All right, we're going to switch some gears and talk more about AI here. So, next up, I want to bring on EG, who's the co-founder of Infura, to share his learnings from just building Infura, but also just what does it mean to design for from an AI native world to cryptonative world to an AI native world. So, please give VG a big round of applause. Thanks, Caric. All right.
Hi, everybody. Uh, my name is EG Galan. I'm one of the co-founders of Infura and I'm going to do a talk today that is different. And it's something I haven't done before it, but given how impactful AI has been in the last couple of years, everybody's been factoring it into their product designs, their strategies, their road maps, and coming from the crypto space, it looks different depending on what layer of the stack you sit at. And so being core infrastructure that's been powering blockchain, in particular Ethereum's ecosystem since 2016, it's something that's kind of been like echoing around in my head when I look at the parallels between AI infrastructure and what we went through building infura.
So, uh, that's kind of what I want to share and hopefully it's helpful for those of you that are either building products as you factor it into how you can improve and better serve users and deliver value or maybe it gives you an idea of what you want to build next. Uh so first inference is the new infrastructure like when we look at models whether it's you know some self-hosted model or the big LLM from the labs that's the new infrastructure that's the gateway that's unlocking all of this new innovation that we're seeing a lot of our new products are built um on top of the inference gateways offered by the big labs and it is very reminiscent of early days of Ethereum where when I got into the case, it was about Ethereum as the world computer and what are the things we're going to build with this new uh compute paradigm. So all of the stuff that you just heard about with DeFi and all of these things that now exist, it it wasn't what was there at the time. Everybody was still trying to decide like what they were going to try to build and what could be done with the technology. And AI is going through that same cycle very rapidly.
And with inference, it starts with just delivering access. So you can look at the big LLM gateways as the same kind of gateway that unlocked Ethereum and web 3 for a lot of people early on. And it starts as like base layer infrastructure that eventually becomes more and more accessible and then uh commoditized. So as you deliver a product at that layer, how do you continuously evolve and improve your product and honestly like stay relevant to delivering value to your users? And so the first thing that happens is there's abstractions.
Uh for RPC it looked like performance improvements, things like indexing and offer a higher level API, something that was focused on maybe like a DeFi API or an NFT indexer so that you can get data faster or more performantly and uh with additional metadata that was useful for a specific use case or application. And you start to see similar types of abstractions built on top of AI inference or the the different models are now being trained for specialized purposes in that same type of abstraction layer. Then they have to deal with scale. So this is talking about how Infura dealt with scaling problems and scaling challenges and how that improved us as a product company. You see the the labs going through that right now.
all of the trillions of tokens per second or whatever the current number is that they're handling. When you try to compare one of the big labs models to a self-hosted model for some workloads, it's good enough. For others, it there's like no comparison. And that become that comes from being able to have that scale. You see these scaling problems and you're able to battle test, harden, and improve your infrastructure faster than everybody else.
But eventually they catch up because it's infrastructure and it just takes time or energy or money or traffic and they can catch up. So I'll go through that with how it affected Infura. Rewinding for a second in case you don't know what Infura is. Infura is a blockchain API. We were the first ones to make it so that nobody had to run a node to connect to Ethereum.
Our core value prop was we wanted people to try to do something with Ethereum and it was going to be a lot easier to do that if you didn't even have to wait a minute for a node to download the latest data so that you can send a transaction that resonated pretty quickly with people. Um, even though there was always kind of people that wanted to run their own node or said that it was a privacy concern, like it ultimately solved a problem very early on and people wanted to just get to building. So that whole thing that you always hear about like who's your target user and solve their problem. It found us like we we weren't really looking to do that. Other people came to us and asked for help and we just naturally started to solve that problem.
So by being that gateway and solving that problem, we unlocked a lot of growth in the ecosystem and that infrastructure wave that kind of crests and then leads to um mass adoption either the user side or innovation on the application side went through this cycle for us where any ICO any token sale that happened in 2017 2018 all of those users didn't have to download and run their own Ethereum client. They used typically MetaMask which was connected to Infura and then we had to scale to handle that and deal with all of the challenges that came and then CryptoKitties the uh wildly successful viral app that launched in 2017 uh battle tested not just us but the network itself. And so both Ethereum and its mempool handling for being able to handle tons of transactions on the network improved as a result of that and our systems also had to improve. We basically spent six months rearchitecting how we handled transactions after that. And that enabled millions of users to start using Ethereum with having a better user experience and not no downtime or outages like we were having in maybe late 16 and 17 because we went through those hardships.
And then all of those users now being on boarded and having a good experience helped with DeFi summer. Not taking credit for DeFi summer. and saying that you needed that infrastructure to not be a problem and it wasn't a problem for a lot of people by that point because it'd been four years that we had been running. So every infrastructure wave has people that want to self-host versus use a managed service. And you hear this with every cycle.
So maybe at one point it was your friend saying run your own email server. Somebody's going to just take all of your data and sell ads to you. They were right, but you still did it just like I did it because it wasn't worth the pain. It wasn't worth the hassle. And then you heard the same thing with Ethereum.
They said, "Run your own Ethereum node. Don't trust Infura. I appreciate it if you did. We didn't sell your data. We didn't give you ads.
Um, we ended up doing a traditional SAS service and giving you customer support and all of that. Pretty straightforward. Uh, local models are now the new thing I hear. Don't use the big hosted LOMs. run your own.
You can download Quinn, use the openw weight ones. Uh, privacy is important and it's always going to exist. That's always going to be an option. It's just not something that you can really evangelize. If if somebody sees value in that for themselves, they're going to do it.
But you can't really design your product like with that in mind first. Uh, really just focus on solving your your core problem. Who is that target user and what's the problem you're solving for them? And for us it was uh minimizing the friction to try things out um instantly. This is a tough one that we learned over time.
The the hard lesson here is we started with 100% market share. We were the first ones to have an RPC service and as the adoption of Ethereum grew, our market share actually declined like month over month. Our users grew. I didn't recognize it at the time. we were getting something like over 200 new user registrations a day.
I didn't know if that was good or not. Uh this was my first startup. Uh now that's amazing for tons of uh startups would love to have that kind of metric and even then there were new services starting up that were also getting new users because the ecosystem was growing faster than that. So our market share was declining a bit. Um, and we could see that happening and we started to do things to keep improving our reliability, our quality of service, trying to offer different features.
But at that base layer infrastructure, just like with inference, at some point it's diminishing returns and it's harder to differentiate as a product when you look at one RPC provider or one LLM provider versus another. So, one of the things that I learned was uh we had a four-year head start before that kind of hit in 2020. And the way that we used to really differentiate ourselves was on performance. And at some point, we hit a a moment where the performance difference was um negligible to uh customers when they're discerning between things. And you can't really fight that at a infrastructure layer.
it eventually happens. So you as a product company have to differentiate and for us it was moving up the stack and saying there's other things in value that we can provide and it it isn't just best performance of our PC. These primitives as they get cheaper enable new opportunities and one of them for us is the ability to go multi-chain. the cost of scaling for Ethereum was so high that it kind of kept us locked in even when newer more or newer chains were launching that people were interested in and we just couldn't afford to split our time and resources on it. But as that performance of our systems improved and the costs improved, it allowed us to then go multi-chain around 2020 2021 support optimism, arbitrum, polygon, things like that.
And as all of that stuff gets cheaper, then you look for better margin products. We started to look at doing NFT APIs and higher um higher quality indexing services that we could sell B2B so that people people could build their own indexing um systems and APIs on top of our data and not just doing the uh lower level core RPC anymore. When you look at those access primitives right now for AI, now people are starting to do the abstractions. You see the labs themselves doing them with like cloud code or open claw and these kinds of things. You're starting to see workflow and orchestration stuff built on top of that.
What's coming next in my opinion is the uh outcomes and coordination layer between these things. Maybe one of those things is agent-to- agent payments, which is something I'm super excited about and have been involved in um trying out early versions of like X42 and some of the other things that enable uh agentic payments. But coordination between these systems is where things naturally go as you start looking at your infrastructure product as a developer platform and product. So AI has that same shape at least right now with inference and models as infrastructure and early early on there's a ton of value in just having a easy user experience for people to try building on top of this new paradigm and then it um it means that the people or the products and the teams that are building that core infrastructure are going to look very different at this point maybe six months from now or a year from now not four to six years from When you look at agentic engineering, this is one of my favorite projects right now if you haven't seen it from Gary Tan at YC where he built this agent stack where it's a role for different people in a traditional company where each agent can help build a prototype. And so he'll um use this to build you know a one one person multi- aent startup.
I love this idea because for those with an entrepreneurial mindset that constantly have these ideas that they want to try out, it's a great way to just build a a proof of concept and nerd out on uh using agents. So, I'm building my own right now. This isn't a product pitch. This is just me geeking out for a second of like what does this enable for us? Infura always made it so that nobody else had to run their own uh run their own node infrastructure.
So what if I built my version of that Gstack and call it the Estack? Maybe it means Ethereum stack, maybe it's EG stack. I haven't decided yet, but it has different roles in here like this is effectively our org chart for Infura of a core development team that manages protocol upgrades. uh keeps an eye out for client deficiencies, bugs, performance improvements, and then actually rolls out the update for that. Uh a software engineer that's actually building the optimizations, caching layers, periodically snapshotting the the state so that you can quickly spin up nodes instead of having to wait for however long.
um uh SRRES for uh responding to incidents and then PMs defining the road map. Support engineers um keeping track of any tasks or tickets that need to be addressed. The fun ones are actually that's valuable having a crypto Twitter lurker because the first time that you detect either like an attack or an incident, it's not an alarm that goes off. It's normally somebody complaining on Twitter. And so actually keeping track of crypto Twitter is part of the job as is being a DGEN because that gives you the signals of what new things you could be providing or what um new products you could be creating to solve one of these API infrastructure problems.
So uh keep an eye out. I'll eventually release this. So when everybody can build anything like there's no moat like that right there disrupts Infura itself and allows anybody to build an Infura type system. So that's me not trying to uh build a moat around like the Infura business. But you if you accept that and say okay that's not defensible it's a waste of time trying to defend that.
How do we compete? How do we differentiate? And I've been looking at this for years now because it's been happening to Infura for like four, five, six years. Unique data, proprietary data and insights, um, relationships with existing customer bases. Um, things that are difficult to copy, uh, difficult to replicate.
You can't vibe code those things. And then network effects within crypto. We know a ton about network effects and community value. That's the stuff that really building products around. Like when you look at the products that actually have like adoption, I can build an open claw equivalent right now.
Might not be as good, but it'd be functional, but there's an ecosystem around open claw now. And that ecosystem is building plugins and skills and all of these improvements and constantly helping improve the product that however good I am as an engineer, you can't keep up with that network effect. So build network effects into your products. And that's what we were doing with Infura. So about two three years ago, I started building uh a marketplace model for Infura where we wouldn't be the only ones serving traffic for RPC.
We would build a protocol that would allow anybody to permissionlessly serve that node RPC traffic. And it had a staking model where we were going to have people stake to join this marketplace. And that state guaranteed an SLA so that you could earn rewards for serving the traffic. If you were going to be launching a new layer 2 or layer 1, you didn't have to go to Infer and say, "Hey, how much do we have to pay you to help support our new network, you could just put it up as a bid on this marketplace and say or an ask on this marketplace and say, uh, we're willing to pay this much for somebody to meet these requirements." And that actually resonated with a lot of people and we're still building a version of this right now with the staking aspects of it.
It's too complicated in where we want to scale out. We want to push a lot of new services onto the marketplace. So we're just doing a more traditional staking model which is just deposit a bond to join the marketplace. All right. Um, the other thing I wanted to say about this is with new APIs being included in this marketplace, you hear a lot about agents and agent a to agent payments and agents um, needing to be discovered.
There's the ERC80004 registry as one of the things where people are registering agents for discovery. Discovery is just one of the things people use a marketplace for. Standardized payments is another, but there's also a reputation aspect of it. Uh, one of the things that we built here is if somebody joins as a uh, RPC provider, how do you know if they're actually performant? You don't necessarily need to know they're trustworthy.
Um, you care about if they're going to deliver you the correct data in a verifiable way because it's blockchain data. We can verify it. And actually building tests that measured that is what we did with Den. So we spent like six months building out test suites that were continuously running on any new endpoint and could say the easy stuff is like latency and uptime. And then we started to create deeper checks that were what's the consistency of this system?
If somebody has some sort of off-chain indexing or performance caching layer, is it introducing bugs because sometimes there's stale data? And that solves this for RPC because now we can give a score to everything on this marketplace. When you go to clawhub or any of the skill hubs or marketplaces out there, there's no reliable metrics like that for these AI equivalents. And so that's where we're focusing now is trying to be this coordination layer for marketplaces for skills AI skills APIs broadly, not just focused on RPC. We always wanted to generalize this solution and go after uh pretty much anything that you wanted to connect to.
I I really believe in using blockchain as that coordination layer, that world computer like I mentioned and uh building a product on top of that delivers a better product experience. Like this is solving a problem that actually exists and is affecting a ton of people out there that are trying to do agentic development or AIdriven development. when you tell your LLM build me this vibecoded app and it says okay I deployed with Versell and then I used superbase for your login and I use this and that what does it say when you ask why did you choose those things what if it could be deterministic or a datadriven decision where it says so I went to this marketplace it had all of these verifiable uh receipts and reviews of all of these other people that are using this service rather than right now my LLM, I don't know about yours, says a bunch of people on Reddit seem to use this one and so that's why I'm using it. I'm like, okay, must be all right then. Uh, but people are now vibe coding or agentically coding to prod to production.
So, uh, let's let's use data back decisions in how we build. So, that's what I'm building. Uh, we'll be releasing more info soon about the the new product that's not live yet, but that's my um X handle. Feel free to follow me, reach out to me. DMs are open and happy to chat with anybody that's kind of working in that space.
Thank you.
Automatic transcript — names and jargon may be misspelled.