Good morning, everybody. So we're going to talk about AI today, but it's going to be a little different from the generic AI conversations about superintelligence. We're going to talk about making AI useful, particularly about making AI useful in large companies, which I think we all believe — whatever you think about superintelligence — is a massive opportunity. And just as cloud took quite a while to diffuse into organizations, it's still happening. We're not done with cloud, and it's created hyperscalers and a ton of value. Enterprise AI will have a similar adoption curve. It's not going to be as rapid as consumer because organizations need to absorb AI, and that takes time and committees and implementation, and it's a decade-plus journey, but incredibly, incredibly valuable because of what AI can do in the enterprise. So we're going to talk about all of that, but also discuss why it's so difficult to deploy AI quickly in the enterprise. And if we have time at the end, we'll ask some questions. So before I start, I think it would be useful for everyone to just introduce themselves and specifically to talk about the problem that your company is solving as a starting point for the discussion. Arvind, we'll start with you at Glean.
Yeah. Hi. So my name is Arvind. I'm the founder and CEO of Glean. Glean is an enterprise AI company. We build a product that looks and feels like Google or ChatGPT, a place where employees can go and ask any questions. And Glean uses all of your company's context and knowledge to help answer those questions for them. So it's a fundamental, basic tool — like we're trying to bring AI into every employee's day-to-day work lives. And that's what the company's about. And Misha.
Hi, I'm Misha. I'm the CEO of Reflection. Reflection is building the US DeepSeek, effectively. Over the last year, you probably observed how DeepSeek came out and then a bunch of other great open models from China came out. Whereas in the West, I would say, especially in the States and really outside of China, AI power and compute is getting concentrated among a few closed lab players, and we aim to step in and build the best open model of US origin and export it to the rest of the world, into large enterprises and sovereigns.
Hi, I'm Martin from Cohere. So we are solely focused on enterprise AI. We're one of the, let's say, outside of China, six companies that create frontier AI tech from scratch, not taking other people's foundational models, but the only one focused on enterprise AI. We don't have a consumer chatbot. We never will. And we're also the only ones independent from US big tech. So we're officially Canadian, but we're a global company. And that means we deploy on any cloud rather than being beholden to one. But actually, well over 90% of what we do is private deployment. So it's deploying on VPC, regional clouds, all the way to on-prem, air-gapped on-prem. A little bit like Arvind was saying, the key here is we ship our tech to your data. And so for critical infrastructure sectors, often highly regulated, where they really care about data privacy, data residency, now we talk about sovereignty — whether that's national sovereignty or within the four walls of a bank itself. It's a very different proposition to what are quite generic global consumer chatbots that now are being deployed through cloud into an organization.
Thank you. And close us off.
Very nice to be here. Great to be on this panel. I'm founder and CEO of Unifor. We are an end-to-end data and AI stack meant for large businesses — businesses who want a sovereign and open data and AI platform in which they can plug in different models, different agentic layers, and things like that. That's what Unifor delivers to these businesses. We sell to about 2,000 businesses across the globe, several of which are Fortune 500 companies. And using our platform, what these companies do is connect up their data sources, which could be fragmented, living in hundreds of systems, use those to fine-tune a model of their choice. We're not a builder of foundational models, and so the choice is up to our customer. And then once they've done that, they can build agentic workflows on our platform. And I think I'm picking up this theme. Our positioning also is we deliver all of the stack in a sovereign and an open manner to our customers. So there's an interesting thread here, which I don't think was actually deliberate in constructing this panel, that these solutions need to be open, they need to be tailored, they need to be flexible, they need to adapt to their customer needs — which is language you typically don't hear on the consumer AI side, which gets all the attention, where all people want is a highly performant, next-generation, out-of-the-box chat and video and voice and text interface. But the enterprise need is quite different.
Martin, I want to — sorry, Misha, I want to start with you on some of these themes, especially around open source. Why is open source such a big topic, especially for companies?
So if we think outside of AI, most of enterprise is built on open source. Linux is the open source OS that powers most enterprise servers. Like most computers — I mean, our individual computers that we use have a closed operating system, but the servers that run compute for enterprises, the majority of them run on Linux. And most technology that an enterprise adopts and builds on top of — it's really for technologies enterprises build on top of them — those are all majority open source. And the reason is that you want a pretty great degree of customizability, and something becomes a really important part of your technology stack, you want to own it, right? And you don't want to be locked into any particular vendor. So open models are kind of the new, let's say, next iteration of open source. If you're a large enterprise or a sovereign, AI is becoming your largest IP asset. It's going to approach, right? It's artificial labor. It's not software. It's artificial labor, meaning that it's going to be the largest line item that an enterprise has. And when something becomes so critical, you want to own it. You want to run it on your infrastructure. You want to customize it for things. You want to keep it in your data plane. And all these things — the right kind of product surface for that is open, right? Closed products just have a hard time adhering to enterprise needs at that scale.
You mentioned lock-in as a big part of your answer and something that companies have been scared of for quite a while. And Arvind, maybe coming to you on this concept of lock-in. Glean is a very agnostic platform in the sense that it sits on whatever data, on whatever models, and is focused on being useful to the enterprise. How do you think about this concept of lock-in in the world of AI? How do you talk to your customers about it? How do you solve to avoid it?
Yeah. Well, I think with AI and AI agents specifically, businesses are increasingly understanding that, you know, now it's not just technology that you're reliant on. Once you start to build agents on AI agent platforms, now in many ways, like your services and your entire business is running on technology, you know, at a much larger scale than before. So lock-in has become a real concern. Like we hear this from enterprises continuously that they want to make sure they're not beholden to one particular vendor, one particular cloud environment, one particular model provider. They want to have that choice. They want to have the ability to actually work with many different vendors and switch them, you know, at will. And so that's certainly a thing which is top of mind for enterprise buyers. And it starts with models first. So there's no large enterprise, you know, that's happy with just relying on one model provider, whether it's OpenAI or Google or Anthropic. They want to make sure that all of the industry's innovation is available to them. So with Glean, you know, that is one of our key selling points is that, you know, we make sure that an enterprise has that option to use any model, you know, that's in the industry, you know, for their use cases. You want to run your AI workloads in different cloud environments. So we also give that flexibility. And then finally, also making sure that customers and enterprises have this ability to build agents themselves on an open horizontal platform, which is what we provide. So instead of having to buy agentic products from many different vertical-oriented companies, they're able to actually standardize on our platform. So it's been a big trend, like especially I would say in the last nine months, like, you know, this year, enterprises are increasingly vigilant about how they can make sure that they have that independence and the ability to sort of, you know, change vendors. One, like maybe last thing to add is we're also seeing this trend of enterprises not wanting to sign long-term contracts. Instead of, you know, signing a three-year, five-year term with a vendor, they are actually, you know, right now trying to actually do just one-year contracts so that, you know, they keep that optionality open in terms of if they want to switch to somebody else.
Thank you. And Martin, maybe an interesting counterpoint because you build models, right? And so obviously you want your customers to use Cohere's models primarily — correct me if that's wrong — but how do you think about this topic of model lock-in in your case, and how do you answer that question to customers who want to use AI?
Yeah, I mean, I'll try and be a bit controversial here to have an interesting panel. Many of these models, they're not open source, they're open weights. Linux was something created by distributed thousands and thousands of developers. The models we're talking about, for example, Llama as one example, was trained with billions of dollars from a company that's literally worth trillions of dollars. They're not doing it as a charity, right? They're trying to control certain elements of that. The question is, if you build on that, will it be available next year? Will there keep being updates and so on? You also have no idea where the data is from. So you, Section 5C of the community license agreement, you accept all legal liability for any litigation or any issues that happen through toxicity or copyright from using that model literally when you click it and download it. Most people don't read, but you're assuming all liability. So it's open weight. It's not open source. But I think more importantly, we believe — and maybe not for everything — but enterprises want ROI, not AGI. I'm talking about some of our co-founders, Nick and Aidan and Ivan, who talk about that. Some of the teams may say we want flexibility, we want to swap things out. The analogy I use, or two analogies, is it's a bit like electricity. It's really hard to make. It is somewhat a commodity. You need a nuclear power plant. We happen to have built one, so have many others, but no one walks around saying, do you want 25 kilowatt hours? Do you want a two-for-one on Thursdays? But that's a little bit how AI has been, which is, how much are your tokens? Let's do an evaluation of 17 models against each other in some generic benchmark. That's not what is going to create value and actually create impact inside enterprises. And what we've seen is a bit like when I was a kid, my dad and I went to RadioShack. We bought some transistors, bought a circuit board, a soldering iron, and we made a transistor radio. And it was super fun when I was 10. But my dad didn't use that to play music that evening to our guests. And right now we have a lot of companies who are, you know, in the tech teams, are building things together. But what we find more and more is that the senior leadership wants solutions. They want things that actually work. And I'm not saying anything against what we're talking about here. I know that Glean is a solution. Obviously what you do is provide solutions. And so what companies are really looking for, enterprises, is a partner to help them create solutions that are tailored, customized, configured for their problems. I used to be at Google. Google Search, 3.3 billion of us use it every single day. It's by definition generic. It has to be generic. Enterprise is the opposite. That's why Google's done zero in enterprise search. It's because it's building to building combat. And we're finding actually, Arvind, that some of our partnerships are long-term partnerships because they're not looking for a vendor they want to swap out. We would never say we're a vendor. You know, we're a partner. And in fact, we do enable our partners to, if they want to use, I don't know, Anthropic for code and plug that into our North agentic AI platform, fine. If they want to use something, you know, I'll sell you a Ferrari. If you only have diesel where you live, you can put a diesel engine in. It won't work quite as well. But if there's a reason to do it, that's fine.
I admire your commitment to not mentioning the other model companies. It's great.
Well, I think that there are many outstanding companies out there, but it's a bit like the electricity to some degree. There's like, are you going to compare five different types of electricity, or are you going to think about how that power is going to be able to create solutions with a partner who will help you deploy that and create things that will drive to impact and so on within what you're doing? Again, very different to consumer, very different to prosumer, and some of the generic things like PowerPoint's great, my competitor uses it too. It's not going to give me alpha. So we focus on those capabilities that companies want that give them alpha, that enable them to do something different with customization and configuration for what they need with the various engines and the various technologies that are deployed.
Thank you. And I want to bring you into this. The notion of electricity is used a lot as an analog for AI, and I'm curious for your take on that because it seems to me that electricity doesn't really improve fundamentally, but AI seems like it does on some cadence. And so as you build your tools, how do you think about this idea of model providers? Is it fungible? Are they commodities, which is another word that's, I think, thrown around without a lot of interrogation? How do you grapple with these concepts as you go to companies and try to make AI useful for them?
So look, I want to make this real for all of us in the audience. You've been talking about a lot of concepts, but out there in the wild, the world of business AI is very complex. Let's take the case of a Fortune 100 company, a telecom major out of France. 40% of their business is in their home country, which is France. They run on a GCP cloud, Google Cloud infrastructure. And by government mandate, their model of choice for their French operation is Mistral. 20% of their business is in some other European countries where, because of partnerships, they run on a custom version of Microsoft Cloud — not Azure, Microsoft Cloud. Here they prefer to work with Gemini as their model of choice. Talking about energy — was there a wave at the back? Sorry. Talking about energy, another 20% of their business is Africa, and because of lack of reliable energy in Africa, they don't have data centers, they don't have a cloud partner, so they run their business in Africa on-prem, bare metal on-prem. Their data is all over the place. In Africa, the data has to sit next to the compute, whereas in France and other parts of Europe, the data platforms are different. The chief AI officer now cares about delivering a consistent AI experience for the end consumer — he's calling the call center, want to experience AI — or the employees who might be trying to use agentic applications. But he cares about delivering a consistent experience in France, in Portugal, and in Kenya. But the infra is unique. The models are unique in each country. The data sits everywhere. This is what business AI is all about. Business AI is more of an architecture challenge as opposed to saying, let me use this model or let me use this agentic layer. It's a complex architecture challenge to bring data, models, agents in a scenario where the business feels like they have choice that they deserve. It feels like they can move to things as they evolve and not feel locked in, but be able to run it in different infrastructure environments, different compute environments, and still deliver a consistent experience. That's what our architecture is designed to help.