theCUBE + NYSE Wired: AI Agent Conference | AGNT Podcast Ep. 1 with Gemma Allen & Raphaëlle d'Ornano

AGNT Podcast Ep. 1 with Gemma Allen & Raphaëlle d'Ornano

Irina Denisenko, chief executive officer of Knox Systems, features in this episode of theCUBE's Mixture of Experts series in partnership with NYSE Wired. As part of the lead-up to the AI Agent Conference, Denisenko shares the journey behind Knox Systems and discusses emerging trends in cloud security and artificial intelligence (AI). Hosted by Gemma Allen, this episode sheds light on applications of AI and cloud innovations in a rapidly evolving technological landscape. In an engaging session with Gemma Allen of theCUBE, Irina Denisenko details their path to establishing Knox Systems as a pioneer in expedited Federal Risk and Authorization Management Program (FedRAMP) cloud services. Denisenko's experience as a co-founder of Class.com propels them to tackle government complexities, providing crucial insights into cloud security standards and the importance of minimizing time and cost investments for Software as a Service (SaaS) companies. The discussion highlights theCUBE Research’s ongoing exploration of AI and technology developments guided by trusted hosts. Denisenko notes that the key to navigating structural challenges in achieving FedRAMP compliance lies in leveraging AI-driven cloud management. They emphasize the critical nature of real-time monitoring and rapid responsiveness in cloud security, which Knox Systems uniquely delivers to its prominent clients such as Adobe and Armis. The conversation underscores the strategic significance of Knox Systems’ services, empowering SaaS vendors with cutting-edge technology and facilitating secure government collaborations.

Share this session

Clips
More from theCUBE + NYSE Wired: AI Agent Conference

Gemma Allen

Host, theCUBE + NYSE Wired

theCUBE

Raphaelle d'Ornano

Founder & CEO

Decoding Discontinuity

play_circle_outline AGNT podcast focus: enterprise tech meeting agentic AI era

play_circle_outline Model Context Protocol (MCP) and Salesforce's Agentforce: Accelerating Agentic Development and Enterprise CRM Adoption

play_circle_outline Agentic AI: How Human-in-the-Loop Systems Turn Company Assets into Workflow, Revenue, and Margin Superpowers

play_circle_outline Context, Not Raw Data: Grounded Memory as the Defensible Moat for Agentic Systems

play_circle_outline Berkeley study: closed models dominate agentic AI deployments in production

Info

AGNT Podcast Ep. 1 with Gemma Allen & Raphaëlle d'Ornano

Gemma Allen

Host, theCUBE + NYSE Wired theCUBE

HOST

Raphaelle d'Ornano

Founder & CEO Decoding Discontinuity

search

Gemma Allen

>> Welcome to AGNT, the podcast where enterprise tech meets the agentic era. I'm Gemma Allen, joined by my co-host, Raphaelle d'Ornano, broadcasting from the New York Stock Exchange. In every episode we unpack how intelligent systems are reshaping companies, markets, and the way real work gets done. From Fortune 500 boardrooms to breakout upstarts, we're digging into the strategies, technologies, and people to find the next chapter of AI. Let's get into it. Raphaelle, so excited to do this with you.

Raphaelle d'Ornano

>> I'm really excited to be here also. Thank you.

Gemma Allen

>> So first off, just to give listeners an understanding, we actually met at Dreamforce in October, an Irish woman and a French woman at a tech conference in San Francisco. And you had had a very interesting morning where you had some very good thoughts on the Salesforce ticker, right? And how, I guess, what it means, what CRM means in 2025. Let's chat a little bit about that. Let's kind of fill people in on the kind of work you've been doing and the thought leadership you have been driving in agentic readiness and how that has, I guess, brought us to this point.

Raphaelle d'Ornano

>> Sure. Well, look, I've been passionate about the field ever since all of this craziness or non-craziness started more than three years ago. But one year ago when Entropic introduced the MCP, the model context protocol, I saw that this was going to go into another dimension. And over this past year, not over the past three years, but over this past year, we have really assisted to this Agentic build out by which the new paradigm of Agentic AI is the one that we are entering. And I think Salesforce really has had the vision. And when Mark decided to go all in into Agentforce at last year's Dreamforce, not 2025 or 2024, I saw that was really interesting, the move that the company was taking. So of course, going in that year, so when we met in this edition, it was super interesting to see how one year into agentic AI, Salesforce had really taken the curve and started positioning themselves to be the adopters and the enablers of what is a fascinating tech transformation.

Gemma Allen

>> Yeah. I mean, it's fascinating, right? Because it's a company that's been an industry leader for so long. The world of enterprise tech is moving so fast and changing so fast. And there's a lot of uncertainty, I think, around what it will mean to be an enterprise player 10 years from now. So, I think the conversation is so incredibly timely. But let's talk about AGNT and what it means. I love it, I love how much you've owned it as well. It's this idea that if we were to have an ETF set up around agentic readiness, we would call it AGNT. And who would be the winners and losers in that space, right? And you definitely have some strong thoughts on who's doing well and who's going to be on a game of catch up. But talk to me a little bit about that too. Explain to the audience how you think about things in terms of Fortune 500 meets agentic readiness.

Raphaelle d'Ornano

>> Sure. Well, so we're in this new paradigm. In this new paradigm by which we have a new technology, which is systems of actions are not possible, we're very far from autonomous actions. And first, we must not confuse agentic AI with, oh, things are fully autonomous running by themselves. No. Thank goodness humans are in the loop and agentic systems are really starting to be built out. So in any case, we have this reality. The question now is, how do companies, tech and non-tech, take their superpowers from the current paradigm and translate that into this new world? How do companies take context, data, physical infrastructure, customer relationships, trust everything that they have built over the last years, decades, et cetera, and turn those modes into superpowers that will unveil and allow them to become the super companies of this new Agentic era, in which systems of action will be one of the defining characteristics, enabling not only strong productivity gains and hence margin improvements, CapEx improvements, but also enabling new revenue sources in ways that I think we have not at all fully imagined right now? But by which companies are going to be able to redesign a certain number of workflows, including their customer workflows, customer facing workflows, and completely redefine how they can serve their customers, monetize that, and doing so, revisit their business model from the top down? And I think that's the exciting opportunity. And the technological state that we are in right now is unfolding at such quick speed-

Gemma Allen

>> Incredible....

Raphaelle d'Ornano

>> that it's just being built day by day. And I'm super excited to see how that transformation is unfolding and it's going really fast.

Gemma Allen

>> So, let's just talk about MCB for a second, right? Because we hear all the time, "You have to get your data in order, you're not agentic ready unless you have your data set up exactly as it needs to be to feed these models.", et cetera. But we also know that the intention and the want to have better streamlined data, it's not like it was a matter of will, right? There was a whole lot of technical roadblocks that happened across all industries for a very long time that made it very difficult for folks to really create effective usage of data full stuff. How has that fundamentally just changed overnight? I think that's what confuses me in the enterprise tech conversation. Is it the companies that are like doing this and creating solutions for that, that are going to kind of be like industry and category leaders? Or how is it suddenly a magical problem that can be solved?

Raphaelle d'Ornano

>> So I think first there's a semantic, I would say, confusion that is allowing some companies to claim they have moat when they don't and vice versa. So in the sense that data is not the moat, I've said that and have surprised many times, context is the moat. What do I mean by that? An agent has to take a decision and for that, it's kind of like a body with a brain. The body needs to have a brain or the body will be moving in all kinds of directions. For an agent to be able to have that intelligence, they need to have access to certain kind of data in the right way, providing them the context. So, what a company needs to actually make agentic work is to have the context that will ground the decisions that the agents are making. That context is deep data, defensible data, data that is not portable from one system to another, like actual data that you have that is structured in the right way with the right memory systems, short term, long term, episodic, semantic, whatever kinds of memory systems, and there are lots are being built right now. And it's actually super complicated. So, an agent cannot work if it's not grounded in the right context. That context does encapsulate data, but data alone is not sufficient. So, the MCP has allowed to solve the fact that these agents are now grounded and have access in a secure way to the data that they need. And I think this is what is being figured out right now. But it's moving. I mean, there's a lot of moving parts in this discussion.

Gemma Allen

>> Okay. So we want to get to a point, and we will get to a point where every two weeks you're going to come on. We're going to record this episode together and we're going to talk about particular companies, right? What's happening in the market, leaders, winners, you know, warts and all in terms of what's happening in the world of Agentic AI. Today I think we're going to talk a little bit about MiniMax because this was obviously a fascinating IPO, a company that went from not to a hundred in what seemed like the speed of sound. They doubled their valuation on day one of their listing on the Hong Kong Stock Exchange. Let's break it down a little bit. Tell me, what are your thoughts on this company? Fascinating player. What does it mean in the race too, right, in terms of the Sputnik moment between the U.S. and China? Give me your thoughts.

Raphaelle d'Ornano

>> So MiniMax is first, I think, a fascinating IPO because it is actually the first LLM IPO with Zhipu, who is the other Chinese competitor. They happened to IPO the same week, so beginning of January. Both have had tremendous success so far on the Hong Kong Stock Exchange where they have listed. But they are the first LLM IPOs. So I think, I mean, we're entering 2026. This is the year where Entropic is expected to IPO, probably OpenAI, though I think later. But anyways, we should have one big U.S. household name IPOding, maybe two. Databricks is also said to IPO. I mean, there's a lot of rumors, so this should be a plentiful year in terms of IPOs. But the first ones that actually went to the public markets were Chinese companies. Now this comes at a time where there are very strong, of course, geopolitical tensions where Chinese open source, which used to be an object of curiosity, is not that anymore and is now, I mean, a core player in this whole market. And so, it comes at an intriguing moment. Now, if we look specifically at MiniMax, MiniMax is having a play by which they're deliberately not playing the frontier game. Their game is, we want to be good enough to be the leading coding agent with their latest model M2.1. And of course, you can debate around what is good enough, but their claim is this is the model that needs to have the most widespread adoption amongst coders for us to lead the Agentic AI revolution with coding as the wedge. They also do have a consumer mode where this is not a pure B2B enterprise LLM. They also have actually a very successful consumer mode with their applications, which are very dominant in the Chinese market right now, and which may lead to, I mean, a mode in itself. But I'm interested in the companies through the coding wedge because we know over 2025 that Entropic, for example, has had tremendous success with this-

Gemma Allen

>> 40... I think they went from 25 to 40% in terms of enterprise adoption, right? Especially for Claude.

Raphaelle d'Ornano

>> I mean, it's huge.

Gemma Allen

>> It's huge.

Raphaelle d'Ornano

>> They're like the adoption of Claude Code, I mean the panic that was created in the markets just last week with Claude Cowork that is directly built by Claude Code. So MiniMax, we're going to have to follow what they do from, again, number one, in aGentic AI workload, are they used as a dominant model in which they are super good? One of the reasons being that they are multimodal from the start. Now, even though this company has very strong attributes, and I think they're a strong contender, and the valuation, not the valuation, the opinion I gave on them was quite a positive one, which is not always the case. Usually I tend to bash a lot of the companies.

Gemma Allen

>> I love critical women. Raphaelle, love it.

Raphaelle d'Ornano

>> I mean, when I actually did the scoring of MiniMax, I was like, "Okay, this is a super interesting company. It doesn't get the best score, but it has a very good score, better than some other companies that I will not name for the purpose of this show, but that are widely known companies operating in the same space."

Gemma Allen

>> You will not name yet for the purpose of the show.

Raphaelle d'Ornano

>> Right. So nonetheless, there are two things that I think warrant caution on MiniMax for investors, of course. The first one is that they're competitor, DeepSeek, which shook the world just one year ago, by the way, is coming out in February with a new model that is built on a new training technique that is gaining a lot of interest. And which has been the object of very profound research. So, they issued a paper just before the Christmas break showing how their mechanism was going to enable training faster by orders of magnitude. While MiniMax has been very efficient on the inference side, but has not fully cracked the lightning attention mechanism that had been at the core of their research back in January of last year, they kind of like backtracked on that. And then the model they released the M2.1 is, I would say, I mean, it's not nice to say that, but in some way it's kind of a patchwork. It's like very good, but it's far from what DeepSeek could actually be proving to us next month.

Gemma Allen

>> Let's talk a little bit about the whole M2 mixture of experts multimodal model, right? Because is it a safe play to claim to be above average or whatever language they use are good enough at everything? Antropic with Claude has obviously become very enterprise bound, right? They've developed something that's very well governed, gives enterprise like a feeling of safety and security. OpenAI is a little bit all things to all people right now. It's an underlying frontier model though for so many others. Google Gemini, Google's already so naturally sticky, right? It's so easily integrated in terms of how you can imagine Gemini just becoming another tool that you use in your day to day lives. What do you think is the safest best for models like that? Do you need to be the best at one thing in 10 years from now or what do you think?

Raphaelle d'Ornano

>> So I would say your question brings two different, like has two sub questions and I'm going to answer both of them. So number one, in terms of what is good in the enterprise, I think to the advantage of closed U.S. models, there's a super interesting study by UC Berkeley that was released in December called Measuring Agents and Production that shows that in a dominant way, agentic AI deployments, over 300 of them, are being done with closed models that are sufficiently good to be not even fine-tuned or with any other techniques that are prompted well and deployed as a genetic AI use cases. So, that's a super interesting finding. And that's Anthropic, that's open AI. Developers, many will prefer like open source techniques. There's much more complexity to that, but today what the study from Berkeley shows is that this was not the dominant way of doing. So, I think that brings a perspective on open source is not necessarily winning. I mean, the game is completely on and to be honest, I myself have no idea. I mean, again, let's see how this unfolds. But that's on the customer side, on the adoption side. The second question that is a key question is, when will these companies and how will they get to like any kind of path to profitability? Again, this is an investment case. If I put a hundred million or more in OpenAI, Anthropic, MiniMax, Zhipu do I get my money back and how? I mean, from like, does this company actually get profitable at one point? I think the U.S. companies haven't cracked that yet. The margins are improving very largely. For example, gross margin and Tropic has shown very strong advancements. I mean, training compute is still an absolute huge bucket of expense, so let's see. I think where the Chinese are interesting, it's that they're showing that there is a path to better dollar efficiency that is already there. And again, if DeepSeek progresses as expected, that gap in training could be very massive.

Gemma Allen

>> And Raphaelle, how do you think right now the financial instruments are looking at this? If you think about Facebook back in the day where it was a $20 per user, right? It's turned out to be something way bigger than that. I think it's like $200 per user or whatever, but the assumption was it was a $20 per user model that Facebook make on mass usage of the product. What do you think is being used to measure the opportunity and the financial probability of these LLMs? Do you think it's the same? It's based on just mass scale?

Raphaelle d'Ornano

>> Look, I think today that... So when you're investing in an LLM as a public investor, you're investing in optionality. I think that investing in a reasonable case by which you have a path to growth, path to profitability that is well sketched out, I don't believe that one second. I obsess around the moat of these companies. I try to like, again, make sense of that. There are too many unknowns to have the slightest idea of... I mean, you can have ideas, but you cannot do like a regular DCF as if you were assessing a classic company, even like a classic high growth technology company. So, you're buying an optionality here. And you're kind of asking yourself, is there like a single point of failure that is going to completely blow up my investment thesis if I invest in company A, B, or C? Is there a sufficiently high probability that this thing could completely blow up, and that I would lose everything even though things go well? I think you're more looking at from that perspective. But no investor today has... It's done that no investor knows, it's that we do not know because from a research perspective, things are being built out. Again, how this is going to play out in terms of, how do you monetize, how is the product adopted both in the consumer and enterprise market? What is your inference efficiency? What is your compute trading efficiency? Not even talking about the rest of the cost structure, too many unknowns. So right now, MiniMax, I think is a play with strong optionality. It's going well. I hope it continues well for them. But it's not like paradoxically, it's not negative too much for the U.S. companies because of that study from Berkeley that brings a light of perspective, I would say.

Gemma Allen

>> So before we talk about... I want to finish up, I want to talk about what it means to really analyze the moat, right? I know you said earlier it's not data. But before we go there, just quickly let's touch on the decision to go public by MiniMax in such a short space of time, right? Do you think that's a CapEx play, a credibility play, funded by Alibaba, the sovereign wealth fund of UAE? What are your thoughts? Why do you think they decided to do that so quickly?

Raphaelle d'Ornano

>> I mean, first I think, I mean, these companies are burning a lot of cash. So the public markets provide, again, this opportunity of financing. I think the public markets are starting to be ripe to, okay, this is a new paradigm. Again, like there's the agentic paradigm, there's the AI companies that enable the whole build out. These companies could be like massive companies and you're getting that optionality, so it depends on exposure. But I think there is an interest from public market investors now to really go into models that three years ago, were seen even by some big private investors as like, oh, there will be no more value. There will be no value in LLMs. I want to be three years back. Remember, companies, well-known VCs in this country were saying, "Oh, we should not be investing in LLMs. Applications is the only layer that is going to capture value from all of this build out." Some VCs are in trouble for having said that because I mean, there's a lot of value in the LLMs right now, hopefully for a long time. So I think they're consuming a lot of cash. There's a window that is right. Let's see how this plays out. And again, this is not even dependent on the companies themselves. This is dependent on do they actually get access to the chips that allow them to train their models efficiently? I mean, this is where geopolitics comes into the picture. This is a super complicated game, but it depends on your exposure.

Gemma Allen

>> Okay. Well, listen, Raphaelle, so excited to have you and to get to talk to you. Like I said, biweekly, we're going to be talking about different companies, what's happening in the markets, Captain Markets meets Silicon Valley, breaking it down from the perspective of AGNT, checking just how ready the Fortune 500 is for the world of Agentic. And excited to see what unravels.

Raphaelle d'Ornano

>> Super excited.

Gemma Allen

>> Thanks so much.

Raphaelle d'Ornano

>> Thank you.

Gemma Allen

>> Okay, folks. Tune in two more weeks when Raphaelle and I catch up again. This is Gemma Allen from TheCUBE. Thanks so much for watching.

AGNT Podcast Ep. 1 with Gemma Allen & Raphaëlle d'Ornano

search

Gemma Allen

>> Welcome to AGNT, the podcast where enterprise tech meets the agentic era. I'm Gemma Allen, joined by my co-host, Raphaelle d'Ornano, broadcasting from the New York Stock Exchange. In every episode we unpack how intelligent systems are reshaping companies, markets, and the way real work gets done. From Fortune 500 boardrooms to breakout upstarts, we're digging into the strategies, technologies, and people to find the next chapter of AI. Let's get into it. Raphaelle, so excited to do this with you.

Raphaelle d'Ornano

>> I'm really excited to be here also. Thank you.

Gemma Allen

>> So first off, just to give listeners an understanding, we actually met at Dreamforce in October, an Irish woman and a French woman at a tech conference in San Francisco. And you had had a very interesting morning where you had some very good thoughts on the Salesforce ticker, right? And how, I guess, what it means, what CRM means in 2025. Let's chat a little bit about that. Let's kind of fill people in on the kind of work you've been doing and the thought leadership you have been driving in agentic readiness and how that has, I guess, brought us to this point.

Raphaelle d'Ornano

>> Sure. Well, look, I've been passionate about the field ever since all of this craziness or non-craziness started more than three years ago. But one year ago when Entropic introduced the MCP, the model context protocol, I saw that this was going to go into another dimension. And over this past year, not over the past three years, but over this past year, we have really assisted to this Agentic build out by which the new paradigm of Agentic AI is the one that we are entering. And I think Salesforce really has had the vision. And when Mark decided to go all in into Agentforce at last year's Dreamforce, not 2025 or 2024, I saw that was really interesting, the move that the company was taking. So of course, going in that year, so when we met in this edition, it was super interesting to see how one year into agentic AI, Salesforce had really taken the curve and started positioning themselves to be the adopters and the enablers of what is a fascinating tech transformation.

Gemma Allen

>> Yeah. I mean, it's fascinating, right? Because it's a company that's been an industry leader for so long. The world of enterprise tech is moving so fast and changing so fast. And there's a lot of uncertainty, I think, around what it will mean to be an enterprise player 10 years from now. So, I think the conversation is so incredibly timely. But let's talk about AGNT and what it means. I love it, I love how much you've owned it as well. It's this idea that if we were to have an ETF set up around agentic readiness, we would call it AGNT. And who would be the winners and losers in that space, right? And you definitely have some strong thoughts on who's doing well and who's going to be on a game of catch up. But talk to me a little bit about that too. Explain to the audience how you think about things in terms of Fortune 500 meets agentic readiness.

Raphaelle d'Ornano

>> Sure. Well, so we're in this new paradigm. In this new paradigm by which we have a new technology, which is systems of actions are not possible, we're very far from autonomous actions. And first, we must not confuse agentic AI with, oh, things are fully autonomous running by themselves. No. Thank goodness humans are in the loop and agentic systems are really starting to be built out. So in any case, we have this reality. The question now is, how do companies, tech and non-tech, take their superpowers from the current paradigm and translate that into this new world? How do companies take context, data, physical infrastructure, customer relationships, trust everything that they have built over the last years, decades, et cetera, and turn those modes into superpowers that will unveil and allow them to become the super companies of this new Agentic era, in which systems of action will be one of the defining characteristics, enabling not only strong productivity gains and hence margin improvements, CapEx improvements, but also enabling new revenue sources in ways that I think we have not at all fully imagined right now? But by which companies are going to be able to redesign a certain number of workflows, including their customer workflows, customer facing workflows, and completely redefine how they can serve their customers, monetize that, and doing so, revisit their business model from the top down? And I think that's the exciting opportunity. And the technological state that we are in right now is unfolding at such quick speed-

Gemma Allen

>> Incredible....

Raphaelle d'Ornano

>> that it's just being built day by day. And I'm super excited to see how that transformation is unfolding and it's going really fast.

Gemma Allen

>> So, let's just talk about MCB for a second, right? Because we hear all the time, "You have to get your data in order, you're not agentic ready unless you have your data set up exactly as it needs to be to feed these models.", et cetera. But we also know that the intention and the want to have better streamlined data, it's not like it was a matter of will, right? There was a whole lot of technical roadblocks that happened across all industries for a very long time that made it very difficult for folks to really create effective usage of data full stuff. How has that fundamentally just changed overnight? I think that's what confuses me in the enterprise tech conversation. Is it the companies that are like doing this and creating solutions for that, that are going to kind of be like industry and category leaders? Or how is it suddenly a magical problem that can be solved?

Raphaelle d'Ornano

>> So I think first there's a semantic, I would say, confusion that is allowing some companies to claim they have moat when they don't and vice versa. So in the sense that data is not the moat, I've said that and have surprised many times, context is the moat. What do I mean by that? An agent has to take a decision and for that, it's kind of like a body with a brain. The body needs to have a brain or the body will be moving in all kinds of directions. For an agent to be able to have that intelligence, they need to have access to certain kind of data in the right way, providing them the context. So, what a company needs to actually make agentic work is to have the context that will ground the decisions that the agents are making. That context is deep data, defensible data, data that is not portable from one system to another, like actual data that you have that is structured in the right way with the right memory systems, short term, long term, episodic, semantic, whatever kinds of memory systems, and there are lots are being built right now. And it's actually super complicated. So, an agent cannot work if it's not grounded in the right context. That context does encapsulate data, but data alone is not sufficient. So, the MCP has allowed to solve the fact that these agents are now grounded and have access in a secure way to the data that they need. And I think this is what is being figured out right now. But it's moving. I mean, there's a lot of moving parts in this discussion.

Gemma Allen

>> Okay. So we want to get to a point, and we will get to a point where every two weeks you're going to come on. We're going to record this episode together and we're going to talk about particular companies, right? What's happening in the market, leaders, winners, you know, warts and all in terms of what's happening in the world of Agentic AI. Today I think we're going to talk a little bit about MiniMax because this was obviously a fascinating IPO, a company that went from not to a hundred in what seemed like the speed of sound. They doubled their valuation on day one of their listing on the Hong Kong Stock Exchange. Let's break it down a little bit. Tell me, what are your thoughts on this company? Fascinating player. What does it mean in the race too, right, in terms of the Sputnik moment between the U.S. and China? Give me your thoughts.

Raphaelle d'Ornano

>> So MiniMax is first, I think, a fascinating IPO because it is actually the first LLM IPO with Zhipu, who is the other Chinese competitor. They happened to IPO the same week, so beginning of January. Both have had tremendous success so far on the Hong Kong Stock Exchange where they have listed. But they are the first LLM IPOs. So I think, I mean, we're entering 2026. This is the year where Entropic is expected to IPO, probably OpenAI, though I think later. But anyways, we should have one big U.S. household name IPOding, maybe two. Databricks is also said to IPO. I mean, there's a lot of rumors, so this should be a plentiful year in terms of IPOs. But the first ones that actually went to the public markets were Chinese companies. Now this comes at a time where there are very strong, of course, geopolitical tensions where Chinese open source, which used to be an object of curiosity, is not that anymore and is now, I mean, a core player in this whole market. And so, it comes at an intriguing moment. Now, if we look specifically at MiniMax, MiniMax is having a play by which they're deliberately not playing the frontier game. Their game is, we want to be good enough to be the leading coding agent with their latest model M2.1. And of course, you can debate around what is good enough, but their claim is this is the model that needs to have the most widespread adoption amongst coders for us to lead the Agentic AI revolution with coding as the wedge. They also do have a consumer mode where this is not a pure B2B enterprise LLM. They also have actually a very successful consumer mode with their applications, which are very dominant in the Chinese market right now, and which may lead to, I mean, a mode in itself. But I'm interested in the companies through the coding wedge because we know over 2025 that Entropic, for example, has had tremendous success with this-

Gemma Allen

>> 40... I think they went from 25 to 40% in terms of enterprise adoption, right? Especially for Claude.

Raphaelle d'Ornano

>> I mean, it's huge.

Gemma Allen

>> It's huge.

Raphaelle d'Ornano

>> They're like the adoption of Claude Code, I mean the panic that was created in the markets just last week with Claude Cowork that is directly built by Claude Code. So MiniMax, we're going to have to follow what they do from, again, number one, in aGentic AI workload, are they used as a dominant model in which they are super good? One of the reasons being that they are multimodal from the start. Now, even though this company has very strong attributes, and I think they're a strong contender, and the valuation, not the valuation, the opinion I gave on them was quite a positive one, which is not always the case. Usually I tend to bash a lot of the companies.

Gemma Allen

>> I love critical women. Raphaelle, love it.

Raphaelle d'Ornano

>> I mean, when I actually did the scoring of MiniMax, I was like, "Okay, this is a super interesting company. It doesn't get the best score, but it has a very good score, better than some other companies that I will not name for the purpose of this show, but that are widely known companies operating in the same space."

Gemma Allen

>> You will not name yet for the purpose of the show.

Raphaelle d'Ornano

>> Right. So nonetheless, there are two things that I think warrant caution on MiniMax for investors, of course. The first one is that they're competitor, DeepSeek, which shook the world just one year ago, by the way, is coming out in February with a new model that is built on a new training technique that is gaining a lot of interest. And which has been the object of very profound research. So, they issued a paper just before the Christmas break showing how their mechanism was going to enable training faster by orders of magnitude. While MiniMax has been very efficient on the inference side, but has not fully cracked the lightning attention mechanism that had been at the core of their research back in January of last year, they kind of like backtracked on that. And then the model they released the M2.1 is, I would say, I mean, it's not nice to say that, but in some way it's kind of a patchwork. It's like very good, but it's far from what DeepSeek could actually be proving to us next month.

Gemma Allen

>> Let's talk a little bit about the whole M2 mixture of experts multimodal model, right? Because is it a safe play to claim to be above average or whatever language they use are good enough at everything? Antropic with Claude has obviously become very enterprise bound, right? They've developed something that's very well governed, gives enterprise like a feeling of safety and security. OpenAI is a little bit all things to all people right now. It's an underlying frontier model though for so many others. Google Gemini, Google's already so naturally sticky, right? It's so easily integrated in terms of how you can imagine Gemini just becoming another tool that you use in your day to day lives. What do you think is the safest best for models like that? Do you need to be the best at one thing in 10 years from now or what do you think?

Raphaelle d'Ornano

>> So I would say your question brings two different, like has two sub questions and I'm going to answer both of them. So number one, in terms of what is good in the enterprise, I think to the advantage of closed U.S. models, there's a super interesting study by UC Berkeley that was released in December called Measuring Agents and Production that shows that in a dominant way, agentic AI deployments, over 300 of them, are being done with closed models that are sufficiently good to be not even fine-tuned or with any other techniques that are prompted well and deployed as a genetic AI use cases. So, that's a super interesting finding. And that's Anthropic, that's open AI. Developers, many will prefer like open source techniques. There's much more complexity to that, but today what the study from Berkeley shows is that this was not the dominant way of doing. So, I think that brings a perspective on open source is not necessarily winning. I mean, the game is completely on and to be honest, I myself have no idea. I mean, again, let's see how this unfolds. But that's on the customer side, on the adoption side. The second question that is a key question is, when will these companies and how will they get to like any kind of path to profitability? Again, this is an investment case. If I put a hundred million or more in OpenAI, Anthropic, MiniMax, Zhipu do I get my money back and how? I mean, from like, does this company actually get profitable at one point? I think the U.S. companies haven't cracked that yet. The margins are improving very largely. For example, gross margin and Tropic has shown very strong advancements. I mean, training compute is still an absolute huge bucket of expense, so let's see. I think where the Chinese are interesting, it's that they're showing that there is a path to better dollar efficiency that is already there. And again, if DeepSeek progresses as expected, that gap in training could be very massive.

Gemma Allen

>> And Raphaelle, how do you think right now the financial instruments are looking at this? If you think about Facebook back in the day where it was a $20 per user, right? It's turned out to be something way bigger than that. I think it's like $200 per user or whatever, but the assumption was it was a $20 per user model that Facebook make on mass usage of the product. What do you think is being used to measure the opportunity and the financial probability of these LLMs? Do you think it's the same? It's based on just mass scale?

Raphaelle d'Ornano

>> Look, I think today that... So when you're investing in an LLM as a public investor, you're investing in optionality. I think that investing in a reasonable case by which you have a path to growth, path to profitability that is well sketched out, I don't believe that one second. I obsess around the moat of these companies. I try to like, again, make sense of that. There are too many unknowns to have the slightest idea of... I mean, you can have ideas, but you cannot do like a regular DCF as if you were assessing a classic company, even like a classic high growth technology company. So, you're buying an optionality here. And you're kind of asking yourself, is there like a single point of failure that is going to completely blow up my investment thesis if I invest in company A, B, or C? Is there a sufficiently high probability that this thing could completely blow up, and that I would lose everything even though things go well? I think you're more looking at from that perspective. But no investor today has... It's done that no investor knows, it's that we do not know because from a research perspective, things are being built out. Again, how this is going to play out in terms of, how do you monetize, how is the product adopted both in the consumer and enterprise market? What is your inference efficiency? What is your compute trading efficiency? Not even talking about the rest of the cost structure, too many unknowns. So right now, MiniMax, I think is a play with strong optionality. It's going well. I hope it continues well for them. But it's not like paradoxically, it's not negative too much for the U.S. companies because of that study from Berkeley that brings a light of perspective, I would say.

Gemma Allen

>> So before we talk about... I want to finish up, I want to talk about what it means to really analyze the moat, right? I know you said earlier it's not data. But before we go there, just quickly let's touch on the decision to go public by MiniMax in such a short space of time, right? Do you think that's a CapEx play, a credibility play, funded by Alibaba, the sovereign wealth fund of UAE? What are your thoughts? Why do you think they decided to do that so quickly?

Raphaelle d'Ornano

>> I mean, first I think, I mean, these companies are burning a lot of cash. So the public markets provide, again, this opportunity of financing. I think the public markets are starting to be ripe to, okay, this is a new paradigm. Again, like there's the agentic paradigm, there's the AI companies that enable the whole build out. These companies could be like massive companies and you're getting that optionality, so it depends on exposure. But I think there is an interest from public market investors now to really go into models that three years ago, were seen even by some big private investors as like, oh, there will be no more value. There will be no value in LLMs. I want to be three years back. Remember, companies, well-known VCs in this country were saying, "Oh, we should not be investing in LLMs. Applications is the only layer that is going to capture value from all of this build out." Some VCs are in trouble for having said that because I mean, there's a lot of value in the LLMs right now, hopefully for a long time. So I think they're consuming a lot of cash. There's a window that is right. Let's see how this plays out. And again, this is not even dependent on the companies themselves. This is dependent on do they actually get access to the chips that allow them to train their models efficiently? I mean, this is where geopolitics comes into the picture. This is a super complicated game, but it depends on your exposure.

Gemma Allen

>> Okay. Well, listen, Raphaelle, so excited to have you and to get to talk to you. Like I said, biweekly, we're going to be talking about different companies, what's happening in the markets, Captain Markets meets Silicon Valley, breaking it down from the perspective of AGNT, checking just how ready the Fortune 500 is for the world of Agentic. And excited to see what unravels.

Raphaelle d'Ornano

>> Super excited.

Gemma Allen

>> Thanks so much.

Raphaelle d'Ornano

>> Thank you.

Gemma Allen

>> Okay, folks. Tune in two more weeks when Raphaelle and I catch up again. This is Gemma Allen from TheCUBE. Thanks so much for watching.