Market Predictions, Rates & Inflation, DOGE, CES, AI Compute | BG2 w/ Bill Gurley & Brad Gerstner

Elon said he thinks that every cognitive task that can be done by a human will be able to be done by an AI within 3 or 4 years.

Yeah. If you believe that to be true, the value of all that human labor that you're replacing is measured in trillions.

Good to see you, Brad. Happy New Year.

Happy New Year. I have on a blue shirt today—not a black shirt, as you probably can't tell, Bill. I have on my green pants, a little Notre Dame spirit for the big Notre Dame game tonight. I know you have a big Texas game coming up.

Yeah, that's tomorrow in Dallas. We'll have to see if both those teams could win. In advance, it would be pretty spectacular.

You and I were talking about some of this insanity going on in L.A., and I know we both have a friend who's lost a house. There's a lot of debate online about whether any of the policies we've had are proximate causes, at least, of the severity of what's going on. One of the things that I know we both appreciate is that, as we're entering 2025, the conversation and debate going on online is more robust than it's been.

I think it rubs a lot of people the wrong way, but I tend to be in the camp that more open debate and accountability is just better. No matter what side of the political aisle you end up being on, we should all be in favor of just doing better. When you look at how horrific these scenes are in L.A., you have to ask the question: What could we have done better?

Yeah, I totally agree with you. I have this framework in my head that I think about in relation to these types of situations. A lot of people evaluate politicians and policies based on what they think the intent of the decision was, but they fail to follow up and see whether the output or the outcome is identical to what they thought the original intent was.

All too often, the original intent of something may have sounded good, or you might be voting for someone because you agree with their point of view. But if the policy fails to achieve that—or, in many cases, achieves the exact opposite of that—then you really have to ask yourself, what's the point?

I hope that this type of accountability and transparency, shining a light on it, could lead to better outcomes in the future. There are always trade-offs, obviously. You can't have everything.

1. Frontline Ideas for 2025

For sure. I thought we could do something unique and different, if you're up for it, related to this time of year.

Obviously, large investment funds like yours typically operate on a calendar cycle. Sometimes the reporting is looked at annually, and sometimes even some of the fees and whatnot are calculated on an annual basis. I'm sure that creates an annual cycle for you.

As you've been going through that process, I thought it would be really cool to expose the listeners to both the analysis that you're going through now and a window into that process. How does someone—a large professional investor like yourself—think about this time of year? Specifically, what are you looking at now? Give everyone a peek inside.

No doubt about it. We've been having this conversation for 20 years, but half of our business—the venture capital part of our business—has a much longer cycle time. What informs how we think about the annual cadence of our public-market positioning, which is the other half of our business, is often these big trends that we're debating.

We ended last year in this great conversation with Dylan about the tension or debate in the world about how much compute the world needs, all the investment going on, and so forth. We'll get into that today. All of those things, including a major political change going on in this country, impact how we think about this year.

At the end of every year, I not only look at errors and omissions—what could we have done better in our public trading and our public investing for 2024? We had a great year, and I'm proud of the team, but there's always a lot we could have done better—but I also try to look ahead.

I try to put myself in the shoes of where I think the world will be in December of the following year. You really have to get in the habit of being an analyst, a forecaster, a prognosticator—thinking about the big trends, but also thinking about all of the competing things going on: interest rates, inflation, and the backdrop.

We go through that exercise. I journal to myself, and I make everybody on the team do this independently, Bill. Then we get together. This started at the beginning of December, but it certainly informs how we're positioned at the start of the year.

What's top of mind? What are the big blocks?

There are a lot of exciting things that can cause you to believe that this is the golden age, and that 2025 could be a really phenomenal year. Lower taxes, lower regulations, GDP growth is strong, and employment looks good.

At the same time, we have these megatrends that have been percolating for a decade and really seem to be bearing fruit. Obviously, AI is the biggest megatrend, but ancillary things like robotics and self-driving cars are yielding productivity improvements to the economy, which, as we know, is a huge driver of GDP.

There's a lot to be positive about. On the other side, you have to look around the world and ask: We have a situation in the Middle East—could that get better or worse? The situation with China—better or worse? The situation in Ukraine—better or worse? Those things can break both ways, but I could make an argument for how all of them get better.

On the one hand, there's a lot of enthusiasm. At the same time, valuations, Bill, are quite high relative to where we started in 2023 and relative to where we started in 2024. The world assumes that things are going to be better.

2. The Bogeyman (Interest Rates and Inflation)

As we look at it, I kept asking the team, “What is the bogeyman? What's the thing that we're not thinking about that could go wrong here?” If you look at the start of this year, I would say the bogeyman that's out there is that you and I have a lot of smart friends, and some of them have been shorting the U.S. 10-year. That is, they expect rates to go up.

I really think that inflation and higher rates, combined with higher valuations, could put a damper on the market, at least in the first part of this year. Again, this is just one of the potential outcomes, but we've seen the 10-year go up almost 100 basis points. People didn't expect that. They thought the Fed was going to be cutting rates, the 10-year would come down, and that would lead to more housing sales and more sales of a lot of things—a tailwind for the economy.

The exact opposite has happened. We have to explore why that's happening and where we expect it to go. That would be the overall framework.

Let's dive in on the positive side, specifically the enthusiasm around AI. I would share with you, just as an observer rather than as a participant, that when I listen to you talk, and when I listened to the short interview that Elon did at CES, I'm hearing a level of enthusiasm that feels somewhat unprecedented to me over the past 3 decades, in terms of how much excitement there is about what's possible and how much investment is going into it.

Why wouldn't that just be—first of all, do you agree with that? Second, how do you interpret it from an investor point of view?

There is certainly enthusiasm, but I would say there's plenty of a wall of worry as well. For every conversation I have with people in Silicon Valley who say they're going to invest more in compute, that this is the year of agentic AI, and that we may see AGI in the next 18 months, I have another call—maybe with a big fund in New York—who says, “Listen, this is going to lead to the telecom bust of 2000. This is Cisco all over again. There's a bubble.”

3. Mega Cap Valuations & Expectations

There are the words that everybody says, Bill, and the byproduct of those words is the valuations for companies. Maybe the place to start is just looking at where the valuations are for big tech as we start the year.

When I look at that, the S&P 500 recently peaked at about 22 times earnings in Q4 2021. It troughed in 2022 at 16 times, and now we're at 23 times. The S&P 500 as a whole is near its recent peak valuation.

If you look at companies like Meta, they're trading at 23 times. Google is trading at about 21 times. Nvidia peaked at 66 times and is now at 36 times consensus estimates. I would tell you it's closer to 28 or 30 times our estimates.

I look at the valuations for these businesses, growing at a rate that's unprecedented for a company of this size—

Correct. At the highest level, I look at that and ask, “Is that a bubble?” I say no. It's not as good a deal as you got at the start of 2023, and it's not as good a deal as you got at the start of 2024. But if we achieve the growth expectations we have for these businesses, then I don't think it's overly onerous.

The question really is: What does the market expect in terms of earnings growth? First, what was earnings growth last year, and then what does the market expect for earnings growth this year?

Here's a chart that shows megacap earnings growth broken out from the rest of the S&P 500. In 2024, the big 5—in this case, Microsoft, Meta, Nvidia, Amazon, and Google—grew earnings by 44%. Stocks were up a lot, but earnings were up tremendously. Forty-four percent is extraordinary.

If you look at the S&P 500 as a whole, because of the tailwind of the Mag 5, earnings for the S&P 500 as a whole grew 9%. If you look at the other 495 companies in the S&P, their earnings only grew 2%. It was pretty anemic earnings growth in the S&P 500 in 2024 relative to the Mag 5.

If you look at what's expected now—the consensus forecast for 2025—we see the Mag 5's earnings growth come down from 44% to 21%.

How much of that has already started, Brad? Do you know what Q4 was—that deceleration? I'm just kidding.

There's certainly a deceleration. Remember, these businesses are coming off easy comparisons, and by the time you get to 2025, they have very hard comparisons. They're still growing 20% on massive earnings and revenue bases. It's pretty extraordinary. Five years ago, nobody thought they would be growing that fast.

4. Big Tech CapEx

The interesting thing here that worries me a little bit about the market, Bill, is that we expect non-tech earnings to grow from 2% to 11%. That's a big acceleration in non-tech earnings. I asked myself, “Where does everybody think this is going to come from?”

The next slide is the answer to that. The yellow-shaded box shows that people expect health care earnings to go from plus 4% to plus 20%. They expect industrials to go from negative 4% earnings growth to positive 16%. Materials are expected to go from negative 10% to positive 17%.

Those are huge turnarounds and huge accelerations in earnings. A big component of that is associated with the tax cuts, but I think that only accounts for about 30% of that bump in earnings growth.

If you said to me again, “Brad, where is there a potential bogeyman?” it would be this: If those earnings aren't delivered—if you don't see this acceleration in non-tech earnings in the S&P 500—then it's hard for me to see how you can have a big year.

I don't think we're going to see a year that looks anything like 2023 or 2024. But if you simply asked whether we can get to 10% or 15% on the market, a couple of things have to happen. Number 1, you have to see this acceleration in earnings. Number 2, you really have to see interest rates not go above 5%.

I think if interest rates go up a lot, that becomes a big albatross on market performance in 2025.

Let's come back to the interest-rate question. I want to stick with the large-cap companies for a second. You and I were having a discussion about their capex trends, and this is remarkably different from any window of past tech investing. This level of capex wasn't part of the equation, other than perhaps in a manufacturing company.

Why don't you set this up? I know you put together a slide. What's happening with capex in these large companies?

Capex was growing before the ChatGPT moment, but nothing like we've seen it grow over the last 2 years, and we're forecasting it to grow for the next couple of years.

This first slide shows the combined capex of Google, Meta, Amazon, Microsoft, Apple, and Oracle—both what they did in 2024, which you can see was a huge step up—

Yeah, from about $160 billion to $260-ish billion.

Exactly. Bill, it consumed the vast majority of the incremental free cash flow of those businesses. They clearly all believe that these are really NPV-positive investments, and we can dig into that.

You heard [likely Satya] say—I think they did about $20 billion in capex in their most recent quarter—that they expected to spend $80 billion in capex this year. That's basically the run rate he was on.

Yeah, basically, it's just the run rate.

But when people hear the number, it's still pretty shocking.

Yeah. In the CES interview, Elon said, “That's a big number in anyone's universe,” which, coming from his point of view, is like, “Yeah, that's big.”

Right. If you go company by company, there was a debate—and we talked about it with Dylan in December—about whether people would actually spend this money. Would they actually buy more compute in 2025?

I think we can put that to bed. Based on the conversations we've had, what people are hearing out there, and the conversations at CES, these investments are going to show up in 2025.

The reason they're showing up, I believe, is that people are still seeing a lot of returns on training. We'll get into the different scaling laws that they're building for, whether that's pretraining, post-training, or test-time scaling. They're investing in making the models better.

5. Scaling Inference

The other thing that's coming on really strong is inference. Remember, Jensen told us on our podcast that 40% of his revenue today is inference, but inference is about ready to—

Because of chain-of-thought reasoning?

Yeah, because of chain-of-thought reasoning. It's about to go up by 100,000 times—a million times, maybe even a billion times.

That's the part that most people haven't completely internalized. This is the industrial revolution. This is the production of intelligence, and it's going to go up a billion times. Jensen expected inference to go up by 100,000 times, a million times, maybe even a billion times.

Inference is scaling very, very quickly, and that's all new compute that has to get built out to support it.

Looking at this chart that you put together, it looks like Meta and Microsoft are moving above 25% of revenue on capex. Oracle and Amazon are in the middle, in the 10% to 15% range, and then Apple, ironically, is falling to 5%, which makes sense because Apple is not investing in frontier models.

The other thing is that those same top 2 are approaching 100% of free cash flow, which is certainly a moment of pause. How do you frame either of those? How do you think about whether percentage of revenue should matter? Is there a limit that's too high? Then we'll do free cash flow.

I would tell you this is a very, very robust debate. I was out in Omaha in December, and I'll tell you they're pretty skeptical about the amount of dollars being spent.

They—the biggest investor in Omaha—are worried, like a lot of other investors who've seen these moments, that this is like the telecom buildout. I will tell you, though, that I have a lot of respect for Elon, Satya, Sundar, and the people who are making these investment decisions.

The numbers are there. As [likely Satya] told us on the pod, they have $10 billion in inference revenue, and I think he expects that to grow significantly. That is the ROI on the dollars that he's putting into the ground.

But they are making a bet, Bill. You can make a bet for 1 of 2 reasons. You can make an offensive bet that this will drive future revenue growth or profit growth in your business, or it could be a defensive bet—a prisoner's dilemma. I can't not do this because my competitors are doing it. I think that's where some of these concerns emanate from.

It's very clear to me that part of the reason the multiples on these businesses have not gotten even higher is this wall of worry that they're spending a lot of money and may not see the return.

Are there any rules of thumb? Does 25% or 30% matter to you as an investor when you're looking at that percentage of revenue? Then we'll do free cash flow.

6. Future of AI

Of course. You and I both know that when we're investing in a startup, we want—here's the irony—over the last 10 years, people celebrated raising bigger, bigger, and bigger rounds. You and I would look at each other quizzically and say, “People lost the script. The goal is the least amount of money in for the most amount of money out, not the most in for the most out.”

Today, the thing that's very clear to me is that over the last decade we've had super-high returns on the incremental capital that got invested. There's no doubt in the short run that those returns are compressing.

As an investor, you have to have an imagination to believe. Just like when Google was investing heavily in the early 2000s, or when Amazon was investing heavily in 2008, 2009, and 2010 in AWS, you have to believe that there's a pot of gold at the end of this rainbow.

I happen to believe it. I see the revenue growth inside some of these businesses—the inference revenue growth inside these businesses. When you replace human labor, Elon said in this recent interview—I think it was with Bill Miller at CES, and it was a virtual interview—that within the next few years we'll be able to do any cognitive task.

It obviously begs the question: What are we all going to do?

Elon said he thinks that every cognitive task that can be done by a human will be able to be done by an AI within 3 or 4 years. If you believe that to be true, the value of all that human labor that you're replacing is measured in trillions.

I think these are things that most of these CEOs have determined, even though that makes them a little uncomfortable. You said it makes the CFOs talk a little bit higher-pitched, and all that's true, but I think they've all determined they can't not make this level of investment in something that has the very high potential to be this big.

Looking at the free cash flow, it certainly causes you to step back and say that the past decade has been an unprecedented period of free cash flow from the Mag 7. The amount of cash that's been put on their balance sheets was so unprecedented, and everyone would talk about what they were going to do with all this money.

To move from that to a place where 100% of incremental free cash flow is now spoken for is certainly a change. My friend Mike Mauboussin would get mad at me for worrying about it because he would say that all that matters is not that they're using up their free cash flow, but what's the return on the incremental investment.

Of course, but that's super hard. That's the great puzzle. This is the thing where I think that, in the early parts of a phase shift—in the first 3 or 4 years of a phase shift—it's really hard for the traditional Wall Street analyst to model this.

Remember, I give this quote: In 2023, you had 26 analysts on Wall Street covering Nvidia, and the consensus forecast of all 26 of them missed by 80%. That's how wrong you can be if you're thinking linearly at a moment of a phase shift.

I think we're still in that moment, and that's why I think it's an advantage to be in Silicon Valley. You're spending all your time with the people who are actually doing this, being exposed to what they're actually building, and seeing what they're actually building.

Once it's well known—think about 2016—the cloud was well known, and everybody could model it reasonably well. But in 2012 and 2013, when you were earlier on in that phase shift, there was a real opportunity in companies like Snowflake and Okta to do things that people thought were not possible at that point in time.

Let's spend a second on what you call the prisoner's-dilemma argument. I guess you might include just the hyperscalers at this point.

As an aside, there has been talk that Meta has been hiring an enterprise group and is structuring new deals around higher-end versions of Llama or unlimited versions of Llama. Do you have a perspective? Are they entering the enterprise hyperscaler business?

I have no evidence that they are. I take Mark at his word from last year, in one of his podcasts, where he said they already charge license fees to the large hyperscalers to use Llama and to provide it to them in certain ways.

I think it's de minimis revenue in the scheme of Meta, but he's smart enough to maintain the optionality. People have found searches for executives to come in who have certain backgrounds. I don't know that that is in fact occurring, but again, you could have revenues of a couple billion dollars and, in the scheme of Meta, that's still not all that material.

They're smart enough to maintain the optionality, Bill. You saw Nvidia release some models, which is the packaging of Llama models to make them easier for some customers of Nvidia to use. It would only make sense to me that Meta is thinking about that.

Let's include Meta for the sake of this question and argument. Let's say there are 4 hyperscalers, and this is the prisoner's-dilemma thing. Every one of them is being forced to announce their capex, or they do it as a matter of their earnings call. They're all going to give us their capex.

They're not necessarily being forced to do it, but they announce it as part of their earnings call. It's become this point of focus. Everyone's paying attention to it.

If Satya were to say $60 billion next year instead of $80 billion, does he have a concern that people would perceive that as meaning they don't believe as much, or that they're losing ground to other people who are taking share?

Do you get into this reflexive argument that, if we want to be perceived as leaders in AI, with confidence that we're going to take our fair share, don't we have to announce that we're betting via capex?

It would be fairly conspiratorial to think that they're doing this just so the optics are that they are leaders. What I will tell you is that [Likely Satya] has said they have been compute-constrained for all of last year, and their inference revenues are going to accelerate in the first half of 2025 because they will be less constrained.

OpenAI has said publicly—Sam has said many times—that the reason they can't release models widely, the reason they couldn't release some of the voice products that you wanted, and the reason they couldn't widely release Sora is because they were compute-constrained. They didn't have the compute they needed to support these models.

Part of the reason they charge higher prices for Pro models is that they have to artificially reduce demand because they don't have the compute. I think across the board there's compute constraint.

That occurs for 2 reasons, Bill. Number 1, you have over 300 million people a week who've decided that they want to use ChatGPT instead of search or something else to answer their question. The usage is bigger than people expected.

On the other hand, these new models require inference-time compute, which is a compute hog. These things require huge amounts of compute. Those 2 things combined mean that we simply didn't have a compute infrastructure in 2024 that kept up with demand.

I think the investments we're seeing in 2025, frankly, aren't even going to get us to the point where we have a compute surplus. I think we're going to end this year compute-constrained, and I think a lot of the frontier labs feel the same way.

One quick aside on that: You've seen Google release a high-end, paid-for consumer product—a product for an individual user, because it could be a consumer at a company as well—with a price point on it.

When Elon was talking about their release of Grok 2 and potentially Grok 3, they said Grok 2 is always going to be free. I infer from that that Grok 3 might be paid for.

This, to me, is a positive outcome for OpenAI because I think when people frame the fight as a competition with search, they assume that the frontier is going to be free and not paid for. If multiple people fall in line behind them with subscription pricing, that could start to get sticky.

It's going to be very, very difficult for any frontier lab to invest at the level that's going to be required and not have a robust business model behind it.

What I would tell you, without saying anything other than what's been said publicly about OpenAI's revenues, is that they ended the year—I think it was rumored—at a $4 billion to $5 billion run rate, growing very robustly. You would expect a company at this stage to be growing at least triple digits.

You start thinking about $10 billion-plus in revenue for this company. Now compare that to some of the other companies. Mistral and a lot of the other model companies have fully pivoted. They've already raised their hand and waved the white flag: We don't have the revenues that can support the spend to keep up.

It's going to be interesting to see what Anthropic does. Their revenues are reported to be less than $1 billion, mostly on the enterprise API side, not from the consumer side. Can they afford to keep up?

As large as Google is, and as large as X is, I don't think you can go out and spend without having a business model to support the spend. What you're hearing would make sense to me.

7. Co-opetition in AI

One other thing I want to say about this, Bill, is that we've spent a lot of time internally trying to model out what we think of that compute demand relative to compute supply. Are we really constrained?

I want to keep coming back to this point that Jensen made. Maybe we'll insert the clip here on the pod, where he talks about how his inference revenue is already 50% of his revenue—already upwards of 50% of his revenue. I asked him whether the mix is going to go up, and he said, “Well, Brad, of course, because I think inference is going to go up by a million times or a billion times.”

What drives that, Bill? I think you've heard a lot of people say that 2025 is going to be the year of agents. You have these o-series models, and now you have deep reasoning out of Google, which launches this whole different vector of scaling intelligence and reasoning.

The thing about those models, as you well know, is that they are compute hogs. The number of tokens that you have to produce and repopulate back into the prompt, and the number of branches of inference that you go down, dwarfs single-shot ChatGPT.

Now add in personalization, long-term memory about each individual user, and actions—“Book my hotel. Book my...” Do things for me. Again, those things are all compute-consumptive.

That's why I think the frontier labs like OpenAI are modeling out the expected compute demand and then looking at what they have and saying, “We're not even close.” We're not even close.

I do want to come back to that, but I want to finish the capex thing real quick. Obviously, one way to look at this—you have a slide with Google, Amazon, Microsoft, Apple, and Oracle, and they're all spending this money—is to ask what it means for those stocks.

But I'm sure as an investor the easier thing to consider is that these companies are telling us and forecasting that they're going to spend this amount of money, and the processes for spending that amount of money are sticky and slow. They're not fast. You can't back out; you pre-commit and go build.

Who are the recipients of this stuff? That's an easy win.

It's a great question. We started the conversation with how we think about the framework for the year. What I didn't say at the start is that we started 2023 with 95% net long. Think of that as all your chips on the table—our longs minus our shorts.

We started last year, I think, around 80% net long. That can go up and down, but it's our chips on the table. You might say, “Well, Brad, if you think it's the golden age and you think all this money is being spent, why the hell do you have fewer chips on the table?”

I just told you valuations are a lot higher, so let's start there. Warren Buffett and Charlie [likely Munger] have told us that the most important thing in investing is the price of entry. The price of entry to play is higher, and the variant perception is lower.

When I look at Q1, when it comes to the Mag 5—or the Mag 6—what I say is that there's going to be a lot of foreign-exchange headwind. The strength of the dollar has gone up a lot. A lot of these companies have revenues denominated in non-U.S. dollars, so their revenues will have headwinds from foreign exchange. I'm not sure that's totally appreciated.

Secondly, their capex is going up a lot. I think Dylan said on our podcast—and I agree with him—that capex is higher than people think. Both of those things aren't particularly good for the biggest companies trading at high valuations.

I don't think it's going to cause some cataclysmic event, but it is a headwind.

But who are the recipients? Nvidia is obviously the biggest one, but who else is a recipient here?

I first want to show a slide on Nvidia because I hear a lot of people say, “Oh my God, it's...” We owned it since it was $1.20 a share. Today, split-adjusted, it's about $145 a share, so it's gone up a lot. But people say, “I remember when it went up 2 times, and Jim Cramer was telling everybody to sell it. It's up 2 times.” Then it went up 3 times, and people said to sell it.

Every time, you have to recalibrate based upon the facts on the field. It went up a lot because its earnings went up a lot. It went up a lot because its revenues went up a lot.

Here's a slide, Bill, that just shows the consensus data-center revenues for Nvidia. The key is to remember that Nvidia's data-center revenues are mostly these chips that are driving training and inference, and this is a business with very high margins.

In 2023, it was $61 billion. In 2025, it's estimated to be close to $200 billion. If you notice the jump between 2024 and 2025, again, this is consensus, it's about a $60 billion increase.

If you simply add up the hyperscalers that we just went through, that's $60 billion. Where else are they spending the money? They're spending some on custom ASICs, but it's tiny on a relative basis. Most of this stuff is going to Nvidia, by and large.

In 2025, Jensen told you at CES that he's got 40 AI factories. He doesn't have 6 people in the world building this stuff; he's got lots of people building this stuff.

I actually think that, if you look at the consensus numbers, they probably imply that Nvidia's market share is going down from 2024 to 2025. I don't think that's true.

So you're comparing expectations for Nvidia with this precommitted capex spend and your expectation of what percentage of that Nvidia will capture?

Correct. Remember, Nvidia's stock has not gone up from June of last year until now. It's basically flat because this debate has been going on.

We had a lot of debates about whether we've hit a wall on pretraining scaling, whether the models can continue to scale. We had the DeepSeek model come out. It was a smaller model and very capable, so people said maybe you don't need all these chips.

There is real tension in the world about this name.

Another question that comes up is whether, if we're moving from more spending on inference than pretraining, you need this big, large, holistic cluster. Can you use a more distributed approach? Would you use products that aren't GPUs—TPUs from Google, or some of the startups' products? Of course, they can't ramp to anywhere near this scale anytime soon.

Google has said, and I'm pretty sure he said it publicly, that they're spending a lot on GPUs, not just TPUs. But you asked the question: Who are the recipients?

Nvidia is clearly high on that list. Remember, Nvidia was up 140% or 150% last year, but Intel was down more than 20%, and AMD was down on the year. This was not a situation where everybody in the semiconductor complex won.

Another group—I know I share this feeling with my friend Gavin Baker—is that we're also investors in the memory space. SK Hynix is providing high-bandwidth memory, which becomes very important, particularly in a world of inference-time compute.

We think we have a memory shortage, which is a key part of the Nvidia supercomputer ecosystem, for as far as we can see. Micron and SK Hynix would be in that memory space.

Let's pause on SK Hynix. When you look at companies like Nvidia, or others that are doing well in this environment, they have multiples that suggest everyone understands and buys into the story.

If I'm reading the numbers correctly, SK Hynix trades at about 6 times forward earnings, which is insanely low for most companies. What's going on here? Do people not believe the number, or is this the commodity argument that happens with hard drives, where you buy it with the highest multiple and sell it with the lowest multiple?

I think the reason it has that multiple is that people do think it's hard drives. They think it's commodity memory, that it's a boom-and-bust business, that you can easily put a lot more supply online, and that supply will ultimately drive down the cost and margins. They also think demand is cyclical.

We believe those things are not true. We think this has become a far more sophisticated product, with lots of software baked into it. We think this is secular, not cyclical.

8. Role of Power in AI Development

There are 2 ways to win. Number 1, you get the earnings growth of the business. The second is that you could have a re-rating higher in terms of the multiple people are willing to pay. If SK Hynix gets treated more as a noncommodity company over time, people could come to believe that.

Of course, you and I talked a lot around the Diablo episode last year. We have power issues in this country. How are we going to power up the 5, 10, or 15 gigawatts of data centers that we need to bring online?

Remember, if they're spending this money, Bill, then they're telling you that they are. You have to have power, you have to have a shell, and you have to fill it with chips.

One of the things I continue to have people tell me is, “I know this is hard to believe, but we are power-limited. We have capex that we would put online that we're not putting online because we can't power it.”

I had a call this week with the COO of PG&E and with the head of Diablo Canyon. I'm very focused on talking to our friend David Sacks and others in this administration. We have to give regulatory relief; otherwise, we're going to have huge hurdles placed in front of our AI industry.

China is building 100 times as much as we are in terms of power, and power is the single most important primitive to AI. I'll give you one example: Gavin [likely Newsom] still has not signed the extension for Diablo Canyon beyond 2029. It's insane. We have to make that happen this year.

It's 10% of California's clean, green power. We have to extend it so that they can begin doing the appropriate planning. You and I think they should probably be expanding it, but at a minimum, the idea that you would take 10% out of the grid at a time when we're already capacity-constrained is totally insane.

Let me ask you this: Altimeter has historically been tech-focused. When your information leads you in this direction, do you start moving in and out of energy companies?

A lot of my tech peers have. If you look at some of the highest-returning companies in what I would call the semiconductor-related complex last year, they were in fact companies in the energy space. We looked at all of those.

Ultimately, I think one of Altimeter's overarching North Stars is essentialism. Don't make anything more complex than it has to be. I would always ask my analysts, when they brought me a power company, “Why is that better than Nvidia?”

It's a derivative Nvidia. It's the exact same bet. Why not just buy more Nvidia? We tend to take bigger bets on our best ideas rather than diversify into things that we know less about.

The truth is, I do know that we need a lot more power, Bill, but I don't necessarily know exactly how the regulation is going to play out, or exactly how small nuclear reactors and all this other stuff will develop. This falls into the Buffett comment: There's a fool in every market, and if you don't know who it is, it might be you.

9. Interest Rates & Economic Growth

One of the things we ought to come back to, though, Bill, is that you asked what the bogeyman could be, and I said interest rates. Maybe we should spend a few minutes double-clicking on rates—what's going on, why rates are going up, and how that may or may not change.

If you look at this chart, the black line shows that the 10-year, over the course of the last couple of months, has been trending up and is now around 4.7%. This is despite the fact that inflation has largely—remember, just a couple of years ago we had a 9-handle on headline inflation.

I remember saying that a couple of years from now we'd be back to a 2-handle, and people laughed. They said, “No way.” In fact, that's where we were. If you look at the Morgan Stanley consensus forecast for this year, they expect inflation to finish the year at 2.2%.

You have to ask the question: Why, then, are people concerned? What's going on? I think there are a few things at play here.

Number 1, the Trump election got people really excited about more growth in the economy. You're going to have $400 billion of stimulus on an annual basis if the Trump tax cuts are passed. That's a lot of stimulus into the economy.

Regulatory relief will stimulate the economy. What is that stimulus going to do? A lot of people are concerned that it could reignite inflation. They're saying there's a chance that inflation goes higher, and if it does, rates aren't going to be able to go lower.

If you look at the next slide, the expectation in the world has gone from a lot of rate cuts to only 1 rate cut in the first half of next year. Basically, the market is saying that there's too much stimulus and we're not going to get rate cuts.

I think in the first part of this year there's going to be a lot of tension around whether core PCE is continuing to roll over. We think that a lot of the components, like rent-equivalent and shelter, will cause it to continue to go down.

10. Federal Spending & DOGE

The other thing going on here—and this is something I know you and I are close to and care a lot about with DOGE—is that the market is saying, “We expect this stimulus to come in, but we do not expect Congress to have the courage to cut an equivalent amount out of the deficit.”

If you think about this, if we have a $400 billion tailwind from the tax stimulus, I think what the market would like to see is $300 billion or $400 billion of cuts come out of the budget. When you look at what Congress has done over the last 5 or 10 years, there's no evidence. We've never done it.

The market is just saying, “I'll believe it when I see it.” Trump has talked about interest rates being too high. He talked at the press conference the other day about wanting them to go lower.

If he talks to his chief economist in the White House, I think they'll say these are the issues at play. We really have to make these cuts.

We have a reconciliation package, so all of this is going to get determined—they've told us—in a single reconciliation package by April or May of this year. This is the backdrop that's going to impact valuations, and that we have to grapple with and try to understand.

Here's my forecast: I think DOGE and this administration are actually going to cut a lot out of the budget. I think you have the leadership in the House and the Senate.

If they cut hundreds of billions of dollars out of this budget, then I think the market will say, “Great, fiscal discipline. We don't have as much stimulus.” I think you'll see the 10-year come in.

If the 10-year comes in, then I think it could be off to the races. But if we do the opposite—if we don't make the cuts, and the 10-year goes to 5.25% or 5.5%—it's going to be an anchor on the stock market and an anchor on the economy.

I did a little analysis that I shared with you, and I want to review it with folks on the pod because, given how important I just said this is, I think it's important to try to get our arms around it. Is it possible—can we even cut $300 billion or $400 billion annually out of spending?

Here's the federal spending chart. We went back to 2019, and you can see the total spending by the federal government in 2019 was $4.4 trillion.

You can see the subcomponents: Social Security, where we spent $1 trillion; Medicare, $644 billion; Medicaid, $490 billion; and the different categories under that. Defense was $676 billion, and so forth. You can also see what our net interest expense was at the time, because interest rates were relatively low and our debt was a lot lower.

The next column is what we actually spent in 2024. This is our best estimate. The CBO has given us an estimate, as you can see in the footnotes. Total spending was $6.7 trillion.

The question is: Is that what we would have expected? Not what we expected. What we did, Bill, is create what we call a fiscal-year 2024 baseline. How did we determine that baseline? We went back to 2019 and grew every category roughly at 2.5% per year.

We said GDP has grown at 2.5%, inflation is growing roughly at that rate, and the population is growing at that rate, so that's what government spending should roughly grow at. If you look there, the 2019 baseline budget adjusted at 2.5% says we should have spent $5 trillion, but instead we spent $6.7 trillion. That's a $1.7 trillion differential.

When you hear DOGE talk about the ability to find $2 trillion in savings, this is, I think, what has people optimistic that there's an opportunity to make cuts. You go through there, and I think there are some interesting ways in which we can go after those savings. There are some proposals, and I'll post one that goes through and says, "Here is $700 billion of easy deficit reduction that both Democrats and Republicans agree on over 10 years." That doesn't get us there, but it is certainly a start.

I think it's going to be critically important that we get our arms around federal spending this year, because that, to me, is the biggest potential bogeyman out there. There are ways that Trump can potentially do some of this unilaterally through a decision challenging the Impoundment Control Act and just refusing to spend what Congress allocates. But given that the Republicans control both houses of Congress, I'm expecting, and I hope, that we see cuts that at least offset the tax cuts.

Yeah, I mean, one of the challenges, obviously, is some of this spending. You take healthcare, for example: those categories, on a per-population basis, have been growing way above your 2.5% baseline assumption and expanding as a percentage of the overall spending for the government. It may require you to attack that specific problem within that industry, not just say, "We're going to spend less."

Listen, no doubt about it, there is a global demographic problem. You and I can look at any population chart of any country on the planet: the number of people over the age of 60 is increasing at a very fast rate, and the number of new people coming into the workforce is slowing down as a total percentage. That means that your working population, which is paying taxes to pay for all of these things, is a much smaller percentage.

Demographically, that is going to be a challenge. But so what? How do you deal with it? You can either just give people fewer things than what you've promised in the past, or you can come up with more efficient delivery mechanisms. Obviously, we know about all of the easy and crazy shit the government spends money on that DOGE has been talking about, which we can cut.

My assumption is we're also going to have to harness technology and AI. Defense spending has got to get a lot smarter. You really have to go through every category, and if you zero-base this, I tend to think there are huge opportunities here. But listen, the track record of Congress, both Republicans and Democrats, has not been good at cutting these costs.

What I would say is that if we don't do it, then there's a real risk that the bond vigilantes come into the bond market. You end up with higher rates because they don't believe that we're on a stable fiscal path.

Look, I think it's an assumption in American life that the government's inefficient. I don't think there's anyone anywhere who would stand up and say, "We think our government is great at execution and spending dollars." There's no one who defends it. They might argue that it has to exist regardless, but they don't argue that it's efficient.

So there's plenty of opportunity. Elon said it's like shooting fish in a barrel. If I'm being asked to look at inefficiency in the government, there will be opportunity. I think there's an even bigger opportunity, which he also hinted at: if you find policy that's restrictive and unnecessary, you could unlock GDP growth, which helps in this equation as well.

I totally agree with you that the inhibitor here won't be identifying or knowing what to do. It will be whether the system can actually reform itself, and we'll have to find out. We'll put in here the charts that we showed last year on how you get to a balanced budget in 2029, but it's just not that hard to get to a balanced budget at $6 trillion.

That's $1 trillion above what we should be spending on the baseline today, but it just gets you back to that 2019 baseline. It assumes that we have some acceleration, but not a lot, in terms of growth. We are going to get acceleration: cutting taxes by $300 to $400 billion a year is going to accelerate it, and we are going to get efficiency and productivity gains from AI. These things are happening.

By the way, relative to Europe, which is a basket case and a disaster, the U.S. is in such an incredible position. We just have to have the courage of our conviction and get rid of some of this wastefulness. That's the only bogeyman I see out there.

11. Regulatory Landscape for AI

But, Bill, there's one other thing. We've talked a lot about spending; let's talk for a second about regulation. I know there's an AI bill brewing down in Texas, akin to California's SB 1047, that we got killed. So, as we have a lot of changing politics, I suspect that Trump is going to rescind all, or at least a major part, of Biden's executive order on AI. What's going on in Texas and around AI?

Well, you and I have talked about this in the past. There had been a huge movement underway, and I would call it an unusual movement, when there's a new market evolving and people are literally begging for regulation. That got a bunch of different parties up on both sides. Part of those efforts led to the Biden executive order, and they also led to this big push in California, which we talked about. What was it—SB 1047? Everyone got on one side or the other, and everyone in our community had a point of view.

You had Scott Wiener as the person inside the legislature in California pushing for that, and it ended up with this huge argument, followed by Gavin [likely Newsom] vetoing it. I suspect that most people see the administration change and our friend David Sacks coming in as the AI czar, and they would expect that this is, at least for now, put to bed.

What people are probably not aware of is that the people pushing for this regulation have moved underground to a certain extent, and they're pushing it in a bunch of different states. I've heard there are as many as 25 state-by-state initiatives. I guess this is how policy works in this country, but there are a bunch of reasons to be really worried about this.

The one that's got the most heat right now is in Texas. I think Governor Abbott, and obviously Elon and all his companies, have had this massive impact on the Texas economy by making it the state with the least red tape. It's the most pro-innovation and pro-business state, and for this thing to pop up there is just so ironic in my mind. It would be literally brain-dead.

When SB 1047 was being proposed, people, including myself, said, "Why don't you just write a sign that says, 'Move your AI company to Texas'?" Well, now those words sound stupid because Texas is doing this. The first thing that people should realize is that there is no reason to do this on a state-by-state basis.

At the end of Obama's term, he had an initiative that he didn't get around to, which I wish he had. He had identified about 100 different industries where there's state-by-state regulation, including hair care and things like that. It's just wasteful red tape.

When so many people in Washington are worried about our competitive position versus China, to build a gauntlet that our companies would have to move through to adhere to state-by-state interpretations and rules is just mind-numbingly—I want to say the word "stupid" out loud. The only reason I fear saying it is that someone trying to write that legislation would use it against me.

I can't imagine it. We have a changing administration, and we have people looking at this. If we feel we just have to do something, God, please, I'd prefer it at a national level with federal preemption. This is the Interstate Commerce Clause. There's nothing more advantageous to us over Europe than the fact that they have this crazy patchwork of regulation across all these different countries, whereas here, if you're operating an internet company, you have one rule of the road.

We need to have one rule of the road around AI. I think it should be promulgated, and it will be promulgated, at the federal level. I've heard this bogeyman out there that people say, "Everybody in Silicon Valley doesn't want any regulation. They're just a bunch of crazy libertarians." No. What I think you're hearing is pushback to uninformed regulation that would slow us down and cause us to lose an important race to China.

If we care about the race with China, then the first thing we need to do is reduce the impediments to running the fastest race we can run, while still caring about AI safety, national security, and all of these things. But if you have this patchwork where these states are onerously regulating all of these companies out of the gates, all you're doing is handing the gold medal to China. I think unintended consequences are really bad.

Let me tell you about some of the details, and I'll include a link in here. Dean Ball wrote a really solid analysis of everything that's wrong with the Texas proposal. One of the things that's in there is that you have to do a risk assessment, similar to creating an audit. For those of you who've run companies and gone through audits, you know how difficult it is.

Not only would you have to do it, but you would have to publish a risk assessment. So, whatever product I'm working on, I've now stepped through these hurdles and written these reports. Then there's liability associated with it looking backward. If I do something bad, I'm liable for it, but I also get a second look backward: did I run the analysis where I should have known about the risk that was there?

I give the people looking for regulatory capture credit. I think they saw that they were losing at the national level and realized that they could create messy anxiety at the state level, which may actually promote trying to get to a national outcome. But this would be horrific. I think they're going to get a lot of pushback. I would be shocked if it happened in Texas.

I think you have a lot of coordination going on in Washington, and I think it's quite smart. I think it's on both sides of the aisle. You had a great report published by Congressman Jay Obernolte in the House, who was co-chair of the AI task force in the House. The Senate has done some work, and hopefully Sacks will get there and provide a really clear coordinating function. That will have federal preemption and get great direction out of Congress.

One other thing, Bill: I'm just thinking about 2025 predictions. Last year, when we were talking about regulatory capture, I think there was a lot of fearfulness, and a lot of it was directed at OpenAI—that these models and companies with a closed model were trying to lock everything down before people could catch up with them, trying to squash what was happening in open source.

I think you're going to see, in 2025, a lot more companies open-source more of their models. We just saw Microsoft put out some open-source models last week, which is really surprising. My own sense is that there's a lot more commonality between all of the major players around AI than there is division. We'll see, but what I'm hearing out of a bunch of them is that we're going to see a bunch more open source.

Yeah, but there are 2 dimensions to this. You could be right that the bigger companies could all agree on a regulatory framework. There's also the issue of small versus large, and there's a lot to pay attention to there.

Before we wrap up, I had 4 or 5 things that I'd like to bounce off of you and get your opinion on as I look forward into the year. The one that I've been thinking about a lot is Google.

We talked a lot about how search is at risk, and clearly, for those of us doing so many searches first on ChatGPT or another product like that, there is a real argument as to what's the point of Google in the long run and whether that still holds. On the other side, when we've looked at them in the past, we've said, "Man, they have an incredible number of assets."

When you look at their Gmail product, their Docs product, their Slack-competitive products, their Zoom-competitive products, and the fact that they own Android and this mobile operating system, I was surprised to hear Jason on All-In get super excited about Google in this way and announce that he had switched to a Pixel phone.

I'm thinking to myself, this could be the perfect AI basket of assets. You look at a company like Glean, which I know your company is doing well, and you can create Glean for the rest of us if a company commits to being on the Google stack and has all these things. It becomes a combination that no one else has. Apple doesn't have it. Microsoft doesn't have it. It requires insanely great execution to get it all right, but I find myself wanting this so badly that if a whole bunch of people said, "Man, I want this so bad that I'm going to switch from my iPhone and get onto Android," that would be a real tell. That would be something special.

Yeah, well, I'll tell you: I don't know what it was—late November, maybe early December, before the recent run-up—but I started seeing a bunch of breadcrumbs around Gemini, Deep Research, and NotebookLM, just a cycle time improving there.

I talked to a lot of companies. A lot of companies were still seeing growth in their search volumes and their lead generation—online travel companies, et cetera—from Google. I started to see a lot more evidence of them just getting fit, getting more efficient, and, frankly, I think a lot of credit goes to Sundar. I think he's starting to get feisty and see the things that need to get done here.

I would say let's keep a few big pieces in perspective. They are facing the largest innovator's dilemma in history, certainly from my perspective—the history of Silicon Valley. I think it's almost impossible to replace a 99% incremental-margin business, i.e., a monopoly search business, with whatever comes next, because whatever comes next, they may be important in it, but they're not going to have a 99% monopoly. OpenAI is going to be there, Meta's going to be there, et cetera, and I don't think their margins are going to be the same.

So that's the first thing. Number 2 is, how could I be wrong? You could envision a world where the pie grows so much, Bill, that Gemini—and if they were able to displace Apple on the phone—could really replace that. I think you have plenty of time to wait and see and pick up the breadcrumbs on whether those things are occurring.

I will tell you this: I suspect Apple, for the first time, is looking at Google's assets and saying, "Okay, this is orthogonal." They were never worried about Pixel over the course of the last 8 years, but I imagine, to your point, this is the first time in a long time that they've worried that Google could appropriately embed Gemini, Deep Research, and a search assistant on the phone that could book my hotel and just make it a 10x better product.

I'll tell you another area where they have an advantage, which is hopefully useful to everyone out there. Gemini is really good at local because they have all the local reviews. I find myself now, if I'm going to a restaurant, asking, "What are the 3 dishes people have loved the most? What are the 2 that you stay away from?" You can get down to a level of detail that you wouldn't have used before, and it's very valuable.

What do they like at Tony Black[?]

Well, you have to go do—

The second one that I think is super interesting to me, that we're going to see play out in the first half of this year, is X.

xAI, Elon’s commitment to the largest cluster—the concerns about running out of data, the concerns about maybe there being a parameter limit, at least against the text dataset—draws into question: Do you need the largest single cluster? Grok 3 has been promised, I guess, in the first half of the year. It would be super interesting to see if something comes out of that product that is competitive with the cutting edge at OpenAI, or maybe above it. I don’t know, but I’m very curious about how that plays out.

I think there’s been a lot of attention focused on that. I’m not sure that I would measure xAI’s success solely on the dimension of, “Is the Grok 3 model better than what OpenAI’s latest is, or Gemini?” Because remember what they’re doing, really, is building a larger cluster on pre-training, so they can do more on pre-training. But remember, these other models are infused with post-training and test-time compute, et cetera.

I think it’s a test of that pre-training argument. Let’s say they do achieve something and come out with something that’s better than everybody else. That would be a new data point relative to where everybody’s thinking.

Here’s the way I think about xAI and Elon: He’s got 3 vectors on which he is leveraging AI. Number 1 is robotics, and he recently said—and I’ve heard Google and others say this—that he expects there to be millions of humanoids alive and in the wild by 2028. Think about that: That’s 3 years away. It’s really incredible.

We’ll post this video of this Chinese humanoid that somebody said passed the visual Turing test because humans were confused about whether or not it was real. It was so damn good. So he’s got that vector. The second vector he has is autonomy, right? I picked up a new Tesla at the end of last year with hardware for, just so I could be on the latest of FSD. It drove me home from San Francisco last night.

So those are 2 real-world—not language models. Now we’re talking in the world of bits and atoms. And then the third one is really around xAI. By that, you’re implying that, just because those are in separate companies, xAI could become the hyperscaler for all three?

I think there’s no doubt about it. It already is, and it will continue to be. He has a keiretsu, as they say in Japan, of companies that can leverage each other’s insights, and I fully expect that they will.

His need to build these large clusters and place these bets is because he’s got to support huge potential businesses—not just competing, not just building Grok to compete with ChatGPT, although I think they’ll try to do that too as part of X, et cetera. He has lots of ways that he can leverage it.

Remember, if you build a large, coherent cluster for pre-training and you decide, “Okay, we’re hitting up against some scaling limits of pre-training,” you can use it for training reasoning models, you can use it for post-training, you can use it for inference. It’s not like you’re wasting this just by building these larger clusters. I think what I’m most convicted in is that these guys are going to spend a lot of money on GPUs because, as Jensen says, more of the world is moving to workloads that demand machine learning and accelerated compute.

Okay, that was number 2. Bill, what’s number 3? One last question on number 2: It seems to me that xAI wants to make the argument that having its own hardware is a competitive advantage versus OpenAI. Do you buy into that argument? How do you think that plays out?

All right, number 3. We’ll do 4 and then we’ll wrap it up. I don’t recall in any of the previous waves—the PC wave, the client-server wave, the mobile wave—the amount of chatter about coopetition that I’m seeing here. People took out of our video with Satya where he said, “Oh, I already have my own models,” and then they released a few, implying that they’re in some kind of coopetition with OpenAI.

You’ve heard all of these hyperscalers talk about doing their own processors of some sort. Both Trainium at Amazon and the TPU at Google compete with Nvidia, but they’re buying Nvidia’s products. Nvidia put out models this week that compete with some of those customers, and then they announced this AI-driven PC unit, which would presumably compete with Dell, even though Dell’s onstage helping them with the Elon build.

I’ve never—I don’t recall, at least I think about the Wintel world, which I was on Wall Street covering—I just felt more like people stayed in their lane and were thankful to be a part of this broad ecosystem that was growing. I just see so much tension. It’s super interesting to me as someone who’s watching it, but I’m curious what your thoughts are on it.

I don’t know that I have any strong perspectives there. Maybe a couple of things. Number 1, the amount of capital that it takes to do these things is not the domain of venture capitalists. If you’re a new entrant—if you’re Anthropic or OpenAI—frankly, they were forced into this because there was no other choice when they realized the amount of capital was going to be nuclear-level capital. You just had to turn to these big companies, so these big companies ended up owning pieces of them, but they couldn’t be totally dependent upon them.

So you had Google, which hedged its bets a little bit with Anthropic, and you had Amazon, which made a bet on Anthropic, and Microsoft on OpenAI. Now, as these companies become big and successful—listen, if they don’t become big and successful, Inflection just gets subsumed by Microsoft and Character gets subsumed by Google. But if they get some scale on their own, like OpenAI, now it probably couldn’t be subsumed, from an antitrust perspective, by Microsoft, and it has standing on its own. I think you have that dimension.

I think the competition is as aggressive as it’s ever been. I think sometimes we make more out of this competition than there is. I don’t think Microsoft or Nvidia launching a few models means they intend to be in the frontier-model game, where they’re competing head-to-head with these other folks. At the same time, there’s a long tail of their customers who are more than satisfied to use a tightly coupled model rather than maybe trying to cobble it together themselves by going to Llama or using a lot of these other plumbing services. So we’ll have to wait and see.

Here’s one thing that I would just point out in that regard: We just looked at that capex chart—almost $300 billion of nongovernmental R&D being spent by the biggest companies in the world to advance causes that are aligned with America. We hearken back to the age of Bell Labs and Silicon Valley, where you had national government spending, and I just think this is so damn bullish for us. We have printing presses; we have companies generating enough cash flow to invest this much money to put us at the bleeding edge. It creates incredible national strategic advantage.

They were getting pushback for doing stock repurchases. Now that people said they don’t have any use for this cash, well, now they do. So I guess that’s a positive.

12. Reasoning Models & Chain of Thought

All right. The last thing that’s on my mind, and that I’m really looking forward to better understanding in 2025: It does appear, for the time being, that the majority of the enthusiasm around the advancement of AI is around this chain-of-thought thing. As you’ve talked about, I think it’s somewhat ironic that it’s less efficient, but now that’s a good thing. If someone had built a model that was just hyper-efficient, where the margins expanded, that would be better. But okay, now the argument is it’s less efficient, but it’s going to consume more compute, so that’s a good thing.

Seeing how well this applies to different uses of AI will be super interesting. What are the coding companies seeing when they use chain-of-thought precisely against that? When this model was first released, even OpenAI said it’s not for every use case, or they hadn’t seen it be successful in every use case. Now maybe they will in the future.

I’ve signed up and paid for all these Pro versions in the past few weeks. I’ve been throwing problems at them. Especially the one that’s interesting is it goes away—both the Google and the OpenAI products—it’ll go away for 10 minutes. How much better is that output? I think this is going to be something that’s really important to watch on the edge this year.

Obviously, if the compute price keeps falling, I think you fall into this “why not?” argument. Why not run 20 passes on something if the marginal cost gets somewhat irrelevant? With the fast followers and the way you’re using synthetic data to create even smaller models, maybe you can get the smaller model to run the second pass, double-checking your work, if you would. Anyway, I think this will be the fun part to watch on the edge this year.

I can tell you a couple of things on that. Number 1, I can tell you that the researchers inside these labs feel like they’re looking into AGI, and they’re surprised by how skeptical the world is. There’s a lot of dissonance that I think they have because they’re like, “Oh, my God, we’re getting close.”

The second thing I would tell you—part of the problem on that is they’re seeing things ahead of time, correct, and they’re unable to release them because they’re not fully baked, 100%, and they don’t have the compute. They don’t have the compute.

But I’ll tell you this: Part of the reason they’re so bullish on these reasoning models is that, from a scaling perspective, reasoning models can get better just because you throw more compute at them as well. So where are we on that scaling curve for reasoning? Most of them tell me that we’re around a ChatGPT 2.0 level from that logarithmic scaling, so they expect a lot of scaling advantages in 2025 and 2026 on the reasoning models.

And then finally, when it comes to utility—where are these things going to be put into practice?—I think you’re going to see some pretty dramatic breakthroughs this year on coding. Think about all the startups that we’ve seen in the coding space. A lot of them have used prompt injection and other techniques in order to get these agents to do things that they want.

I think you’re going to see real coding agents backed by Google and OpenAI, driven by their most sophisticated inference-time reasoning models, that are going to be meaningfully better than what’s in the market today. That’s 1.

And I would say, secondly, if at the enterprise level you think of coding as the tip of the spear as to what these agents can do, remember that you and I started last year talking a lot about the consumer—memory and actions. I think you’re going to see ChatGPT imbued with these capabilities, and I imagine the same happens with Gemini, et cetera.

We’ve seen it. We in fact did a demo using Booking.com and Anthropic’s computer use, where booking a hotel and doing other things—it’s still fairly embryonic, but again, that’s not built on the back of o3. Now build that on the back of o3, and it becomes very, very capable.

So if you said to me, “How wide is the use case going to be?” I think now the use case is very narrow: It’s researchers. But I think you’re going to see the aperture on the reasoning models expand pretty dramatically, and I think you’re going to see increasingly a blending of the pre-trained and the o-series model, because ultimately it’s 1 model that’s going to help you do all the things that you want to have done.

I’ll close with this. You didn’t ask for my input on it, but I’m not a fan of macro analysis. I think it falls in the “too hard” bucket because of all the different variables at play.

Great seeing you, and I look forward to a fun year of doing these with some incredible guests. I learn a lot kicking around with you. It’s a lot of fun. All right, take care.