Latest / The Neon Show / Why OpenAI Has No Moat & LLM Prices Will Collapse | Avinash Lakshman, Weilliptic

Transcript

Why OpenAI Has No Moat & LLM Prices Will Collapse | Avinash Lakshman, Weilliptic · The Neon Show

• 0:00
  SPEAKER_00: Open weight models will be LLMs will become hard disk drives or SSDs, they're all commodity today.
• 0:06
  SPEAKER_00: If the infrastructure costs are so high and if everyone is demanding for LLM costs to go super low, how are you going to bridge the gap?
• 0:14
  SPEAKER_00: If you raise the cost astronomically, are people still going to pay?
• 0:18
  SPEAKER_00: See, there's this whole talk about circular financing.
• 0:21
  SPEAKER_00: Nvidia is loaning money for someone to buy GPUs and then they become influence companies.
• 0:26
  SPEAKER_00: I don't think we should be afraid of that.
• 0:28
  SPEAKER_00: There is a very dangerous trend that I see with AI, where everyone's moving agentic and they're building agents.
• 0:35
  SPEAKER_00: Guessed in what language?
• 0:37
  SPEAKER_00: In my opinion, that language needs to be debreeded from the data center.
• 0:41
  SPEAKER_00: But you still didn't answer my question.
• 0:43
  SPEAKER_00: Intentionally, maybe.
• 0:48
  SPEAKER_01: Hi, this is Siddhartha Lualiya.
• 0:50
  SPEAKER_01: I'm your host at Neon Show and founder of Neon Fund, a fund that has invested in some of the best companies between US-Asia Corridor in enterprise AI, like Spotraft, CloudSec, Atomic Work.
• 1:02
  SPEAKER_01: Today I have with me Avinash Lakshman.
• 1:04
  SPEAKER_01: Avinash, welcome to the podcast.
• 1:06
  SPEAKER_00: Great to be here.
• 1:07
  SPEAKER_01: So right now you are seeing right what's happening in AI infrastructure.
• 1:11
  SPEAKER_01: A$25 billion hedge fund run by a 25-year-old, went Forex leveraged, then crashed, right?
• 1:18
  SPEAKER_01: So do you think this there is signs of bubble in AI infrastructure right now?
• 1:26
  SPEAKER_00: Man, that's a very hard question to answer.
• 1:28
  SPEAKER_00: There are people far more well-versed in this in the economic aspects of it who have failed to kind of answer this question historically.
• 1:40
  SPEAKER_00: What I feel will happen, where I this is based on what I see today.
• 1:46
  SPEAKER_00: Of course, it will change as the information changes tomorrow.
• 1:51
  SPEAKER_00: But I think what is happening is uh I'm gonna get a lot of probably criticism for saying this, but but I'll say it out.
• 2:02
  SPEAKER_00: I'll say it out aloud.
• 2:04
  SPEAKER_00: I think open weight models will happen.
• 2:07
  SPEAKER_00: I I believe that LLMs will become a commodity if they have not already become.
• 2:16
  SPEAKER_00: LLMs to me, the way I look at it, since you alluded to a lot of storage, LLMs will become what came of specialized servers or hard disk drives or SSDs, they're all commodity today.
• 2:33
  SPEAKER_00: If you remember in the late 90s, early part of this century, you know, there the hardware vendors would pound their chest, oh, we have this new server with this RAID 10, pack 10.
• 2:46
  SPEAKER_00: So they were selling at 10 extra price.
• 2:47
  SPEAKER_00: Exactly.
• 2:48
  SPEAKER_00: Exactly.
• 2:50
  SPEAKER_00: Look at what where they are today.
• 2:52
  SPEAKER_00: Nobody cares.
• 2:53
  SPEAKER_00: It's just a dumb server.
• 2:54
  SPEAKER_00: All the intelligence moved up into the software stack.
• 2:58
  SPEAKER_00: I think LLMs will go the same path uh if they are not already there.
• 3:02
  SPEAKER_00: And I think uh you already see a lot of uh, you know, this there's a lot of uh outrage, uh perhaps too strong a term, but uh asking for the ban of open weight models and things like that.
• 3:17
  SPEAKER_00: I'll tell you, I'll give you an analogy for a historical analogy.
• 3:21
  SPEAKER_00: And we add this in Hedwig too.
• 3:23
  SPEAKER_00: Every time you take proprietary infrastructure software, they'll be like, hey, we have this in open source.
• 3:28
  SPEAKER_00: How do you compare and contrast with them?
• 3:31
  SPEAKER_00: Right now, as an entrepreneur who's trying to sell that, it is my responsibility to come up with a cogent reply.
• 3:40
  SPEAKER_00: Right?
• 3:41
  SPEAKER_00: If I come with proprietary software that does n number of things, you cut you have something in open source that probably does a subset or maybe does a bit of everything.
• 3:53
  SPEAKER_00: And if someone asks me, you know, Siddhartha does this, you do this.
• 3:58
  SPEAKER_00: I cannot say you need to ban Siddhartha.
• 4:00
  SPEAKER_00: Yeah, I think that's a very uh narrow-minded response.
• 4:04
  SPEAKER_00: You need to take, you I mean, all proprietary software vendors had to come up with an argument to say, make the case for why they are better than an equivalent open source.
• 4:16
  SPEAKER_00: Uh or and right now it is playing out as open-weight models, and I think we will get over that hump.
• 4:22
  SPEAKER_00: Uh, because the reason why this becomes critical, and to answer your to to tie it back to the bubble, if this is going to become commodity, and if the infrastructure spend is so astronomically high, how are you going to bridge that gap?
• 4:41
  SPEAKER_00: There's a lot of people who talk about there's a lot of podcasts.
• 4:44
  SPEAKER_00: I think there's this one guy called Ed Zitroan, you might have heard of him.
• 4:48
  SPEAKER_00: He is just beating the drums so hard, my my ears are hurting.
• 4:52
  SPEAKER_00: He's saying that this this whole these companies should not be allowed to IPO.
• 4:57
  SPEAKER_00: That's the take he has.
• 4:58
  SPEAKER_01: Open AI and SL.
• 4:59
  SPEAKER_00: Yeah, he says they should not be allowed to go public.
• 5:02
  SPEAKER_00: Why?
• 5:02
  SPEAKER_00: He says that they have no clear path to profitability.
• 5:06
  SPEAKER_00: And it's impossible to get to profitability.
• 5:08
  SPEAKER_00: That's what that's the argument he makes.
• 5:10
  SPEAKER_00: Now, if the infrastructure costs are so high, and if everyone is demanding for LLM costs to go super low, how are you going to bridge the gap?
• 5:20
  SPEAKER_00: You bridge the gap only by increasing your costs, and I think to you will have to increase your cost astronomically to bridge that gap.
• 5:28
  SPEAKER_00: Now, if you bridge if you raise the costs astronomically, are people still going to pay?
• 5:34
  SPEAKER_00: Do have they already evaluated the ROI?
• 5:36
  SPEAKER_00: These are things that be they are unknowns.
• 5:39
  SPEAKER_00: So I believe the bet, hence, will be going a layer above and providing that operating system that can drive that ROI and things like that.
• 5:51
  SPEAKER_00: And this will probably play out into becoming commodity.
• 5:58
  SPEAKER_01: So so your so so you still didn't answer my question.
• 6:02
  SPEAKER_00: Intentionally, maybe.
• 6:04
  SPEAKER_00: Your question is very pointed.
• 6:06
  SPEAKER_00: Is it a bubble?
• 6:07
  SPEAKER_00: Yes or no?
• 6:07
  SPEAKER_02: Yeah.
• 6:08
  SPEAKER_00: Uh it's I think the answer is yes or no.
• 6:13
  SPEAKER_00: Uh and the way and uh I I mean uh the reason the reason I hesitate to answer, it's not that I don't want to give you an answer, it's it's that I don't know the answer because it's kind of very nebulous.
• 6:31
  SPEAKER_00: The LLM players I think will become commodity.
• 6:37
  SPEAKER_00: Does that mean that they will implode?
• 6:40
  SPEAKER_00: Uh perhaps.
• 6:42
  SPEAKER_00: Uh but the thing that will not go away, uh I don't think AI will go away.
• 6:48
  SPEAKER_00: That I do believe will still hold up.
• 6:50
  SPEAKER_00: There's a lot, it's a very powerful tool, and I think it will continue to evolve, but the way it is consumed, the layer on top, is where everything will move towards.
• 7:00
  SPEAKER_00: So maybe it will play out by all the AI hype going into those layers while this becomes commodity.
• 7:07
  SPEAKER_00: So it may come out even on the other side.
• 7:09
  SPEAKER_00: That's the way I probably think it will play out, if that makes sense to you.
• 7:13
  SPEAKER_01: So you are saying unlike the the other five.
• 7:17
  SPEAKER_01: See, today all the focus is on AI infra.
• 7:21
  SPEAKER_00: AI infra.
• 7:22
  SPEAKER_00: Now, again, AI infra, you've got to be very careful about how you define that, right?
• 7:25
  SPEAKER_00: Because there are there are uh these inference providers, just pure inference providers, like the core weaves of the world or the DBs of the world, whatever.
• 7:35
  SPEAKER_00: Um that's AI infra.
• 7:38
  SPEAKER_00: Would you call the guys who build the LLM AI infra?
• 7:44
  SPEAKER_00: I would.
• 7:45
  SPEAKER_00: So which AI infra are you specifically referring to, right?
• 7:48
  SPEAKER_00: Uh, because uh uh, for example, the open weight models, they're available.
• 7:54
  SPEAKER_00: Anyone can run them.
• 7:56
  SPEAKER_00: So you can take them and you can run them.
• 7:58
  SPEAKER_00: So then you become the inference provider, but you have nothing to do with the actual LLM.
• 8:03
  SPEAKER_00: You're only providing the infrastructure for it to run on.
• 8:06
  SPEAKER_00: So it depends on who you're uh specifically uh referring to.
• 8:11
  SPEAKER_01: So let's talk about each layer of AI infrared, where you think value will continue to get built and where you think it will be commoditized.
• 8:19
  SPEAKER_01: Just start from the the the basic uh data centers, GPUs, every layer.
• 8:25
  SPEAKER_00: Uh I I think where things will will actually shine is in the infrastructure that one puts together for running agents and for monitoring the activities that AI provides.
• 8:39
  SPEAKER_00: That's where I think everything will move.
• 8:41
  SPEAKER_01: Can you give examples from the past which it happened so?
• 8:46
  SPEAKER_00: Oh, yeah, yeah, very easily.
• 8:48
  SPEAKER_00: Um, it's a very easy example.
• 8:50
  SPEAKER_00: If you think about, like I said, the specialized servers and running the way you ran just say storage infrastructure, right?
• 8:58
  SPEAKER_00: You just people were paying an arm and a limb to buy specialized servers with specialized hardware and specialized networking, etc.
• 9:07
  SPEAKER_00: etc.
• 9:08
  SPEAKER_00: All that became is now commodity.
• 9:11
  SPEAKER_00: No one uses uh SANS anymore.
• 9:16
  SPEAKER_00: Um, it's all running on Ethernet.
• 9:21
  SPEAKER_00: 10 gig, now you have, I think, much more than that.
• 9:24
  SPEAKER_00: 25 gigges is becoming like the uh has become commodity.
• 9:29
  SPEAKER_00: So where did all that it's so you commoditize that layer?
• 9:32
  SPEAKER_00: So what happened?
• 9:34
  SPEAKER_00: Specialized servers became commodity servers.
• 9:36
  SPEAKER_00: You can buy a server, you can put in any SETA drive in there, put SSDs in there.
• 9:40
  SPEAKER_02: Yeah.
• 9:41
  SPEAKER_00: And but how do you protect it?
• 9:42
  SPEAKER_00: Where did the intelligence move?
• 9:44
  SPEAKER_00: It moved to the layer on top in the form of software.
• 9:46
  SPEAKER_00: That's where software systems.
• 9:48
  SPEAKER_00: But I think the the maximum value got created in cloud rather than yeah.
• 9:52
  SPEAKER_01: I mean, the cloud guys did the same thing, right?
• 9:55
  SPEAKER_01: Uh but cloud never got commercialized by that logic.
• 10:00
  SPEAKER_00: Uh how so?
• 10:02
  SPEAKER_00: Because what did the cloud do?
• 10:04
  SPEAKER_00: The cloud's not running any sans, cloud's not running any specialized uh you know networking capabilities that that were actually powering storage infrastructures in the past till the end of the last century.
• 10:15
  SPEAKER_00: Cloud basically came in and actually did what I'm actually telling you.
• 10:19
  SPEAKER_00: Take S3.
• 10:20
  SPEAKER_00: S3 is running on pure commodity servers.
• 10:22
  SPEAKER_00: Where does the storage happen?
• 10:24
  SPEAKER_00: It happens on local disks that are right there.
• 10:28
  SPEAKER_00: There is no hardware RAID, so to speak.
• 10:32
  SPEAKER_00: It's all software-based.
• 10:34
  SPEAKER_00: All the intelligence moved into the software stack.
• 10:36
  SPEAKER_00: That is where the commoditization happened.
• 10:38
  SPEAKER_00: So the cloud did commoditize a lot of these things.
• 10:41
  SPEAKER_00: Now, if you're asking me about did the cloud get commoditized?
• 10:44
  SPEAKER_00: Now, when you say that, you gotta be, can you clarify what exactly you mean?
• 10:49
  SPEAKER_00: Here is why I the way I would elaborate on the cloud is when you the cloud itself actually today is kind of a commodity because it is like electricity.
• 11:03
  SPEAKER_00: But there is still more intelligence that you can build.
• 11:06
  SPEAKER_00: Because what do you what happens when you have regional cloud outages?
• 11:10
  SPEAKER_00: How is your company protected against that?
• 11:13
  SPEAKER_00: Where does that infra live?
• 11:15
  SPEAKER_00: So you keep moving it about, that's where we come in.
• 11:18
  SPEAKER_00: We come in and we run as an overlay on top of all these clouds.
• 11:21
  SPEAKER_00: Why?
• 11:21
  SPEAKER_00: A few months ago there was an AWS East Coast outage, it took out large portions of the internet.
• 11:27
  SPEAKER_00: Were all those guys prepared for that?
• 11:30
  SPEAKER_00: No.
• 11:30
  SPEAKER_00: Do they care?
• 11:31
  SPEAKER_00: I hope they do because downtime means significant loss in revenue.
• 11:34
  SPEAKER_00: So that's where these kind of solutions come in, where you run as an overlay on top of all these cloud providers.
• 11:40
  SPEAKER_00: So we are running on top of something that is already commoditized, which is the cloud.
• 11:45
  SPEAKER_00: So that's the same trend I see playing out here too.
• 11:48
  SPEAKER_00: But the the commoditization doesn't show in the pricing of the cloud.
• 11:53
  SPEAKER_00: Hey, something's got to give.
• 11:54
  SPEAKER_00: Someone's gonna charge you for what you use, right?
• 11:57
  SPEAKER_00: I mean, uh the cloud guys also have to recover the costs for what they do.
• 12:02
  SPEAKER_00: At the end of the day, you know, we complicate this a bit too much.
• 12:05
  SPEAKER_00: At the end of the day, it's business 101.
• 12:07
  SPEAKER_00: I have a commodity, which I believe you need.
• 12:10
  SPEAKER_00: You've got to pay the right price for it.
• 12:12
  SPEAKER_00: What is the right price from my perspective?
• 12:15
  SPEAKER_00: It's becoming like a physics, yeah.
• 12:17
  SPEAKER_00: Uh inertial frames.
• 12:19
  SPEAKER_00: Which frame of observance are you at?
• 12:21
  SPEAKER_00: For me, as a provider of the capabilities, I have cost my cards for that solution.
• 12:28
  SPEAKER_00: The money I make should be should at least make me break even.
• 12:32
  SPEAKER_00: Right?
• 12:33
  SPEAKER_00: So that's where my price should come in.
• 12:35
  SPEAKER_00: And for you, you have a need.
• 12:38
  SPEAKER_00: Now, you one would have to figure out if I were to run this by myself, what would my costs be as opposed to if I run it over here and do the economics.
• 12:47
  SPEAKER_00: I think, and that's what people are trying to do.
• 12:49
  SPEAKER_00: And I think even today there's an argument that cloud is extremely expensive and you're better off running it in your own data centers.
• 12:55
  SPEAKER_01: Yeah, that's that's what I'm saying, right?
• 12:57
  SPEAKER_01: Cloud has never become a commodity because it's still very expensive for any, let's say, SaaS company or today, uh, any digital native business, let's forget ISV or digital native business, 10 to 20% of their revenue directly goes to the cloud.
• 13:12
  SPEAKER_01: And what is happening now, it's become worse for them.
• 13:14
  SPEAKER_01: 10 to 20 percent of revenue is also going to the LLM.
• 13:18
  SPEAKER_01: So they are left with 60% gross margins where they can't account their custom customer service, sales, marketing, RD.
• 13:26
  SPEAKER_00: See, there is that is also see what has happened since you threw the LLM in there.
• 13:32
  SPEAKER_00: Let's leave the LLM aside for one second.
• 13:35
  SPEAKER_00: Cloud is something that is very easy, you can get started with.
• 13:38
  SPEAKER_00: Yeah, right.
• 13:39
  SPEAKER_00: But pretty soon, as you would see with any business, your success becomes your enemy.
• 13:46
  SPEAKER_00: Now, with success comes more spent in the form of cloud spent.
• 13:51
  SPEAKER_00: So there will be a point in your in the evolution of your business where you would need to make you need to take a decision.
• 14:01
  SPEAKER_00: Do I need to run this elsewhere?
• 14:06
  SPEAKER_00: Oh, and if I were to run it, if I were to run and manage my own environment, what will my costs look like?
• 14:14
  SPEAKER_00: It's not just costs.
• 14:16
  SPEAKER_00: If you start running in the cloud, it's very easy to be seduced into using the services that they offer.
• 14:24
  SPEAKER_02: Yeah.
• 14:24
  SPEAKER_00: Right?
• 14:26
  SPEAKER_00: It's not easy to just, you know, uproot all that and take it somewhere else because you need to have those services available to you.
• 14:35
  SPEAKER_00: So there's a lot of things you need to think through, and that's the calculations that every, I guess, business will have to go through.
• 14:44
  SPEAKER_00: If you start using, let's say you start using S3, and your applications are deeply entwined in that.
• 14:53
  SPEAKER_00: If you want to get away from AWS and you decide, okay, you know what, I'm not going to the cloud.
• 14:58
  SPEAKER_00: I need to figure out how I'm where are you going to get your equivalent of S3 from?
• 15:02
  SPEAKER_00: I'm just giving one example, but there could be many such systems.
• 15:05
  SPEAKER_01: So I think that's why Databricks is becoming valuable because a lot of now applications are running on Databricks rather than uh on AWS.
• 15:17
  SPEAKER_00: So Databricks is trying to take away value from see the thing is you moved uh a problem from one place to another place.
• 15:26
  SPEAKER_00: That's all you've done.
• 15:27
  SPEAKER_00: You've just shifted the problem.
• 15:29
  SPEAKER_00: Tomorrow, I'm not saying Databricks will do that.
• 15:33
  SPEAKER_00: I don't want anyone to misconstrue what I'm saying.
• 15:35
  SPEAKER_00: So let's replace Databricks with XYZ corp.
• 15:38
  SPEAKER_00: Yeah, right.
• 15:40
  SPEAKER_00: XYZ corp could come and tell you, hey, you know, run on our infrastructure, we have one-tenth the cost.
• 15:45
  SPEAKER_00: How do you know that's not just a lure to get you there?
• 15:49
  SPEAKER_00: And then at a certain point in time, they could just go 10x, and then you're caught in the same situation.
• 15:56
  SPEAKER_00: So that's that's the evolution that every company will have to go through.
• 16:01
  SPEAKER_00: And that's why companies like you know, like the Googles of the world or the Amazons of the world, they built their own infrastructure because they realize that at the scale at which they operate, it's just not financially viable.
• 16:16
  SPEAKER_00: Imagine if Google were running on the cloud.
• 16:18
  SPEAKER_00: Imagine they were not in GCP and they were running on AWS.
• 16:20
  SPEAKER_00: It would have gone billion.
• 16:22
  SPEAKER_01: All their profits would have exactly.
• 16:24
  SPEAKER_00: In fact, they're getting wiped out by LLM.
• 16:27
  SPEAKER_00: Yeah.
• 16:27
  SPEAKER_00: Inference.
• 16:28
  SPEAKER_01: Yeah, another thing that I want to discuss with you is today there are hundreds or probably thousands of inference providers.
• 16:35
  SPEAKER_01: What will happen to them?
• 16:38
  SPEAKER_00: You're going to get me in a world of trouble.
• 16:41
  SPEAKER_00: See, there's this whole talk about you know circular financing and things like that, right?
• 16:50
  SPEAKER_00: But I don't think we should be afraid of that.
• 16:53
  SPEAKER_00: You know, people draw, people say, oh, Nvidia, Nvidia is loaning money for someone to buy GPUs and then they become inference companies.
• 17:03
  SPEAKER_00: What do you think has been happening in the car industry for the last 50 years?
• 17:07
  SPEAKER_00: When you go buy a car here, who's financing it for you?
• 17:12
  SPEAKER_00: They are allowing you to buy the car by giving you this financing, right?
• 17:16
  SPEAKER_00: So they're basically the bank is financing it.
• 17:18
  SPEAKER_00: Yeah, but you but they own the car, right?
• 17:20
  SPEAKER_00: So this kind of circular kind of financing, I think they're kind of equivalent in a lot of ways.
• 17:27
  SPEAKER_00: But the car industry never went belly up.
• 17:30
  SPEAKER_00: Yeah, we don't this isn't got belly up either.
• 17:33
  SPEAKER_00: So maybe it'll just evolve the same way.
• 17:36
  SPEAKER_01: Uh but coming back to my original question, there are hundreds and thousands of inference providers.
• 17:40
  SPEAKER_01: Oh, are there?
• 17:41
  SPEAKER_01: Okay, I uh but I know of half a dozen, maybe.
• 17:44
  SPEAKER_01: Yeah.
• 17:45
  SPEAKER_01: They're now coming with specialized inference, voice specialized, image specialized, video specialized, specialization for every use case.
• 17:52
  SPEAKER_00: Okay, I think they will all get consolidated.
• 17:54
  SPEAKER_00: Uh uh that's how the industry has always operated, right?
• 17:59
  SPEAKER_00: This all tick search.
• 18:02
  SPEAKER_00: There were many search providers.
• 18:04
  SPEAKER_00: Yeah.
• 18:05
  SPEAKER_00: Uh and then Google came and then it all got consolidated.
• 18:09
  SPEAKER_00: And practically now there are maybe two.
• 18:11
  SPEAKER_00: Yeah.
• 18:12
  SPEAKER_00: Google and maybe Bing.
• 18:13
  SPEAKER_02: Yeah.
• 18:15
  SPEAKER_00: That was consolidated.
• 18:16
  SPEAKER_00: Look at social media.
• 18:17
  SPEAKER_01: A lot of social.
• 18:18
  SPEAKER_01: But those are consumer problems, those are not enterprise problems.
• 18:21
  SPEAKER_01: Inference is not serving to consumers, inference is serving to B2B.
• 18:26
  SPEAKER_00: Uh well, yes and no, in the sense that uh at the end of the day, it is a business.
• 18:35
  SPEAKER_02: Yeah.
• 18:35
  SPEAKER_00: Who you serve to is immaterial to me, right?
• 18:38
  SPEAKER_00: A business is a business.
• 18:39
  SPEAKER_00: Whether you're you might be, but this whole concept of consolidation is not new.
• 18:44
  SPEAKER_00: So I think all those guys will get rolled up into someone, and over time you might have a handful of inference providers providing all these capabilities.
• 18:53
  SPEAKER_00: And maybe all these guys are coming up now because they see a value for a quick turnaround and an exit.
• 18:58
  SPEAKER_00: You know what I mean?
• 18:59
  SPEAKER_00: So that could also be the way it will play out.
• 19:02
  SPEAKER_01: So you have said the maximum value uh that is to be captured is in the layer that is managing your agents.
• 19:12
  SPEAKER_00: That's selfishly, I would want it to be that way.
• 19:17
  unknown: Okay.
• 19:18
  SPEAKER_00: And uh logically, I think we can make a case for that.
• 19:20
  SPEAKER_01: But but let's say if you are not running that comp the company in the same space, would you still have the same view?
• 19:26
  SPEAKER_00: Maybe I wouldn't have even thought about it.
• 19:28
  SPEAKER_00: So uh that's uh that that it uh I wouldn't be able to answer the question because you don't think about things that you're not involved in.
• 19:36
  SPEAKER_00: But uh I definitely would not have.
• 19:37
  SPEAKER_00: Uh but I think just it doesn't have to be necessarily AI, right?
• 19:42
  SPEAKER_00: I mean, these are problems.
• 19:44
  SPEAKER_00: If as agents explode, the problems that people have been talking about, things like multi-cloud, being able to protect yourself from regional cloud outages, those things also get exacerbated.
• 19:58
  SPEAKER_00: So you will be forced to solve those problems.
• 20:01
  SPEAKER_00: And hence the focus will be more on those kinds of infrastructures.
• 20:06
  SPEAKER_00: That's probably the case that I'm I think that's the case I should be making more than anything else.
• 20:11
  SPEAKER_00: Uh it would be it would absolutely be needed, and that's why I say that everything will gravitate towards.
• 20:18
  SPEAKER_00: Now, would that get commoditized over time?
• 20:20
  SPEAKER_00: Would that get consolidated?
• 20:22
  SPEAKER_00: Obviously.
• 20:22
  SPEAKER_00: That's the way the business is always kind of played out.
• 20:26
  SPEAKER_01: So coming back to the, I think our listeners are pretty much interested in the AI economy, uh right, where the value gets accrued.
• 20:34
  SPEAKER_01: All right.
• 20:34
  SPEAKER_01: So a lot of investments are happening today in the energy space powering AI, in the data center space, uh, right.
• 20:44
  SPEAKER_01: And uh uh so uh do you still think like we are short of AI infrastructure globally?
• 20:51
  SPEAKER_01: And this the the shortage, uh let's keep the the financial bubble aside.
• 20:58
  SPEAKER_01: Uh we are still short of AI infrastructure for the next 10 years.
• 21:02
  SPEAKER_00: It depends on what you mean by infrastructure.
• 21:04
  SPEAKER_00: Yeah, but since you mentioned power, uh there is a very dangerous trend that I see with AI where everyone's moving agentic and they're building agents.
• 21:17
  SPEAKER_00: Guess in what language?
• 21:19
  SPEAKER_02: Yeah.
• 21:20
  SPEAKER_00: Python.
• 21:20
  SPEAKER_02: Yeah.
• 21:21
  SPEAKER_00: In my opinion, that language needs to be debreeded from the data center.
• 21:25
  SPEAKER_02: Okay.
• 21:26
  SPEAKER_00: It's the most inefficient language ever made by mankind.
• 21:31
  SPEAKER_00: The last 30 years, forget AI, last 30 years, yeah, the industry has done a lot of work in driving efficiency in the data center.
• 21:42
  SPEAKER_00: How?
• 21:44
  SPEAKER_00: As the number of cores in the servers increased, they started thinking in terms of how many more VMs can I pack within a server?
• 21:52
  SPEAKER_00: How many more containers can I pack within the server?
• 21:55
  SPEAKER_00: How much juice per core can I extract so that My footprint goes low, so that my data center costs go low, my power costs go low, cooling costs go low, etc.
• 22:06
  SPEAKER_00: etc.
• 22:07
  SPEAKER_00: So a lot of work happened in that manner.
• 22:10
  SPEAKER_00: And here we are with AI.
• 22:12
  SPEAKER_00: We are claiming that this is the next wave of how applications should be built.
• 22:17
  SPEAKER_00: We are also claiming that the number of agents will be exponentially more than the number of applications we have today.
• 22:23
  SPEAKER_00: More than the number of humans, also.
• 22:24
  SPEAKER_00: Exactly.
• 22:25
  SPEAKER_00: But what language did you choose?
• 22:28
  SPEAKER_00: The most inefficient language.
• 22:30
  SPEAKER_00: Now, why?
• 22:32
  SPEAKER_00: It is, in fact, Python is even worse than JavaScript.
• 22:39
  SPEAKER_00: Why do you say so?
• 22:41
  SPEAKER_00: It's an interpreted language.
• 22:43
  SPEAKER_00: So there are studies that show that if you take a non-trivial agent, okay, the power utilization of that agent written in Python is about 15 watts.
• 22:56
  SPEAKER_00: Okay.
• 22:58
  SPEAKER_00: The equivalent agent implemented in say Rust consumes 0.7 watts.
• 23:07
  SPEAKER_00: How the world doesn't know that?
• 23:09
  SPEAKER_00: There are studies out there.
• 23:11
  SPEAKER_00: And why are still people building in Python then?
• 23:13
  SPEAKER_00: The trend is shifting now, but it's shifting slowly.
• 23:16
  SPEAKER_00: So I'll get to that.
• 23:18
  SPEAKER_00: So think about these numbers.
• 23:21
  SPEAKER_00: 15 watts, 0.7 watts.
• 23:24
  SPEAKER_00: Now, just to make the computations easy, the back of the envelope computations, let's make 1510, and let's make this one watt.
• 23:33
  SPEAKER_00: Which means I'm at one tenth.
• 23:35
  SPEAKER_00: Okay.
• 23:36
  SPEAKER_00: It doesn't seem very large if it is one agent I'm talking about.
• 23:41
  SPEAKER_00: Now scale that up to 10 million agents.
• 23:44
  SPEAKER_00: At 10 watts, you're looking at 100 megawatts as opposed to 10 megawatts.
• 23:52
  SPEAKER_00: Let me put it in perspective.
• 23:54
  SPEAKER_00: If you take a forget the hyperscalers, okay.
• 24:00
  SPEAKER_00: Take medium-sized data centers.
• 24:03
  SPEAKER_00: They typically operate at 100 to 150 megawatts.
• 24:08
  SPEAKER_00: Okay.
• 24:09
  SPEAKER_00: Say 100 megawatts.
• 24:10
  SPEAKER_00: Okay.
• 24:11
  SPEAKER_00: That's their total draw from the grid.
• 24:14
  SPEAKER_00: Now there is there is something called PUE, but basically what it is, it's a ratio between the total draw from the grid and what you need to power your computing.
• 24:23
  unknown: Okay?
• 24:24
  SPEAKER_00: Now if you have a medium-sized data center which is drawing 100 megawatts from the from the grid, and if they have a need, if they have to run 10 million agents, they're all written in Python, the entire grid is used to power just your agents.
• 24:39
  SPEAKER_00: It's impossible.
• 24:40
  SPEAKER_00: Now what do you what will they have to do?
• 24:41
  SPEAKER_00: They have to spin up another data center.
• 24:44
  SPEAKER_00: But imagine if you did that in a different language, like Rust.
• 24:47
  SPEAKER_00: You're going to use only 10 megawatts and you're good.
• 24:50
  SPEAKER_00: Power usage effectiveness.
• 24:52
  SPEAKER_00: That's what it stands for.
• 24:53
  SPEAKER_00: So now, which basically so apparently for uh for the hyperscalers, which are very efficient, the PUE is 1.1, 1.2.
• 25:01
  SPEAKER_00: Okay.
• 25:02
  SPEAKER_00: Even if it is one, one is to one, let's say.
• 25:04
  SPEAKER_00: Yeah.
• 25:04
  SPEAKER_00: If it is one, that means whatever your draw is, you're using all that.
• 25:07
  SPEAKER_00: Yeah.
• 25:08
  SPEAKER_00: But uh it's about 1.5 for other data center vendors.
• 25:12
  SPEAKER_00: So which means that uh if you're using if your draw is 100 megawatts, what you're using is about 67 megawatts.
• 25:21
  SPEAKER_00: Yeah.
• 25:22
  SPEAKER_00: You can't even run this.
• 25:23
  SPEAKER_00: But if you change the language, you can.
• 25:27
  SPEAKER_00: So that's so when you talk about infrastructure and power, we already know what the LLMs are doing with the inference engines, right?
• 25:35
  SPEAKER_00: But you've made it worse.
• 25:36
  SPEAKER_00: If you you know that that cannot be reduced.
• 25:39
  SPEAKER_00: Maybe it can, I don't know how, but let's not go there.
• 25:43
  SPEAKER_00: Let's take that as the invariant.
• 25:45
  SPEAKER_00: You should start driving efficiency in the above layers.
• 25:48
  SPEAKER_00: But we have gone crazy by taking this approach.
• 25:51
  SPEAKER_00: And that's where I think we come in too.
• 25:55
  SPEAKER_00: And we, if you use our SVKs to build agents, you build them using Rust.
• 26:00
  SPEAKER_00: And so right off the bat, it's not just power.
• 26:03
  SPEAKER_00: Power has direct correlation with water consumption.
• 26:06
  SPEAKER_00: I think there are metrics for this.
• 26:07
  SPEAKER_00: I think for every uh one kilowatt hour, you use two gallons of water.
• 26:16
  SPEAKER_00: So now imagine if you scale it up, one kilowatt hour is one kilowatt use per hour.
• 26:24
  SPEAKER_00: So if it is 100 megawatt hours of energy consumed, you can emit you can compute how many gallons of water you need.
• 26:32
  SPEAKER_00: That has direct correlation to carbon footprint, uh, how much uh carbon dioxide emissions are on a per kilowatt.
• 26:40
  SPEAKER_00: These are all numbers well studied.
• 26:42
  SPEAKER_00: So if you're going to make all this inefficient without any, and also the whole concept of agents being able to do anything and out of whack, and then there's no guardrails, nothing.
• 26:56
  SPEAKER_00: First, you need to sandbox that too.
• 26:59
  SPEAKER_00: What is the perfect sandbox?
• 27:00
  SPEAKER_00: One of the best sandboxes that are available today is WebAssembly.
• 27:04
  SPEAKER_00: So if you run your agents as WASM-compatible modules, you not only get these guardrails in place, and if you can implement them in a language that keeps them all, keeps that all nice and tight from a power utilization perspective.
• 27:21
  SPEAKER_00: Those are the kind of things that I think you should start focusing on from an AI infrastructure layer.
• 27:25
  SPEAKER_00: The shift I see is already happening.
• 27:27
  SPEAKER_00: People are, there's a lot of people now talking about implementing agents in Rust, but a lot needs to be done.
• 27:33
  SPEAKER_00: And I think we are way ahead of the game in terms of providing an infrastructure where we can provide all that.
• 27:38
  SPEAKER_00: Because when I described the wheel chain for you initially, I said globally distributed cryptographically secure compute platform.
• 27:46
  SPEAKER_00: And you deploy what we call applets on them, and they execute in a sandbox environment on that platform.
• 27:52
  SPEAKER_00: So we provide you all these capabilities today.
• 27:55
  SPEAKER_00: So coming on us, you're extremely future-proofed.
• 27:58
  SPEAKER_00: That's the that should be the biggest takeaway.
• 28:00
  SPEAKER_00: And I think coming back to AI infrastructure, this is where the innovation will start shifting towards, in my opinion.
• 28:09
  SPEAKER_00: I could be wrong, but I'm willing to stand corrected.
• 28:13
  SPEAKER_01: Avinash, you have built two formidable companies, right?
• 28:16
  SPEAKER_01: Uh the first company has been in storage space, and the second company in AI governance space, uh, right?
• 28:24
  SPEAKER_01: But before that, uh, you know, you're a legend in yourself.
• 28:27
  SPEAKER_01: You worked, you created Dynamo DB.
• 28:29
  SPEAKER_01: Uh Dynamo.
• 28:30
  SPEAKER_01: Dynamo.
• 28:31
  SPEAKER_00: One of the guys.
• 28:32
  SPEAKER_01: Yeah, yeah.
• 28:32
  SPEAKER_01: You created Cassandra.
• 28:34
  SPEAKER_01: Uh, right.
• 28:35
  SPEAKER_01: What are the some of the other creations on your which are well known, uh, and the thought process behind it before you started your first company?
• 28:44
  SPEAKER_00: Well, my son is one of my creations.
• 28:46
  SPEAKER_00: Uh but uh no, technologically speaking, uh, one correction.
• 28:50
  SPEAKER_00: Umliptic is not just AI governance, but of course, we'll talk about it more.
• 28:58
  SPEAKER_00: No, I think these uh four that you have captured has aptly captured what I would actually stress on when I look back at what I've done.
• 29:07
  SPEAKER_00: Uh and I always like to do something different.
• 29:11
  SPEAKER_00: Like I don't like if I do something now for a period of time, I don't want the next thing that I do to be an extension of whatever I did before.
• 29:21
  SPEAKER_00: Lessons learned, I would take, absolutely, but uh otherwise it gets too boring.
• 29:26
  SPEAKER_02: Yeah.
• 29:26
  SPEAKER_00: So if you look at the systems, this is a fourth real large system that I'm I'm actually building right now.
• 29:33
  SPEAKER_00: Dynamo solved a particular class of problems.
• 29:38
  SPEAKER_00: Then I did Cassandra, which solved an even greater set of problems in my mind.
• 29:43
  SPEAKER_00: And then I did Hedwig.
• 29:45
  SPEAKER_00: Hedwig, you know, the software, in my opinion, was far more complex and far more intricate than what I ever did in the previous two.
• 29:55
  SPEAKER_00: And one way to kind of see that is Cassandra was a Cassandra-like system, it was part of the stack of the entire Hedwig stack.
• 30:06
  SPEAKER_00: So you can that gives that should give you a pulse for how big the Hedwig stack was.
• 30:11
  SPEAKER_00: And now what I'm doing is completely different.
• 30:13
  SPEAKER_00: It's uh what we're trying to accomplish is build the operating system for bringing together autonomous AI and tokenized economies.
• 30:20
  SPEAKER_00: So it's it's a very different space.
• 30:22
  SPEAKER_00: There's a lot of new things that I had to learn.
• 30:24
  SPEAKER_00: Of course, you always take the lessons that you learn from whatever you built in the past, but uh it's a very it solves a completely different setup.
• 30:31
  SPEAKER_01: Yeah, so I would love to learn, you know, first the three years of journey of building veloptic.
• 30:35
  SPEAKER_00: No, great.
• 30:36
  SPEAKER_00: Um see what what drew me to what we are doing right now, I started studying, I I I read something about what is called threshold signatures in cryptography.
• 30:51
  SPEAKER_00: And uh it was pretty fascinating.
• 30:54
  SPEAKER_00: And I I have always had a deep mathematical background.
• 30:58
  SPEAKER_00: Uh when I say mathematics, I'm not talking about uh trigonometry calculus, I'm talking about abstract, pure mathematics.
• 31:05
  SPEAKER_00: So a lot of abstract algebra, topology, those are things I've done.
• 31:10
  SPEAKER_00: But let's be clear, I'm a failed mathematician.
• 31:13
  SPEAKER_00: I'm not like a trained career mathematician, but I understand those conceptually reasonably well.
• 31:22
  SPEAKER_00: And so when I started reading what was happening with the cryptography, I found something really powerful could be done if you marry that with the systems knowledge that I had.
• 31:34
  SPEAKER_00: And that's how Valiptic actually came into existence.
• 31:39
  SPEAKER_00: I start, I also looked at the world of blockchains.
• 31:43
  SPEAKER_00: And if you actually look at blockchains, a blockchain is a monolith in my mind.
• 31:48
  SPEAKER_00: Yes, it's a large distributed system, but it's largely a monolith.
• 31:52
  SPEAKER_00: And what I mean when I say that is when you disseminate state, as your network grows, your throughput should decrease and your latency should increase because the system is growing.
• 32:06
  SPEAKER_00: How do you address that problem?
• 32:07
  SPEAKER_00: How do you contain that?
• 32:09
  SPEAKER_00: That was an interesting problem for me to look at, and that's how everything came to be.
• 32:16
  SPEAKER_00: One thing led to the other.
• 32:18
  SPEAKER_00: AI became like the top of the town when Chat GPT was announced.
• 32:24
  SPEAKER_00: And over time we looked at agentic AI and things like that, and we felt like whatever we were building was a very natural fit in that ecosystem.
• 32:35
  SPEAKER_00: So that's how we are where we are.
• 32:37
  SPEAKER_01: Got it.
• 32:38
  SPEAKER_01: And what are some other problems that you have solved in the liptick?
• 32:41
  SPEAKER_00: Very good question.
• 32:42
  SPEAKER_00: So, see, it goes back to what we were trying to do with uh marrying cryptography and systems.
• 32:49
  SPEAKER_00: And the way you solve that monolith problem is by partitioning.
• 32:53
  SPEAKER_00: Okay, so the first thing we did, uh and and also the other problem that exists in kind of traditional blockchains is there's always one entity that is generating a block, which is again very slow.
• 33:08
  SPEAKER_00: But there's been a lot of research and a lot of work that has happened in that space, uh, you know, with DAC-based consensus and things like that, where in a round everyone generates a block.
• 33:19
  SPEAKER_00: So by design, that throughput is unparalleled when compared to traditional blockchains.
• 33:26
  SPEAKER_00: So we built a system, and so the foundational offering that we have is what we call the whale chain.
• 33:32
  SPEAKER_00: And what you the way you want to think of the whale chain is it's a globally distributed, cryptographically secure compute platform.
• 33:40
  SPEAKER_00: Okay, and it manifests itself as a collection of what we call whale pods.
• 33:45
  SPEAKER_00: Now, these whale pods can be globally distributed, we can spin them up at will.
• 33:51
  SPEAKER_00: And with that, we solve a problem that we saw even in my previous company with you know, the current geopolitical environment, data sovereignty, compute sovereignty becomes very critical.
• 34:04
  SPEAKER_00: So you can spin up these whale pods in a manner where you can very intrinsically solve the data sovereignty and compute sovereignty problem.
• 34:11
  SPEAKER_00: Like if you if you tell me you're a customer in, say, Singapore, and you tell me I don't want my data or compute to ever leave the territory that is under Singapore jurisdiction.
• 34:24
  SPEAKER_00: It's very easy for me to do that.
• 34:25
  SPEAKER_00: I'll spin up a part in Singapore and I'm done.
• 34:27
  SPEAKER_00: But I'll manage everything in a way where it looks like one large globally distributed platform.
• 34:33
  SPEAKER_00: So once you do that, there's a lot of problems you solve.
• 34:37
  SPEAKER_00: And you solve the throughput problem, you solve the data sovereignty problem, uh, and because it's cryptographically sound, you solve the uh, you know, it's all based on Byzantine fault tolerance.
• 34:51
  SPEAKER_00: So you you're inherently solving problems that were plaguing due to uh cybersecurity and things like that.
• 35:00
  SPEAKER_00: So that's where all these things fall into place.
• 35:05
  SPEAKER_01: And and Enveloptec, do you have any current customers, pilots?
• 35:10
  SPEAKER_00: Yeah, yeah, we have a lot of design partners and a lot of pilots and customers uh at the moment.
• 35:14
  SPEAKER_00: Uh but you know, we are still a very small company.
• 35:17
  SPEAKER_00: So there's a lot of interest, but we got to balance the interest with what we have right now.
• 35:23
  SPEAKER_00: Uh and I think we are at a point where we need to start expanding more.
• 35:27
  SPEAKER_01: But when you talk about data sovereignty and compute sovereignty, you are not providing the data infra and the compute.
• 35:35
  SPEAKER_00: Uh not necessarily.
• 35:36
  SPEAKER_00: Yes, that's a good that's a good point.
• 35:38
  SPEAKER_00: So think about it this way.
• 35:40
  SPEAKER_00: Uh let's say we are within the US, and I want to provide data sovereignty for US West Coast.
• 35:49
  SPEAKER_00: Yeah.
• 35:49
  SPEAKER_00: Meaning there are companies in the West Coast that they want their data and their compute only to be on the US West Coast.
• 35:55
  SPEAKER_00: Now, what will I do?
• 35:56
  SPEAKER_00: I would spin up a whale pod that will I could do that just by spinning it up inside US West uh AWS West Coast region, for instance, right?
• 36:07
  SPEAKER_00: Then you're guaranteed that your compute and your data will be within AWS West Coast.
• 36:14
  SPEAKER_00: Can we agree on that?
• 36:16
  SPEAKER_00: Now, if you're in the US East Coast, I could do the same and I could say, okay, I'll bring it up on GCP East Coast, and I'm guaranteeing you that everything will be contained within that region.
• 36:28
  SPEAKER_00: Now, could we declare victory there?
• 36:30
  SPEAKER_00: I don't think so, because if you have a regional cloud outage, then all the applications that you have deployed on this pod that was in that West Coast region will be unavailable.
• 36:44
  SPEAKER_00: So, how do you solve that problem?
• 36:46
  SPEAKER_00: What if you could have your pod kind of be an overlay on top of AWS West, GCP West, and Azure West?
• 36:55
  SPEAKER_00: Now you can even withstand regional cloud outages.
• 36:59
  SPEAKER_00: So that's the way we tend to think, and that's the way the systems evolved.
• 37:04
  SPEAKER_00: And that's how we guarantee data and compute sovereignty.
• 37:07
  SPEAKER_00: So I could do the same thing in India.
• 37:08
  SPEAKER_00: Like in India, you have AWS is in India, GCP is in India, Azure is in India, they have two, three regions.
• 37:15
  SPEAKER_00: So we could spin up something across those regions.
• 37:18
  SPEAKER_00: And it's not just about the cloud.
• 37:20
  SPEAKER_00: Any data center provider can also be part of this ecosystem.
• 37:23
  SPEAKER_00: So we could have a pod that spans AWS, GCP, Azure, some private data center, like in India, it's Control S, right?
• 37:32
  SPEAKER_00: It could be Control S or uh any such provider, any Equinix data center can be pulled into this mix so you can get a very good spread.
• 37:40
  SPEAKER_00: So it truly addresses multi-cloud and coupled with that the data and compute sovereignty.
• 37:47
  SPEAKER_01: And are any governments using the solution today?
• 37:50
  SPEAKER_00: Uh not yet, but that's where we want to go.
• 37:53
  SPEAKER_00: But I think we are pretty close to getting some things lined up on that on that front.
• 37:57
  SPEAKER_01: Got and uh uh uh is this a solved problem by any other provider?
• 38:04
  SPEAKER_00: I don't want to say no.
• 38:05
  SPEAKER_00: Uh the way I would position it is in the following manner.
• 38:09
  SPEAKER_00: And since you brought up blockchain, let me draw the parallel between the two.
• 38:13
  SPEAKER_00: See, this is just for illustration purposes.
• 38:19
  SPEAKER_00: If you look at something like Ethereum or Solana, you get full insight into what are all the transactions that are taking place.
• 38:26
  SPEAKER_00: How do you do that?
• 38:27
  SPEAKER_00: Through the block explorer, right?
• 38:30
  SPEAKER_00: So you could now think of that as an audit log, so to speak.
• 38:34
  SPEAKER_00: It's a it basically shows you what is a ledger, right?
• 38:37
  SPEAKER_00: You can see who did what when at any given instant of time, right?
• 38:42
  SPEAKER_00: Of course, they do it in a manner where you mask it, but be that as it may, you can still tell who is who who did what when.
• 38:49
  SPEAKER_00: Now, in the world of AI, that's largely unsolved.
• 38:54
  SPEAKER_00: If you if you are in an enterprise, for instance, or if you're the government and your employees are using Chat GPT or Cloud or Gemini or Copilot, or if you're in India and if you're using Sarvam AI, right, then it's completely shadow AI.
• 39:12
  SPEAKER_00: You have no insight what they are looking for, you have no insight what information is going out to the LLM.
• 39:19
  SPEAKER_00: Could it be proprietary information?
• 39:21
  SPEAKER_00: You have no, absolutely no clue.
• 39:23
  SPEAKER_00: Unfortunately, the audit log is the chat history, but you can always delete your chat history and it's gone.
• 39:30
  SPEAKER_00: What if you could provide a ledger for that, right?
• 39:33
  SPEAKER_00: Where, and when I when I say ledger, I just use it loosely, you don't have to immediately think it is Ethereum or Solana.
• 39:40
  SPEAKER_00: But if I can provide an immutable trail where you can always tell Siddharth did this search on ChatGPT at this instant of time.
• 39:53
  SPEAKER_00: Now it's not just, and people could use different tools.
• 39:56
  SPEAKER_00: Like you might use ChatGPT, someone else might use Gemini.
• 39:59
  SPEAKER_00: Yeah.
• 39:59
  SPEAKER_00: But I still need to know who did what when.
• 40:01
  SPEAKER_00: So if I have to lock it down, I need to have the power to lock it down.
• 40:05
  SPEAKER_00: That visibility is completely absent.
• 40:08
  SPEAKER_00: Now, there are people who I believe solve slivers of these problems, but no one in my mind does it as comprehensively as we do, based on the research I have done.
• 40:19
  SPEAKER_00: As always, I'm willing to stand corrected.
• 40:22
  SPEAKER_01: So I want to touch upon the journey of Hedwig.
• 40:25
  SPEAKER_01: Like, how did you come up with Hedwig?
• 40:26
  SPEAKER_01: And secondly, if you can share about the growth, you mentioned that companies like Airbus were one of the largest customers uh in Hedwig, and how did it get acquired?
• 40:37
  SPEAKER_00: Yeah, absolutely.
• 40:38
  SPEAKER_00: So Hedwig was uh, you know, post uh my Facebook stent.
• 40:45
  SPEAKER_00: And you know, uh if you're here, I was young, and if you're here, everyone wants to start something.
• 40:52
  SPEAKER_00: And I felt like uh with the knowledge that I had garnered having built systems like Dynamo and Cassandra, and you know, when you're at Facebook, we operated the system at scale.
• 41:01
  SPEAKER_00: So and we also did that with Dynamo.
• 41:03
  SPEAKER_00: So you also tend to get a good pulse of how to run these systems at really large scale.
• 41:09
  SPEAKER_00: And uh we had a I had a beautiful idea, I believe, to start Hedwig.
• 41:15
  SPEAKER_00: And at that time it did seem like a problem that was worth solving.
• 41:19
  SPEAKER_00: So I founded Hedwig.
• 41:20
  SPEAKER_00: And uh we built a storage fabric that solved the hybrid multi-cloud problem.
• 41:28
  SPEAKER_00: So it's like invisible infrastructure that applications don't even have to care where things are, but everything seamlessly works.
• 41:37
  SPEAKER_00: And how do you make this work in a manner where existing workflows don't change?
• 41:44
  SPEAKER_00: And the other big problem that I had seen at that time was people were procuring different uh storage artifacts for different workloads.
• 41:55
  SPEAKER_00: For block, they were going with one particular vendor, they would go with a totally different vendor for any file-related access and a totally different vendor for object storage like access.
• 42:05
  SPEAKER_00: We felt that all that could be collapsed on one singular platform, and that's what we brought to the to the market.
• 42:10
  SPEAKER_00: So complete operational efficiency could be delivered because you're collapsing multiple different protocols on one singular platform, and uh it manifests itself in a way where applications don't need to change.
• 42:23
  SPEAKER_00: Because if you have to go to every enterprise and tell them to change the application, your system is dead in the water.
• 42:29
  SPEAKER_00: It's not gonna happen.
• 42:30
  SPEAKER_00: So it has to sneak in underneath, which means that it has to plug in into existing protocols.
• 42:35
  SPEAKER_00: And so you have this new protocol that's plugging into an existing protocol and providing this seamless integration.
• 42:42
  SPEAKER_00: So nothing changes, but something beautifully has changed underneath.
• 42:45
  SPEAKER_00: That's what we wanted to bring to the table, and that's what we delivered.
• 42:49
  SPEAKER_00: But along the way, um HPE got interested in what we were doing, they invested in us.
• 42:55
  SPEAKER_01: How much did you raise overall as a company?
• 42:58
  SPEAKER_01: Yeah, it was a long time ago.
• 42:59
  SPEAKER_00: Uh I think total we raised about 52 million.
• 43:03
  SPEAKER_00: So actually, if you know, if you if you go back and look as a storage infrastructure company, we were the least capitalized.
• 43:12
  SPEAKER_00: Um, and I ran a very frugal shop.
• 43:14
  SPEAKER_00: One thing that I did learn was that you know, keep the teams small.
• 43:18
  SPEAKER_00: You don't really need large teams.
• 43:20
  SPEAKER_00: And I think that thinking has now been uh substantiated by the way Elon Musk runs his companies.
• 43:27
  SPEAKER_00: You know, his teams are very small.
• 43:29
  SPEAKER_01: Uh and that's a very maximum size of Hedway?
• 43:33
  SPEAKER_00: Number of engineers, I don't think we were ever more than 15 uh over the lifetime of the company.
• 43:41
  SPEAKER_00: I think even 15 might be pushing it, but let's go with 15 as a number.
• 43:44
  SPEAKER_00: Uh that was the size of the company.
• 43:47
  SPEAKER_00: And uh so total, we raised about 52 million toward throughout the entire lifetime of the company.
• 43:54
  SPEAKER_01: So you did Hedwig from 2000 to 2019?
• 43:57
  SPEAKER_01: Yes.
• 43:58
  SPEAKER_01: In 2019, you got acquired by Common.
• 44:00
  SPEAKER_01: World.
• 44:00
  SPEAKER_00: Yes.
• 44:01
  SPEAKER_01: So so when did the majority of the revenue come in?
• 44:04
  SPEAKER_01: Which years?
• 44:05
  SPEAKER_00: See, these the kind of systems that I work on, it takes at least three to four years to build.
• 44:11
  SPEAKER_00: Um, and you have to keep in mind infrastructure is the bottommost layer of the stack.
• 44:18
  SPEAKER_00: You can't there's a lot of talk you will hear about, hey, you know, we had these design partners right from the beginning.
• 44:24
  SPEAKER_00: No one works from you right from the beginning.
• 44:26
  SPEAKER_00: Because if you make mistakes there at the infrastructure layer, you're done.
• 44:31
  SPEAKER_00: So you need to get the product to a point, find a sweet spot, you know, when you're confident that you could take it to market.
• 44:38
  SPEAKER_00: Uh, take it and put it in the hands of some other customer because you cannot have it break frequently.
• 44:45
  SPEAKER_00: At the application level, it's fine.
• 44:47
  SPEAKER_00: So there's always this dichotomy you will hear about, you know, when is the perfect time to go to market?
• 44:52
  SPEAKER_00: Do you have to wait?
• 44:53
  SPEAKER_00: Do you have to make it perfect?
• 44:54
  SPEAKER_00: It actually depends on the product that you're building.
• 44:57
  SPEAKER_00: There's no blanket answer.
• 44:59
  SPEAKER_00: I believe if it is an application, if it breaks, people will be more tolerant.
• 45:03
  SPEAKER_00: It's fine.
• 45:04
  SPEAKER_00: So you could go to market, you could take it to a customer even before you have hit what you would perceive as a, you know what, I'm comfortable giving it to somebody.
• 45:13
  SPEAKER_00: But with infrastructure, it's slightly harder and it's a longer cycle.
• 45:17
  SPEAKER_00: So we started getting to revenue, I would say, about three and a half years post the founding of the company.
• 45:24
  SPEAKER_00: Yeah.
• 45:25
  SPEAKER_00: It's been over a decade now, so I don't remember the numbers, but we were growing year on year, let me put it that way.
• 45:31
  SPEAKER_00: And uh so we were at a good spot from that perspective.
• 45:35
  SPEAKER_00: And 2019, we got acquired, I think one month prior to the pandemic, causing the entire shutdown globally.
• 45:42
  SPEAKER_02: Yeah.
• 45:43
  SPEAKER_00: Uh so I I felt it was a good timing, and I think even the acquisition, which I'm sure we'll talk about, yeah, uh, they had a need, and we felt that we could fulfill that need, and it seemed to be work out really well.
• 45:56
  SPEAKER_01: And how did the acquisition happen?
• 45:59
  SPEAKER_00: So uh Sanjay, uh Sanjay Mitschandani, who's the current CEO of Comvault, was good friends with uh one of the board members of EdBrick uh from True Ventures because True had invested in another company.
• 46:15
  SPEAKER_00: I forget the name of that company, but I think Sanjay was CEO there, and they made the introduction.
• 46:20
  SPEAKER_00: And Commvault also needed something to compete more aggressively with some of their other newly upcoming competition.
• 46:30
  SPEAKER_00: Like I think it was Rubrik and Cohesity and people of that ilk.
• 46:34
  SPEAKER_00: And we filled up we filled a big uh gap in their portfolio in order to make them compete more effectively with them.
• 46:41
  SPEAKER_00: And that's how the whole thing came to be.
• 46:43
  SPEAKER_00: And that was the reason for the entire acquisition, too.
• 46:46
  SPEAKER_00: So we I got introduced to Sanjay, and I think uh the whole conversation was done, dusted, locked, and loaded in a couple of months.
• 46:55
  SPEAKER_01: I think I had like a couple of meetings with them, and wow it and I think the public numbers really 200 million plus acquisition.
• 47:03
  SPEAKER_00: Yeah, it's a 225 million dollar acquisition.
• 47:05
  SPEAKER_00: Uh because see, the way I look, I think of deals, uh a deal is a good deal when both parties walk away thinking that they could have gotten more.
• 47:21
  SPEAKER_00: Yeah, but that's that's a good deal.
• 47:23
  SPEAKER_00: I'm a great believer in the karmic cycle.
• 47:26
  SPEAKER_00: You know, let's say if you and I were to get into a deal, if I hoodwink you, and if I get a great deal, you realize later that you got the long end of the stick, or the short end of the stick, rather.
• 47:39
  SPEAKER_00: Uh there's a negative cycle that gets created for you mentally.
• 47:44
  SPEAKER_00: I believe that affects me.
• 47:46
  SPEAKER_00: I don't want deals of that nature.
• 47:48
  SPEAKER_00: Yes, financially it could be rewarding, but I think the other negativity that comes with it might overwhelm the financial benefits that you get.
• 47:58
  SPEAKER_00: So a good deal is always where I feel I think I could have gotten more, and you feel that I think I could have gotten it for less.
• 48:05
  SPEAKER_00: Yeah.
• 48:06
  SPEAKER_00: That's a perfect deal.
• 48:07
  SPEAKER_00: Everyone walks away happy.
• 48:09
  SPEAKER_00: Um, and I think I accomplished that.
• 48:12
  SPEAKER_00: Could I have done more?
• 48:13
  SPEAKER_00: Probably yes.
• 48:15
  SPEAKER_00: Um, uh that that was the acquisition was never in my mind.
• 48:18
  SPEAKER_00: That was never a number that I might have felt was a great acquisition, personally for myself.
• 48:26
  SPEAKER_00: But I think it worked out really well.
• 48:28
  SPEAKER_00: Given even the the uh global environment at that time, and I look back, you know, everything was shut down for two to three years.
• 48:35
  SPEAKER_00: Anything could have gone south.
• 48:37
  SPEAKER_00: So always be content.
• 48:40
  SPEAKER_00: That's a lesson I learned.
• 48:41
  SPEAKER_00: I mean, not that I didn't have that in my DNA before, but it just reaffirmed that.
• 48:47
  SPEAKER_00: So that grounded you as an entrepreneur?
• 48:49
  SPEAKER_00: Tick time.
• 48:50
  SPEAKER_00: It's very reaffirming.
• 48:51
  SPEAKER_00: You know, you hear all these, you know, people always give you advice.
• 48:54
  SPEAKER_00: You know, I can give a lot of advice.
• 48:57
  SPEAKER_00: But again, advice, you know, blanket advice doesn't really help anybody.
• 49:01
  SPEAKER_00: Because advice also depends on the DNA of the person you're giving it to.
• 49:06
  SPEAKER_02: Yeah.
• 49:06
  SPEAKER_00: Right?
• 49:07
  SPEAKER_00: How I internalize things and how you internalize things could be very different.
• 49:11
  SPEAKER_00: So you it's the experiences that make you who you are.
• 49:15
  SPEAKER_00: Always listen, internalize, throw away what you think is useless, and take the ones that you think you could really use.
• 49:22
  SPEAKER_00: So the the reaffirming part for me is just be content.
• 49:26
  SPEAKER_01: Yeah.
• 49:26
  SPEAKER_01: So so and I I believe in uh not giving advice.
• 49:30
  SPEAKER_01: I think advice don't work for a simple reason, unless somebody has experienced what I have experienced, uh they'll never get my perspective.
• 49:39
  SPEAKER_01: So let's say if I tell some of my founders to move to the US, it won't work.
• 49:44
  SPEAKER_01: So rather, what I try to do is I set up 30 to 50 meetings for them in Bayeria.
• 49:49
  SPEAKER_01: So let them experience that and and decide for themselves whether for them is the right choice or not.
• 49:55
  SPEAKER_00: Got it.
• 49:56
  SPEAKER_00: No, it makes sense.
• 49:57
  SPEAKER_00: That's I think you probably articulated it better than I did.
• 50:00
  SPEAKER_00: Uh that's exactly where I was going with it.
• 50:02
  SPEAKER_00: So when people, you know, any advice you get, I mean, it's you know, things like if you waited too long for perfection of your product, then it's too late.
• 50:14
  SPEAKER_00: And no, it depends on the context and which product you're building.
• 50:18
  SPEAKER_00: If you're building a website, probably yes, because the the success of the website depends on the feedback and the user experience.
• 50:26
  SPEAKER_00: But if you're building infrastructure, it's very different.
• 50:29
  SPEAKER_00: So you gotta put things in perspective.
• 50:32
  SPEAKER_01: Yeah.
• 50:33
  SPEAKER_01: My experience has been, you know, at NEON we have invested in 71 companies roughly.
• 50:38
  SPEAKER_01: Uh, eight shutdowns till now, and six good exits.
• 50:42
  SPEAKER_01: I think and most of the exits happen in infrastructure space.
• 50:45
  SPEAKER_00: Got it, got it.
• 50:46
  SPEAKER_00: So infrastructure is hard, but you know it is hard, nobody cares, but everybody needs.
• 50:56
  SPEAKER_02: Yeah.
• 50:57
  SPEAKER_00: So if you are uh see in business, you need to find what is needed, you know.
• 51:05
  SPEAKER_00: So that's why I guess infrastructure always wins.
• 51:07
  SPEAKER_00: It's hard, but everybody needs it.
• 51:10
  SPEAKER_01: Yeah.
• 51:11
  SPEAKER_01: And if you can share, you know, what what rough era you were at when you got acquired at that point in time.
• 51:19
  SPEAKER_00: With Hedwig?
• 51:20
  SPEAKER_00: Yeah.
• 51:20
  SPEAKER_00: Jeez, I think it was double your double digits for sure.
• 51:24
  SPEAKER_02: Yeah.
• 51:24
  SPEAKER_00: I think the if I remember, uh again, don't quote me on the numbers.
• 51:28
  SPEAKER_00: Uh, I mean, if I'm wrong, I'm wrong.
• 51:30
  SPEAKER_02: Yeah.
• 51:31
  SPEAKER_00: Okay, that's fine.
• 51:32
  SPEAKER_00: I think the first year of revenue we went to about 2 million ARR.
• 51:37
  SPEAKER_00: Then the second year of revenue, I think we went close to anywhere between six and eight.
• 51:44
  SPEAKER_00: And then the third year, I think we went to 15, 15 plus.
• 51:49
  SPEAKER_00: So we were on a good trajectory.
• 51:50
  SPEAKER_00: And we had a very small team, by the way.
• 51:52
  SPEAKER_00: I had maybe two sales guys and uh two, maybe three.
• 51:58
  SPEAKER_00: And we had HPE as a partner who was pushing us aggressively.
• 52:02
  SPEAKER_00: So I thought that was a pretty good number.
• 52:05
  SPEAKER_00: Uh, we were on the cusp of breaking past that, but then the acquisition came and we had to look at things more holistically.
• 52:12
  SPEAKER_00: And to give you some numbers, if you actually you can do this research too.
• 52:16
  SPEAKER_00: We got acquired in 2019, and I think we were probably the biggest or the only acquisition that Commvault had made.
• 52:28
  SPEAKER_00: They had made one other in the past, but it was, I don't think it was anything worth talking about from their by their own volume.
• 52:36
  SPEAKER_00: So we were the first largest uh acquisition that Comvault made.
• 52:40
  SPEAKER_00: When we got acquired, Comvault had a market cap of about 1.8 billion.
• 52:45
  SPEAKER_00: Um this was 2019.
• 52:48
  SPEAKER_00: So these are numbers one can easily verify.
• 52:51
  SPEAKER_00: I had looked at things in 2024, so it's a five-year period, and you look back, and their market cap more than quadrupled.
• 53:01
  SPEAKER_01: So right now their market cap is four, four billion.
• 53:05
  SPEAKER_00: Maybe about five, I think 2024 they have touched up to eight billion.
• 53:08
  SPEAKER_00: Okay.
• 53:08
  SPEAKER_00: Right.
• 53:08
  SPEAKER_00: So I would like to believe we had something to do with it.
• 53:12
  SPEAKER_01: I uh I think in in Hedwig, one thing that it seems uh attractive from outside is that the revenue could be very large per customer.
• 53:23
  SPEAKER_01: In infrastructure, it it usually is like if it's a Yes, because storage is a very very strange beast.
• 53:31
  SPEAKER_00: It depends on the capacity that you sell.
• 53:35
  SPEAKER_00: And you know, and storage is a very old industry.
• 53:38
  SPEAKER_00: You know, things are very well understood, the pain points are well understood, the use cases are well understood, and the pricing, whether you like it or not, is well understood.
• 53:49
  SPEAKER_00: It's cost per terabyte, no matter what capabilities you bring to the table, however, however intelligently it's woven, it has to be boiled down to cost per terabyte.
• 54:01
  SPEAKER_00: So the more capacity you sell, the more money you make, and uh larger the enterprise, greater the capacity.
• 54:10
  SPEAKER_00: So that's where I think you get the kick.
• 54:12
  SPEAKER_01: Yeah.
• 54:13
  SPEAKER_01: So so is the current company also positioned in in that way that you you you capture a lot of value like a storage would do.
• 54:23
  SPEAKER_00: Good question.
• 54:23
  SPEAKER_00: I'll give you a couple of insights that we that I had that I personally learned through my Hedwig journey.
• 54:30
  SPEAKER_00: In Hedwig, we were a software company, which basically meant we built software, and let's say you were a potential customer, you want to use our software, we will recommend hardware SKUs for you.
• 54:43
  SPEAKER_00: Yeah, right?
• 54:44
  SPEAKER_00: And then you might come back to me and say, hey, you know what, we are a Dell customer.
• 54:49
  SPEAKER_00: So can you tell me what hardware spec and which class of Dell service I should procure?
• 54:57
  SPEAKER_00: So we would we would know exactly what to recommend to you.
• 55:01
  SPEAKER_00: So what's the first step in this journey?
• 55:03
  SPEAKER_00: You're gonna do a POC before you procure it, right?
• 55:06
  SPEAKER_00: But even for you to do a POC, I need to tell you this is the hardware you need to procure.
• 55:12
  SPEAKER_00: So you go procure it.
• 55:14
  SPEAKER_00: There's a lead up time.
• 55:16
  SPEAKER_00: Typically, we were seeing six weeks to eight weeks for you to just procure the servers, rack and stack them in your data center.
• 55:25
  SPEAKER_00: Okay.
• 55:25
  SPEAKER_00: Once you do that, we come in, we install the software.
• 55:29
  SPEAKER_00: Okay, and then the testing process begins.
• 55:33
  SPEAKER_00: But if your lead up time to procure the hardware is six to eight weeks, you have almost one quarter is gone.
• 55:41
  SPEAKER_00: That means any revenue that I see from you, there's a minimum of a two-quarter lead up, right?
• 55:48
  SPEAKER_00: And one of the reasons for that is because the time taken for you to procure stuff.
• 55:52
  SPEAKER_00: One thing I did was I decided I'm not going to be caught in that model.
• 55:57
  SPEAKER_00: So I'm going to offer this up as a service, so to speak.
• 56:00
  SPEAKER_00: So for me, the cost to enable you or the time taken to enable you, if you want to try us right now, is I'll say, depending on the use case, less than 24 hours max.
• 56:12
  SPEAKER_00: Okay.
• 56:13
  SPEAKER_00: Um, and we'll talk about why I say give such a large number.
• 56:18
  SPEAKER_00: But that was one thing that I learned.
• 56:21
  SPEAKER_00: The cost to get someone just plain onboarded.
• 56:28
  SPEAKER_00: And the second thing is the bigger problem that I a problem that I just did not want to live through again.
• 56:37
  SPEAKER_00: We build software.
• 56:39
  SPEAKER_00: So we are deployed at you, one customer, X, then we go to customer Y.
• 56:44
  SPEAKER_00: Customer Y hits some bug.
• 56:47
  SPEAKER_00: Now, what do I do?
• 56:49
  SPEAKER_00: I have to come back to you.
• 56:51
  SPEAKER_00: I may not necessarily tell you what the bug is, but I I know that I if it happens there, Murphy's Law, it could happen here.
• 57:00
  SPEAKER_00: So I want to preemptively protect you, protect myself, so to speak.
• 57:04
  SPEAKER_00: So I want to push the patch to you.
• 57:07
  SPEAKER_00: Now, we saw enterprises come in and say, hey, you cannot patch willy-nilly whenever you want to.
• 57:13
  SPEAKER_00: We have a cycle, and you have to fall into that cycle.
• 57:17
  SPEAKER_00: Now, what does that do?
• 57:18
  SPEAKER_00: That means that I hope and pray that you don't hit that problem till that particular time.
• 57:23
  SPEAKER_00: And I have to maintain my this is a software engineering nightmare.
• 57:28
  SPEAKER_00: Now I need to maintain so many different versions to be backwards compatible and make sure that every time I build a new feature, do I need to roll it into all these things and make sure that all those work?
• 57:41
  SPEAKER_00: And any change that I make, is it going to be backwards compatible?
• 57:45
  SPEAKER_00: I need to worry about a lot of such problems.
• 57:48
  SPEAKER_00: I did not want to be in that business.
• 57:50
  SPEAKER_00: If I run it as a service, I can patch it whenever I want.
• 57:53
  SPEAKER_00: But this leads to another problem, which is very important and very important to actually think through and solve.
• 58:01
  SPEAKER_00: How can you do software upgrades in a manner where there is absolutely no downtime?
• 58:07
  SPEAKER_00: Right?
• 58:09
  SPEAKER_00: That again goes back to software engineering, the way you build your systems.
• 58:12
  SPEAKER_00: Like if I take one server down, things should continue to work like as if nothing happened.
• 58:18
  SPEAKER_00: I patch that, I bring it back up, then I rinse and repeat that.
• 58:21
  SPEAKER_00: So those kind of capabilities also have to be thought through and built into your system.
• 58:25
  SPEAKER_00: And I decided offering this as a service, if I hit a bug there, I can patch it and everybody just gets it.
• 58:32
  SPEAKER_00: So those are some of the lessons that I learned.
• 58:35
  SPEAKER_00: And that's why I did Wailuptic the way I did it, uh, where everything will be offered up as a managed service, like a snowflake or an S3 or what have you.
• 58:45
  SPEAKER_01: How do you capture the same amount of value?
• 58:48
  SPEAKER_01: Okay.
• 58:49
  SPEAKER_00: So the value now manifests itself as it really depends on what problem you're solving.
• 58:56
  SPEAKER_00: So when it comes to AI, the way in my simplistic explanation of how I see the problems, you want to view it this way.
• 59:06
  SPEAKER_00: You have AI.
• 59:08
  SPEAKER_00: There are two ways in which you can access it through the browser, and then you have a non-browser environment.
• 59:14
  SPEAKER_00: Okay.
• 59:15
  SPEAKER_00: When I say browser, what do I mean?
• 59:17
  SPEAKER_00: Chatbots.
• 59:18
  SPEAKER_00: Yeah.
• 59:18
  SPEAKER_00: Right.
• 59:19
  SPEAKER_00: It could be cloud, it could be Gemini, chat GPD, you could have implemented your own chatbot.
• 59:24
  SPEAKER_00: Now, how does a chatbot matter?
• 59:29
  SPEAKER_00: It matters because you could bring context to the LLM by talking to your proprietary systems.
• 59:36
  SPEAKER_00: How do you talk to your proprietary systems?
• 59:39
  SPEAKER_00: Through MCP.
• 59:40
  SPEAKER_00: Okay.
• 59:41
  SPEAKER_00: So now you have the opportunity to implement this MCP, but now where do you run it?
• 59:49
  SPEAKER_00: You need infrastructure for it.
• 59:52
  SPEAKER_00: Now, when you run this infrastructure, you have to worry about the auto-scaling of it.
• 59:58
  SPEAKER_00: You have to worry about protecting it from cybersecurity issues.
• 1:00:02
  SPEAKER_00: One of the value propositions we bring to the table are don't worry about infrastructure.
• 1:00:10
  SPEAKER_00: Implement the MCP, run it on top of our infrastructure.
• 1:00:14
  SPEAKER_00: All those extraneous problems are orthogonal to the problem you want to solve.
• 1:00:18
  SPEAKER_00: We'll take care of it.
• 1:00:20
  SPEAKER_00: Okay?
• 1:00:20
  SPEAKER_00: So now that's one problem.
• 1:00:23
  SPEAKER_00: Second, in browser-based AI, which we talked about backstage, you have no idea who is asking what question, what systems they're touching, what data is going to the LLM, nothing.
• 1:00:33
  SPEAKER_00: So all that, you need, in my opinion, a ledger to track all that activity.
• 1:00:38
  SPEAKER_00: We can track all that activity seamlessly on our platform.
• 1:00:42
  SPEAKER_00: Now, the moment I do that, my pricing is very simple.
• 1:00:47
  SPEAKER_00: For every audit, I will charge you X units.
• 1:00:51
  SPEAKER_00: Yeah.
• 1:00:51
  SPEAKER_00: Okay.
• 1:00:52
  SPEAKER_00: This is fiat currency, no cryptocurrency.
• 1:00:56
  SPEAKER_00: So to keep calculation simple, I could say for every audit, I'll charge you one cent.
• 1:01:02
  SPEAKER_00: And I can track it at an organizational basis.
• 1:01:06
  SPEAKER_00: It's very easy to do that because if everything is an audit, then all you need to do is to track the number of audits.
• 1:01:13
  SPEAKER_00: You don't have to track anything else.
• 1:01:15
  SPEAKER_00: An organization, there is a counter.
• 1:01:17
  SPEAKER_00: At the end of the month, this month, you did 1000 audits.
• 1:01:21
  SPEAKER_00: Your price, your price for that month is 1,000 cents.
• 1:01:26
  SPEAKER_00: Simple.
• 1:01:27
  SPEAKER_00: We'll worry about the scaling, everything.
• 1:01:29
  SPEAKER_00: Now, everyone is in AI.
• 1:01:32
  SPEAKER_00: There are billions of chatbot users.
• 1:01:34
  SPEAKER_00: How many people, if you go into an enterprise, is larger the enterprise, the more the chatbot uses.
• 1:01:40
  SPEAKER_00: So the same philosophy applies, right?
• 1:01:42
  SPEAKER_00: What capacity was becomes now number of people using chatbots.
• 1:01:46
  SPEAKER_00: So that's a browser environment.
• 1:01:48
  SPEAKER_00: If you go into a non-browser environment, that is, you can break that into two categories: coding agents, non-coding agents.
• 1:01:58
  SPEAKER_00: Okay.
• 1:02:00
  SPEAKER_00: Let's think about coding agents.
• 1:02:03
  SPEAKER_00: Typically, the harness is, I think, the more popular ones are just codecs and cloud code.
• 1:02:08
  SPEAKER_00: Yeah.
• 1:02:08
  SPEAKER_00: Okay.
• 1:02:09
  SPEAKER_00: Forget the LLM that is used because you can use open wing LLMs for that.
• 1:02:13
  SPEAKER_00: But do you have insight into what it is doing?
• 1:02:17
  SPEAKER_00: Absolutely none.
• 1:02:18
  SPEAKER_00: All you get is an end product.
• 1:02:21
  SPEAKER_00: But tomorrow, typically the way things work, you run this product in production, there are any issues.
• 1:02:27
  SPEAKER_00: You always try to narrow it down to the offending git commit.
• 1:02:32
  SPEAKER_00: Right?
• 1:02:32
  SPEAKER_00: That's how engineering works.
• 1:02:34
  SPEAKER_00: And then you look at that git commit to figure out what went wrong.
• 1:02:38
  SPEAKER_00: But now that git commit is not authored by a human, it's authored by an agent.
• 1:02:44
  SPEAKER_00: How does the agent author it?
• 1:02:46
  SPEAKER_00: You would have given it prompts.
• 1:02:47
  SPEAKER_00: Yeah.
• 1:02:48
  SPEAKER_00: But you have no insight.
• 1:02:49
  SPEAKER_00: No one's tracking what prompts you give over time before you committed it.
• 1:02:54
  SPEAKER_00: We track all that.
• 1:02:55
  SPEAKER_00: Now, again, that for us is an audit.
• 1:02:58
  SPEAKER_00: So now then, if you have a user spin up five instances of cloud code, that one becomes five.
• 1:03:07
  SPEAKER_00: So if you are an organization, even with a hundred employees, and if each one spins 10, it's like thousand employees.
• 1:03:15
  SPEAKER_00: So you can see the multiplicative scale applying even there, right?
• 1:03:19
  SPEAKER_00: So we bring that to non-coding agents.
• 1:03:22
  SPEAKER_00: We but to coding agents.
• 1:03:23
  SPEAKER_00: We can bring the same to non-coding agents.
• 1:03:26
  SPEAKER_00: What do I mean by that?
• 1:03:27
  SPEAKER_00: Autonomous agents.
• 1:03:28
  SPEAKER_00: When agents do work, today you have no insight.
• 1:03:32
  SPEAKER_00: Who did what when?
• 1:03:34
  SPEAKER_00: Were they authorized to do it?
• 1:03:36
  SPEAKER_00: Who authorized them?
• 1:03:37
  SPEAKER_00: These are the questions that need answering.
• 1:03:41
  SPEAKER_00: Now, we capture all that.
• 1:03:43
  SPEAKER_00: Again, everything is an audit.
• 1:03:45
  SPEAKER_00: So you can see that already playing itself out.
• 1:03:47
  SPEAKER_00: So if you spawn a thousand agents, you might be a 10-person company, but you have thousand agents.
• 1:03:52
  SPEAKER_00: To us, it's like 1,000 employees.
• 1:03:54
  SPEAKER_00: So your number of audits are going to just increase and increase over time.
• 1:03:58
  SPEAKER_00: And that's what we monetize.
• 1:04:00
  SPEAKER_00: That's the equivalent of the capacity problem that we were talking about in the storage space.
• 1:04:06
  SPEAKER_00: And you had asked me a question earlier.
• 1:04:09
  SPEAKER_00: Do you have any competition?
• 1:04:11
  SPEAKER_00: You know, I spoke about all these audits and all these money.
• 1:04:14
  SPEAKER_00: There are people who do little things.
• 1:04:15
  SPEAKER_00: There are people who do it only for coding agents, there are people who do it for agents, they do a sliver of it.
• 1:04:20
  SPEAKER_00: But no one does it as comprehensively as we do.
• 1:04:23
  SPEAKER_00: Okay.
• 1:04:24
  SPEAKER_00: We provide the infrastructure, which means we take the burden of you thinking about infrastructure off the table.
• 1:04:29
  SPEAKER_00: The mantra is focus on what you need to build.
• 1:04:33
  SPEAKER_00: And just let's worry about everything else.
• 1:04:36
  SPEAKER_00: Everything else is orthogonal to the problem you want to solve.
• 1:04:39
  SPEAKER_00: Let us take care of it.
• 1:04:41
  SPEAKER_00: Drop it on top of our platform.
• 1:04:43
  SPEAKER_00: And then you get complete insights.
• 1:04:45
  SPEAKER_00: You can even control at the granularity of like tool calls or agent calls as to who can make this call at any given instant of time.
• 1:04:53
  SPEAKER_00: And everything will manifest itself as an audit, which we will monetize for you.
• 1:04:58
  SPEAKER_00: That's in a nutshell what we are trying to do.
• 1:05:01
  SPEAKER_01: Thank you so much.
• 1:05:06
  SPEAKER_01: Oh, thank you.
• 1:05:07
  SPEAKER_01: You know, uh and I would love to do part two.
• 1:05:11
  SPEAKER_01: Oh, absolutely.
• 1:05:13
  SPEAKER_01: Absolutely.
• 1:05:14
  SPEAKER_01: You know, but but thank you so much.
• 1:05:15
  SPEAKER_01: It's been such a pleasure and to host you on the podcast.
• 1:05:18
  SPEAKER_02: Thank you for having me.
• 1:05:19
  SPEAKER_02: Thanks.