（预览）Meta计划投入1350亿美元，“AI泡沫”的泡沫？为什么超大规模云厂商不应该投资TSMC

Hello, and welcome to a free preview of Sharp Tech. Hello, and welcome back to another episode of Sharp Tech. I'm Andrew Sharp, and on the other line, Ben Thompson. Ben, how are you doing? Are you surviving winter out there?

I'm feeling pretty touched right now, actually.

Okay.

I warned you before we came on that my voice is not right. It was much worse. I just recorded Dithering, and it was much worse there. I had sort of just woken up. But you expressed your sympathy and concern about winter, and then you jumped on. I think your voice sounds terrible, too.

Mm-hmm.

You made it sound bad for me.

My voice isn't in a great place either. 'Tis the season. January and February.

That's great. Yeah, this might be a rough listen. Hopefully the content's good because the sound-generation capabilities appear to be quite low on both sides.

There you go. So with that caveat and/or warning out of the way, we can dive into the news of the day. We're going to start with Stratechery's bread and butter here, Ben: a little company out there in Menlo Park.

I will read from Bloomberg, which wrote this week: “Meta's better-than-expected sales outlook helped ease Wall Street concerns about plans for unprecedented spending on artificial intelligence this year. The social networking giant topped projections for holiday-quarter revenue and gave a strong forecast for the current period during its earnings report on Wednesday.”

And then later in the story they write: “Meta projected record spending for 2026, driven by Mark Zuckerberg's aggressive campaign to amass the infrastructure, computing power, and talent that he deems necessary to win a competitive AI race. Zuckerberg has said his strategy centers on front-loading computing capacity in preparation for reaching the company's goal of superintelligence, a theoretical milestone at which AI can meet or outperform humans at many tasks.”

What are you laughing about?

Just that they have to add that it's a theoretical milestone at which AI can meet or outperform humans at many tasks.

So much for being a neutral sort of arbiter on these points.

Yeah, I'm sort of giving away the game here. To get there, Meta is spending aggressively. The company estimated that full-year capital expenditures will be between $115 billion and $135 billion at the upper limit.

So it's now time to play the most dangerous game: What is Wall Street thinking, reacting to these earnings? Why do investors like these results and up to $135 billion in projected CapEx this year after getting squeamish in the middle of last year's $72 billion in AI spending? What do you think?

I don't know. That's a good question. It's pretty interesting because last quarter I thought, once again, Wall Street totally overreacted.

Mm-hmm.

And this quarter, it's like an overreaction in the opposite direction. Last quarter was, I think, actually a very impressive quarter from Meta. They—

Yeah.

They delivered increased impressions for reasons that I sort of identified back in Q2.

Mm-hmm.

They clearly can increase engagement, and they can increase ad load, and both seem to be happening.

Yeah.

And they didn't really cop to the second one last summer. They did this earnings call, which I thought—

Mm-hmm.

Was very interesting. And you could argue that's actually a very bullish signal for Meta's business. You could absolutely make the case based on the last 6 months that Meta has been under-monetized throughout its entire history.

Yeah.

I remember back in— For a long time, they just sort of got growth for free as more people joined social, joined Facebook, and then Instagram came along. And I made this point last summer: Investors for ages gave them insufficient credit for how organic their growth was.

Yeah.

And so it wasn't just that they underestimated the extent to which the entire world was Facebook's audience and would grow into that—and, by the way, was never shrinking. It's still growing, even on its sort of old properties.

Mm-hmm.

But also, every time they would introduce these new inventory surfaces, that was a huge opportunity to increase ads even more, whether that be Stories or Reels or whatever it might be. And the concern I expressed, I think a year or a year and a half ago, is: Where's the new inventory going to come from, right?

Right.

Like, yes, this Meta AI thing could be a thing. It could be an app that people use. But it wasn't like a Reels or a Stories thing where there's this obvious new format with obvious new potential.

There was a limit to how much they could grow the monetization pie in that scenario. That was at least the concern.

No, it's just that they need— Like, back in the day, Stories clearly worked, so—

Right.

Of course they would get money from ads. Reels clearly worked, so of course they would get money from ads. Does Meta AI work?

Well, no, I know, in terms of—

And is it a good advertising surface? These were—I think it was fair to be somewhat skeptical.

Mm-hmm.

It turns out that Meta does have more inventory available, which is all their existing products—

Shoving ads in.

They can shove that many more ads into.

Exactly. That's what we've learned.

And—

We've also learned that they can do that without bleeding users because engagement is up. I mean, it really does—

No, right.

Underscore how dominant the company is.

No, it's both. They're dialing both. They're increasing engagement. Now, of course, they want you to believe that both of these are downstream from AI and investments.

Mm-hmm.

And they're not wrong. They do have new recommendation capabilities. They do have a new ad model, apparently this one called GEM, that is sort of— It's not an LLM per se. It is similar to an LLM. There are aspects of it that are transformer-based.

And the key part of that is that it does scale. You throw more compute at it, and it gets better, which is great. And that is a core validation of the thesis that Meta should be spending more on compute, because there's a direct link between spending more on compute and getting better ads.

Mm-hmm.

I don't think that takes $135 billion.

That's what I was going to ask: Can you actually solve that without spending $130 billion? Yeah. Well, it seems like what Wall Street is doing is buying this story that the spending to date—$72 billion last year—is accretive to the core businesses.

No, here's the deal. They grew 26% last quarter, which again, I think last quarter is more impressive than this quarter. This quarter they grew 24%. It was actually less.

Mm-hmm.

And they had more of a currency advantage this quarter. Next quarter, they're talking about growing 30%, or as much as 30%. I think it's 26% to 30%. Again, that's with roughly 4%, or 400 basis points, of that from currency, with the U.S. dollar weakening. But on a $200 billion company, that's incredible.

Mm-hmm.

That's like their best growth in 5 years.

Yeah.

So that, more than anything, I suspect, undergirds why suddenly everyone's piling back into Meta. And this is part of the Meta pattern. They continually surprise Wall Street and surprise investors with their ability to grow from ever-larger bases. Every time this happens and they show these incredible growth numbers, it's off of a base that is dramatically larger than the last time there was a crisis, right?

Yeah.

And so I think this fits in—

So every 18 months people say, “Oh, turns out that's actually a pretty good business.”

Yeah, it's more like 3 to 4 years, but yes. It happened with Stories. It happened with Reels. And now it's happening with, actually, we can just shove ads everywhere as much as we want.

Yeah.

And—

And there's nobody that can touch us. Our users aren't going anywhere. So it's all really, really impressive.

What did you think of the results, though, and the idea— I was struck by a note in your daily update on Thursday. Almost the entire free cash flow that Meta has is going to go to AI CapEx in 2026.

Yeah, my takeaway is that the extent to which Mark Zuckerberg is burning the boats, as it were, is actually underappreciated. I don't know that people have fully accepted or internalized the implication of this spend.

How do they come up with $135 billion at the top end? I actually think it's pretty simple.

Okay.

$135 billion is how much free cash flow they're projected to have this year. And so we're talking about putting basically every single dollar of cash they have toward buying data centers and chips and all the sorts of things that go into that.

Mm-hmm.

And to put that in context, a few years ago they were spending—I don't have the number in front of me, but I would say $12 billion to $15 billion—in CapEx.

The increase from 3 years ago to this year's projected number is basically more than they've lost in Reality Labs over the last 13 years.

Wow.

This supposed albatross where they're spending all this money—

Excess.

—sort of excess—

And Mark asks, yeah.

And Mark Zuckerberg, on a lark—and fine—

It's a fantasy.

—if he keeps turning out results, we'll let him spend money—is a footnote compared to the amount of money they are projecting to spend.

It's crazy.

And that is—

Yeah.

Yeah. What it is is—and this earnings call, actually, I think I might have literally laughed. I laugh at very stupid things because I'm a business nerd. But the very first analyst question is, “Can you give us some examples of how you expect to get a return on invested capital over the next 3, 5, 10 years?”

This was my favorite part of your daily update, just for the record.

And Mark Zuckerberg—

And what it does say—

He’s like, “No.”

Yeah, exactly.

Yeah.

It's like a paragraph-long “no” from Mark Zuckerberg is probably going to be the beginning of that earnings call.

Right, which I appreciate the honesty.

Uh-huh.

It was very honest, the part that I quoted. Oh, my voice is fading. You can hear it.

I can't wait to see—

There was another few hundred words—

—how it holds up over 90 minutes here.

—of him actually trying to come up with examples of stuff they might do.

Yeah.

But it should be appropriately taken with a grain of salt. The grain of salt is that Mark Zuckerberg is like that meme. Remember the salt guy meme of the guy—

Mm-hmm.

—just flinging salt? That was Mark Zuckerberg at the beginning of this answer, basically.

Flinging—

Yes.

—details and ideas.

Everything I say should be taken with a grain of salt, and let me throw the salt on the stake for you because you're gonna need a lot of it.

Well, we got a question from Adam that tracks with some of my instinctive skepticism that I gave away when I was reading the Bloomberg article. Adam says, “Ben, in your update Thursday, I would have liked some discussion of whether Zuckerberg ought to come to terms with the idea of being just an app or a set of apps. I'm skeptical of Meta's right to win in Gen AI even as a capability. Isn't all the ad revenue just funding Zuck's ongoing yearning to be more than a set of apps?” What do you think of that?

I do think it continues to be underrated and underappreciated, the extent to which Apple saved Meta from itself, saved Mark Zuckerberg from himself.

How so?

People forget the Facebook platform, the idea that we're going to be this place for other apps, and we're gonna have Facebook Credits, FarmVille, and all these sorts of things. That was the stated goal in what the company was doing. What happened was the phone comes along, they underinvest in it, and they actually—

Mm-hmm.

—try this HTML5 app approach that would theoretically give them more flexibility to do this sort of stuff in the future, and they got killed for it. Then, in 2012 or 2013, they just restart everything and redo the app. It's gonna be a great app on the phone.

It's gonna be about mobile, yeah.

And as an app on the phone, you don't get to be a platform. The phone is the platform. You don't get to do a platform on top of a platform, not just because of Apple's policies, which are certainly part of it—

Mm-hmm.

—but also because you're dealing with a 3.5-inch screen. What are you gonna actually do with it? The reality is that forced Facebook into their advertising business, which is incredible.

Yes.

I pointed this out way back in 2013 when I first started. I think I wrote an article, “Mobile Makes Facebook Just an App,” basically making this case: Facebook's being saved from itself.

Mm-hmm.

What was incredible even back then—and this point is even more true today in 2026—is that, from a business perspective, what is incredible about Facebook is that they somehow have gotten permission from users to continually show them full-screen ads.

Yep.

You scroll the feed, and for some fraction of the time, every single pixel on this device in your hand, on which you're focused, is an ad.

Is an ad, yeah.

And users preferred it, liked it, and responded to it dramatically better than they did to the banner ad idea.

In-line advertising—

That's right.

—on newspapers. Yeah, totally.

The banner ads were actually—and this is a great example of how people think about these things so wrong, and what actually happens can completely upset your assumptions.

Mm-hmm.

People would think, “Oh, the banner ads. At least then they can see their content, and they're not so bothered by the…” This is the negative view of advertising that I think is persistent. People are ashamed of advertising. It's like, “Well, we'll just put it on the side. We're not gonna bother you too much, but can you look over here?” This is basically the newspaper industry in a nutshell. What Meta—Facebook back then—achieved with the feed and Apple forcing them into the constraint of a 3.5-inch screen is that they figured out how to not tiptoe around ads, but shove them in your face—

Yeah.

—to the extent that people weren't bothered, and they actually liked it.

Well, it's really true. As a user, it is a preferable experience to see a full-screen ad and then go back to the full-screen content as opposed to seeing—

It's all content.

—split-screen ad.

That's the thing.

Yeah, exactly.

It's all content.

Well, that's true.

Well, and the Instagram ads are—

The whole point is—

—good content as well.

Well, this is what happened last quarter. Last quarter basically was a revalidation of the point I made in 2013. Not only are the ads way better now, because they're all video, they're immersive, and you can click through and do all these things you couldn't do back then, but actually we have way more capacity to shove this in people's faces than we thought we did.

Yeah.

We can put more ads taking over the full screen. By the way, now it's not like a scroll when you're going past it. It's a whole video that's locked. When you're in a Reel and you flick and it's going through a video, you're like, “Well, I just spent 20 seconds watching an ad.”

Mm-hmm.

When it was actually totally in my control to skip it the whole time, but the ad was compelling and interesting, and actually I kind of wanna buy this thing. This remains underrated, and I think it remains underrated by Facebook itself. They are insufficiently grateful to Apple for forcing them to do what they didn't wanna do. Now, I think they have much more valid points later on. We've talked about ATT. I think what Apple does in terms of APIs for devices—

Mm-hmm.

—where they favor their own stuff is clearly problematic, particularly for a company trying to build a device business.

Facebook's not Facebook without Apple and that—

Apple is—

—forcing function. Yeah.

—is arguably the most important company to their success, which they will never, ever want to admit.

Well, and look, in terms of what Facebook/Meta is now trying to do in AI, the reason I'm skeptical and the reason I identify with some of the questions that Adam is asking—I'm not a Meta investor, and so far be it from me to tell Mark Zuckerberg how to spend $135 billion—but on its face, Meta's AI strategy looks like a business that's been sketched on the back of a napkin where Zuckerberg—

No, it doesn't look like a business at all. I think you're actually giving them too much credit. But here's the defense of it.

Okay.

If you truly believe that AI is a total paradigm shifter—

Mm-hmm.

—and that everything that—yes, we'll look back on the smartphone; the smartphone will still be around, like the PC's still around, but not actually pertinent to the way people interact in the future—and this idea of just-in-time AI, whether it be delivered via glasses, little handheld devices, earpieces, or whatever it might be, then—

You'll need the infrastructure. You'll need the muscle groups—

You'll need to have actually—

—otherwise you won't matter.

That's right.

Yeah.

And it's actually interesting. The contrast between Meta and Apple right now is extremely striking.

No kidding.

Apple has basically abandoned this. They're outsourcing to Google. John Gruber on Daring Fireball just pointed out that, if you go back to the Tim Cook doctrine, one of the core things is that we have to own and control the technologies that are critical to our business.

Apple is basically either giving up on that doctrine or stating that AI isn't actually going to be that big of a deal in terms of disrupting their core business. Which, by the way, might all be true, right? Maybe—

So far, it has been true.

They can just rent AI all out.

Totally.

That's right. And—

I mean, it's fascinating to me watching Meta because we've kind of raked Apple over the coals for its complacency over the last two and a half years even as every conversation does include the caveat, "This might be smart, but it's less interesting from a technology standpoint."

Right. So that's why I'm going to push back on you laughing at this.

Okay.

Like, if you actually—if you're going to rake Apple over the coals—shouldn't you be lauding Mark Zuckerberg—

Well—

—for basically saying, "We are going to push and squeeze our core business, and we're not going to dilly-dally around."

Mm-hmm.

"We're actually going to take every single dollar from our existing business and put it into being competitive and controlling our own destiny in the future."

Totally. You interrupted me before I could say, "Actually, I've come to respect Apple's stance more when I look at what Meta's doing," because I look at the napkin and it says, "CapEx plus superintelligence equals profit." That's basically what I'm taking away from a lot of what Mark Zuckerberg has had to say over the last year or so.

I find myself sitting here being like, "All right, so what's the product, and why is that product going to be differentiated from any of the companies that are already dominating in this space?" I think one argument is that performance will be correlated to compute and the infrastructure that you have, and so they'll eventually take the lead that way or get near the top that way.

But they seem to be operating from the assumption that none of their present-day advantages will matter in an AI-dominated future, which itself is a pretty speculative premise. That's why a lot of this is such a fascinating bet from Zuckerberg.

I'd push back on that. I actually think it is a defense of Meta's spending that AI, even as it is today, has tremendous upsides for their business. I wrote this 2 years ago: Meta should be the biggest beneficiary of this.

We talked about generating ads just in time, better targeting, or, in the long run, every single pixel on Instagram should be monetizable. Everything should be an ad. That is viable.

With the technology as it exists today, you can identify everything on screen, link it to something else, and have a buy button. Right now, when I take a screenshot of our podcast, I'm looking at you and thinking, "What kind of mic does Andrew have?" I should be able to click on that.

I should buy that. Yeah.

I should be able to buy that. That should be monetizable. There's no reason, given the technology as it exists right now, that couldn't be done.

So Meta, more than any other company, has the product overhang from the model capabilities and all the things that could be built given what exists today, never mind what might exist in the future. I actually buy the case that they—

I do too. I just don't know that you need to spend $130 billion to get there. Do you?

But why not? We're talking about AI as it exists today. What is AI going to look like in 2 to 3 to 4 years? If you're not investing today, then all you're doing is making it that much harder to catch up and falling behind.

I guess my defense of this is, what do you want them to save the money for?

That's a good question, too. Again, I'm not a Meta shareholder, and this makes for more interesting podcasting, but I'm just not sure where the money leads. I think that's a fair question to ask at this point, given the lack—

Well, but—

—of clarity from Meta and Zuckerberg.

Yeah, but if it was super clear, everything... Again, I do think there are real things in terms of their current business. Yes, we laughed at his sort of, "I don't know, you know, I'm not gonna give you a clear answer here."

Mm-hmm.

But he did say in the answer, "It'll benefit our current business," and I do think that's super clear. Actually, if anything, I think he understates this and, I would argue, doesn't fully understand this, in part because one of my critiques of Zuckerberg is that he doesn't really like ads either—

Yeah.

—and doesn't fully understand or think about the potential there. I feel I am much more enthusiastic about Meta ads than Mark Zuckerberg is, which is kind of a problem.

I mean, you could go back to—

I'm a better advocate for them than he is.

—you're enthusiastic about all the AI opportunities, going back to the interview with Zuckerberg about a year or a year and a half ago.

Right. No, I think it's more this ad angle in particular. I think he could actually sell this much better if he were deeper into and cared more about ads, because I just think the opportunity here is astronomical.

You look now—yes, there is an aspect of their growth that is from shoving more ads in front of people. But number one, you get away with that because your recommendation gets so good that people don't mind. Also, they talk about a 1% increase, a 5% increase. That level of increase in terms of effectiveness in targeting and click-throughs on a $200 billion business is astronomical.

These are very real gains that they are achieving today, and I think it's completely legitimate. When they talk about building new targeting built around LLMs, which is basically what they're saying is, "We can understand the content," and then, "We're going to build new ones where you can go back further in people's history to understand what is interesting to them," that's about increasing context windows and understanding all the stuff that's going on. All this is totally legitimate. These are legitimate arguments.

Yeah.

I actually think my critique of Zuckerberg is that he doesn't make these arguments well enough because he doesn't care. I think Mark Zuckerberg just wants to control the future technology, which he believes is AI. It's working out because he has smart people who do care about ads and build them out, but at least he's better than OpenAI in that he accepts the reality of ads and lets those people do what they need to do.

Right. We do need money in order to sell any of this. That makes sense.

Right. So I do think the future opportunity is real, but why do you need... This is, I guess, the critical thing. Let's consider regulation.

Okay.

You love regulation. I don't.

It depends on the regulation, for the record. A lot of bad EU regulation out there, but sure.

What is my core concern with regulation?

The core concern, drawing on about 50 different Sharp Tech episodes over the last couple of years, is that regulation stifles innovation, and we fail to account for the products and businesses we don't see in an overregulated environment. Is that right?

Excellent. Great job. A+ for you. You know, it's always tricky at the end of the semester. You're tired, worn down, and you have your final, and you're a little sick.

Nope. That's what the Starbucks is for, Ben.

Can you actually come through in the clutch?

That's what the Starbucks is for.

Can you come through in the clutch and give the right answer? And you did it. I'm extremely proud of you.

Okay.

So the point is, it's a hard argument to be on the antiregulatory side because on the pro-regulatory side, you get to point to actual harms that happen. On the antiregulatory side, you have to point to things that didn't happen, that were foreclosed.

Yeah.

Which, by definition, you don't actually know what they are. That's what is so frustrating and also concerning: we don't actually have a measure of everything that didn't happen.

Mm-hmm.

And it could actually be way larger than we think it might have been. But then people like you who want more laws are like, “Well, give me an example. Give me—what is it?”

Yeah.

But that’s what you’re doing to Facebook and AI.

What didn’t we get? Okay.

That’s what you’re doing to Facebook and AI, right? There is an aspect where the more transformative you think AI is, almost by definition, the less you can say what the transformative aspects are going to be.

Mm-hmm.

Simply because what is possible in the future wasn’t even imaginable previously.

Sure.

And so, I have sympathy for Mark Zuckerberg in that he clearly believes this is that level of transformative. He believes it to the extent that they’re going to spend every single dollar of free cash flow they have.

Mm-hmm.

But implied in that level of belief is that we’re pursuing things that are definitionally impossible to say what they are because that’s how transformative this technology is.

Yeah.

Now, maybe that’s an insane perspective for a public company CEO, but that’s the whole point of having founder control.

Totally.

So, I mean—

I mean, I respect where he’s coming from. I do think Adam nailed it, that Zuck at his core is a technologist and a dreamer, and he feels like this is the next big thing, and he wants Meta to be at the center of it, and it’s his company. So it’s his prerogative to spend. I just am not sure it’s an existential concern to the extent that you need to spend every last dollar.

I know, but the point where you validate whether it’s existential or not is too late.

Mm.

That’s what Apple’s—

Yeah.

We’re running an experiment. Let’s schedule our podcast for 2036 and see who was right, Apple or Meta.

Right. And I think ultimately my skepticism here is a vote of confidence in the moat that Meta has with Instagram and Facebook and how durable those businesses are going to be, even in a paradigm-shifted environment 10 years from now. But maybe I haven’t read enough Dario essays to fully internalize the revolution that’s coming for all of us. Who can say? All right, and that is the end of the free preview. If you'd like to hear more from Ben and I, there are links to subscribe in the show notes, or you can also go to sharptech.fm. Either option will get you access to a personalized feed that has all the shows we do every week, plus lots more great content from Stratechery and the Stratechery Plus bundle. Check it out, and if you've got feedback, please email us at email@sharptech.fm.