Good afternoon, everyone. I’m Jessica Lessin, the founder of The Information, and I couldn’t be more excited to be here this afternoon with an all-star panel on AI infrastructure and the future of this important area. So I’ll introduce our panelists and really look forward to the conversation.

We will leave time for questions at the end, so please be saving those. To my left, we have Peng Xiao, who is the head of G42 out of the UAE and pretty much has his hands in every interesting AI project of the moment, so we’re thrilled to have him. Sarah Friar, the CFO of OpenAI, who has been doing business deals with everyone in the ecosystem and, of course, is leading one of the world’s most important AI companies.

Michael Intrator, the CEO of CoreWeave, which also needs no introduction, and Rob Goldstein, the Chief Operating Officer of BlackRock. Rob, I’m going to start with you, because I’m going to mix it up, and please, you are all experts on every topic, so feel free to chime in. Even if a question is not directed at you, Rob, just minutes ago, you were telling me that on the topic of AI and infrastructure, the future is limitless.

Let’s elaborate, and then maybe poke holes in that afterwards, but let’s start with that. What is it, when you stare at 2026, that leaves you with that feeling?

Sure. Well, what is remarkable to me, and I think people very quickly lose sight of, is that three years ago at Davos, the conversation was, you have to go check out this thing called ChatGPT, which was a brand new thing, and everyone was playing with it, like, how do I drive from X to Y?

The most basic, basic, basic things. And I think we forget, people talk about what inning this is, to use a U.S. analogy of baseball.

I think the national anthem is still happening. I don’t even think this has started yet. And if you think about the cycle as being sort of buildout, then adoption, then transformation, we’re still early in the buildout stage.

And if anything, I think there’s so much dialogue right now on bubbles, when in reality, the bigger issue, at least for the next one, two, three years, is rationing capacity. So to us at BlackRock, as the COO, when I say the opportunity is limitless, I just see if you’re someone who’s involved with operating a company, for example, we’re just starting to recognize the transformative capabilities of these technologies.

And I think that when you really think about the company, there isn’t any aspect of the company that can’t be transformed. You know, we were talking earlier about Davos and the nature of Davos and the amount of organization that’s required in terms of the many meetings that you have. Think about in a few years, everything will be automatic.

You’ll know that you bump into someone in the street. You’ll know who they are, the business that you do with them, what you should ask them. So the productivity cycle, I don’t even think has started yet.

And I think it’s limitless in terms of at least my career. The opportunity for transformation will be enough to keep me busy for the next 10, 20 years easily.

So job security.

Easily, at least for me.

So if we talk about, let’s break it down a little bit. We’re talking about infrastructure here. We’ve got to build things.

We’ve got to finance things. We’re seeing this gold rush, but there’s also been talk about, you know, power. How can we support this?

Are data centers going fast enough? Maybe to you, Michael, I mean, what are you, are there some bumps along the way in this moment of scaling? What needs to happen in 2026 to get to the next step?

And then I’ll come to you, Peng, on the same question.

So first of all, thanks for the opportunity to speak. Look, it’s such a fascinating business because the way the funnel works is like everyone sees these unbelievable products that are coming to market that have changed the world and will continue to change the world over the foreseeable future in ways that we haven’t even contemplated yet.

And that is one part of the market. Underneath that, it is physical, right? The defining characteristics of the boundary that exists right now is a physical boundary.

And those physical boundaries occur in a bumpy world, right? And it’s about power, it’s about concrete, it’s about copper, it’s about human resources. And when we think about human resources, you’re talking about the trades.

You need plumbers, you need electricians, they need to be trained. Four years ago, we were building a data center. You had 100 electricians, 80 of them were experts, and 20 of them were beginning their career.

And now you have 2,000 electricians and 80 of them are experts and everyone else is beginning their career. There’s a physicality to the business right now that the world is struggling to translate into valuation, into public markets, into the path to delivering these products that are generationally, and I don’t even mean like one generation, this in my mind goes down as a stepping, you know, the wheel, intelligence, sort of like that scale of change that’s occurring and behind the scenes of that it’s so, you know, like to go to one of these data centers to walk into like what is functionally the Death Star and you walk through these things and you’re just like, wow, there’s a lot of building, there’s a lot of cables, there’s a lot of optics, there’s an incredible amount of carpenters that are working here.

It’s wild. Right?

Peng, how do you feel about sort of the pace of these projects and are they at the pace that we need to be meeting the demand? And how, and then across border as well, what is the state on that?

Well, let me first say that the top expert here in building is Mike, who just spoke to you on the subject very eloquently. We do have a G40 to a branch of our business also focused on building what we call token factories. This is basically core and shell and also GPU management business.

To put this in context to answer your question, right now, some of you already know this, Abu Dhabi is undertaking the build out of a five gigawatt AI campus, which was announced last May when President Trump visited Abu Dhabi.

I have the simpler view, again, because I’m not a technical expert in building centres. My view, which I discussed with Sam multiple times, and you also, sorry, is eventually In the long term, the cost of intelligence will equal the cost of energy, eventually. This is a unique advantage why we are drawn, G42 and UAE Abu Dhabi, drawn into this business because we have a national advantage in energy production.

To have five gigawatt on the grid and no permitting issues to be able to build, it’s a blessing. And this is why we’re working with OpenAI, Microsoft, many others to provide capacity to them. As we speak right now, there are over 7,000 construction workers in the desert with over 100 cranes, building about 250 megawatt per quarter to deliver on this five gigawatt project plan we have.

So we’re obviously very bullish. And we believe UAE will consume a big chunk of that, but we also will be able to export those tokens as packaged intelligence to the rest of the world. So in UAE, to give you perspective, you mentioned that I seem to be doing everything.

Because we are serving as a fairly large company in a fairly small country, we’re serving the entire society using AI. And one of the tasks we have, KPIs this year, is to produce over one billion AI agents to boost our GDP. These agents range from everything on coding agents, to petroleum engineers, to cybersecurity analysts.

So we did our calculation. If we actually can deliver one billion agents, AI agents, by the end of this year, and they are actually working even just 12 hours, because they can work nonstop, just 12 hours a day, they’ll be consuming probably close to one gigawatt of AI infrastructure.

This is how bullish we are, and why we’re building this infrastructure in UAE.

So Sarah, you need to buy a lot of this capacity and have done so through just a number of different partnerships. I think it’s the one-year anniversary of Stargate. It is, yeah.

Today, which, how’s that project going?

So incredibly, this time last year we announced Stargate to the world and then very quickly actually had our president alongside Masa San, Sam, and Chuck Robbins actually from Cisco, and Larry Ellison from Oracle, stand up in the Roosevelt Room and talk about the importance of that build.

And what’s wild today, our Oracle campus, for example, we said at the time we’d do upwards of $500 billion and we’re already well over halfway in terms of getting that built up. We’re actually training models in that Oracle campus today on the latest chips. So it’s gone kind of, I think, better than we all dreamed a year ago.

It’s come almost a pinch-me moment. I think more broadly, like to kind of pull the threads together, a year ago or a year and a half ago, OpenAI felt like a little bit one-dimensional. We worked with one CSP, Technic or Weave under Microsoft, one chip provider, NVIDIA.

We had one product, ChatGPT, and one business model, a subscription for $20. Today, if I think of it as like a corner of a Rubik’s Cube almost, today we have a base of infra that is almost every CSP you can mention. We’re diversifying our chip portfolio, including our own inference chip that just taped out.

On the product side, we’ve gone from a consumer product that was really a chatbot to now be a task worker that can do things like healthcare for you. On the enterprise side, something that can go from a simple ChatGPT wall-to-wall deployment to APIs to now agentic behavior throughout the enterprise, we believe we’re there. And then an API platform.

And of course, Sora, because now we have multimodal. So the product platform is multidimensional. And then finally, the business model is becoming multidimensional, from simple subscription, through SaaS-based pricing, through enterprise licenses, through credit-based pricing, through commerce into advertising.

And then ultimately, I think we can do some real value sharing. For example, in drug discovery, what if we took a license down the line to the drug that is discovered and used that as a way to pay for it? So if I think of that Rubik’s Cube analogy, now you spin it, and you say, okay, I’m gonna take a low-latency chip, like Cerebris, that we announced a few weeks ago, I’m gonna create a high-end, really fast coding SKU, literally best in the world, and I’m gonna have a high-end subscription price for it.

It’s like I’ve turned all the yellows, I’ve created a block. And that brings you back to this conversation about build. We are just getting started, right?

You used baseball, since I’m a kind of American, but a Brit under the surface. I’ll go to electricity. I’m not gonna do a sports analogy.

But I think we’ve wired the house and turned on the lights, but we have not explained to people that now they could heat the home, they could cook in that home using electricity, they can have entertainment.

And so the capability overhead is massive. Even if models improved zero from today, there is still so much productivity to be had just with what’s in people’s hands.

I think OpenAI has also said you’re building your own data centers. Is that part of the mix in terms of your infrastructure footprint?

Today, I’d say we’re on a journey. We really are utilizing our CSP partners because it’s a way to stay lighter weight on the balance sheet, and frankly, to be able to work with folks, like for example, when we work with G42 and UAE, they have the local expertise.

They know how to do the land power shell. When we get inside the data center, as Mike knows, we will tend to bring to bear a lot of a very strong point of view. on what the kit and the rack and so on looks like sitting in that data center.

We have a whole scaling team whose job is to make sure that frontier models can get trained on that large fabric and therefore what does that fabric need to look like? What is the chipset? What’s the cooling?

What’s the power? How is it just connected? And so that’s a place where we’ve created a lot of, I would say, our own IP.

So we’re on a journey. We made an announcement with SoftBank Energy to do our first colo kind of build-to-suit. So think of that more as the next step.

We’re not all the way to build our own today because frankly, we have great partners to work with. But, you know, where will we be in three years? I think you said it well when just three years ago child TPT was just a new thing.

So we’re still young in our journey.

But your new chip is being tipped out already. That’s impressive progress. That’s the heart of data center.

Yeah, that new chip has, you know, it’ll be an inference-specific chip and it’s all about how do we keep driving down cost on the inference side, right? I find it amazing that from chat GPT-4 where it cost about $33 per token, per million tokens, today chat GPT-5 mini costs nine cents. Like $33 to nine cents is a 99% reduction in costs in about two, a little over two years.

That kind of blows my mind back to the point about just getting started because now you’re creating a cost backdrop that actually allows you to give access to everyone.

The other thing that it does is, and it doesn’t do it yet because we are still in a market that’s pinned, right? Like OpenAI can consume the capacity of the world. There’s five other models that can consume the capacity of the world, which gets to part of what you were saying.

But the drop in cost being as precipitous as it is, how many more ideas are going to get an opportunity to… come into existence and that’s what I mean by like we don’t, collectively, everyone on the stage, everyone in this room, I would argue, we don’t know what our clients or who our clients are yet because there is a, it is a natural, you know, like supply and demand don’t work yet, right, because the market is literally pinned to the red line and so it’s always up against the top but as we begin to see, and markets are good at this and so, you know, like they will ultimately bring supply and demand into some sort of balance which will allow for the creation of tons of new things that we just don’t know about yet and that’s really exciting also, matter of fact, it’s probably the most exciting part of it for me because I just think it’s, you know, just as people conceived of what exists right now that blows us away, they’ll have an opportunity to do it again in the future.

The demand is still very much in the early discovery stage.

Yeah.

And I think that one of the things that often gets overlooked, going back to the reinforcing nature of creating this intelligence, the number of lines of code in the world is going to increase exponentially.

We were talking about this yesterday and just from a BlackRock perspective, our ability to leverage our engineers to just be exponentially more productive, the net result of that is we need more compute for like the old stuff.

Yeah. So, I think that we are going to have a phase where just the old cloud requirements are going to grow even faster than people expect. In addition to this, and then you add on it, it’s very clear almost all the software in the world in the next 20 years, you can figure out if it’s in the next two or the next 18, almost all the software in the world is going to be interchangeable.

with the AI capabilities that we’re describing. So everything will be inference in terms of how it actually operates. So I think the nature of the technology and the innovation it enables winds up being self-reinforcing in a way that’s only gonna accelerate.

I had the opportunity to interview Andy Jassy yesterday and he was, that’s the AWS argument about the growth in cloud. He called it the middle of the barbell right now. Staying on chips for a second, he also talked about Tranium, which he also believes will be necessary to get the costs down as well as the individual efforts.

I wonder, in the infrastructure stack, how much change do you think we’ll see on the chip side of it? Obviously NVIDIA is the leader, but Michael, maybe to you, I know you’re, I think you’re all NVIDIA maybe at CoreWeave right now, or do you see a world in which you’re working with other chip providers?

So the way that we’ve built our business is we are led by clients and the best solution in market is and has been the NVIDIA technology set. And we can’t keep up with our demand for NVIDIA. It’s hard to allocate resources to, for a company that’s growing as fast as we are, and we are, with the exception of everyone else on this stage, unique, right, in terms of our growth prospects.

And that is a, that is said with the, it’s an incredible set of companies that are doing incredible things and going through a growth profile that is, you know, at any other point in my career would have looked at and would have like caused my head to break.

So. Look, you know, our belief and the way that we have constructed the company and our North Star is take the best computing infrastructure, build the best software solution to be able to orchestrate and deliver it, which will lead to the best product to enable the most companies to be as productive as they can in bringing their products to market.

You know, from where I sit, there’s going to be lots of different alternatives. There’s a lot of different ways to define best, whether it’s cost, whether it’s capacity, whether it’s what you can do with it, flexibility. There’s, you know, like the concept of best is not that simple in this space as we look forward in time and think about the different ways that compute is going to be consumed.

So from my seat and from the business that we are building and the scaling that we are going through, we think that the world will have lots of alternatives, but the best infrastructure out there is going to be built by NVIDIA.

It’s going to be built on a platform that is like ours, if not ours, and it’s going to have a software layer that’s able to deliver it to the market in its most performant configuration. And until the market explains to me otherwise, and markets have a good way of doing that, that’s our plan, that’s our strategy, and we’re going to keep our head down and keep hammering.

So another key component to all of this is capital, of course, and I think last year brought a variety of new flavors of financing deals from off-balance sheet deals to all sorts of new partnerships. I mean, maybe, Rob, from the investor side, are you seeing, I don’t want to call them novel because they’re not novel in the history of finance, but what are you seeing in terms of vehicles for financing data centers that’s catching your eye?

Well, I think this is a generational capital opportunity. where if you think about the numbers we’re talking about, and if you think about, you know, we have this term within BlackRock, the fast river, and you want to put your boat in the fast river to have it help you along the way, and this is one of the fastest rivers I think any of us will experience in our careers.

I think we’re seeing creativity, as there always is in the financial engineering side, but at the same time, I think we’re seeing more and more people looking to create the ecosystem in terms of not only having it as capital partners, but having it as true strategic partners.

And if you go back, for example, to BlackRock and some of the things we’ve done, we’ve actually looked to create investment vehicles that bring together as GPs, people like MGX, so leveraging the incredible innovation of Abu Dhabi, people like Microsoft, people like Nvidia, and bringing that ecosystem together so that way you’re able to prioritize focus in terms of actually bringing these projects to completion.

Sarah, what are you seeing and how do you think about this?

Yeah, so I mean, I think it starts with compute. So compute is the defining characteristic of what we need to supply the demand. It is the binding constraint today.

And I think it is a core competitive advantage because there’s not enough of it. And so if you don’t have it, you get slowed down. And we have faced that a lot in the past year.

People always ask like, well, what we know if you’d had more compute, what would have happened? I’m like, it’s quite simple. We have more compute, we’d have more products, we’d have more revenue.

We would literally have had frontier models pulled in by six months, 12 months, 18 months, because we gotta know what we wanna do. And it’s not just revenue. I’m a CFO, I care about our business model.

I’m gonna come back to it in a second. But we’re talking about breakthroughs in areas like healthcare. And if I’m a cancer patient who with that breakthrough actually gets my life saved.

In some ways, speed is even more of the essence right now. So I think we’re holding stuff back that just would help the world writ large. That said, from a CFO sitting in my seat who now has to work through how to pay for that, I mean, there’s nothing beats a good business model.

First and foremost, cash flow is king. And so what we have seen is as we have invested in compute, our revenue has kind of risen to meet the challenge, right? I use, often in these rooms, talk more about ARR because it’s the true load of the business.

It’s the compute I need. It’s the business I need to sustain it. So ARR has gone from $2 billion to $6 billion to over $20 billion just in the last, so in that moment where Chats GPT burst on the scene.

There’s never been a company like it, right? To Mike’s point, I was a research analyst at Goldman Sachs for over a decade of my career. If someone had showed me that model, I would have said, that’s crazy.

Like, you’re wrong. And we see just this continued momentum as we hit 2026, by the way, it’s not stopping. Our consumer business is hitting daily highs, which is super fun to see.

And our enterprise business is kind of shooting through the roof, right? Jessica’s not wrong. Davos is just an incredible moment to go talk to CEOs and folks who are making these business decisions, right?

I was saying to her as I hit the floor, I was just with the CEO of a large, like if I said it, like a world-famous bank who, you know, is feeling a little bit behind, had just spent time yesterday with Carlos Torres, who’s the chairman of BBVA.

BBVA wants to be an AI native bank. They’ve gone from 10,000 seats deployed to over 120,000 seats deployed. They’re rethinking their call centers.

They’re rethinking their credit, how they do credit decisioning. They’re in 25 countries, so they have to do multilingual. I mean, totally new way of thinking about enterprise growth.

To finance it, we are trying to be creative. Both, we’ve done equity-based financing. We did the largest equity round ever, $41 billion last year.

We have done unique partnerships. Our AMD warrant structure, I’m really proud of because it’s a great alignment of incentives. If we buy those AMD chips, we could get all the way up to owning almost 11% of the company.

And we think that if we do buy all of those chips, we really help create a lot of market cap for Lisa and team. And they deserve it. They’re an incredible company as well.

We have talked about our partnership that’s brewing with NVIDIA to fund us as we buy gigawatts of their chips as well. And so what we see is the ecosystem kind of rising to work with us, and that includes private equity shops as well, have been very, very interested in how can we deploy into their portfolio companies and with that create a business model around it.

So we are just getting started. But I do think it’s important that the ecosystem rise together, otherwise, we’re just going to go slower. And back to where I started, that means that that patient, that child in school or whatever is not going to get access to intelligence.

There’s a habit in my industry, and I’m not saying I believe this, but I’m interested in your reaction to it, to calling these circular financing deals in the sense that the sense is that all these deals basically amount to some amount of risk in the system.

Sarah, how do you think about that?

I mean, there’s an implication behind it, though, that the demand is not real somehow. And I think, I mean, we can all just start at the beginning of the panel again. I don’t think that we can deny that the demand isn’t here.

And I think in particular, people are still have, like a lot of the folks who are working on Wall Street or working in your world, Jessica, lived the bubble bursting before in the internet generation. And that’s because when I think about what the internet was like at the time, the point of email, if there were only like three people on the thing, right, you still had to send letters in the mail. I would say to all of you right now, if I went around the room, is ChatGPT not making a difference in your life?

Like, do you not immediately see value in it? And again, like for our frontier users, they’re using effectively, if you look at tokens as a measure of intelligence usage, they’re at 7x what just the average user today is using because they’re coding, they’re doing deep research, they’re in a university lab.

Countries, we’ve looked at it on a country level. The frontier countries are using 3X the amount of tokens already. And so that’s where I think the idea of kind of circularity and so on, it’s that implicit like, oh, the demand’s not there, they’re all just like trying to shore it up, that I completely refute.

And so instead, I just view it as the ecosystem, you know, the folks on this DS right now, right? We all feel like we’re short the thing. And so if we can push because Abu Dhabi and Peng are the most AGI-pilled of them all, or have the power.

And Mike over here knows exactly how to build a data center and to really think about the tech behind it. And BlackRock is doing incredible things to create financing vehicles, and actually has gotten very deep in data center technology itself, right? Those are all good things because everyone is bringing their expertise to the table.

I’ll add one more thing too, if you don’t mind, maybe zooming out even further. And Robin and I had a discussion about this during lunch, which is I’m a bit worried about many of our conversations here in the West becoming self-referential. I want to point out another country called China.

They may not have the best models in the world, the best computing in the world, but they’re doing what you termed as ruthless adoption. They are driving use cases in every aspect of the society. In fact, they’re a model for many of us to emulate.

They are charging ahead to use AI. And guess what? Being a user, influencing traffic, getting the intelligence out, can circle back to create better models.

Test, time, and compute, yeah.

Exactly. So study the Chinese usage adoption of AI technology. It’s a lesson for all governments.

Sarah, I want to ask one more follow-up on the business model question you talked about. Some of your competitors here at Davos have raised questions about why you guys are going to advertising so soon. I don’t know if you saw those tweets earlier today, but you’ve talked about your plans too and outlined your principles, but I don’t know if you have a reaction to them saying, oh, it’s a little bit early, they must need the money.

How do you think about the ad opportunity?

So, I mean, first and foremost, like the why, it goes back to my Rubik’s Cube. Remember at the top layer, I’m trying to create as much optionality as possible. What I know today is that 95% of our users for ChatGPT are free users because our mission is AGI for the benefit of humanity, not for the benefit of humanity who can pay.

And so in order to create access, I have to create a really strong business model. Early is a weird word because in ad models, you have to be at scale. Subscale ad models don’t work.

So that would be early. But when you have 800 million weekly active users, you’re already far beyond the scale of many companies who started in that model. But I think we have to be principled.

Number one, we are not gonna change the output of a model based on advertising. We have to make sure users remain fully trusting us that you always get the best answer, not the benefit of humanity and an advertiser who sponsored them, right? Second, we are going to be very careful about not sharing your conversations to advertisers nor selling your data.

And number three, we always wanna make sure there’s ad-free on-ramps, if that is your preference. But for many folks, and I particularly find it at Davos because you are talking about a globe where what I love about technology, it is the ultimate democratizer because I’m from Northern Ireland. I already said this once.

If you take a farmer in Northern Ireland, they can literally have the same phone with the same chat GPT level of intelligence as Bill Gates or Elon Musk or pick the richest man in the world at the moment, hopefully a woman in the world soon too.

Literally the same. They can’t drive the same car probably. They can’t have the same size of house.

They don’t go on the same fancy vacations, but they can literally have the same technology at their fingertips. But it is our mission to make sure that that can be the case. And so we want to be able to pull all levers along the way.

We know our users are using ChatGPT, not just for things like health, but also for commerce. It’s very natural to come down that funnel because you’ve had a very deep conversation. You might have started by saying, Hey, I’m expecting a new baby.

I really need a new baby stroller. What are the best out there in the market? But here’s my price point.

With memory, ChatGPT actually really understands a lot of this already. It might give you some interesting other ways to think about that purchase. And then what we hear our users saying is, Help me just consummate it.

And that’s where a very interesting conversation with an advertiser can begin. But I think appropriately so because you’re adding value to the end user. And I think that, again, we have to remember our North Stars.

That’s actually why we started by publishing our principles first before we even started testing.

I’m going to go to questions in a moment. But, Peng, you brought up China. And it reminded me of a conversation I had 15 minutes before I walked into this room.

I was talking to an entrepreneur who’s, I don’t want to give too much away, but building an AI company that’s not based in the U.S. And I said, Where do you train your models? And this person says, It doesn’t matter because I can do it so much more efficiently than everyone else.

And I didn’t have time to grill this person on what they were using or how they got it, per se. But it did make me wonder, I guess, but specifically, has China figured out something around training from a sort of research side to fundamentally change this equation? I know this came up in DeepSeek.

There was a little speculation. And then more broadly, if all of a sudden breakthroughs in AI itself lead to training models much, much more efficiently, is the infrastructure opportunity limitless? Or do we kind of hack that somehow on the model side?

Maybe Michael, maybe Peng.

I’ll comment on the first. I’m not an expert in this area. But I can tell you from my personal point of view, China is very good at engineering.

I think they’re superb in squeezing every bit of engineering advantage all of the infrastructure they do have. But I believe they are certainly behind on fundamental research and model breakthrough. That’s my view right now.

This is why earlier I mentioned where China leads today, in my opinion, is adoption. It’s in fact ruthless adoption, as Rob said earlier. This is where they potentially can gain an edge.

If we’re still debating, can we use it, should we use it, how do we use it, will the other nation or nations that are moving ahead to adopt it will be falling behind, even with the best model in the world.

And in terms of shortcuts to AI infrastructure consumption, I think when you look at China’s usage today, it’s really actually consuming more inferencing capacity. It actually needs more capacity to meet the demand of billions of people. Training, I think they are limited.

But on inferencing, I believe, globally, there will be a lot more demand.

I want to reframe the question. Sure. I think that it is important that we all take as a basic truth that there will be order of magnitude step functions and improvement and efficiency within this technology within the next five years.

Pick your flavor. Like the DeepSeek step function shocked everybody except anybody who even remotely touches AI. This technology, the way that we’re using it, what we’re doing, when DeepSeek came out, there’s two things that matter about when DeepSeek came out.

It was, what, two years? after Chad GPT-3?

Yeah, that’s right.

So like, the world kind of picked its head up and said, oh, AI, I get it, at Chad GPT-3. And two years later, like the technology is going through this incredible ramp of maturation. You have to work from the assumption that we will make step functions on the physical infrastructure, we will make step functions on the logical infrastructure, there are going to be unbelievable breakthroughs, and the whole world trembled when that happened.

Certainly the financial world, right? The other thing that mattered is every single one of my clients picked up the phone, started screaming at me, get me more GPUs now. I need more tokens, I need more tokens, get me more GPUs now, right?

And so that’s a indication from where the rubber hits the road, right? Where people are building products, where people are serving inference, where people are monetizing AI. That is telling you that that was an acceleration of the business, not a fundamental change of the business.

The improvements that were exhibited in the technology from DeepSeek were impressive. Full stop. But we also took tokens from $30 down to $0.09, okay?

So there’s a countervailing component of this, which is, okay, hit me with another DeepSeek, let’s take the $0.09 down to $0.0001, and tell me what gets built. The world is going to absorb the tokens. The world is going to consume…

It’s not the infrastructure that’s endless, it’s the voracious appetite for intelligence that is limitless. That’s the part that we have to focus on. And every single step function in the foreseeable future, certainly within the horizons we’re working on…

are going to do nothing more than accelerate the business, right? And that’s the way I look at it. That’s the way I’ve positioned myself to build my company as I look forward in time.

And I think, you know, whether or not people are working through it to the same conclusion, if you’re in this business, you have somehow, someway gotten to that point where you’re saying, yeah, intelligence is a net positive.

It will continue to be a net positive until human extinction.

I mean, could I just, though, just really fast on the training? Because I think we’re oversimplifying as well. Remember, what we know today, because again, three, four years ago, but there’s pre-training, which still needs large compute fabrics, lots of data and incredible algorithms coming from the smartest minds on the planet.

But then post-training is where we got to the reasoning paradigm. And our research should tell you post-training, so kind of what we call our O-series of models, is still probably more akin to chat GPT-3. The GPT model, which is the large, is already we’re at five.

But O, which is then when we do the post-training, is still, we have a lot that we can still do there. And then there’s exactly what Peng said, test-time compute. So at the moment where you’re making a call, there actually can be real training done on the inference side, like maybe even down at a device level.

And it’s combatorial. It’s not like one plus one plus one is three. It’s combatorial.

And so again, I think there’s a reductionism that the world likes to use. Markets love to use it, right? A single thing.

It’s also complicated to understand, but yes. Totally. But it’s much more rich and diverse than I think the market often gives it credit for.

And that’s why I think you get these wild swing moments, like DeepSeek, or before it was actually the laws of scaling are dead. Then there was the DeepSeek moment. And so market loves these moments.

But I would say if you look back at a stock chart or a market chart, it’s like it looks like the biggest divot of all time. And in reality, what we’ve just seen is this ongoing kind of motion for more.

Catalytic. I think if you zoom out… What you have to believe is that nothing is more positive for economic growth than productivity step functions.

And if you believe, which I strongly do, we strongly do, that AI presents a generational opportunity for productivity step function, whatever accelerates that is positive overall, even if in the short term there’s a little bit of disruption.

So I’m going to take the last question and, oh, sorry, I’m going to ask one more and we’ll do it fast because we have to end on time because some people have some places to go. So what worries you all? This is a very rosy, I mean, it’s a very exciting picture, I think, but what could go wrong or what could throw a bigger speed bump in the way?

What we’re doing exists in a broader backdrop and there are geopolitical, there are, you know, like, there are a lot of things that occur outside of our important but narrow, or at least narrow and expanding, let’s put it that way, line of business.

So what worries me? There’s, you know, geopolitical interactions that can distort the market, they can, you know, in terms of protectionism, they can distort the market in terms of cost of capital, they can distort the market in terms of accessibility, it can do all kinds of wild…

generations. And if I were to put on the top of my list what I think about from a how I manage risk at our company, which I spend an enormous amount of time on, that’s got to be first and foremost.

Okay.

I jumped on that because I knew everybody else was going to say it.

A very quick question because I want to make sure to get to one and we’re going to end right on time.

Okay. So AI is scaling much faster than planetary systems can regenerate. So my question is why will heat first regulation or water, energy and carbon?

And who should be responsible for that?

So I’ll start, but the builders of the underlying should also definitely chime in. So I was going to, to answer both your questions, I was going to say speed and the ability to bring people along so that they trust, right? If we all, like we live it, we breathe it every single day, but if you are just a kind of normal person living your life out there in the world, this can all feel like kind of tech gobbledygook is one of my favorite words.

And then there’s a lot of trust that still has to be earned. And I do worry that we talk in tech talk, not in like real people speak, like what does this actually mean for me? If I’m a mom of a diabetic kid, I work hard all day, I come home, I’ve got 30 minutes to cook dinner and it needs to be healthy for my child, how are you going to help me?

And the answer has to be, actually, I’ve got this thing on your phone called Chet-GPT, you could take a photograph of your fridge right now, we’ll give you a recipe, you can cook in 30 minutes or less, that is exactly what you can, your kid can have.

Like you’re like, hallelujah, thank you. And so then to come to your question, I think we also have to build trust in communities. Communities who are going to have a data center springing up, is that a good thing or a bad thing?

It’s going to bring me jobs or is it going to ruin my community? Is it going to use my water or are they going to be thoughtful about it? Is it going to raise my electricity prices or are they going to help protect my community?

Because I still have to put bread on the table for my child. And so I think we all have a real onus up here. to be bringing along communities and talking in their language, not at them, but with them, taking their feedback and really thinking about how we put that into how we build our businesses because we can get a little bit in our ivory tower.

I’m going to end it there for time. Thank you all for the fascinating conversation. Good to see you.