Microsoft CPO on AI Strategy, Anthropic’s Acquisition Plans, Gaming Startup Trains AI | Oct 16, 2025

0:13Welcome, everyone, to the Informations TI TV. My name is Akash Basricha. It is Thursday, October 16th. We have got a great show planned for you today. We have got Microsoft's chief product officer of AI at work coming on the show momentarily to talk about the current state of play in their AI business. You do not want to miss that conversation. We're also talking about how one company is using video games to train AI models and separately about how Anthropic is planning to go hunting for some more acquisitions. And finally, we've got a big story publishing today about how XAI plans to outfit the big data centers it is planning with enough AI chips, which, of course, are very expensive to come by.

0:53It's going to be a great show, so let's get right on into our first guest. It's a very interesting time for Microsoft in the AI race. The company, in many ways, had a big head start with its early investment in OpenAI, but now it's up against some of the same challenges that many other enterprise companies are facing, like convincing businesses that their products can actually demonstrate real ROI. I want to bring on someone who is working on that challenge every day. Aparna Chenapurghata is the Chief Product Officer at Microsoft of AI at Work. Aparna, it's great to have you on the show. Welcome to TITV.

1:26Thank you, Akash. It's slow news. week, month and a year and a decade. Always, always. Okay, well, given how slow things are, let's talk about all the exciting things that are happening in the world of AI. You know, one of the things that we've been talking about this week on show is Dreamforce. And Mark Benioff came out saying, you know, hey, AI, it might take a little longer than we thought in terms of integrating it into enterprises. And that was a bit of a change of tune from last year, but it kind of connects to this theme overall that people are sort of coming to grips with the fact that, hey, AI has taken a while.

2:05Uptake is a little slower than we may have thought. Are you seeing that in your business with the customers that you're working with? First of all, for perspective, I think this is the fastest diffusion of technology that ever happened, let alone consumer on the enterprise side, right? So I've been through the internet shift. I was early at an internet content delivery system at Akamai, and obviously through mobile and the web growth and cloud, and here we are at AI, the compression of diffusion is what used to take decades is taking years. And of course, we are saying, why isn't it happening in months?

2:40But I think that's one thing for perspective. Second thing I'd say is we are actually seeing quite a bit of adoption in our customers. 90 % of our Fortune 500 customers are actually using a co-pilot studio, co-pilot building agents. We have more than 100 million customers across consumer and commercial for Copilot. And then we're building a lot of agents that are focused at work. So what I would say is, look, you can't do the wild, wild west thing. You have to build. For us, the way I think about it is, I tell my team's tool, we are operating like a three-legged stool, right? So one leg of the stool is, of course, the best intelligence.

3:19And that momentum is incredible. Even as we speak, I'm sure there's a new model dropping. So that's one, the advanced intelligence. But the second one is like actually figuring out how people work, meet them where they are. And that's where the way we are looking at it is, of course, integrate in the tools that people actually already use. Hundreds and millions of people already use Word, Excel. We launched the agent mode in Excel. And then third leg of this tool is trust and compliance and security and all that. You can't compromise that. I want to talk about the trust element because one thing that I've been interested in with your products is as you seek to integrate more AI into products like Excel and the Microsoft 365 Office Suite broadly, I mean, it takes a lot of trust for people to trust AI with producing sensitive documents or complicated spreadsheets.

4:11I mean, these are big financial models. How do you actually accelerate the adoption and actually embed trust into customers when a lot of the time it's, I asked the AI to do it, it was fine, but I still have a lot of work to do to sort of make it what I need to do in the first place. And it's also not a point in time frozen question, Akash. So if you look at the first wave of AI, it was mostly these instruction following models, right? It's like basically question answering. It was like, that's why, no surprise, chatbots were the first wave of AI products. And we did that too, like where you can ask a question about a document, but that only goes so far.

4:50It's an improvement in productivity, but it's not in order of magnitude more. The second wave that we started really zeroing in on in the earlier part of the year is going beyond answering to reasoning, right? Once you actually have reasoning models, then you say, no, no, actually, give me the entire sales briefing that I need. to go into this prospective customer. So we launched an agent called researcher, and it's mind-blowingly good at connecting the dots between your meetings, your emails, and the world information, right? That's not about saving time anymore. That's about, you know, giving you superpowers.

5:25Like I couldn't do it as a human. And now you have an agent that can do that. Now the third wave that's coming is, okay, you have answering, you have reasoning, you're doing, right? And that's what you're pointing out, which is how do you kind of, and the trust levels are increasing as you go. But it takes a lot of, I mean, you're handing over control of work, essentially, to this agent. I can only anticipate there's a huge cultural shift that has to happen in the way people do work. And so that's what, you know, I wonder how you think about accelerating that. Is it just educating people? Are there products that they are asking for that you still have yet to develop?

6:05I mean, how do you think about accelerating? There's three things we're doing, right? So first one is saying, look, what is the objective way that we can measure how good the products are? So we have all these benchmarks externally, right? Like about how good AI is at competitive coding. Turns out, newsflash, most people are not doing competitive coding when they come to work. I don't know, Akash, maybe you are. I'm not doing that. I am not doing competitive coding, no. Right, so what we said is, okay, let's come up with workbench, which is what are the set of tasks? We have a spreadsheet bench, we have a PowerPoint bench, we have a benchmark for how meetings and effective communication happens.

6:45So what we are saying is taking a very hard-nosed look and saying, how well are the models and the products doing on that front? So let's be objective and get some data on that. The second thing we are doing is, of course, keeping the human in the center. So I'll give one example. We launched a project agent in our team's product. So if you're in a channel, you can add this project agent, you can ask it questions, you can take notes, it follows up stuff, note taker, all sorts of stuff. But in that case, we are kind of trying to say, hey, the human is in the center and the loop. You're not trying to kind of like build an agent that runs away and does a whole bunch of things without the human approving.

7:25Let me ask you about some of the technology under the hood. So we've reported about how Microsoft is now leaning more on Anthropic for the underlying models for several of its co-pilot products. Have you seen customers react better to those particular models as they're being used? I think the way to think about this is, you know, people in the industry are constantly talking about, oh, is the model the product? Is it app first versus model first? And we've seen this movie before. The stance and the posture we're taking is it's value first, right? For the customers, like in what situations, first of all, giving them choice in terms of the models that are accessible to them.

8:05That's the first step. But then what we are trying to figure out is, hey, in this case, when we say we want to build this agent that does, that is almost like the CEO's chief of staff for everyone, right? Like this researcher agent I talked about. The goal is to help you think and act like a CEO. Models are a means to that, right? And so we're looking at everything left to right and saying, hey, what are some of the technology bits that are useful for giving the best report as I walk into this meeting, right? What are the tools that are best for code generation and visualization and so on? So I think that's the approach we're taking, the customer and the value first.

8:42What about pricing? You have, right now, the pricing model is on a license basis, a per seat basis. You also have consumption-based pricing elsewhere in the suite of products. Which of those are you seeing customers respond better to right now? It's an end. I mean, if you think about like the Azure AI platform model or the Copilot Studio model, we are seeing incredible usage of like people building these agents. Now, the agents themselves, they're not going to be like, you know, what we internally see is it is a power law, right? There are a few things that a lot of people use, but the beauty of it is there's agents that like people have built that three people use, right?

9:24And so consumption makes a ton of sense there. And for the main M365 office suite and like the co-pilot usage there, it is where you are meeting users where they are in their workflows. So, of course, like a per user model makes sense for us. So, it's an and. Right. And last question for you, you know, as you seek to build your product suite out, what is it that customers are asking for most right now that you are trying to deliver on? What's coming next? I would say two things. One is this idea of, look, individual productivity is really important. And we are really racing to figure out, hey, what are the set of agents that come to work for you, right?

10:05Like if you had a team of virtual agents, today you get a badge and a PC when you join a company. What are these five, like if an employee is an agent boss, what are the five digital workers that should come to help work? That's number one. I would say the second one is more around multiplayer. Like individual productivity is one thing, but how do we make teams and organizations? A lot of the CIOs and a lot of the business folks, as well as kind of like actually developers are looking at, I'm not in a, you know, this is not a single player sport. I'm working, coordinating, a whole bunch of effort goes into it.

10:42Can AI make that go away? How do you do that? I mean, is this more ways of like sharing documents? What does it look like on the ground? A great example of that is the agents that we are launching in now that can attend meetings, that can be part of the group channels with you. And if you look at people working across time zones, incredible amount of time goes into working harder and keeping track of things and connecting the dots, et cetera. Forget 996, we have 24-7 with AI. And so we should put that to work. So that's the perfect example for me of AI that actually amplifies and not just automates.

11:21Great. Well, Aparna, thank you so much for coming on the show. It is certainly an interesting time for the company, and we are closely watching to see what you and your team are working on. That is Aparna Chennepurghada, the Chief Product Officer of AI at Work here on TITV. We've talked a lot on this show about Elon Musk and XAI's grand data center ambitions. And one of the most important questions has been how XAI is going to fund the construction of these data centers and also buy all the chips that they need for it. I want to bring on Myles Krupa, our AI and finance reporter, and Theo Waite, our Elon Musk reporter, to tell us more about a story that they are publishing today that sheds some light on how exactly XAI is going to address that challenge.

12:03Theo and Myles, welcome back to the show. It's great to have you. Thanks for having us. Let's talk the wonderful business of data centers of Colossus. Theo, we're going back to the city of Memphis, which I know is a city that is very close for you. Remind us about the scale of the data center that XAI is trying to build in Memphis, and then talk a little bit about how the company is looking to outfit this data center with chips and actually finance all those investments. So XAI already built one data center in Memphis called Colossus that's been up and running for over a year now. They have the second one called Colossus 2 that they're still finishing up and filling up and powering that they say is going to be more than one gigawatt of power with about 550 ,000 NVIDIA chips in it, which would make it the first data center of that scale in the world.

13:02It's really unprecedented in size. There are lots of other companies that have announced projects that would be that big, but it seems like XAI will probably be the first to actually finish building one if, you know, if they can get the money to do so. Buying that many chips obviously costs billions of dollars and XAI needs a ton of money. So to buy those chips, XAI is doing a deal with Valor, which is an investment firm that has very close ties to Elon and is raising a ton of equity and debt for a SPV, which is like, you know, essentially a subsidiary company that will buy NVIDIA chips and lease them to XAI.

13:44Miles, let's go over to you. Talk to me about this SPV. You were talking to me before the show. I mean, SPVs are not rare. They are pretty common in many cases in Silicon Valley. But what is unique about about this deal in the way in which it's structured yeah that's right we've seen other companies like meta turn to spbs to raise money for data centers and effectively get these costs off their balance sheet and we've also seen debt backed by nvidia gpus in the past what's a bit different here is that um the firm that's raising the money for this is valor an investment firm It's not a core weave like a cloud provider.

14:26So effectively, this is going to turn Valor, an investment firm, as Theo said, that has very close ties to Elon Musk, turn them into a cloud provider for Elon Musk's AI company. And that's sort of an interesting position for Valor to be in. They're also an investor in XAI. And, you know, they're basically asking investors to make a bet that OpenAI, excuse me, XAI is going to be able to make its lease payments on these chips. And then that the chips will be valuable years down the road. So this is turning a venture fund into kind of like a cloud provider player? Is that the idea? Yeah, that's right.

15:11I mean, it will be leasing the chips to XAI, much like CoreWeave leases chips to OpenAI. So, you know, however, CoreWeave has an operating history as a cloud provider, and Valor does not. Valor just woke up and said, hey, you know what? This cloud business sounds pretty good. Why don't we take a shot at it? Is this common in Silicon Valley? I mean, have we seen other firms do this? We've talked about venture funds taking on more roles in the technology ecosystem than just providing venture funding. Is this happening with other firms too, or is this unique? Yeah, well, this structure is a bit different than others that have been done in the past.

15:56But yes, it is basically part of a bigger trend in which, given the funding costs of the AI boom, the amount of money it takes to purchase hundreds of thousands of NVIDIA chips, you're seeing companies trying to shift those costs off their balance sheet, out of their CapEx totals into other areas. So it's to sort of lower their overall cost of funding, be able to raise debt more easily, and just kind of keep this whole thing going, turning to other partners to basically finance a lot of these really, really expensive purchases. Theo, I want to come back to you. As it relates to Elon Musk running XAI, I'm wondering to what extent this feels like a deal that only Elon Musk could have sort of orchestrated, or is it something that people say, you know, this is actually something that we could see more of?

16:57I guess what I'm getting at is, is this an Elon Musk deal, or is this an AI data center deal that we could see more of? So I think it's an Elon Musk deal in the sense that he's, you know, calling on a very close ally to kind of arrange this whole thing, essentially. But, you know, I think, and Miles pointed this out while we were reporting this story, you know, there are parallels to this kind of structure. There's a meta data center in Louisiana that's under construction, or is, I think, maybe starting construction soon, that is also being financed by an SPV. But with that deal, there's only$3 billion in equity and$26 billion in debt, which essentially means investors see it as less of a risk because the debt component is so high.

17:46So it's not like XAI can just get unlimited money or anything. They are this kind of unproven player that doesn't have a lot of revenue, doesn't have a core business that's throwing off money the way Meta does. So there is certainly a premium for the amount of risk here that XAI does have to pay. And Miles, what about the power component to these data centers? That is obviously something else that companies have to figure out how they're going to pay for. How are they planning to finance all of that investment? Yeah, that's right. So XAI is using power resources that are actually located across the state line in Mississippi.

18:27and it's a joint venture with a power company where XAI is actually the minority owner. And so that power company they're working with owns a slight majority of those power generating assets tied to natural gas. And what they've done is they've gone out and raised a$550 million loan to go purchase more power assets. So in that aspect as well, XAI doesn't quite own this really critical component of the Colossus 2 data center. And the fact that they don't own any of these assets, Miles, and I'll close with this, does this pose any kind of a risk to the fundamental business model that XAI is trying to develop here?

19:14Is there a way that this could sort of end up coming to hurt them in the long run? It's obviously a great way to get the assets in the short term when you don't have to put up all the cash. But I'm just thinking sort of five, 10 years down the line here? What do you think the implications are? Yeah, well, I guess one consequence is that if XAI were in a position where it needed to sell some of these assets, it wouldn't really immediately be able to do that because it wouldn't have control. XAI does, as far as we understand, have a right to purchase the chips at the end of the term. That's kind of an interesting part of the deal.

19:52So it could, after it's finished its lease payments, and buying these chips. But as we know with NVIDIA chips, they depreciate over time. And there's new ones coming out too. They'll actually be worth. Right, and there's new ones coming out always. And, you know, it's a very competitive space and I can certainly see Elon saying, look, we got to get the new ones. I mean, you know, the old ones were the old ones. Well, look, it's a fast moving story and I appreciate both of you coming on. And Theo covers all things Elon Musk and Miles covers all things AI and finance here at The Information. Thank you both for sharing more of those details with us.

20:33And with that, let's get to our next guest. Dealmaking, M &A, and acquihires have been one of the biggest elements of the AI boom as many companies try to find ways to grow. And this week, the information published an inside look at Anthropics dealmaking strategy and how that stands to change in the near future. I want to bring on Valita Pau, our deals reporter, to talk about what she found and what she's hearing. Valita, welcome to TITV. It's great to have you. It's great to be here. First time on the show. We'll have to get you on more often. Let's talk about Anthropix M &A strategy. Tell us about what the M &A strategy has been traditionally and then how that's looking to change based on your reporting.

21:15So with a 40-year-old startup, they haven't really done any M &As. So the closest things they've ever done in the past was like this aqua hire back in August when they kind of hired co-founders and most of the staff of this human loop, which is a really small startup that evaluated the LN models. And before, their CEOs really focused on like hiring researchers or like bringing on specific peoples with like expertise to add on to the teams. But recently they have been putting together like a corporate development team and they brought on like two former bankers and a venture capitalist to build out their like corporate development team.

21:53So Andrew Sloto, who was a former partner at SoftBank Mission Fund, is like heading up that team. And they are kind of like telling outside bankers that they are ready to do like acquisitions. And that's like kind of a break from their strategies in the past when they're like focusing on talents. So they're gearing up to do more acquisitions. They've got all this talent. Did you get a sense at all from what they're looking to buy at all? So Anthropic is really successful when it comes to selling AI to power all these coding tools they were selling to developers. And that made up a huge chunk of their revenue.

22:27So they could want to do some deals that bring on other parts of that whole software development space for other tools that developers can use other than coding. they can also do like acquisitions to focus on like specific industries like financial services healthcare like cyber security that is like a more enterprise focus and but like they can also do that with like a partnership or like a deal so I think they're still evaluating you know what's like the best for them I wonder if you got a sense what is actually driving this desire to pursue M &A more you talked about a little bit at the start but on the other side you know we have open AI they're doing a lot of deals.

23:09Is this a movement to stay competitive? Is it something that is more growth focused for them? Talk a little bit about what you think is driving this new focus. So I felt like it's kind of have to address like, I know the arrival, OpenAI has been doing a lot of like deals using like the highly valued stocks to equity deals. So they did three stock deals since like this June 2024, when they buy like Rothset and analytics startups, they bought IO. They also have recently staffed it as to join for the application site. And they tried to buy Winsurf and also have acquisition talks with the Cursor Maker and Sphere.

23:49So they have been very motivated to using their highly valued equity to do deals. And that could prompt Anthropic, which closed a big funding round this September. so they could use some of those stocks to do the same as well. And I guess people are like AI Labs and trying to find new ways to monetize. So they could also look for these kind of more AI application sides. And certainly you have all of the talent that comes with all of these new companies too. I mean, if you're just trying to get your hands on researchers or expertise in a certain area, OpenAI has done deals like this. I can certainly see Anthropic doing deals like that.

24:32Thank you, Belita, for coming on the show. It was a really interesting story, and I'm excited to have you back on more as we find out more about the deals that these companies are doing. That is Belita Pow, our deals reporter here at The Information. Speaking of deals, late last year, OpenAI quietly tried to buy a video game company called Metal for the modest price to$500 million. Ultimately, that deal fell through, but it has inspired Metal, a 10-year-old company out of the Netherlands to overhaul its ambitions entirely. The company has now spun out a new startup called General Intuition, which aims to use video game data to train AI models.

25:08They raised more than$130 million in seed funding today from Coastal Ventures and General Catalyst. And I want to bring on CEO Pim DeWitt to talk more about where he hopes to take the business. Pim, welcome to TITV. It's great to have you. Likewise. Nice to meet you. So let's get to the brass tacks here. What was metal and what is general intuition? So view metal as a way to capture and share memories while people are physically apart. And usually that's while playing video games. So you're playing a game, something interesting happens, and then as a result, you hit a button. It then syncs that video to your phone and you can share it with your friends.

25:48And I think what we realized is by doing so, we've sort of built up the equivalent of the episodic memory of billions and billions of hours of humans observing simulation, which, and when you're playing video games, you're sort of transferring your perception to whatever embodiment you're playing the game as, right? It could be horror. It's basically like storing video game footage, essentially. Yep, yep. Okay, okay. So you play the game, something happens in the game, you want to save it, you save the footage. And then there was an aspect about this that I found interesting, which is that video game data is actually very helpful for training AI models.

26:28Yeah. Why? That's right. So think about the type of data that AI really needs, right? So in particular, let's talk about adverse events. So car crashes, things that have a really high cost in the real world. You want to get this data from simulation because you don't want to actually put people in danger, right? And so in addition, when you're playing video games, you're essentially transferring your own perception into whatever device you're using to control your game character. And that game character can sort of be referred to as an embodiment. And so as a result, it is this very, very good representation of spatial temporal reasoning over usually long horizons, which is games also have this very nice character of having a very clear reward signal so you can understand whether you're progressing correctly or not correctly as you're doing things, which is very, very important for training to get that type of feedback, right?

27:32And so it is a, similarly to how programming, for instance, really accelerated LLMs, right? Games are the equivalent of verifiable domain for spatial temporal reasoning. And so as a result, it's incredibly important. Yeah, well, it's funny. We play all these video games. We don't think they're as complicated when we play them, but there's certainly more thought behind them. You know, the thing that I want to ask you about, and maybe this relates to your new startup that you've spun out now, General Intuition, is I feel like this gets a little bit meta and circular, right? Because, I mean, on one hand, we have AI now that helps us build or generate videos.