Scott Shi - Kite AI: The Economic Layer for Agentic AI - TOKEN2049 Dubai 2025

Before I start, I want to do a quick survey. How many of the audience is actually developer? Could you raise your hand if you're a developer? Okay. I I need to change my pitch to a non-developer audience, I guess.

Um, my name is Scott. Um, I come from Silicon Valley, so this is actually 20 degree hotter than Silicon Valley, San Francisco. Um, kitei is building an economic layer for agentic AI. A little bit about myself, I spent last 15 years building foundational data and AI infrastructure. Before Kite AI, I co-ounded a streaming DB company.

Before that, I work at Uber. You know, build Uber first skill AI. You can think of that's actually provide label data for AI and then grow a team from myself to 200 people. I also was a founding engineer at the southwest einstant which is a first south instant AI and data cloud. Um that's my background.

I spent last 15 years in foundational data and AI infrastructure. My co-founder she got her PhD from Berkeley. She co-ounded uh she was one of the founding engineer data one of the largest autoML company before she was a lead product manager at um data brick before which is also a data infrastructure. So we have been building kite since 2022 and we are backed by the worldclass investor like general catalyst, PayPal, HVC, hash and hashki and Samsung next and then our work our infrastructure and our work has been approved and used by our partners for example on three layer and story. So we have been this ecosystem for three years.

Why now? For last month actually I spent two days build a very simple personal agent help me do all the work including order Uber order Uber Eats do the deep research in this case everything run locally my personal agent is actually a very simple agent but in power by kite service agent not only now I can use my agent to help me do all the work I can say exactly end to end transparency as well so agent AI will be super cycle. This will be AI super cycle not five years from now exactly today. This is not feasible six months ago but you know since four months ago we have a deepse release this logical chain and then now you have all this open AI open anthropotic released a model and today last week we have this Q tan released the newest model which is already reach a closed house model performance. So what's is actually easier for AI agent in the future?

I think every single action is actually going to be done by AI agent. You want your AI agent do a research for example which token going to rise in this case. If you want your AI agent help you do a research maybe in diabetes. It should be done the work for you. It will call different research on paper.

It will call API to do you know a summary. Maybe do some inference and upload the result in Google doc. The key part is that you don't want to stand in the way of the AI. You want the AI do all the work for you as a human. The only thing you want to do for AI is that agent is that you want to authorize the agent and you want to make sure prove the spending if the agents um you know spend under your budget.

So that's why there is a huge gap for AI to to use agentic AI to operate to transact. The gap is that the AI can need to have a ID to code accountable. Without ID AI cannot be held accountable. So that's why first for agent you need a ID. Once you have ID you need to have a wallet for AI.

They can spend autonomously and you know transactionally. All the existing infrastructure are not built for a genetic AI. If you think about our currently the credit card, it's either three percentage of credit card fee or 35 cents minimum charge. But in the work of agentic AI, most of the action going to be just maybe a small API call to sensor the data research, one sensor inference. So that's why all the existing infrastructure are not built for agentic AI.

Additionally, I would argue the hardest problem for adoption of the agent is trust. How do you make sure as a user I trust my agent? How do I make sure agent to agent call I trust in this case? So that's arguably is one of the hardest problem to uh prevent the mass adoption of Gentic AI. So that's why Kite is here to not only assign ID to the agent but end to end verify and end to end transparency provide end to end transparency for the agent.

So kite is basically economic layer for the agentic AI. Use blockchain word is blockchain for AI. It's very different than AI for blockchain. AI for blockchain maybe you're building a AI agent to your particularly deep you know um um Dex or particularly use case. But blockchain for AI is that we want to provide a foundation for AI.

In this case AI agent may come here looking for the service agent that they need exchange for service and get for the exchange payment for the service. So in this case you have value chain which as a agent you may not actually know what's going to do next because the next step maybe actually becomes depends on previous step. You don't actually know the dynamic attribution beforehand. So you exchange for service and for payment. So that's why you have a AI value chain AI payment chain.

So that's what kite is for. We want to make sure hold every single agent accountable. So we will assign ID to agent and I would arguably like similar to 15 years ago with mobile 8 you actually want to put all the difficult logic and the heavy computing heavy computation in the server side. So in this case the personal agent is actually more like a privacy um you know you make sure your money is secure but the heavy lifting the actual work is done by the service agent. So, kite is a platform to provide agent to agent trust focus on provide a service agent to serve your personal agent.

The service agent can be just data provider. For example, if you want to read um if you want to book a holiday holiday party with your teammates. In this case, you may want to read some Yelp data, find the best hotel for you in Dubai. Um it could be like GPO. I may need to do scientific research.

I need two hour H100 GPO. In this case, I can give my requirement to the GPU provider and say, "Hey, help me find two hour GPU or cheapest time." In this case, notice that each of this requirement the service level agreements different. For GPU, it's probably time to first token or token throughput. For this Uber ordering, it's more about Uber order completion, order accuracy.

So each of the service agreement between your personal agent to the service agent is different. All this need to be programmable. So that's why kite provide a layer one infrastructure for you to deploy customized smart contract between your personal agent to the service agent. So kite if you look at architecture for kiteai we actually be very modular and composable. In this case you can reuse the data for different purposes.

For example, I gave you a concrete concrete example before in Uber. There are only probably 30 base tables. You have a table to to have, you know, trips related data. You have a table with rider data. You have your users payment data.

Then use a combination of different table. It's a composable. It's like you you use it like a Lego piece. And for different use case, you can use trips data and then writers data and say if this rider is actually first time catch rider. It could be like you have a first-time cash writer trying to ride the, you know, rob the Uber in this case.

In this case, probably need trips data, writers information data, credit card data to make sure you don't have malicious rider trying to rob the driver in this case. And then but you can use the same trips data and GPS data to to build a collision detection. This is actually true project I built at Uber. If I'm taking an Uber from here to the airport and I I have a car accident, Uber will call me Uber will call my driver and call Nan at the same time to make sure I got you know I got um um Nan the first first hand you know immediately. So in this case is basically use a composable the same data trips data rider information data for completely different use case.

So that's why kite is built as a ecosystem or site. So the same data can be used in completely different use case. So we have a subnet or data subnet and we also have a agent and service subnet in this case. So we want to have a a rich ecosystem subnet. So you have this 1 plus one larger than three effect.

Why do we need a purpose-built layer one? So kite is actually a layer 1 plus additional infrastructure. We need to calculate for attribution um contribution and then u verification. You cannot first is that air compute share very different than financial transaction. If you think about financial transaction the single car bought is that ms asset security how much asset is transfer and in this case you you can wait for few seconds or even few minutes for the asset to be finalized and the finality you know finality time can be anywhere between 60 minutes for bitcoin or few seconds in this case but using AI is sharing completely different traits for AI inference you want to get a less than 100 milliseconds and also you you probably only take advantage or 100s of the transaction financial transaction cost.

You know, you don't want to pay for $2 just doing a inference in this case because each task actually going to to anywhere between you know my build my local agent each task can anywhere between you know 20 steps to 50 step in this case you want to actually get it affordable price you don't want to actually get it a very expensive price. Second is second is that not only you want to execute this inference on execution layer you also want to have end to end transparency and verifiable you want to make sure all the inference provide my service agent be able to verify and then this should be done by settlement layer again very different than transaction financial transaction you want to verify it there in this case and at last that the consensus layer is very different if you think about Ethereum Today Ethereum is giving $4 billion block rewards to a million validator doing exactly the same thing exactly same thing you know just a test station verify test station but we want to redirect 90% of this reward block reward to AI contributor in this case that's why we need a consensus layer first is that end to end calculate what's attribution or contribution from the AI contributor second is that redirect 90% block reward to the AI contrib contributor. So that's why Kite actually going to build AI compute as a first class citizen. This including add AI compute as a first class computer citizen along with the transaction. We also track the AI compute volume as a receip to receive this um block rewards.

If you look at the data structure from transaction and AI compute is very different as transaction you mainly spend on who is spending what money but if you look at AI compute you have a large size of input and it could be like a marty model it could be the text to video video to video it could have video model and today now most of the most of the models support long context you can put this whole Shakespeare in the input and then let your model you know produce output so in This case AI compute as a first class citizen will be very different than financial transactions. All the existing layer one all the existing blockchain is built as the for financial transaction. They're not beautiful AI related compute. So at time if from if you look at from product perspective so kiti unlock composibility monetization and liquidity or AI set in this case we we're going to have a marketplace or AI service agent service provider so if you're building agent if you're building a service for agent if you're providing data service you know come to talk with us we already have top builders from near ecosystem um beat tensor joining us in this case and then provide a service agent to pro to serve AI user. Second is that every single component in kite is built as composable.

So in this case you may use data from vendor A you may use some data from another provider B and have this one plus larger three effect. The last part is that we want to host a vibrant D5 ecosystem. And then this is actually um an AI platform we're going to release in in less than two months. So in this case you can see if you're a service provider for example I build my personal Uber agent service. It help me call Uber.

It help me order Uber is uh in this case we're going to release a demo and a website for this. In this case, if you for example, if you want to build a agent service, if you want to build a data service, you can come to us and then build this service. And all this service is actually to target agent to agent service or agent to data service. At hand, I think not in five years, not in 10 years. In next 24 months, most of work probably going to done by your agent.

So in this case, you just need to authorize and and proof for the action for the agent. Kite released our first version of Tisnet less than three months ago. We already have five millions of wallet actually on top of us interact with us and we have a more than billions inference call agent inference call on top of us. So um yeah so um again so if you're AI builder if you know someone actually building service for agent uh if you're interested in agent to agent talk or if you if you know for example if you just have some data you think is going to be very valuable for the agent come to talk with us and then please join Kite AI ecosystem. Um that concludes my presentation.

Thanks folks.