lot of us in the infrastructure industry are providing our technology stacks in open source. And so I think that, and, you know, but the pace of innovation is really significant. So I think, you know, what people really need right now is advice on what to do.

Yeah.

And then once they have a measure of advice, you know, the building part of it isn't, isn't dramatically different. And, you know, advice is— I don't know that you need in-house advice necessarily, depending on the size of your company.

Right.

Of course.

You know, we see one thing that we see is because that stack is the capabilities coming from you is so sort of, you know, even though you're humble to say not full stack, but it's really a very complete, that if someone wants to actually have the data lake, they can actually come to you. They want to integrate the data into this app, they can do that as well. One thing I think what's interesting is then the moat, right? Going back to the moat conversation is what's the proprietary data set that, the individual customers have that can make it, that could feed into the machine learning model that will give them an edge on whatever they can do for their customers.

That's actually a really important point because I think a lot of people, when they set up their initial contracts with their customers, you know, they don't include, you know, what rights that they have to the customer's data. And that's an important question. And I think it's that we all sort of, end up on these systems, whether they're consumer or business, where every so often the EULA changes and you just have to hit accept after paging through, you know, 8 pages before you can do anything. But, and that, I think a large part of that is because, you know, people need better access to the data, not because they're interested in invading your privacy, just because they want to share it, aggregate it, figure out how, you know, how to provide a better experience. Experience. And so I think the degree to which we can, you know, a company can set those things up easily, quickly, and so forth, removes those internal barriers because the value really is in aggregating a multitude of people together, right?

I mean, I think we used to think about what's the IP rise, right? The software, you know, what is the IP stack? Nowadays we ask a lot of questions to the companies that we fund. What's your data rights, right? Do you have the rights to, when your customers give you data or when you're collecting data from different sources, do you have rights to that? We've been surprised by some, how some of the large SaaS companies, as much as the talent-based machine learning talent that they have applied to it, they actually can look into their own customers' data to actually apply intelligence.

I mean, you have to be very careful about it, right?

Because I mean, it's the privacy things here.

Well, and you know, it's a trust thing, right? You can't afford to break customer trust ever, right? I mean, and if you're a customer-centric company, like we aspire to be at Amazon, you know, I think you have to be very thoughtful about what you're doing and why you're doing it. And that you're doing it in such a way that it really benefits the person who's providing you that information, not, you know, doing it in any sort of negative way.

Yeah, totally. Shifting a little bit where, you know, we've seen so much of the things is global in nature now, right? Just curious about what are you seeing on the, from that perspective, you know, a lot of customers want to have these, you know, multi-master datasets everywhere. How do you, historically that's been a really difficult thing to do. I'm just curious about what are you guys doing on that front?

So, you know, we're working on it. So, you know, projects are underway, we announced them in the most recent re:Invent, but it is a challenging problem. I think people have an expectation for low latency communication and, If I'm playing Angry Birds in San Francisco and I go to London, I expect to still be able to start that game from where I left it off. And what's true in consumer is equally true in enterprise. And it's a funny thing that enterprise tends to trail consumer. Is that right? Think about messaging apps. The cost of them on the enterprise side versus the value of you get just by using, say, Facebook.

Yeah.

Right. And the pace of innovation on the consumer and the simplicity of the experience and all the rest of those things.

Yeah.

And, you know, the fact that, you know, it's free, right? For me as a consumer at least.

Right. The, when you think about your product and, you know, people are thinking about, you know, we always talk about having product strategy, right? And now we think about coming back to the data rights a little bit. Even like globally in Asia, how do you think about data strategy? Like what's the tactical advice that you give to people when they like, is it just sort of even having a conversation on data strategy? Like when you get into sort of talking to, when you meet your customers, what is it that you want them to think about?

Yeah, it's, that's a really interesting question. I think I frankly don't know. I think it, you know, I'm not, deeply strategic guy. I mean, my basic notion is do what my customers want and just like aggressively go and keep iterating because, you know, that way I don't have to think forward 5 years. I can think forward one scrum, you know, and, you know, as, or, you know, maybe a little bit longer than that if it comes to the question of something that's just fundamentally hard to do.

Yeah.

But, you know, I think it's It's really just a question of understanding what your customers want from that side. From a data rights side, I think it's really just a question of understanding what one might want to do and making sure that you handle the privacy considerations, particularly with the European Union, GDPR, and all of that. But, you know, which I think is not just an EU issue, right? You want to be able to do that for any customer anywhere.

Well, maybe another way to come at the question is like, if you think about the capabilities, the range, like the 12, 15 products you have, what is it that people— and we've gone from batch to real-time, we've gone from sort of global in nature and the depth of the pipeline and data lake. What is it that people don't appreciate sort of how they should use these products yet? That would turn that sort of that business advantage to them.

So, you know, I think what's increasingly important, I think, is bringing simplicity for people. And so I announced a bunch of stuff at the past re:Invent, AWS's user conference in November. And one of, I was pretty surprised that the thing that really most attracted people in the data space was serverless. This notion that, I'm just going to send requests, the server spin up, spin down, resize, whatever. And why would someone care about that? You can kind of build that kind of thing on your own if you had to. But I think the benefit is that we kind of are increasingly in the world— like 10 years ago, we were in the world where you had 9 backend engineers to every frontend engineer. That's flipped over. Right? So now you've got— like one backend engineer, you know, two or three ops engineers and the rest of 'em are front engineers, which is awesome if you're an apps person because the frontend is where you're providing value, right? And you know, so the degree to which—

You take the complexity out already on the backend.

That's right. And the degree to which we can provide that simplification, I think, matters a great deal. And it's, you know, it's not just about AWS, right? There's so many app comp— sorry, API companies that are out there now, right? Like, Previously, if I were creating an application, I'd have to burn like 4 engineers just on how to do payments. And now I do an API call and that's amazing, right?

And we like to think of it as the application, the API application network there, right? How do you—

that's right.

MuleSoft is a— before coming with how they tied everything together.

I think the next interesting thing is gonna be how the SaaS companies themselves become API companies, right? Got it. Like, you know, how do I integrate my expense reporting thing directly into somebody else's application? So it's largely invisible. And, you know, it's, you know, we think that having, you know, the, you know, the time in front, you know, the screen time in front of the customer is valuable. But it's really the data that's valuable. And, you know, attention's a scarce resource and there are only so many applications I'm willing to learn. Totally right. Like, I think we all probably, installed 50 applications on our iPhones or whatever when we first got them, and now we're using 4 or 5, right?

Do you— so when you think about that, right, so coming back to that, how do you add intelligent back into the software stack, right? Um, the notion of having that, you know, sort of that transformation where we're talking about like going from SaaS to that intelligent SaaS, and how do you build that, you know, mode, that data mode in there, are there things that people, the good practice that you're seeing people do right now?

So I think one thing is that the world moves to SaaS, right? I mean, it maybe arguably has already, but there's an awful lot that's still sitting in enterprise applications sitting on-premise and people don't want to work and be, those applications basically treat you like a like a clerk, right? You have to spend a month learning how to do order entry. And I don't want to spend a month learning how to use order entry, or doing time off, or any of the things that I end up having to do, right? I really want small apps that are single purpose. And I want them to have a trivial interface. I want them to be gesture or voice-oriented to the degree possible. And I want them not to bother me if they can do the thing themselves. Yeah. Right? And so that all turns into AI and so forth on the backend. That's right. And also automated integrations into APIs between applications so that they're not asking me to carry a piece of data from one app to another. Yeah. I mean, you know, almost like the iPhone used to be, right? You know, it was transformative to say like, oh, I just read an email or I press on a, you know, something that it recognized as a, a phone number and I'm placing a call.

Totally.

Right. And I, you know, but that doesn't seem to exist in applications today, right? That deeper integration in part because I think everyone wants to own the screen.

They do. And that creates it really, I mean, I think it creates a really interesting opportunity even in enterprise app space, right?

And particularly if you, you know, try to monetize the API call.

That's right.

Right. Because, and that's sort of where serverless comes in. If you can go to something where, I'm not charging you on a per-month basis or on a per-year subscription basis. I'm charging you for what you use. And, you know, I think that's been an interesting move, right? I mean, a large part of the innovation, I think, in the cloud was this move from 3-year TCO, you know, pay provision to peak to pay by the hour. And if you look at something like Lambda, which is our serverless compute offering in AWS, You know, it's pay per request in 100-millisecond chunks. And if you think about 100 milliseconds, that has the same ratio to an hour as an hour has to 3 years.

That's right.

Roughly. That's right. And that's, you know, I think if you, you know, and that's the model. I mean, that's, you know, so we talked a little bit about non-consumption as a place to go. The other interesting thing to do from a business model perspective is overconsumption.

Yeah.

Right? People using something that they're not paying for, something they're not using.

Overprovisioning.

Yeah, it's just useless, right? And so for example, if I had an ETL tool and I pay for it and I'm running that ETL, my ETL jobs, you know, like 8 hours a day, there's 16 hours where it's sitting idle. Why am I paying for it? And, you know, I think SaaS vendors have the opportunity to interleave usage. You know, I mean, yeah, I mean, you have to provision to peak just as I have to provision to peak, but my customer doesn't. And, you know, your customer doesn't, which is awesome.

Yeah, totally.

I mean, the challenge becomes how to make money at those price points. But, you know, that's all about volume, right? One of the things we talk about a lot in AWS, in Amazon, is how to run high-volume, low-margin businesses, right? And to, you know, really pay attention to the underlying efficiencies And you know, that also builds you a moat, right? Because the larger you are and the more you can take out of the efficiencies, you know, that only really comes with scale.

That's right.

And so, you know, that's another positive advantage, just as data is.

You guys have brought the pricing to a level that's so much lower than everything else and made it on demand, that sort of— but it created volume.

And you know, it's based in some sense on the intrinsic sense of optimism that the demand is there. If I can find a price point point and a product that people can afford.

Yeah. Cool. Well, thank you so much for coming here and sharing your thoughts. I mean, round of applause for Anurag here.