Dan Niles for Niles Investment Management of Twitter for AI speed bumper. Token minimization, low cost.
Dan, we've known each other for quite a while. I know your German is impeccable at this point. [laughter] Great to have you here. Before we dive into your three main concerns, let's have a quick look at the markets. You recently this week posted on Twitter and you also looked at the Morgan Stanley momentum index, or here in this case the Goldman Sachs high beta momentum index. They're both pretty much telling the same story. You're seeing quite a dramatic drawdown in momentum stocks generally. The Goldman Sachs momentum index down about 33%. Now looking at the history, that usually are levels where you can start expecting a bounce in the sector. How are you reading this?
Yeah. So I put out a Twitter post this last Sunday basically talking about that, where if you look at the Morgan Stanley momentum index, and so that's you take the longs, momentum longs, you subtract the momentum shorts. Since I put out that post on June 20th, that momentum index from June 22nd through last Thursday was down 28%, which, and it happened over 24 days. If you look at history, whenever that momentum index has been down over 10%, you've had a drawdown. The average is 20% only. This time it's down 28 and it's taken 48 days to do that. This time it's 24 days. So the velocity of this move is greater and it's happened in a shorter time than the 60-plus occurrences since 1996 when I went back and examined this. So you would think, unless there's something structurally wrong, so you're going into a recession, the Fed's hiking rates, you have a global pandemic, you know, something that's really problematic, you should be pretty close to where you want to start adding back in exposure in momentum names and specifically AI infrastructure related names that have gotten just absolutely clobbered, which they rightly should have, that they got overheated. But I think now is a period of time when you can start to do that. Obviously, the biggest risk is you're going into earnings season coming up with Google on Wednesday night and, you know, that's going to be a huge deal for everything AI related in terms of what they have to say.
We're going to dive into that in a minute. Of course, you also had a dramatic rise, almost a parabolic rise in the SOX Philadelphia semiconductor index. Isn't that a bit of a risk regarding a potential bounce and how sustainable is such a bounce going to be?
Sure. I mean, so just if you look back at history and you say, okay, what happened during the internet infrastructure buildup? So from the end of '94, which is when Netscape Navigator was introduced, to its peak in March of 2000, the semiconductor index went up about 850%. Right now it's up less than 400%. And if you look at valuations and you say, okay, well, the poster child for AI is obviously Nvidia, it's trading at about 24 times. The S&P is at 22. So, and Cisco back in its heyday back in 2000 when it was the most valuable company in the world was trading at over a 100 times PE. So it's hard to argue this is overvalued and the rise is not even nearly as great as it was back then. The second part of that is that back, if you look back at the internet infrastructure buildout, there really wasn't new quote-unquote technology per se that happened during kind of that five-year-plus expansion. If you look at AI though, you've gone through three distinct phases. You went through the training phase in 2023. You went through inference during kind of '24 and '25. Now on January 30th, you started the Agentic AI phase and that requires 10 to 100 times more tokens. The only way you get those tokens is with more semiconductor chips. So, you know, unless you believe Agentic's not real, if you need 10 to 100 times more token production, you're going to have a continued expansion, I believe, over the course of the next year at least.
We're going to dive into that in a second, but are you indicating that even if token prices end up getting commoditized, the revenue will still rise faster than the current estimates on Wall Street?
Well, this is where it gets complicated. So it depends on what layer you're talking about. I believe that the large language model layer is going to get commoditized. So the Anthropic, the OpenAI, because you've had, as you saw last week at the World Artificial Intelligence Conference in Shanghai, that was the ninth one. You saw, for example, you know, new models coming out and specifically Moonshot's Kimi K3. You know, in the past, people said, 'Oh, you know, the models out of China, they may be cheap like DeepSeek, but they can't compete at the leading edge.' Now you're seeing these models compete against the best models coming from closed-source OpenAI or Anthropic. So I think at the model layer you're going to have commoditization and that's going to slow down the revenue growth, to your point, which is one of my concerns because you are getting that commoditization. Having said that, that frees up more money, if you will, for companies to say, 'Okay, I don't have to spend a bazillion dollars on my large language model because I can use an open-source version that's 90% cheaper, but I still need to produce the tokens.' And so that leaves money to actually spend at the infrastructure layer, which I think is going to continue to grow. And you can't, you know, you can't open source that, right? You can't go, you need the chips to actually produce the tokens. And so I think now is a phase where you have to get a lot more selective around which portions of the AI trade you want to be involved with.
So if the LLM models are getting commoditized, is that already reflected in investors' expectations? And if not, which companies will see the most dramatic impact? Also looking at the IPOs of Anthropic and OpenAI coming up in the next months or potentially early next year.
Yeah, I don't think that's reflected in people's thoughts yet because a lot of people are still sort of fighting that thesis that large language models will get commoditized. And obviously OpenAI, sorry, Anthropic going public soon, that'll be very interesting to see what they have to say for their revenue growth because they've seen the fastest revenue growth this year. I mean, we've never seen something at this scale. And to give you an idea, at the beginning of 2025, the two companies OpenAI and Anthropic together, they had an annualized run rate revenue of about $7 billion. At the end of, sorry, what did I say, 1995? I meant 2025. Gez. At the beginning of 2025, they had a combined run rate revenues of $7 billion. At the end of 2025, they had a combined run rate revenues of $29 billion. Right now, the combined run rate revenues of those two companies combined is over 70 billion. So we've never seen companies at this size grow this rapidly. Now, the question becomes, as you pointed out, if you've got some pressure on revenues at the model layer, those two are at the epicenter of that. And the best example of that is Coinbase's CEO had a post on June 26th where he talked about, 'Yeah, I'm using these open-source models and doing a bunch of other stuff. And even though my token production is going up, my AI bill or what I'm spending on AI has dropped by almost 50%.' And so that's the part that gets very tricky because, you know, if you have an issue with Anthropic, because we're dealing with stock prices, right? And stock prices are two things. It's earnings times the multiple. What multiple are you willing to put on it? And that's where this is going to get tricky because if you do have an issue with the revenue growth at a company like Anthropic or OpenAI, and OpenAI I think is the one that's going to really have the problem because they're stuck between Google and consumer AI and Anthropic, which from the very beginning focused on enterprise AI because corporations will pay for it. Only 5% of OpenAI's consumer customers actually pay for the advanced versions, right? They use the free stuff. So I think OpenAI's got the huge problem, which is why they've postponed their IPO into next year and we don't know what, you know, the numbers are, etc. We haven't gotten an update very recently, but I think part of that is tied to the fact that their growth compared to Anthropic has really slowed down.
I mean, you've been a, you've been a skeptic about OpenAI and the whole OpenAI universe for quite a while. I think since I've pretty much known you, and a big fan of the Google universe with Gemini capturing more market share as well. But let's, let's go back to Coinbase for a second. So if Coinbase manages to reduce the token budgets by 50%, is this a one-time phenomenon or do you see the top 1% of the dominant token buyers in the United States doing the same thing? What would the consequence be for the sector?
That's a great point and this is why on June 20th I said, you know, I'm expecting an AI speed bump because if you think of the timeline of this, what happened was the following. January 30th, you get the formalization of OpenClaw, which you could say kicked off the whole Agentic AI move, which demands 10 to 100 times more tokens. By March, you had companies like Meta talking about token maxing. We have token leaderboards and employees could see how many tokens they were actually using up. Then Uber came out in April and said, 'Hey, we blew through our entire AI budget for the year in four months. That's a problem obviously.' And then you've then switched in June to companies going from token maxing in the month of March. In June they were talking about token minimization, which is why I put out, hey, I think we have a speed bump problem here. And then you had Coinbase CEO coming out on June 26 talking about, hey, we managed to cut our AI bill by almost 50%. And I think to your point, the top 1% of corporations that spend, they account for probably over 50% of the total AI spend, they're going to try to do what Coinbase is doing. Now the question is, can the other 99% grow enough to offset that 1% slowing down? And even though token, you can see it in the data, the token production has skyrocketed since January 30th when OpenClaw got formalized to now, you've also seen the price per token dropping since about June, which coincided with when you started to hear about all this token minimization stuff. So this is probably one of the most complicated environments to deal with. And what I'm trying to think through is, can the stocks of semiconductors perform really well if companies like Anthropic and OpenAI are seeing slowing rate of revenue growth?
And we're about to find out very soon. [laughter] That's a, and you know, we have the earnings of Google Wednesday night. We have the earnings of Microsoft and Amazon next week. What are you looking for? What do you think the market will care about the most?
Yeah, I mean, the market's going to care about a lot of things. I think with Google, Google's really tricky for the following reasons. Google is picking up the fastest growing customers in AI. They're the smallest of the big three. Amazon's the biggest. Then comes Microsoft. Then comes Google. Google's cloud platform revenue growth, that was growing 34% in the September quarter. That accelerated to 48% in the December quarter and then accelerated further to 63% year-over-year growth in the March quarter. I think that's going to accelerate again to over 70% because remember in the June quarter you had token maxing going on. Google though doesn't guide. So you're not going to get any guidance. Now you may get some qualitative commentary around what they're seeing in Google Cloud Platforms, but unlike Microsoft and Amazon, which will guide to the September quarter, Google won't. I think Google's revenues and EPS forecasts after they report the quarter are going to go up for the future quarters because their core business, which is search obviously, and you know, YouTube and all the other stuff, that's doing quite well. So it's really to some degree going to come down to the commentary and then what do they have to say about things like capex, because Google's obviously trying to figure out how to control their capex budget, which is why you've heard, you know, yesterday there was some talk that they're trying to design a chip where they're etching into the actual silicon parts of the Gemini LLM model to cut the, you know, energy and the cost to produce these tokens. And so everybody's trying to figure out ways to control these bills because the free cash flow for all of these companies has dropped dramatically and in some cases going negative for some of the most cash generative companies in the world, which is why you saw Google do that $85 billion, you know, equity offering, which I never thought we'd ever see them having to raise money, but it tells you how, you know, cash consumptive this AI infrastructure spend is and how much they're trying to get this under control.
And there are rumors that Meta might follow in the footsteps of Google also with a capital raise. Of course, you had also a lot of IPOs, SpaceX IPO, which is now below the initial IPO price. I know it's a different ballgame with SpaceX. You can't really compare those companies in many ways. But you know, I don't know if you saw the most recent Bank of America fund manager survey. They had an interesting question in the survey and asked, do you expect any hyperscaler to announce a cut in capex in 2026? And 61% said no, such a cut will not happen. Yet, when you argue that you're going to see LLM models being commoditized, wouldn't that, wouldn't that be an indicator that capex expansion at the current rate might run out of steam?
Oh, 100%. And that's arguably the wrong question. The right question is, what do you think the capex expansion is going to be for 2027? Because right now for 2026 you're running capex growth of over 70%. Right? And last year it was about the same, you know, levels. And to your point, Marcus, which I think is a great one, the survey should have been, well, do you think growth in 2027 is 50% over, 30% over, 20%? And what I would argue is if that growth rate is much lower than what people are thinking, then you've got a potential problem. And so that's more what I'm focused on because you're seeing all of these companies trying to design custom ASIC chips and, you know, like Google, Google's results yesterday or the news yesterday about them wanting a specialized chip with the LLM infrastructure kind of etched into part of the model, and they're already doing ASICs with a specialized chip for inference versus a specialized chip for training. They're all trying to figure out how do we get these costs down because they're exploding higher because I'm sure they're seeing this too, because if people want lower-cost tokens because they can't afford the higher-priced ones and quite honestly they don't need it. And so, you know, for your viewers, right, it'd be like saying, okay, what's 3 plus 3? And you're asking a PhD in physics for that answer. Well, that's going to cost you way too much, right? They can ask you or me that question and pay a whole lot less. And so there are certain things going on underneath the covers that we've just started to see this play out. And so that's why you've got these three different vectors happening at the same time. 1% of corporations trying to cut their AI bill, the other 99% growing their AI bill. You've got token production going up, but you've got the price per token dropping. And you have those three things happening at the same time. And then of course when it comes to stocks, you have to add in, well, what multiple do you want on that? And that's where it gets really problematic.
I'm really curious, hypothetically, if one of the hyperscalers should do the unthinkable and actually signal that capex expansion is reaching its max. Will those hyperscalers...
It'll be Armageddon the next day.
But you see, that's exactly my question because the stocks of the hyperscalers did not really perform very well this year and there was a time where people wanted to see ever higher capex spend to make sure that they're going to basically win the race. But now if commoditization is becoming an increasing topic, Alex Karp of Palantir criticizing the LLM dominant players for the pricing models they have. Wouldn't a potential reduction of capex or even the signal of just keeping them steady not lift the stocks of the hyperscalers?
Yeah. I mean, so a couple of things. During this momentum drawdown, the Magnificent 7 didn't actually perform that poorly, right? And part of that is people as they thought about, well, if I do have, you know, less money being spent on this, well, that'll be great for the hyperscalers because what people really hate is the fact that they're having to raise money and their cash flows are dropping or going negative. So for the hyperscalers themselves, depending on who they are, the stocks may actually work well for that, right? Remember the day, I'll give you a good example of it, you know, the day that Meta came out or there was this news story that Meta would go ahead and, you know, potentially be launching their own public cloud and that they might end up, you know, being able to monetize that. The stock was up, I think, 9% that day. And so what people want to see is better cash generation from all of these companies. And so what you may see that day is the hyperscaler companies, their stocks go up, but the semiconductor companies get hit at least for a day as people worry about, well, you know, where are you, you know, where exactly is the spending getting cut. I mean, this is probably the most complicated... I mean, I lived through the late 90s. This is probably the most complicated situation, you know, you've been in for a very long time because if you do get the poster childs for this trade in trouble like OpenAI or Anthropic with what rate of revenue growth there is, then people are going to start to think all the way down the stack, down to, you know, through the hyperscalers to the semiconductors, etc., like what are the implications all the way through. And that's, you know, that's what I'm trying to think through right now. I think it's way too early for these guys to say, 'Hey, we're, you know, slowing down capex right now,' because there's no reason they'd all be raising money, right? Amazon went ahead and did a big bond offering, multiple bond offerings already this year. Google did that equity offering. Meta's raised a bunch of money from bonds. That doesn't seem to indicate, based on what they're doing, that there's any slowdown in the near term because the riches for the company that survives is so great that none of them want to lose that race right now. Like Larry Page, there was that famous quote of his: 'I'd rather go bankrupt than lose this race.' So, you know, I don't see that in the near term. Do I think in 2027 there's going to be a plateauing of this? I absolutely do.
