Tokenized Podcast
From Shopify to Sapiom: Building the OS for Agentic Commerce
🇬🇧 EN🇪🇸 ES
Pregúntale al tutor sobre este vídeo
Entra con un email verificado y pregúntale al tutor lo que quieras sobre este vídeo.
TL;DR
Resumen de investigación
▶ Presentación y patrocinadores
El fragmento inicial plantea que, a largo plazo, los agentes podrían dejar de usar interfaces diseñadas para personas y comunicarse mediante APIs entre máquinas, cuando estas se expongan. Simon Taylor presenta Tokenized, centrado en stablecoins y adopción institucional de activos tokenizados, junto a Bam Azizi y el invitado Ilan Zerbib, fundador y CEO de Sapiom. La lectura de patrocinio describe la plataforma de stablecoins de Visa para emitir, mover y gestionar stablecoins, acceder a Open USD y mantener la autocustodia; también menciona tarjetas vinculadas a stablecoins aceptadas donde opera Visa. El anuncio de Mesh afirma que su red conecta más de 300 wallets, exchanges y plataformas de pagos y permite pagar y cobrar con distintos activos. Esos datos aparecen en los mensajes promocionales.
▶ Trayectoria de Zerbib y origen de Sapiom
Zerbib cuenta que empezó a desarrollar productos a los 12 o 13 años, trabajó en seguridad para el Departamento de Defensa francés y se mudó a Estados Unidos en 2014. Fundó Ernie, un producto de consumo que usaba agentes para recuperar dinero cuando bajaban los precios: según él, millones de bots rastreaban la web y el servicio devolvió dinero a 5 millones de usuarios en EE. UU. Dice que dirigió Ernie durante seis años, que la empresa fue adquirida en 2021 y que después pasó 5 años en Shopify, donde lideró Shop y trabajó en identidad y confianza. Sitúa la idea de Sapiom tras la aparición de ChatGPT y los modelos de lenguaje, y dice que tomó forma con avances en razonamiento, MCP y protocolos de pagos para agentes. Menciona x402 de Coinbase, Agent Pay de Mastercard y un protocolo de pagos entre máquinas de Stripe, de la que afirma que Sapiom es socio. Presenta la empresa como infraestructura para habilitar agentes en la economía.
▶ Identidad de agentes, financiación y pagos con stablecoins
Zerbib describe tres componentes de identidad para agentes: la identidad del usuario o negocio que está detrás, la autoridad delegada para actuar en su nombre y la reputación del agente. Dice que Sapiom monitoriza y autoriza acciones, como llamadas a APIs y transacciones, para construir una especie de puntuación de crédito del agente. Considera que la identidad compartida aún no es crítica y estima que podría serlo en 12 meses; menciona una inversión de Okta Ventures en Sapiom. En respuesta sobre sus inversores, dice que Dragonfly lideró una ronda Serie A de 35 millones de dólares, anunciada dos días antes de la grabación. Caracteriza Sapiom como infraestructura de IA, aunque cree que los pagos de agentes probablemente usarán stablecoins, micropagos y liquidación en tiempo real. Distingue las compras de consumo, que podrían usar raíles existentes, de pagos máquina a máquina por SMS, búsquedas o fracciones de segundo de cómputo, para los que defiende stablecoins programables con liquidación inmediata y comisiones muy bajas.
▶ Stablecoins y plataforma de Sapiom
Bam sostiene que, en una economía global de agentes, los medios financieros tradicionales no bastarían para micropagos y cobertura global; menciona un pago de «0.00005 005 cents» y argumenta que reglas de autorización podrían implementarse con contratos inteligentes y blockchain. Opina que buena parte del rechazo del entorno de IA se dirige a cripto como activo, no necesariamente a blockchain o stablecoins como tecnología. Zerbib presenta tres productos anunciados dos días antes: Agent Studio, para pasar de una idea a un agente listo para producción en minutos; Agent Runtime, que aporta entorno de ejecución, sistema de archivos, sandbox y gestión de secretos; y una capa de gobernanza y acceso que permite a los agentes comprar servicios de API mediante un ledger con uso medido y controlado. Añade Sapiom Router, que selecciona servicios y modelos según coste, latencia y calidad.
▶ Coste de inferencia y GPUs dedicadas
Zerbib explica que Sapiom integra la inferencia como una capacidad gobernada y cobrada desde el saldo del agente. Afirma que un cliente de consumo masivo pasó de cero a 1,2 millones en costes de inferencia tras 3 meses, mientras aún tenía miles de usuarios y no millones; atribuye el coste a agentes autónomos que consumen inferencia continuamente, incluso 24/7. Dice que muchas aplicaciones eligen por defecto el modelo más caro y estima que los modelos de pesos abiertos pueden cubrir el 95 % de los casos de uso, con un coste de inteligencia de «two pennies» frente a «dollars» para un resultado comparable. Según él, el enrutamiento y la infraestructura propia reducen los costes de inferencia entre 60 y 90 % manteniendo calidad comparable. Sapiom cobra una tarifa fija por ejecución del agente que incluye infraestructura, capacidades e inteligencia, en vez de cobrar el modelo por separado. Explica que opera GPUs dedicadas porque los modelos abiertos servidos sin servidor podían enviar datos a terceros países y la capacidad disponible en EE. UU. o Europa era reducida, con poca resiliencia; dice que los proveedores priorizaban contratos anuales de GPUs para entrenamiento.
▶ Usos actuales y horizonte de la economía de agentes
Ante la pregunta del presentador, que cita 270 millones de transacciones y 100.000 ejecuciones diarias de agentes, Zerbib no confirma esas cifras y describe usos como agentes de programación, marketing y ventas, publicidad automatizada en Instagram y Facebook, y atención al cliente. En el ejemplo de publicidad, el agente recibe un presupuesto, interpreta el contexto de la empresa, crea un video, compra anuncios y evalúa resultados. Zerbib dice que el volumen de comercio agéntico aún es pequeño, pero prevé que la economía de máquinas alcance billones de dólares en los próximos 3 a 5 años y cree que el mercado podría despegar en 6 a 12 meses. Identifica como barreras actuales el coste y llevar agentes a producción.
▶ Adopción de compras autónomas y papel de los intermediarios
El presentador recuerda una apuesta de 1 dólar sobre lo que ocurrirá en los próximos 12 meses y pregunta por los pagos con tarjeta de agentes. Zerbib responde que la tecnología y los protocolos ya existen; cita la colaboración de Shopify con OpenAI para compras y el protocolo GCP de Google, pero atribuye el freno a que los consumidores no quieren delegar decisiones personales, como hoteles o el color de unos zapatos. Dice que Sapiom se centra más en agentes que compran cómputo, inferencia y servicios de API. Simon Taylor afirma que datos de Shopify muestran que, para comercios de larga cola, quienes llegan tras descubrir productos en un LLM serían entre 4 y 6 veces más propensos a convertir; añade que, en una prueba de Walmart con checkout dentro de ChatGPT, la conversión fue peor que en su sitio web. A propósito de wallets para agentes de Cloudflare, Zerbib describe a la empresa como una puerta de entrada a internet y dice que lanzó stablecoins unos meses antes y que ahora invierte en infraestructura de wallets. Ve a Cloudflare, Stripe y otros como habilitadores de comercios, aunque espera competencia por protocolos, wallets y porciones de las transacciones.
▶ Barreras empresariales y evolución de las interfaces
Zerbib atribuye el despliegue limitado de agentes a falta de educación e infraestructura: construir productos seguros, funcionales y rentables obliga a integrar componentes que no fueron diseñados para agentes. Espera conversaciones distintas y despliegues empresariales a mayor escala en 6 meses, pero dice que hacen falta control del gasto, costes de inferencia menores y gobernanza con trazas de auditoría para que cada acción respete los límites autorizados. En el diálogo sobre compras, explica que los compradores también valoran plazos y opciones de envío, impuestos, descuentos y ofertas adicionales del checkout. Considera que navegar la web seguirá teniendo mercado a corto y medio plazo, pero predice que a largo plazo los agentes usarán APIs máquina a máquina. Dice que los servicios tendrán que exponer APIs y controlar mediante protocolos de pago el acceso a recursos de pago.
▶ Clientes objetivo, organización y cierre
Zerbib dice que Sapiom se dirige primero a desarrolladores que construyen productos autónomos y, a la vez, ofrece apoyo a empresas desplegando ingenieros para ayudarlas a adoptar IA y agentes. Describe su objetivo organizativo como una empresa muy automatizada, con personas propietarias de métricas o componentes y equipos que usan IA para resolver incertidumbres; afirma que aún necesita un equipo grande y quiere mejorar cientos de métricas de negocio. Simon Taylor comenta que el trabajo y los flujos en empresas estudiadas han cambiado notablemente en los últimos 18 meses, especialmente en los últimos 6 meses. Para probar Sapiom, Zerbib ofrece 50 ejecuciones de agente gratis al día y 15 dólares en créditos para desarrolladores. El cierre comparte los canales de contacto de Sapiom, Mesh y Simon Taylor.
Transcripción completa
[0:00] I think there's a massive market in the short to midterm. I think if you think long term I don't see why an agent needs to access an interface that has been designed for human to be able to do an action that is provided by a machine like already communicating with API through an interface. So over time I think the interface for agent will disappear and it would just be machine to machine APIs because those APIs are here on both end [music] it just are not exposed. >> Welcome to tokenized the show focused on stable coins and the institutional adoption of tokenized real world assets. in this show. Well, it's a little bit
[0:30] more agentic commerce themed and all things agentic and AI. Of course, my name is Simon Taylor. I am your head of market dev at Tempo [music] and author at Fintech Brainfood. And I'm joined by my agentic commerce co-host Bamzy, CEO Mesh. How you doing, Bam? >> I'm doing great. Super excited for this episode. >> Joining us today is a phenomenal guest, Elan Zabib, who is founder and CEO of Sapium. How you doing, sir? >> Amazing. Thanks for uh having us today.
[1:00] >> I'm excited to learn more about what you do because it looks pretty cool. But before we do, I just got to remind uh viewers and listeners that views and opinions of contributions today are their own and might not reflect those of companies they represent. And please don't take anything we say as tax, legal, or financial advice. This is forformational purposes only. And of course, this episode is brought to you by Visa and our friends at Mesh. Stable coin operations usually mean a wallet from one vendor, an on-ramp from another, and then your controls are stitched together across both of them.
[1:30] The Visa stable coin platform fixes this handily. You can mint, move, and manage stable coins, access Open USD, and remain your own custodian all in one single shared environment. Then, stable coin linked cards let you spend balances anywhere Visa is accepted. And we know stable kind linked cards are in hyperrowth as our good friend Kai likes to say. That is Visa of course the global leader in payments and of course sponsor of this show. You can find out
[2:00] more at visa.com/crypto. This episode is brought to you by Mesh, the first global crypto payments network. Connecting over 300 wallets, exchanges, and payments platforms. Mesh enables anyone to pay and get paid instantly anywhere in any asset. Mesh makes digital transactions seamless, secure, and universal, fueling the next era of aentic commerce. Learn more at meshpay.com/ai.
[2:33] Ellen, let's start with you before we get into Sapium because uh you're doing some pretty cool things in AI, but how do you end up uh kind of knowing what you do about payments and AI? What's the story there? What's the background? >> Yeah. So maybe I will start uh uh from from France. So I'm a software engineer. Started building product when I was uh you know 12 or 13. I started my career in Saga security. Uh so I worked a bit for the French department of defense but
[3:04] quickly I moved into building startups. Uh so I joined uh I moved to the US in 2014 build the first company called Ernie. So Ernie was a consumer advocate product. We're actually building uh autonomous agents that were getting refunds to people. So if you buy something on the price drop Ernie was able to get your money back in the first three months. So it was actually millions and millions of bots that were scrapping the web extracting data and actually getting money back to to 5 million users in US. Uh so I faced actually what agents are facing today
[3:34] like 12 years ago. uh where the actually the internet infrastructure has been designed since day one to prevent automation and actually today automation it's kind of something that we want to enable AI agents but nothing works because from like the infrastructure the DNS the like the CDN uh the the fraud models everything actually has been here to block those agents and if you want to enable agents you need to have a proper infrastructure for them uh so I was uh running Ernie for six years uh we then
[4:05] got acquired in 2021 and after the acquisition I spent 5 years at Shopify. So I was director of engineering leading a global product called Shop. So I was leading the shop product the identity layer the trust layer build a product called Shop. I've been very deep into everything that is related to uh payments identity uh and trust uh and uh same problems everything is designed actually to make sure that there's a human between every transaction between every actions and now we're in a world
[4:35] where actually want to enable agents to access the economy to be impactful and there's so many barriers that are blocking them to get to production or be economic actors. So that's kind of the history of Sapion. uh both internet infrastructure and financial infrastructure has been designed to block them. Sapium is designed actually to enable them or to build the right infrastructure for those agents to try >> well and and shop shop pay's great heritage um Kaznet and a good old friend and um I am such a fan of that product and I think the thing it's done really
[5:06] well is dramatically shrink the time it takes to check out with identity solving it really it's an identity product masquerading as a as a payments product that sort of remembers you and remembers your shipping details details remembers your card details across multiple merchants around the internet. And so really curious as to uh what you saw in there that made you start to think about Sapium you 12 months ago or so when when
[5:36] you started that like what what is it at Shop Pay and and your career history that makes you think I need Sapium now. We need to go build this thing. Is it that sort of point you were making there about the like the whole internet is designed to block bots? >> Yeah. And like the idea of sapium actually came out like a bit more than two years ago like actually end 20232 when we had ch that went live uh it's the first time that actually people experience like AI the first time that we actually cracked uh large langu large language models uh and actually that we
[6:07] could get value from AI uh and for me it was very obvious because those models were already very smart uh but they were very smart in a chatbot you can ask any question, they can give you any answer, but it's limited to a prompting on a a chat interface where uh the next step were quite clear actually those models would become smarter. Meaning it's not about just giving you an answer to a question. It's as well being able to reason. You give them a problem and they will figure out how to solve it. Um and then being able to actually do things in
[6:38] the background because at the end an agent what we call an agent before like the LM before a chatbot it's just a piece of software that is smart uh and it doesn't have to live on in a chat interface it can live in a background or scripts like any type of software and then that's where you what that's what we call agent really those piece of software that can uh make autonomous decisions on their own when I started to play with CH GP uh in 2022 that's where I realized actually that when that will be a thing that those agents will be
[7:08] able to operate on their own, everything will break because the internet infrastructure everything has been designed to actually prevent those behaviors from the software. And then it took me a while to find the right moment to build. Uh I think the uh uh it started to really take shape early last year when we started to see more reasoning models, chain of thoughts and like really the early of like early last year we had like a massive step change in in the reasoning models and then in the first quarter we had MCP which is like okay now you have very smart model
[7:39] and they can communicate with other systems. uh and then in May June we started to see payment traders emerging for agent like X42 from Coinbase and since then we had Mastercard that build agent pay we are partner Stripe they build machine payments protocol so there's all the infrastructure pieces that are coming together to enable this new world of trillions of agents operating in the economy and Sapium is the platform to enable that and make that a thing possible. >> Yeah. So what is different between agent
[8:10] because my background is identity too. I want to know your take on agent identity or KYA and end user or consumer identity or KYC. Uh what does it work and what does it doesn't work when you move from end user to agent as an end user? >> Yeah, I think on the agent identity there is like a uh three components. Uh first we have like with the entity behind the agent um it might be a user
[8:40] or it might be a business. So the Q the QC or the QYB. So when we talk about know your agents like agent identity the agent identity needs to carry the consumer or the business identity because when an agent make a purchase decision he needs to prove actually with kind of behind the agent uh making sure that there's delegation of authority. Okay, as a business or as a consumer, I give authority to that agents to act on my behalf. And the third, it's actually what's the reputation of that agent? Like on the merchant side, can I trust
[9:11] this agent? Because I might trust the business or the customer behind, but how do I know that this agent actually is not an agent frog that he has a proper reputation? So there is three components. Um the key way like the the uh uh authorization on the carrying the business identity is quite easy you know it's just we need to have a standard that carry those things the agent reputation is much harder because you need new datas that don't exist actually almost like credit score data uh and that's what we have at Sapium because we enable agents to access the economy we
[9:42] actually monitor and authorize every action might be an API code it might be a transactions everything is actually monitored by Sapium and authorized by the by our platform so we as the agents operate actually we build the credit score of the agent which is kind of the missing piece to provide like the reputation of the agent um and those are amazing topics I think it's still 12 months away before the identity is becoming something that is critical for the machine economy to uh thrive we're not there now the problem is at least
[10:12] getting to production but those are very interesting problems uh for example octa uh they invested to sap through octa ventures those are kind topics that we're thinking as eventually we need a shared identity that is that is standard that is uh recognized by everyone. >> Yeah. So you do have a very interesting cap table. You mentioned octa there um but also like a real mixed bag. You've got um anthropic um but then also dragonfly and coinbase and vanex. So there's like digital asset heritage,
[10:43] there's AI heritage, there's identity heritage. Why that cap table? like what is it that you're sort of thinking by having that group of people on you know backing you >> and Excel as well from the previous round with the long-term infrastructure play uh actually I ced the cap table like uh I'm curating a team uh how we build the best capables to win that uh that market and we really can enable that trillion of agents to exist it's
[11:13] very interesting because the way I see it like uh so dragonfly they led uh our cedar sorry our series A that we just announced uh two days ago. So we raised 35 million series A from Dragonfly and when I had that discussion with AI Beron like his on the machine to machine economy is very similar to artsis. So actually we're an AI infrastructure play. We're not a crypto play but actually behind the scene those SHs are emerging because if you want agents to access the economy most of those transactions are going to probably be
[11:45] powered by stable coins by uh microtransaction by realtime settlements and those are kind of the crypto rails that are enabling that. So there's two side of the agent economy. There is the consumer side an agent buying a flight ticket or a pair of shoes which will most likely rely on existing rails. But if you want an agent that buy in real time an SMS a search request I don't know 1 second or 1 millcond of a compute those are only possible truth stable coins like programmatic money uh that can settle in real time uh and that can
[12:17] have extremely low fees and that's where actually those two are merging it's actually to enable those of agent we need an extremely solid uh crypto like our blockchain ecosystem behind the seal um that is powering that power. >> Fascinating. Um, Bam, you and I have spoken a lot about, you know, sort of where's the logical entry point for stable coins, and I think we've had a lot of, um, stable coin fans say, you know, oh, well, obviously it's it's AI, but the AI folks have always been, I
[12:47] don't know, a little bit anti-crypto. How are you thinking about this one? And what's your reflection? Yeah, my reflection is um to to Elon's point um we will have a lot of mission to mission payments that no matter if you you can hate it, you can love it. What whatever your feeling is, you need to put it aside because the traditional financial tools and means uh will break in that new ecosystem or that new framework or the new modern agentic word. Right? So if an agent wants to pay 0.00005 005
[13:19] cents. You cannot wire, you cannot use Visa card or Mastercard. You have to use a stable coin. If you want to have a global coverage, you have to use a stable coin. If you want to have a through authorization and um and uh add rules and limitations and regulations around what agent can do, what cannot, you need to use a smart contract and blockchain. You really like if you want to do it in a truly global sense, there is no other choice. It's not like I have multiple choice. I just go with what I
[13:50] love. Like no matter if you love it or hate it, that's the only choice you have to use it. And I think a lot of uh anti-crypto type of u uh comments you're hearing from the AI world is mostly around uh crypto as as an asset, not blockchain or stable coin as a technology. So I I think a lot of uh thought leaders in in in agentic commerce and especially in AI uh they're they're actually pro crypto and prostable coin. Look at like for
[14:20] instance Sam Alman they have another company called Ward or the the Ward app probably everyone knows to humanity. So everything is on built on top of blockchain and they're like also thinking about like moving identity on blockchain as well. So I don't think anyone hates blockchain itself. I think it's just more of a like crypto NFT like Bitcoin versus like AI. Which one is cooler and hotter? >> Yeah. And uh we can certainly argue that um the Bitcoin folks help build out a lot of NeoClouds that have become very
[14:50] very useful for for infrance compute at this point with with power. Um but I I'm interested to learn you were sort of talking about there's a future state of identity, but right now you're sort of um my understanding is you're sort of helping people root models really effectively and efficiently. um you've got an agent studio and an agent runtime. Uh I look at the difference between the companies that are shipping faster and getting the AI dividend and everybody else as being the people who you like the the two or five companies that have managed to build this all
[15:21] inhouse. You know, you ramp Block Robin Hood. But is SAPM kind of this for everybody else? you're the platform to build agents, but also um the smart model rooting is a really interesting idea which is like, hey, maybe I don't want to go all in on anthropic or um open AI or pay their fees. I want the the task to go to the right model at the right time. So, how how does all of that work in inside the Sapion platform? >> Yeah, it's a it's a it's a great question. So, yeah, we have like what we announced two days ago. It's really
[15:51] three new three new product that we uh brought to the market. really the agent studio which is getting from an idea to a production ready agents in like minutes. So we have our own opinative frameworks on how agent should be designed and how they should be shipped. Uh so we are actually helping builders uh uh uh thinking about agent product and helping them build the right way agents. Uh then we have the agent runtime which is really the deployment environment the production environment for agents to run. It's kind of their world. We provide the file system the sandbox secrets management. So
[16:22] everything is is related to the turnout to the agent. So when an agent runs they can have their own uh environment which provide infinite scale uh and as part of the environment we actually give access to the agents to the economy. So that's the first product we build which is the governance and then the access layer really giving agent a ledger to buy any type of API services. This is baked into our runtime. So running on sapium actually they they can have access to any type of capabilities in a way that is fully metered and governed. Um and
[16:52] then on top of that that's the what we call the sapium router which really for every agent action every aent steps we decide what's the best services and what's the best model. So we optimize based on cost latency quality of the output. So we are constantly optimizing uh the path of like the agent to make sure that the agent always gets the most out of every action. I love that ledger piece because uh you know you sort of think about uh auditability but also a baseline of knowing what these things are doing and then their ability to move
[17:23] money, their ability to pay for things like that that as a foundation really sets you up for for a future. Um and so many companies are trying to do this at the moment. The other thing I saw like you're running your own inference stack on dedicated GPUs. Why? >> Why? Uh hey [laughter] >> um I believe uh uh inference like inference is key like if you need an agent actually the way we see it is like the intelligence is just a capability like any other one like an agent sending an SMS only needed to run a prompt it
[17:54] needs to get access to a capability um so we integrate that into the platform the capability as well like the inference is kind of preparious so every prompt is mered and authorized actually by the platform uh and it's coming actually from the balance of the agent And why it's because like we need to optimize the cost of intelligence. Uh we've seen with some of our customers uh that we're innovating rebuilding like consumers or B2B product where uh if you default to the frontier models for everything you can't run a business
[18:25] actually you have you go back as soon as you >> negative ROI. >> Yeah. The it's it's it's negative margin by hundreds of percents. Um and we've seen that with uh one of our customers that was mass market consumer product automating like businesses for people to really build businesses. Uh but the cost of inference grew from like zero to like 1.2 million after 3 months of operation. Um they were getting users but they were not at million user scale. they were at like thousands of user scale and just the cost of intelligence when you give
[18:56] autonomy is just too high because it's very different to a product where you prompt because when you print you have a human that decide every actions when you have an autonomy it's an agent that work whenever like 247 that consumes intelligence and that intelligence is actually not scalable so today there's a massive barrier to innovation which is actually the cost of shipping those agents when you're an indivi an individual running Europe and clue you manage your spend and it's fine but actually if it's a product you're trying to sell to thousands of customers you
[19:26] you have to control the inference cost because if not you can't even build your product and ship it to your customers so that's what we've seen and then we say actually like the intelligence is should be part of the platform today we are pushing to every customer's consumer B2B like model selection even if you go to the mainstream consumer app builders they ask you to pick your model no one knows what's the best model so everyone defaults to most expensive one where actually for 95% of the use cases now with the new open weights model we've seen over the
[19:57] last few months actually open weights model can get you the cost of the intelligence two pennies versus dandas uh delivering the same output um and that's what we provide. So the fact that we internalize the inference stack is because we're able to ensure we deliver the same quality of an output for a frontier model leveraging open weights model meaning we are able to optimize the cost of the inference by 60 to 90% providing the same uh uh quality of what the agent is trying to achieve uh because we automate uh we do the
[20:28] routting and we uh have all the data to understand actually what the agent is trying to achieve >> and and I imagine as well this sort of um means that the experience for somebody coming to your platform is a little bit simpler and that they're not having to sign a a contract with open router behind the scenes or b like a bunch of enterprise agreements with all of the labs in order to have that inference time compute or the the the models served to them. You're just you're full stack. You they they sign a
[20:58] deal with you, the agent runs inside the runtime and you take care of everything else below it. So fascinating. >> Yeah. And we really believe that um uh to prove that that we believe we're the best at optimizing the the model is actually we don't even charge for the model. When you run an agent on Sapium, we charge you a flat fee per agent execution and that includes the infrastructure, all the capabilities that your agent needs and the intelligence um because we know that we are actually able to optimize your intelligence to get you the best agent with the best outcome, but every steps
[21:29] along the way are optimized by by our router. >> Yeah. What is it? mostly. So if the unit economics goes down or the unit economics makes sense for you, does it does it make sense for you to move away from owning your own GPUs and your own stack or working with others? >> Um we decided to operate our own uh uh GPUs to uh first have dedicated capacity because early on we tried actually to use serverless uh uh uh open weights
[21:59] models open weights models. Uh there's two problems there. Uh on one end you have like most of the open weights models are not served uh in US or Europe. So you're actually sending the data to a third party countries that most of the users don't want to send their data. uh and on the other end for the only few that are US- based or Euro based there are such a tiny capacity where actually you always facing you have zero resiliency in your inference and most of the time they will not be
[22:29] able to serve your inference because there is massive demand uh and they don't allocate enough GPUs to those models because usually they will rather sell year contract GPUs to companies trying to do funing or model training. Uh so we decided to okay we need to provide that capability which is core for our customers to succeed. We need to make sure that our customers have a margin that is good enough for them to even be able to scale because it doesn't make sense to have customers and then after 3 months they go bankrupt because
[23:00] they pay frontier models. Uh so we decided to internalize that and really had the customer succeed. >> Excellent. >> So fascinating. You don't see you don't get a lot of startups coming out the gate at the moment building sort of uh model routting and uh kind of an agent building stack that also has gone all the way vertically integrated. Um, and you know, Elon's getting applauded for having built uh giant uh stacks and data centers, but actually the pragmatism of what you're doing here with like owning a stack for agent workflows at an ultra
[23:32] low cost is is just like diametrically opposed to that sort of scale and and really pragmatic. Uh, which is which is kind of refreshing in a lot of ways. But talk to me about what people are using Sapium for today like 270 million transactions, 100,000 daily agent runs. What what are all those agents doing? >> Yeah, so we have we have seen the long tails of agents. Uh there is the obvious one like a lot of uh coding agents really building products and shipping products. We have a lot of like go to marketing agents like agents being able
[24:04] to find leads and sales and like customers to businesses or like those agents. uh we even have agents that run uh end to end ads for customers uh Instagram on Facebook and all those ad are fully automated by some of the agents running our platform which is fascinating. So customers are giving a budget and then the agent just figure out the context of the company build an asset a video with models and then is able to buy actually ads on ads platforms uh and then close the loop on performance. So it's really the long
[24:35] tail. We have a lot of agents doing customer support, chat. Um, we are really the platform that enables builders to build and innovate and scale. So, we are seeing like any type of uh uh uh agents getting into production on our platform. >> So, we alluded a little bit earlier on to um sort of stable coins and aentic commerce and and today like the volume's super small and super early, but you know, you've talked about identity, you talked about a 12-month time horizon. When does Aentic commerce start to get
[25:05] real in your mind, Elon? >> I don't think we're that far. Uh because as as you shared like the numbers, we're already seeing like meaningful volume. Um I still think it's very tiny like the main barriers people are facing today cost and even getting agents to production. Uh so those are the things that we are solving and that's why we are here to enable the next trillion agents because we believe we give enough advantage for builders that build on our platform by abstracting all the things that are needed to actually build the product but that economy is going to be
[25:35] massive like we're talking about trillions of dollars in the machine economy in the next 3 to 5 years uh and we are like a fraction of that. So uh but when I think we're like six to 12 months away from that market to really uh explode. So you know like Simon is the bad cop I'm the good cop who's more realist and more optimist. So and we have a $1 bet what's going to happen in next 12 months. So what is the hold up for like big companies let's say
[26:08] uh entropic that is on your cap table or open AAI or others to enable their agents to pay with just credit card and what is the hold up there why it's not happening there is no hold up actually the technologies are there uh even Shopify they partner with open AI to deploy like shopping to open AI so uh agents can actually transact there's protocols like GCP from Google there's many protocols that are enabling that the hold up is actually the consumers people they don't really want to uh
[26:39] delegate purchase decision for most of the things um and you know there were a lot of learnings from like the ecosystem on the market on that when CHP was able to buy things on behalf of consumers uh actually they pull back a lot on that direction I think it's a it's a matter of consumers not willing to give away because actually shopping is fun people wants to decide uh their hotel, they wants to decide the color of their pair of shoes like so all those things are kind of very a taste that are personal.
[27:11] So I think that's one of the blocker in terms of technology like the credit card infrastructure is good now like there's protocols that are designed actually to enable those kind of consumer purchases. That's not really the area that we're focused on. We're really focused on like the machine economy, the APIs, agents buying compute, inference or API services. And this is like already there. We're already seeing that aents operating autonomously is really starting and I think the opportunity on that machine economy is so much bigger than on the consumer economy because on
[27:42] the consumer economy it's a market that is existent people are not going to get more money because agents are being on their behalf. However, if you think about the next trillion agents that needs to compute and communicate with the rest of the world this is a massive economy that is starting from zero. >> Makes sense. you're more bullish than on agent tech agent to agent transaction or mission to mission transaction than human to mission. >> Yeah, humans do like shopping. Um there's some good uh you know Alamata
[28:12] Shopify has put out some really good data that for longtail merchants they're like funnel has really changed with AI. um people arrive later in the funnel, they're four to six uh times more likely to convert because they've done all their discovery already inside the LLM and they just show up ready to buy. Um but Walmart when they used uh uh sort of tried to make checkout work inside of chat GPT uh with either Spark or Chat GPT's old version, conversion got worse
[28:43] um than on the.com. And so I mean merchants are going to make the obvious choice there. where's the conversion better on my.com or this other place and on this other place I'm giving up some margin to acquire the customer and the conversion's worse. It's just it's just a bad deal for the merchants as well. So until that changes, which it might um I wonder if we're in sort of this like um WAP phones era of consumer agentic commerce where you know the experience just isn't there yet. So, but but I'm interested just a couple days ago as we
[29:13] record this, we saw Cloudflare launch their uh their wallets for agents. So, what do you think about companies like that and incumbents from the big tech side, the CDN's, you know, um the monetization gateway? How how's all that going to play into this? >> Yeah, that's a that's interesting. Cloudflare are definitely in an interesting position as really being the gate of internet. So, they have the power that not a lot of company have in in this ecosystem. So that's why they put their own stable coins a few months ago and now they're investing into a
[29:44] wallet infrastructure. Obviously I think like uh uh if you sit for us it we're actually working with all those partners because those partner are actually great for the ecosystem. If it's Cloudflare or Stripe or any one of those companies they're actually enabling merchants to sell to a um which is what's lacking today. There's not a lot of merchants that are enabled to be able to sell to agents. On our end, we're the platform where the agents run and we actually want our agents to have access to millions of services. Um, so we actually want uh uh some of those uh uh uh uh
[30:17] players to actually have a big role and enable those merchants. I think it's going to be extremely competitive on the merchant side because everyone is trying to ship their own protocols uh their own wallet and everyone is trying to take a cut on the transaction coming from agents. Um but we'll see where it lands there. >> We shall we shall indeed. Uh look, you you're trying to remove uh anything standing between builders and the next trillion agents. Like what has to be true for those trillion agents to exist? Because I get called in by boards of
[30:48] large organizations all the time and you know they're still trying to figure out how to be better than just using C-pilot. Like is it an education gap? Is it arming the builders themselves? Like what's the gap here? >> Education um on what's possible with the agents. But I think infrastructure that's why actually we exist. It's today if you want to build an aentic product in production or even internal for to your team in a way that it's safe uh that just works and that's cost effective it's extreme because you have to stitch together a lot of
[31:18] infrastructure pieces that were not designed for agents. Um so or you have to build all the stack yourselves or you have to rely on product like sapium but all that ecosystem I think is just new like we've been talking about agents for like two years but actually the day one of agents I would say was kind of open when everyone realized that actually okay we're there and agents can do things and that was just four months ago so I think there is just a matter of timing where in six months we have a very different conversations agents will be deployed at scale much at a much
[31:48] larger scale enterprise will really start to deploy for the agent but I think what's needed for enterprise it's first making sure that the cost in wants to be the agent is not as high as it is today uh that you can control the spend uh and that's what shared uh uh with us it's actually today like the CTOs are becoming CFOs of companies because the cost of inference is extremely high and like the the the cost for a company is massive um and the third is the governance uh audit trails making sure that every agent action is approved
[32:18] making sure that it stay within the boundaries of what a company authorize. So there's a lot of barriers that we're removing over time. Uh and that's why I think we're still a few months away from a massive adoption. >> So uh on on the Cloudflare side um so historically they were like famous for blocking bots and so there was a lot of capture to capture basically bots and typically bad bots. I think it is
[32:49] changing to like prove that you are a bot. So I give you different type of access right and um the way that like we were transacting like human were transacting would just go and fill in a form put your shipping information and you built it as shop pay right so that is very different uh for for agents and how do you see that that uh the agentic commerce would change the interface of how you're browsing the internet for
[33:19] instance you mentioned that people don't want to give away their um their purchase capability. So they they wouldn't authorize their their agent to just go shop on their behalf. But um what if your agent learns from you and every time it comes back with top three options, those are the exact top three options that you pick. At some point you would give away um that uh the last mile delivery or the last part of the purchase transaction to your agent,
[33:51] right? How does that change the interface of the internet as we know today? Yeah, I think for the last mile uh there there are more than just the item selection like what we've seen uh uh from my previous experience is uh people they want to understand the shipping windows u what are the different shipping options they want to understand what are the taxes what are the different upsell because if you go to the website you get actually discounts and like you have bunch of upsell there's an experience that is built on check out that is not present when you let your a buy that people
[34:21] really care about but oh does it show internet. To be honest, I'm not really bullish on um I know there's a lot of investment on a lot of company building like browser use uh and like you know agents browsing the web uh to um do things. Um I think there's a massive market in the short to midterm. I think if you think long term I don't see why an agent needs to access an interface that has been designed for a human to be able to do an action that is provided by a machine like or already communicating
[34:52] with API through an interface. So over time I think the interface for agent would disappear and it would just be machine to machine APIs because those APIs are here on both end it just are not exposed. So there's work for like every service to expose the APIs for the agents and gate them with a payments protocol if it's a paid resources. Um it will take time to adopt. So that's why there's a lot of need today for agents browing the web. I think that's very short to midterm if we think about long term like there's no reason to for an agent to go to website. >> So fascinating. I almost think about
[35:22] Stripe Atlas and Stripe Connect together but for agents, you know, like it's it's the place that you form the new agent identity and then it it's alive but now it needs to do things. So you need this entire dashboard and this entire set of capabilities and uh the ability to build checkouts and carts and optimize like what it's doing and order it and make sure it's performing and make sure things deal with edge cases like returns and refunds. It's like all of that hidden pain that unless you've built in
[35:54] e-commerce before you don't know exists uh also exists for agents too. Uh so I I'm interested in sort of what what the the focus is now. You've got your series A. Um, you know, you've obviously had some transactions on the platform, but who's your ideal client here? Is it the builders of agents? Is it the uh people building uh like the vertical AI agent stacks and them selling out to enterprise? Is it the enterprises themselves? Like where are you where are you aiming here? >> Yeah. So, the opportunity actually is
[36:24] developer first. Like we want to enable the builders building autonomous products because they're the most advanced in their use cases because they're starting from zero. They have no legacy. They can move extremely fast and we are actually enabling them to move much faster and those company are going extremely fast as well because they're quickly replacing incubants with product that are fully autonomous. Uh so on one end we have like the developers the builders. Uh but on the other end we have a lot of value to create and to provide uh to enterprises as well. Um so
[36:54] we have two motions. We already have an enterprise support motion where we deploy engineers inside enterprises to help them adopt AI and deploy agents using Sapium. Uh but we are as well extremely focused on the developer community. We have a very generous feature to actually help builder builders like really build and ship agents to production >> and in order to enable the tribents you need actually to power both end of the ecosystem. Um and that's what we do and that's why we arise this uh this series there. >> Elon this has been an education. What
[37:24] didn't I ask you that um that that I should have asked you? What's the best question you've had recently? >> Maybe an interesting one that is not related about that is not related to payments, but how do we build a company in this new AI world which is a >> top of my front and I'm always even brainstorming with my cloud code on my GP on what's the best way to run uh an AI native company that is enabling trillions of agents as a CEO. Uh and it's a it's a great topic. It's not easy because I want the company to be
[37:54] automated as much as possible and really having uh uh uh uh uh uh people to have eye ownership and run armies of agents. So we still have a big team and I still believe we need a very big team to succeed and to really capture the market because there's so much things to do. Um but the way I design the organization is very different when I was running my previous uh company like a few years back. uh it's less about teams of people working on one product on one problem
[38:25] and the way we design a sepium it's every engineer every team member is owner of a key metric of a key component of the product and then they figure out with AI how to make the right trade-off or to figure out the unknowns uh and actually we scale team horizontally rather than vertically so I still need a lot of people because I have hundreds of business metrics I want to drive on things I want to uh improve on but it's less about the team of 10 working on one thing and like each individual is another. Yeah, I did a report called um
[38:55] the uh the enterprise AI playbook um from studying ramp and block and anthropic and and kind of uh just the the cadences inside the company. the the rituals that you would have seen in 2020 um and what the weekly workflows look like and the job titles have just completely shifted over the last 18 months and in the last 6 months especially if you peer inside these organizations uh what the agentic tooling has done to the day job is is dramatic. So I can only imagine how top
[39:26] of mind that is for you. Well we'll have to get you back and uh talk about that specifically at some point and and see how it's doing. But um that's all the questions I had for now. Um, I want to thank you so much, Alan. If people want to learn more about Sapium and uh get their hands on it, give it a try or they want to speak to you, how do they get hold of you? How do they get hold of Sapium? >> Uh, I'm on X on the uh LinkedIn um and as well of Sapium.ai. You can sign up. We give 50 free agents run a day. We give as well $15 credits
[39:56] for like builders to really experience the product on on cheap agents. Um and we love feedback like there's still a lot of things to uh many barriers to remove and so any barriers that builders are facing we want to know them >> work on them. >> Excellent. Bam. Um if people want to find out more about mesh. >> Yeah. Bam. As easy mesh on X or LinkedIn or mesh pay on X or LinkedIn. >> You'll find me at sytail on all of the socials and at tempo.xyz where we're trying to make machine payments uh an
[40:26] everyday thing. And uh you'll also find a lot more of this show if you go ahead and subscribe. Um you can find it on Apple Podcast, Spotify or on YouTube. And if you leave us a review or a comment, that really helps others find the show. So do engage and I think conversations like this make that more than worthwhile. Thank you very much and we'll catch you next time.
[0:30] more agentic commerce themed and all things agentic and AI. Of course, my name is Simon Taylor. I am your head of market dev at Tempo [music] and author at Fintech Brainfood. And I'm joined by my agentic commerce co-host Bamzy, CEO Mesh. How you doing, Bam? >> I'm doing great. Super excited for this episode. >> Joining us today is a phenomenal guest, Elan Zabib, who is founder and CEO of Sapium. How you doing, sir? >> Amazing. Thanks for uh having us today.
[1:00] >> I'm excited to learn more about what you do because it looks pretty cool. But before we do, I just got to remind uh viewers and listeners that views and opinions of contributions today are their own and might not reflect those of companies they represent. And please don't take anything we say as tax, legal, or financial advice. This is forformational purposes only. And of course, this episode is brought to you by Visa and our friends at Mesh. Stable coin operations usually mean a wallet from one vendor, an on-ramp from another, and then your controls are stitched together across both of them.
[1:30] The Visa stable coin platform fixes this handily. You can mint, move, and manage stable coins, access Open USD, and remain your own custodian all in one single shared environment. Then, stable coin linked cards let you spend balances anywhere Visa is accepted. And we know stable kind linked cards are in hyperrowth as our good friend Kai likes to say. That is Visa of course the global leader in payments and of course sponsor of this show. You can find out
[2:00] more at visa.com/crypto. This episode is brought to you by Mesh, the first global crypto payments network. Connecting over 300 wallets, exchanges, and payments platforms. Mesh enables anyone to pay and get paid instantly anywhere in any asset. Mesh makes digital transactions seamless, secure, and universal, fueling the next era of aentic commerce. Learn more at meshpay.com/ai.
[2:33] Ellen, let's start with you before we get into Sapium because uh you're doing some pretty cool things in AI, but how do you end up uh kind of knowing what you do about payments and AI? What's the story there? What's the background? >> Yeah. So maybe I will start uh uh from from France. So I'm a software engineer. Started building product when I was uh you know 12 or 13. I started my career in Saga security. Uh so I worked a bit for the French department of defense but
[3:04] quickly I moved into building startups. Uh so I joined uh I moved to the US in 2014 build the first company called Ernie. So Ernie was a consumer advocate product. We're actually building uh autonomous agents that were getting refunds to people. So if you buy something on the price drop Ernie was able to get your money back in the first three months. So it was actually millions and millions of bots that were scrapping the web extracting data and actually getting money back to to 5 million users in US. Uh so I faced actually what agents are facing today
[3:34] like 12 years ago. uh where the actually the internet infrastructure has been designed since day one to prevent automation and actually today automation it's kind of something that we want to enable AI agents but nothing works because from like the infrastructure the DNS the like the CDN uh the the fraud models everything actually has been here to block those agents and if you want to enable agents you need to have a proper infrastructure for them uh so I was uh running Ernie for six years uh we then
[4:05] got acquired in 2021 and after the acquisition I spent 5 years at Shopify. So I was director of engineering leading a global product called Shop. So I was leading the shop product the identity layer the trust layer build a product called Shop. I've been very deep into everything that is related to uh payments identity uh and trust uh and uh same problems everything is designed actually to make sure that there's a human between every transaction between every actions and now we're in a world
[4:35] where actually want to enable agents to access the economy to be impactful and there's so many barriers that are blocking them to get to production or be economic actors. So that's kind of the history of Sapion. uh both internet infrastructure and financial infrastructure has been designed to block them. Sapium is designed actually to enable them or to build the right infrastructure for those agents to try >> well and and shop shop pay's great heritage um Kaznet and a good old friend and um I am such a fan of that product and I think the thing it's done really
[5:06] well is dramatically shrink the time it takes to check out with identity solving it really it's an identity product masquerading as a as a payments product that sort of remembers you and remembers your shipping details details remembers your card details across multiple merchants around the internet. And so really curious as to uh what you saw in there that made you start to think about Sapium you 12 months ago or so when when
[5:36] you started that like what what is it at Shop Pay and and your career history that makes you think I need Sapium now. We need to go build this thing. Is it that sort of point you were making there about the like the whole internet is designed to block bots? >> Yeah. And like the idea of sapium actually came out like a bit more than two years ago like actually end 20232 when we had ch that went live uh it's the first time that actually people experience like AI the first time that we actually cracked uh large langu large language models uh and actually that we
[6:07] could get value from AI uh and for me it was very obvious because those models were already very smart uh but they were very smart in a chatbot you can ask any question, they can give you any answer, but it's limited to a prompting on a a chat interface where uh the next step were quite clear actually those models would become smarter. Meaning it's not about just giving you an answer to a question. It's as well being able to reason. You give them a problem and they will figure out how to solve it. Um and then being able to actually do things in
[6:38] the background because at the end an agent what we call an agent before like the LM before a chatbot it's just a piece of software that is smart uh and it doesn't have to live on in a chat interface it can live in a background or scripts like any type of software and then that's where you what that's what we call agent really those piece of software that can uh make autonomous decisions on their own when I started to play with CH GP uh in 2022 that's where I realized actually that when that will be a thing that those agents will be
[7:08] able to operate on their own, everything will break because the internet infrastructure everything has been designed to actually prevent those behaviors from the software. And then it took me a while to find the right moment to build. Uh I think the uh uh it started to really take shape early last year when we started to see more reasoning models, chain of thoughts and like really the early of like early last year we had like a massive step change in in the reasoning models and then in the first quarter we had MCP which is like okay now you have very smart model
[7:39] and they can communicate with other systems. uh and then in May June we started to see payment traders emerging for agent like X42 from Coinbase and since then we had Mastercard that build agent pay we are partner Stripe they build machine payments protocol so there's all the infrastructure pieces that are coming together to enable this new world of trillions of agents operating in the economy and Sapium is the platform to enable that and make that a thing possible. >> Yeah. So what is different between agent
[8:10] because my background is identity too. I want to know your take on agent identity or KYA and end user or consumer identity or KYC. Uh what does it work and what does it doesn't work when you move from end user to agent as an end user? >> Yeah, I think on the agent identity there is like a uh three components. Uh first we have like with the entity behind the agent um it might be a user
[8:40] or it might be a business. So the Q the QC or the QYB. So when we talk about know your agents like agent identity the agent identity needs to carry the consumer or the business identity because when an agent make a purchase decision he needs to prove actually with kind of behind the agent uh making sure that there's delegation of authority. Okay, as a business or as a consumer, I give authority to that agents to act on my behalf. And the third, it's actually what's the reputation of that agent? Like on the merchant side, can I trust
[9:11] this agent? Because I might trust the business or the customer behind, but how do I know that this agent actually is not an agent frog that he has a proper reputation? So there is three components. Um the key way like the the uh uh authorization on the carrying the business identity is quite easy you know it's just we need to have a standard that carry those things the agent reputation is much harder because you need new datas that don't exist actually almost like credit score data uh and that's what we have at Sapium because we enable agents to access the economy we
[9:42] actually monitor and authorize every action might be an API code it might be a transactions everything is actually monitored by Sapium and authorized by the by our platform so we as the agents operate actually we build the credit score of the agent which is kind of the missing piece to provide like the reputation of the agent um and those are amazing topics I think it's still 12 months away before the identity is becoming something that is critical for the machine economy to uh thrive we're not there now the problem is at least
[10:12] getting to production but those are very interesting problems uh for example octa uh they invested to sap through octa ventures those are kind topics that we're thinking as eventually we need a shared identity that is that is standard that is uh recognized by everyone. >> Yeah. So you do have a very interesting cap table. You mentioned octa there um but also like a real mixed bag. You've got um anthropic um but then also dragonfly and coinbase and vanex. So there's like digital asset heritage,
[10:43] there's AI heritage, there's identity heritage. Why that cap table? like what is it that you're sort of thinking by having that group of people on you know backing you >> and Excel as well from the previous round with the long-term infrastructure play uh actually I ced the cap table like uh I'm curating a team uh how we build the best capables to win that uh that market and we really can enable that trillion of agents to exist it's
[11:13] very interesting because the way I see it like uh so dragonfly they led uh our cedar sorry our series A that we just announced uh two days ago. So we raised 35 million series A from Dragonfly and when I had that discussion with AI Beron like his on the machine to machine economy is very similar to artsis. So actually we're an AI infrastructure play. We're not a crypto play but actually behind the scene those SHs are emerging because if you want agents to access the economy most of those transactions are going to probably be
[11:45] powered by stable coins by uh microtransaction by realtime settlements and those are kind of the crypto rails that are enabling that. So there's two side of the agent economy. There is the consumer side an agent buying a flight ticket or a pair of shoes which will most likely rely on existing rails. But if you want an agent that buy in real time an SMS a search request I don't know 1 second or 1 millcond of a compute those are only possible truth stable coins like programmatic money uh that can settle in real time uh and that can
[12:17] have extremely low fees and that's where actually those two are merging it's actually to enable those of agent we need an extremely solid uh crypto like our blockchain ecosystem behind the seal um that is powering that power. >> Fascinating. Um, Bam, you and I have spoken a lot about, you know, sort of where's the logical entry point for stable coins, and I think we've had a lot of, um, stable coin fans say, you know, oh, well, obviously it's it's AI, but the AI folks have always been, I
[12:47] don't know, a little bit anti-crypto. How are you thinking about this one? And what's your reflection? Yeah, my reflection is um to to Elon's point um we will have a lot of mission to mission payments that no matter if you you can hate it, you can love it. What whatever your feeling is, you need to put it aside because the traditional financial tools and means uh will break in that new ecosystem or that new framework or the new modern agentic word. Right? So if an agent wants to pay 0.00005 005
[13:19] cents. You cannot wire, you cannot use Visa card or Mastercard. You have to use a stable coin. If you want to have a global coverage, you have to use a stable coin. If you want to have a through authorization and um and uh add rules and limitations and regulations around what agent can do, what cannot, you need to use a smart contract and blockchain. You really like if you want to do it in a truly global sense, there is no other choice. It's not like I have multiple choice. I just go with what I
[13:50] love. Like no matter if you love it or hate it, that's the only choice you have to use it. And I think a lot of uh anti-crypto type of u uh comments you're hearing from the AI world is mostly around uh crypto as as an asset, not blockchain or stable coin as a technology. So I I think a lot of uh thought leaders in in in agentic commerce and especially in AI uh they're they're actually pro crypto and prostable coin. Look at like for
[14:20] instance Sam Alman they have another company called Ward or the the Ward app probably everyone knows to humanity. So everything is on built on top of blockchain and they're like also thinking about like moving identity on blockchain as well. So I don't think anyone hates blockchain itself. I think it's just more of a like crypto NFT like Bitcoin versus like AI. Which one is cooler and hotter? >> Yeah. And uh we can certainly argue that um the Bitcoin folks help build out a lot of NeoClouds that have become very
[14:50] very useful for for infrance compute at this point with with power. Um but I I'm interested to learn you were sort of talking about there's a future state of identity, but right now you're sort of um my understanding is you're sort of helping people root models really effectively and efficiently. um you've got an agent studio and an agent runtime. Uh I look at the difference between the companies that are shipping faster and getting the AI dividend and everybody else as being the people who you like the the two or five companies that have managed to build this all
[15:21] inhouse. You know, you ramp Block Robin Hood. But is SAPM kind of this for everybody else? you're the platform to build agents, but also um the smart model rooting is a really interesting idea which is like, hey, maybe I don't want to go all in on anthropic or um open AI or pay their fees. I want the the task to go to the right model at the right time. So, how how does all of that work in inside the Sapion platform? >> Yeah, it's a it's a it's a great question. So, yeah, we have like what we announced two days ago. It's really
[15:51] three new three new product that we uh brought to the market. really the agent studio which is getting from an idea to a production ready agents in like minutes. So we have our own opinative frameworks on how agent should be designed and how they should be shipped. Uh so we are actually helping builders uh uh uh thinking about agent product and helping them build the right way agents. Uh then we have the agent runtime which is really the deployment environment the production environment for agents to run. It's kind of their world. We provide the file system the sandbox secrets management. So
[16:22] everything is is related to the turnout to the agent. So when an agent runs they can have their own uh environment which provide infinite scale uh and as part of the environment we actually give access to the agents to the economy. So that's the first product we build which is the governance and then the access layer really giving agent a ledger to buy any type of API services. This is baked into our runtime. So running on sapium actually they they can have access to any type of capabilities in a way that is fully metered and governed. Um and
[16:52] then on top of that that's the what we call the sapium router which really for every agent action every aent steps we decide what's the best services and what's the best model. So we optimize based on cost latency quality of the output. So we are constantly optimizing uh the path of like the agent to make sure that the agent always gets the most out of every action. I love that ledger piece because uh you know you sort of think about uh auditability but also a baseline of knowing what these things are doing and then their ability to move
[17:23] money, their ability to pay for things like that that as a foundation really sets you up for for a future. Um and so many companies are trying to do this at the moment. The other thing I saw like you're running your own inference stack on dedicated GPUs. Why? >> Why? Uh hey [laughter] >> um I believe uh uh inference like inference is key like if you need an agent actually the way we see it is like the intelligence is just a capability like any other one like an agent sending an SMS only needed to run a prompt it
[17:54] needs to get access to a capability um so we integrate that into the platform the capability as well like the inference is kind of preparious so every prompt is mered and authorized actually by the platform uh and it's coming actually from the balance of the agent And why it's because like we need to optimize the cost of intelligence. Uh we've seen with some of our customers uh that we're innovating rebuilding like consumers or B2B product where uh if you default to the frontier models for everything you can't run a business
[18:25] actually you have you go back as soon as you >> negative ROI. >> Yeah. The it's it's it's negative margin by hundreds of percents. Um and we've seen that with uh one of our customers that was mass market consumer product automating like businesses for people to really build businesses. Uh but the cost of inference grew from like zero to like 1.2 million after 3 months of operation. Um they were getting users but they were not at million user scale. they were at like thousands of user scale and just the cost of intelligence when you give
[18:56] autonomy is just too high because it's very different to a product where you prompt because when you print you have a human that decide every actions when you have an autonomy it's an agent that work whenever like 247 that consumes intelligence and that intelligence is actually not scalable so today there's a massive barrier to innovation which is actually the cost of shipping those agents when you're an indivi an individual running Europe and clue you manage your spend and it's fine but actually if it's a product you're trying to sell to thousands of customers you
[19:26] you have to control the inference cost because if not you can't even build your product and ship it to your customers so that's what we've seen and then we say actually like the intelligence is should be part of the platform today we are pushing to every customer's consumer B2B like model selection even if you go to the mainstream consumer app builders they ask you to pick your model no one knows what's the best model so everyone defaults to most expensive one where actually for 95% of the use cases now with the new open weights model we've seen over the
[19:57] last few months actually open weights model can get you the cost of the intelligence two pennies versus dandas uh delivering the same output um and that's what we provide. So the fact that we internalize the inference stack is because we're able to ensure we deliver the same quality of an output for a frontier model leveraging open weights model meaning we are able to optimize the cost of the inference by 60 to 90% providing the same uh uh quality of what the agent is trying to achieve uh because we automate uh we do the
[20:28] routting and we uh have all the data to understand actually what the agent is trying to achieve >> and and I imagine as well this sort of um means that the experience for somebody coming to your platform is a little bit simpler and that they're not having to sign a a contract with open router behind the scenes or b like a bunch of enterprise agreements with all of the labs in order to have that inference time compute or the the the models served to them. You're just you're full stack. You they they sign a
[20:58] deal with you, the agent runs inside the runtime and you take care of everything else below it. So fascinating. >> Yeah. And we really believe that um uh to prove that that we believe we're the best at optimizing the the model is actually we don't even charge for the model. When you run an agent on Sapium, we charge you a flat fee per agent execution and that includes the infrastructure, all the capabilities that your agent needs and the intelligence um because we know that we are actually able to optimize your intelligence to get you the best agent with the best outcome, but every steps
[21:29] along the way are optimized by by our router. >> Yeah. What is it? mostly. So if the unit economics goes down or the unit economics makes sense for you, does it does it make sense for you to move away from owning your own GPUs and your own stack or working with others? >> Um we decided to operate our own uh uh GPUs to uh first have dedicated capacity because early on we tried actually to use serverless uh uh uh open weights
[21:59] models open weights models. Uh there's two problems there. Uh on one end you have like most of the open weights models are not served uh in US or Europe. So you're actually sending the data to a third party countries that most of the users don't want to send their data. uh and on the other end for the only few that are US- based or Euro based there are such a tiny capacity where actually you always facing you have zero resiliency in your inference and most of the time they will not be
[22:29] able to serve your inference because there is massive demand uh and they don't allocate enough GPUs to those models because usually they will rather sell year contract GPUs to companies trying to do funing or model training. Uh so we decided to okay we need to provide that capability which is core for our customers to succeed. We need to make sure that our customers have a margin that is good enough for them to even be able to scale because it doesn't make sense to have customers and then after 3 months they go bankrupt because
[23:00] they pay frontier models. Uh so we decided to internalize that and really had the customer succeed. >> Excellent. >> So fascinating. You don't see you don't get a lot of startups coming out the gate at the moment building sort of uh model routting and uh kind of an agent building stack that also has gone all the way vertically integrated. Um, and you know, Elon's getting applauded for having built uh giant uh stacks and data centers, but actually the pragmatism of what you're doing here with like owning a stack for agent workflows at an ultra
[23:32] low cost is is just like diametrically opposed to that sort of scale and and really pragmatic. Uh, which is which is kind of refreshing in a lot of ways. But talk to me about what people are using Sapium for today like 270 million transactions, 100,000 daily agent runs. What what are all those agents doing? >> Yeah, so we have we have seen the long tails of agents. Uh there is the obvious one like a lot of uh coding agents really building products and shipping products. We have a lot of like go to marketing agents like agents being able
[24:04] to find leads and sales and like customers to businesses or like those agents. uh we even have agents that run uh end to end ads for customers uh Instagram on Facebook and all those ad are fully automated by some of the agents running our platform which is fascinating. So customers are giving a budget and then the agent just figure out the context of the company build an asset a video with models and then is able to buy actually ads on ads platforms uh and then close the loop on performance. So it's really the long
[24:35] tail. We have a lot of agents doing customer support, chat. Um, we are really the platform that enables builders to build and innovate and scale. So, we are seeing like any type of uh uh uh agents getting into production on our platform. >> So, we alluded a little bit earlier on to um sort of stable coins and aentic commerce and and today like the volume's super small and super early, but you know, you've talked about identity, you talked about a 12-month time horizon. When does Aentic commerce start to get
[25:05] real in your mind, Elon? >> I don't think we're that far. Uh because as as you shared like the numbers, we're already seeing like meaningful volume. Um I still think it's very tiny like the main barriers people are facing today cost and even getting agents to production. Uh so those are the things that we are solving and that's why we are here to enable the next trillion agents because we believe we give enough advantage for builders that build on our platform by abstracting all the things that are needed to actually build the product but that economy is going to be
[25:35] massive like we're talking about trillions of dollars in the machine economy in the next 3 to 5 years uh and we are like a fraction of that. So uh but when I think we're like six to 12 months away from that market to really uh explode. So you know like Simon is the bad cop I'm the good cop who's more realist and more optimist. So and we have a $1 bet what's going to happen in next 12 months. So what is the hold up for like big companies let's say
[26:08] uh entropic that is on your cap table or open AAI or others to enable their agents to pay with just credit card and what is the hold up there why it's not happening there is no hold up actually the technologies are there uh even Shopify they partner with open AI to deploy like shopping to open AI so uh agents can actually transact there's protocols like GCP from Google there's many protocols that are enabling that the hold up is actually the consumers people they don't really want to uh
[26:39] delegate purchase decision for most of the things um and you know there were a lot of learnings from like the ecosystem on the market on that when CHP was able to buy things on behalf of consumers uh actually they pull back a lot on that direction I think it's a it's a matter of consumers not willing to give away because actually shopping is fun people wants to decide uh their hotel, they wants to decide the color of their pair of shoes like so all those things are kind of very a taste that are personal.
[27:11] So I think that's one of the blocker in terms of technology like the credit card infrastructure is good now like there's protocols that are designed actually to enable those kind of consumer purchases. That's not really the area that we're focused on. We're really focused on like the machine economy, the APIs, agents buying compute, inference or API services. And this is like already there. We're already seeing that aents operating autonomously is really starting and I think the opportunity on that machine economy is so much bigger than on the consumer economy because on
[27:42] the consumer economy it's a market that is existent people are not going to get more money because agents are being on their behalf. However, if you think about the next trillion agents that needs to compute and communicate with the rest of the world this is a massive economy that is starting from zero. >> Makes sense. you're more bullish than on agent tech agent to agent transaction or mission to mission transaction than human to mission. >> Yeah, humans do like shopping. Um there's some good uh you know Alamata
[28:12] Shopify has put out some really good data that for longtail merchants they're like funnel has really changed with AI. um people arrive later in the funnel, they're four to six uh times more likely to convert because they've done all their discovery already inside the LLM and they just show up ready to buy. Um but Walmart when they used uh uh sort of tried to make checkout work inside of chat GPT uh with either Spark or Chat GPT's old version, conversion got worse
[28:43] um than on the.com. And so I mean merchants are going to make the obvious choice there. where's the conversion better on my.com or this other place and on this other place I'm giving up some margin to acquire the customer and the conversion's worse. It's just it's just a bad deal for the merchants as well. So until that changes, which it might um I wonder if we're in sort of this like um WAP phones era of consumer agentic commerce where you know the experience just isn't there yet. So, but but I'm interested just a couple days ago as we
[29:13] record this, we saw Cloudflare launch their uh their wallets for agents. So, what do you think about companies like that and incumbents from the big tech side, the CDN's, you know, um the monetization gateway? How how's all that going to play into this? >> Yeah, that's a that's interesting. Cloudflare are definitely in an interesting position as really being the gate of internet. So, they have the power that not a lot of company have in in this ecosystem. So that's why they put their own stable coins a few months ago and now they're investing into a
[29:44] wallet infrastructure. Obviously I think like uh uh if you sit for us it we're actually working with all those partners because those partner are actually great for the ecosystem. If it's Cloudflare or Stripe or any one of those companies they're actually enabling merchants to sell to a um which is what's lacking today. There's not a lot of merchants that are enabled to be able to sell to agents. On our end, we're the platform where the agents run and we actually want our agents to have access to millions of services. Um, so we actually want uh uh some of those uh uh uh uh
[30:17] players to actually have a big role and enable those merchants. I think it's going to be extremely competitive on the merchant side because everyone is trying to ship their own protocols uh their own wallet and everyone is trying to take a cut on the transaction coming from agents. Um but we'll see where it lands there. >> We shall we shall indeed. Uh look, you you're trying to remove uh anything standing between builders and the next trillion agents. Like what has to be true for those trillion agents to exist? Because I get called in by boards of
[30:48] large organizations all the time and you know they're still trying to figure out how to be better than just using C-pilot. Like is it an education gap? Is it arming the builders themselves? Like what's the gap here? >> Education um on what's possible with the agents. But I think infrastructure that's why actually we exist. It's today if you want to build an aentic product in production or even internal for to your team in a way that it's safe uh that just works and that's cost effective it's extreme because you have to stitch together a lot of
[31:18] infrastructure pieces that were not designed for agents. Um so or you have to build all the stack yourselves or you have to rely on product like sapium but all that ecosystem I think is just new like we've been talking about agents for like two years but actually the day one of agents I would say was kind of open when everyone realized that actually okay we're there and agents can do things and that was just four months ago so I think there is just a matter of timing where in six months we have a very different conversations agents will be deployed at scale much at a much
[31:48] larger scale enterprise will really start to deploy for the agent but I think what's needed for enterprise it's first making sure that the cost in wants to be the agent is not as high as it is today uh that you can control the spend uh and that's what shared uh uh with us it's actually today like the CTOs are becoming CFOs of companies because the cost of inference is extremely high and like the the the cost for a company is massive um and the third is the governance uh audit trails making sure that every agent action is approved
[32:18] making sure that it stay within the boundaries of what a company authorize. So there's a lot of barriers that we're removing over time. Uh and that's why I think we're still a few months away from a massive adoption. >> So uh on on the Cloudflare side um so historically they were like famous for blocking bots and so there was a lot of capture to capture basically bots and typically bad bots. I think it is
[32:49] changing to like prove that you are a bot. So I give you different type of access right and um the way that like we were transacting like human were transacting would just go and fill in a form put your shipping information and you built it as shop pay right so that is very different uh for for agents and how do you see that that uh the agentic commerce would change the interface of how you're browsing the internet for
[33:19] instance you mentioned that people don't want to give away their um their purchase capability. So they they wouldn't authorize their their agent to just go shop on their behalf. But um what if your agent learns from you and every time it comes back with top three options, those are the exact top three options that you pick. At some point you would give away um that uh the last mile delivery or the last part of the purchase transaction to your agent,
[33:51] right? How does that change the interface of the internet as we know today? Yeah, I think for the last mile uh there there are more than just the item selection like what we've seen uh uh from my previous experience is uh people they want to understand the shipping windows u what are the different shipping options they want to understand what are the taxes what are the different upsell because if you go to the website you get actually discounts and like you have bunch of upsell there's an experience that is built on check out that is not present when you let your a buy that people
[34:21] really care about but oh does it show internet. To be honest, I'm not really bullish on um I know there's a lot of investment on a lot of company building like browser use uh and like you know agents browsing the web uh to um do things. Um I think there's a massive market in the short to midterm. I think if you think long term I don't see why an agent needs to access an interface that has been designed for a human to be able to do an action that is provided by a machine like or already communicating
[34:52] with API through an interface. So over time I think the interface for agent would disappear and it would just be machine to machine APIs because those APIs are here on both end it just are not exposed. So there's work for like every service to expose the APIs for the agents and gate them with a payments protocol if it's a paid resources. Um it will take time to adopt. So that's why there's a lot of need today for agents browing the web. I think that's very short to midterm if we think about long term like there's no reason to for an agent to go to website. >> So fascinating. I almost think about
[35:22] Stripe Atlas and Stripe Connect together but for agents, you know, like it's it's the place that you form the new agent identity and then it it's alive but now it needs to do things. So you need this entire dashboard and this entire set of capabilities and uh the ability to build checkouts and carts and optimize like what it's doing and order it and make sure it's performing and make sure things deal with edge cases like returns and refunds. It's like all of that hidden pain that unless you've built in
[35:54] e-commerce before you don't know exists uh also exists for agents too. Uh so I I'm interested in sort of what what the the focus is now. You've got your series A. Um, you know, you've obviously had some transactions on the platform, but who's your ideal client here? Is it the builders of agents? Is it the uh people building uh like the vertical AI agent stacks and them selling out to enterprise? Is it the enterprises themselves? Like where are you where are you aiming here? >> Yeah. So, the opportunity actually is
[36:24] developer first. Like we want to enable the builders building autonomous products because they're the most advanced in their use cases because they're starting from zero. They have no legacy. They can move extremely fast and we are actually enabling them to move much faster and those company are going extremely fast as well because they're quickly replacing incubants with product that are fully autonomous. Uh so on one end we have like the developers the builders. Uh but on the other end we have a lot of value to create and to provide uh to enterprises as well. Um so
[36:54] we have two motions. We already have an enterprise support motion where we deploy engineers inside enterprises to help them adopt AI and deploy agents using Sapium. Uh but we are as well extremely focused on the developer community. We have a very generous feature to actually help builder builders like really build and ship agents to production >> and in order to enable the tribents you need actually to power both end of the ecosystem. Um and that's what we do and that's why we arise this uh this series there. >> Elon this has been an education. What
[37:24] didn't I ask you that um that that I should have asked you? What's the best question you've had recently? >> Maybe an interesting one that is not related about that is not related to payments, but how do we build a company in this new AI world which is a >> top of my front and I'm always even brainstorming with my cloud code on my GP on what's the best way to run uh an AI native company that is enabling trillions of agents as a CEO. Uh and it's a it's a great topic. It's not easy because I want the company to be
[37:54] automated as much as possible and really having uh uh uh uh uh uh people to have eye ownership and run armies of agents. So we still have a big team and I still believe we need a very big team to succeed and to really capture the market because there's so much things to do. Um but the way I design the organization is very different when I was running my previous uh company like a few years back. uh it's less about teams of people working on one product on one problem
[38:25] and the way we design a sepium it's every engineer every team member is owner of a key metric of a key component of the product and then they figure out with AI how to make the right trade-off or to figure out the unknowns uh and actually we scale team horizontally rather than vertically so I still need a lot of people because I have hundreds of business metrics I want to drive on things I want to uh improve on but it's less about the team of 10 working on one thing and like each individual is another. Yeah, I did a report called um
[38:55] the uh the enterprise AI playbook um from studying ramp and block and anthropic and and kind of uh just the the cadences inside the company. the the rituals that you would have seen in 2020 um and what the weekly workflows look like and the job titles have just completely shifted over the last 18 months and in the last 6 months especially if you peer inside these organizations uh what the agentic tooling has done to the day job is is dramatic. So I can only imagine how top
[39:26] of mind that is for you. Well we'll have to get you back and uh talk about that specifically at some point and and see how it's doing. But um that's all the questions I had for now. Um, I want to thank you so much, Alan. If people want to learn more about Sapium and uh get their hands on it, give it a try or they want to speak to you, how do they get hold of you? How do they get hold of Sapium? >> Uh, I'm on X on the uh LinkedIn um and as well of Sapium.ai. You can sign up. We give 50 free agents run a day. We give as well $15 credits
[39:56] for like builders to really experience the product on on cheap agents. Um and we love feedback like there's still a lot of things to uh many barriers to remove and so any barriers that builders are facing we want to know them >> work on them. >> Excellent. Bam. Um if people want to find out more about mesh. >> Yeah. Bam. As easy mesh on X or LinkedIn or mesh pay on X or LinkedIn. >> You'll find me at sytail on all of the socials and at tempo.xyz where we're trying to make machine payments uh an
[40:26] everyday thing. And uh you'll also find a lot more of this show if you go ahead and subscribe. Um you can find it on Apple Podcast, Spotify or on YouTube. And if you leave us a review or a comment, that really helps others find the show. So do engage and I think conversations like this make that more than worthwhile. Thank you very much and we'll catch you next time.