Intel and the US Semi Future, Will Any Big Tech Incumbents Lose in an AI World?, Questions on Tea, Grok, and The Ringer

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?

I'm doing okay, Andrew. I'm doing okay. How are you?

I'm doing well. We're rounding third on the end of July, headed into August later in the week, and I'm excited to be back. We're going to kick things off—

Well, that was just a really—

—with some chip talk.

—great articulation of where we're at in the calendar. Thank you for that.

Sleepwalking through—

You know—

—the intro here.

That was the most summer-end-of-July intro ever. So I guess you captured the spirit—

I'm doing fine.

—both literally and figuratively.

How are any of us doing? But no, I'm in a good mood. It's good to see you, and we are going to begin with a chip check-in, long overdue for a chip check-in.

1. Intel's Foundry Dilemma

So last week on Intel's earnings call, Intel CEO Lip-Bu Tan said the following: “Up to and through Intel 18A, we could generate a reasonable return on our investments with only Intel products. The increase in capital cost at Intel 14A makes it clear that we need both Intel products and a meaningful external customer to drive acceptable returns on our deployed capital, and I will only invest when I'm confident those returns exist.”

So, Ben, as we take a look at the state of Intel, can you translate that message into English for people who care about the future of US semiconductor manufacturing?

Let's step back from the specifics of particular node sizes. What Tan is referring to, in broad strokes, is basically what I was warning about 12 years ago when I started Stratechery and first wrote about Intel.

The problem with missing mobile is that you're missing out on volume, and if you don't have sufficient volume, you're eventually not going to be able to invest in the leading edge as costs continue to go up. As costs go up, you need to spread those costs over an increasing number of chips, and if the only chips you're making are Intel chips, you're not going to have enough chips to absorb all the costs.

Justify the investment.

And that's—

Yeah.

Right. This articulation is basically just what I said 12 years ago, and I think the challenge with chips and writing about chips, as we've discussed, is that you can see stuff like this coming 12 years ahead of time.

Mm-hmm.

But it takes a long time for it to play out, and there's lots of money to be made or lost in the meantime. This is where you get a real separation between the stock price and what's actually happening.

Tyler Cowen—one of his favorite critiques is always, “Well, are you shorting it?” This is one of those cases where I would use it as a counterexample to Tyler's favorite critique, which is, yeah, if I had shorted Intel, I would've been both right—

Taking a bath.

—and bankrupt.

Yeah.

Right? So, there is a challenge there.

Exactly. It would've been a tough decade.

Yeah, there is a challenge there. Some things are just not very investable in the long run.

Ultimately, probably the way to have approached Intel in 2013—and I think a lot of smart investors did this—would've been to plow your money into TSMC, the company that did win mobile.

So, in broad strokes, whether this is actually true or not, whether this is a negotiating tactic, or whether Lip-Bu Tan just always wanted to get out of manufacturing in general—

Yeah.

—which, by the way, is interesting, because I think on one hand, I and others probably gave Pat Gelsinger too much benefit of the doubt in his statements. I will say my natural inclination is to look at Lip-Bu Tan with a bit of a side eye, at least from a manufacturing perspective, because when you go back to the stock market perspective, the most cogent thing for Intel to do has been to just get out of manufacturing.

Try to preserve the slow decline, become fabless like AMD did previously, outsource to TSMC like everyone else, and maintain and double down on the product advantages that do keep people in Intel. It's not just that manufacturing has become a millstone around their product group. Their products have been bad—not as good as AMD's—but they were also on slower processes, which exacerbated the problems that their product group had.

Right.

Now AMD's over 50% server market share, and that was the segment that kept Intel so profitable and attractive as a stock throughout the 2010s.

And because he was sort of a software guy, I always had a bit of a side eye for him, as someone who cared about the manufacturing side—not just from an Intel perspective, but from a national security perspective. I would say that was true from the beginning.

A tell in terms of the direction this might go. Stepping back, you mentioned the national security side. The rational business choice for Intel may be to step out of manufacturing, but for the last several years, there's been talk about how urgent it is to save Intel because it's important to have an American semiconductor manufacturer. Now it looks like they're one foot in, one foot out on the leading edge with 14A.

This is why nine months ago I wrote the article saying, “Look, I had Intel right pre-Gelsinger.” This company needs to split. If you want to make this work, it's implicit in this statement: Intel products do give Intel Foundry the volume, or at least a huge portion of the volume, they need to function.

Mm-hmm.

If Intel went completely to TSMC, this foundry effort wouldn't just need one external customer; it would need a bunch of external customers.

But my contention then was, yeah, that would be very hard to split them for lots of good reasons. It seems clear to me—and this was my take pre-Gelsinger; then I gave Gelsinger a run, and last fall I was like, “Nope, my initial take was right.”

Intel's culture is fundamentally incapable of becoming a foundry. They cannot not prioritize themselves, and no one is going to want to partner with them and go with them given that reality.

Right.

It has to be split up, which is why I'm saying, look, the US government, if it actually is a national security concern, is going to have to just suck it up and fund it.

Mm-hmm.

Maybe my side eye with Lip-Bu Tan is actually a side eye of, “Yeah, we're on the same page here.” Maybe for slightly different reasons.

Yeah.

From a shareholder perspective, it would probably be best for Intel to dump the manufacturing.

Right.

Well, that's debatable, but there's certainly a case to be made from that perspective. I would argue that if you want this foundry to function, it has to—

Mm-hmm.

I just—Intel's not a customer-service organization, and that's what a foundry needs to be at the end of the day. It's not just whether they have an advanced process. It's whether they're actually partners you can work with and deal with.

2. Intel Optimizes for Performance

There was a lot in this call from Tan, with very legitimate critiques of Intel, that were hilariously directly in line with the critiques of Intel for the last 30 years.

Okay.

All they care about is performance.

Give me some examples.

I think the big one is that they were insufficiently focused on yield and only cared about performance, which is the way Intel's always been. They never really had to be as concerned about yield.

Yield is, of the chips you make, how many are useful?

With logic, it's very challenging because in a logic chip, almost all parts of the chip are distinct. This is different from memory. We'll get to memory in a little bit, because I think we're talking about Samsung.

In memory, you're making the same sort of memory cell again and again and again. So you make a huge sheet of memory, and if a bunch are defective, it doesn't really matter that much. I mean, it matters, but not that much, because you just disable the ones that are defective, and then you still have a useful memory chip.

Right.

In logic, if part of it's defective, the whole chip is defective because every part is much more unique. So, even though making memory and logic both involve fabbing, there's a reason why you generally have different memory companies—SK Hynix, Samsung, and Micron—than you do logic companies: Intel, TSMC, et cetera.

Mm-hmm.

They're very different skill sets, and there's a very different appetite for risk. The boom-bust cycle in memory is insane. It's much more of a commodity. It's really a money game in many respects.

In logic, it's really—what's the word I'm looking for? It's a precision-discipline game because you have—

to sort of get it right. And if you're TSMC, you really have to get it right because your viability is a function of keeping those lines as full and as fast-moving as possible.

Yes.

But Intel's been unique because Intel is fully integrated. They're their own only customer for manufacturing, and so they've always had a little bit more leeway because they had larger margins, in part because they were getting all the differentiated pricing. Intel was the NVIDIA back in the day because they made their own manufacturing. NVIDIA has to pay TSMC a bunch; Intel had it all internal.

Mm-hmm.

That gave them a bunch of leeway. So they could tolerate worse yields, and they could also tolerate just being slower. One of the areas where TSMC and ASML leapt ahead was in the shift from 200-millimeter wafers to 300-millimeter wafers. When you get a larger wafer, the physics of it are different—the way it's moving through machines and around. You kind of have to reengineer from the beginning to get the full benefit of 300-millimeter wafers, which is that the more area you have, the more chips you get per wafer. That means every wafer going through, you're making that much more money because there are that many more chips on the wafer.

Nikon and Canon, at the time, dominated lithography, and they just wanted to adjust their machines: “Okay, now you can put a bigger wafer in.” But that meant it was slower, because controlling the centrifugal forces of rotating the wafer around was difficult, and things like that. Intel was fine with that. Their focus was not on speed and efficiency.

Right.

What they wanted, what they liked, was that they had full control over Canon and Nikon, and they were hand in glove and working together. ASML and TSMC worked together to completely reengineer lithography around 300-millimeter wafers.

Mm-hmm.

That's really when ASML became a force in lithography. This was before EUV. It was because they worked hand in hand with TSMC to reengineer the entire process around the physics of a larger wafer so that you would get the same throughput as a 200-millimeter wafer, but with a larger wafer.

Right.

That meant you got much more efficiency in your line. This is just a mindset of TSMC. It was a mindset of ASML. It leapt them ahead of Intel and Canon. What it speaks to, and why it's pertinent to this discussion, is that Intel, by virtue of its integrated model, has never had to be concerned about efficiency.

Move with the urgency of somebody who's trying to serve customers as quickly as possible.

That's right, and the same thing with yield. They could be less concerned about yield because they had such huge margins. And they had a super-aggressive binning strategy, right? That's why every Intel processor had about 8 versions, which was really just compensating for their yield issues. But that was fine because they were serving—if you got your mini PC, or what was the crap?

Netbooks. The netbooks.

If you got a netbook with an Intel processor, that processor was just riddled with errors, and they're turning off huge chunks of it. But that's fine. They're still making money off it, selling it for $50 or whatever it might be.

But it speaks to the cultural point, and you raised that. You drew a distinction between Intel and Samsung's culture with the news that Tesla's signing a $16.5 billion deal with Samsung, and Samsung's going to fab AI chips for its autonomous-driving dreams. Obviously, TSMC has taken over the world in some ways by serving customers better than anyone else. Essentially, it looks like Intel just doesn't have those muscle groups in order to thrive in this world.

Right. No, that's my big takeaway. Setting aside everything else, Intel just doesn't have it in them to serve customers.

Mm-hmm.

You go back to Lip-Bu Tan's comment, and he's talking about how they didn't really care sufficiently about yield; they cared about performance. That goes back to the Tony Fadell discussion that I had ages ago. The reason why I got that interview in the first place is that I wrote about Paul Otellini's comments about turning down the iPhone order. He emailed me, and he's like, “No, I'm so sick of this. They had no chance.”

That's great.

“Their chips were fundamentally uncompetitive because all Intel cared about was performance. They did not care about efficiency.” Even though they had an ARM division at the time called XScale, their ARM chips were focused on performance, not on efficiency.

Mm-hmm.

This has been the Intel mindset. And this goes back to one of those things where the things we view as rotten culture today were great culture in the past.

Yeah.

You go back to the 1980s and 1990s, and a couple of things stand out. Number 1, you go back to what I brought up before: Pat Gelsinger was the one who fought for CISC over RISC, a less efficient architecture, because his point was, “We have a software moat, and by the time we rebuilt everything for RISC, number 1, we're opening ourselves up to competition. And number 2, our manufacturing, if we had just stayed the course, would've pulled us ahead so that our CISC-based chips would be faster anyway.” It's core to Intel's culture that we'll make it up in manufacturing by just being faster. It's performance.

Mm-hmm.

Number 2 is the entire ethos of computing in the 1980s and 1990s. We talk about this in the context of AI, where you shouldn't be optimizing your AI apps for the models of today because you assume the next model is going to be way better and is going to solve a lot of your problems. That was computing in the 1980s and 1990s: any time you spend optimizing your application to run fast on current chips is a waste of time because by the time you've finished optimizing, new chips will come out—

There will be—

—that will solve your problems for you.

Faster, better chips, yes.

Right. Now, this is how we got bloat in software, but it was a very rational way to build software in a world where chips just got massively faster year after year after year, and Intel was the foundation of this. They were making their chips faster and faster, and this was in an era when all the computers were plugged into the wall.

Mm-hmm.

So Intel was fine. But this is just—Intel is historically self-serving, totally optimized for performance over everything else, not sufficiently focused on yield. They're used to being fat.

Right.

They're used to having big margins, and they're used to just making whatever chip they want and people having to come and get it. This whole backside-power thing is another example. This is where I almost lament the loss of Intel to a certain extent.

Mm-hmm.

I'm sorry, am I monologuing too much, Andrew?

No, it's good.

I'm on a bit of a rant. I feel like I should've written a big article this week to summarize all this together, so you're getting it right now.

There's never a time we discuss Intel without you launching into some 15-minute digression on Intel history and Intel's complicated future.

Yeah.

I mean, I'm looking at Intel, and this is not a specialty of theirs. They suck at customer service. It's not surprising to me in the slightest that they're struggling to find an external customer for the 14A node. I'll let you finish, but I do feel like some finality in terms of Intel's manufacturing shops and Intel's manufacturing prospects on the leading edge might actually be healthy at this point in the process, so that everybody could just move forward—

I know, I should've written a big article—

—with clarity.

This was a mistake on my part. If I'm writing an update over 2 days and then doing a 20-minute digression on a podcast, I clearly have it in my system. I need to get it out.

It's late July.

So sorry.

Let the people have it.

Everyone might get an article next week whether they want it or not. The backside-power angle is, I think, very interesting.

Uh-huh.

Generally speaking, the way chips are laid out is that there is the logic layer at the bottom, then there is a communications layer that sort of links all the logic stuff together, and then there's power at the top. The power part is electricity, and it has to be piped down to the logic layer. It's laid out this way for lots of historical reasons, but one very good reason is that the logic layer is very complicated. That's where you could screw up.

Mm-hmm.

If you screw up, it's easier to screw up earlier in the process than later, and you toss the chip. If you remember, a lot of the cost in chipmaking is the time. The cost is not the wafer itself. A wafer is a few hundred dollars.

Right.

The cost is the time. And so the sooner you can discard a wafer because it's screwed up, that is actually yield-accretive. Even though you're losing a wafer, you're getting a newer wafer in that much sooner. Yield is a function of the number of good wafers out of total wafers. If you have a bad wafer, you want to increase your total wafers sooner rather than later. That's how you get your yield back up.

So where you do the logic—the hard and complicated part, where it's different—you would do that first and then do the other layers on top. The problem is that as these chips get smaller and smaller and need more and more power, you start losing electrons and having RF interference, along with all these sorts of issues from electricity, especially going through this communications layer.

Mm-hmm.

The solution, from a physics perspective, is to separate the communications and the power. This is what is called backside power. You put power on one side of the logic, then you have the logic, then you have communications on the other side, and they're totally separated.

This makes it easier to design the chip because you don't have to worry about trying to fit all this stuff in a small space. It reduces interference issues. There are lots of benefits to doing it, but the cost is yield.

Intel's solution is the most technologically forward solution: They're laying down the power, then they're putting on the logic, then they're putting on the communications, and it's all a fully integrated chip. The risk is that if you screw up the logic layer, you're throwing away a lot more work, so it's going to be worse for yield. It's going to be more expensive overall.

What TSMC is doing is logic and communications, and then for N2, which Apple is using, they're putting the power on top, like it used to be.

Mm-hmm.

And for A16, which Nvidia is going to use, they're making a separate chip with power and bonding it to the bottom. So it's literally making 2 chips for 1.

That approach is inelegant. It's kind of clunky. It's more expensive than the other approaches because you're making 2 distinct chips. But if you step back and zoom out, the end result is that Apple, which is making chips for a phone where the total power is gated—

Mm-hmm.

—you don't want to go above 7 or 8 watts because you're running on battery—

Right. You don't want to drain the battery.

—if you're a computer, a distinct device.

From Apple's perspective, they get no benefit from backside power because they're not amping up the power enough to encounter all the issues that backside power is meant to address. They don't need backside power, and so they don't want to pay the yield penalty or the extra-chip penalty of getting it.

Mm-hmm.

So just from an apples-to-apples basis, TSMC's approach is better for Apple than Intel's approach.

Right.

But what about the Nvidias of the world? They're selling chips for $30,000 or $60,000, or whatever the price might be. So if TSMC jacks up the price, Nvidia's not happy about it, but they have plenty of margin to pay for it.

TSMC chose a clunkier but ultimately more flexible approach to N2. Nvidia's chips are pulling 700 or 800 watts per chip. Just massive amounts of power. They really have a power issue, and they want efficiency, but that efficiency is gated by cooling. Given that you're gated by space in a data center and your total electricity generation, it's really interesting because both Apple and Nvidia care about efficiency, but in totally different ways.

TSMC is delivering 2 totally different approaches using the same logic layer for N2 that serves both.

Right.

Intel's stuck in the middle. It's a bit of a tragedy because Intel remains, and always has been, more technologically interesting and creative than TSMC.

This is a great solution. Intel's 18A and 14A are all backside power. It's a completely new way of building chips. It's very technologically innovative. It's classic Intel.

The problem is it's not meeting anyone's needs.

Yeah.

It doesn't meet Apple's needs. It theoretically meets Nvidia's needs, but this is the problem of not being in the market and not having proven themselves to date.

Well, that's what I'm saying. There's no proven track record here. There's also, now with these comments from Lip-Bu Tan, no clarity about how committed Intel actually is to staying in this world. Then you layer on the lack of customer-service chops and the lack of a track record where they are working with companies and moving with urgency to meet their needs—meeting anyone's needs other than Intel's—and it's not surprising in the slightest that there aren't people beating down the doors to work with Intel.

That then creates a situation where leading-edge manufacturing in the US looks like it's now a 2-player game, with Samsung and TSMC. Would that be a reasonable assessment going forward?

3. The US Foundry Future

The Samsung angle is pretty interesting. I haven't covered them enough. They really ran into some problems, first with EUV and then with gate-all-around.

Part of the 3-nanometer transition for all these companies is a new kind of transistor that becomes even more 3D than FinFET was previously. These transitions are always super difficult and hard. It's usually at these transitions that companies fall off because they just can't do it. The money's too great, someone else does it, and goes forward.

Samsung was making Tesla's chips before, on 7 nanometers, and then they really fell apart on 3 nanometers. They just couldn't get it working, so Tesla went with TSMC for the AI5 chip.

The news here is that the AI6 chip, which will be on 2 nanometers, is coming back to Samsung. They're the first and best customer for this new Texas fab, which was delayed and now is clearly going to be built. It seems like they're the dark horse that came in, and they're going to be the second supplier.

Yeah.

You just zoom out. Number 1, Samsung has some of the same concerns as Intel, particularly around IP theft and things along those lines. That's part of why Apple left them. Qualcomm expressed concerns about this. They actually separated out their silicon division from the rest of the company because there were real concerns from customers about them, particularly about a decade ago.

Some murkiness there, yeah.

Yeah. Intel has all the same concerns, to be clear. This has always been one of TSMC's trump cards: “We're never going to compete with you.”

Right. We're not making phones.

“We're never going to sell a product.”

Yeah.

That's right. We're not designing our own chips. We're only making the chips.

Mm-hmm.

You're doing the design. So that's always been a bit of a concern.

But Samsung is also a memory maker, and part of that is, I think, there's always a cultural challenge for memory makers, as I mentioned before, in making logic because it's a different skill set. Then again, memory also throws off a ton of cash in the boom cycles, which can help fund these transitions.

Memory's also very lithography-intensive, which is, again, why Samsung's EUV struggles set them back in memory also. But it also means they did get over the hump, and now they've sort of figured it out when it comes to logic. Now they have a core customer who's making a lot of chips.

You think about Intel needing a core customer. Who's left? Apple and AMD are not going anywhere. They're locked in with TSMC. Apple is locked in on the leading edge because they help fund TSMC. AMD is locked in because TSMC really saved their bacon. They're really loyal to them. They're not going to go anywhere else, even though they're a node behind.

Mm-hmm.

I just think that's a very tight relationship that's not going to change. Nvidia has always been willing to dual-source, but Nvidia hates Intel with a passion. This is 20 or 30 years of history. If there's any company in the Valley that never, ever wants to get in bed with Intel, Nvidia is arguably at the top of the list.

They believe Intel tried to kill them. Well, they believe it, and it's true, again and again, through both shady and legitimate and illegitimate reasons.

Okay.

It's hard to see Nvidia ever going with Intel. But they would want a second source, right? I mean, because—

They do want a second source—

That's leverage with TSMC.

Right, well, wait, but if they want a second source, Samsung—they've gone with Samsung previously. I think they'd be willing to do it again.

Mm-hmm.

And so then there's Qualcomm. Qualcomm's a little iffy to go with Samsung because they compete in modems, but then again, Qualcomm has kicked the tires the most with Intel and has said no.

Yeah.

They just don't trust that Intel can meet their needs. So Qualcomm's probably the biggest wild card out there as a potential Samsung or Intel partner. I still think the sort of—I think Doug Olofson put this out, and I sort of co-signed Broadcom acquiring Intel, even though they probably don't want the manufacturing part. But Broadcom has a lot of volume.

Did Broadcom try to acquire Qualcomm? That was blocked. There's somewhere in there maybe a path forward. But if Samsung becomes a viable second source, is there room for 3?

Mm-hmm.

I think that's a very fair question to ask.

Are there enough customers—

And Samsung does have—they've done the customer service thing. Again, I went back to that Fadell interview. They worked hand-in-hand with Apple before things sort of fell out because of product arguments.

The good old days.

And Samsung does have experience with that. They worked hand-in-hand with Apple before things fell out because of product arguments.

Well, I mean, that was going to be my final question here because it occurs to me that maybe 3 years ago we were recording in the wake of the passage of the CHIPS Act. We hadn't gone live yet with the Sharp Tech Podcast, but we were recording one of the early episodes that we put in the bank and launched with.

We were talking through the CHIPS Act, and we were talking through Intel as a national champion. Every Intel conversation we've had since, I've been looking at our national champion here and thinking, "God, it's not looking great for the prospects at Intel."

It's so true.

And there are layers and layers of problems there.

It's funny—

Okay.

There are layers and layers of problems. It's like the layers and layers of a chip. The problem is that the base layer of Intel is their culture.

Yeah.

That's rotten. We just have to throw the wafer out.

That's the thing. As an American looking ahead at the future of the U.S. semiconductor supply chain, should I feel good about where things stand right now? TSMC is fabbing in the U.S., and Samsung looks like there's some promising progress on that front as well. If Intel is dying on the vine and we're giving up on the dreams of a national champion emerging here in the next 5 to 10 years, is that okay?

I think things are better than they were. I am somewhat encouraged. I think I was a little skeptical of the TSMC Arizona promises, and maybe I was right to start with. I don't know. It's hard to say.

Where they're at now, I think, is pretty promising. Just in general, Arizona being a validation of some of the stuff we talked about with trade and things along those lines, where just getting a foundation, an anchor—

Mm-hmm.

—starts to pay off over time. There's an aspect from TSMC, maybe the first fab was just to mollify Trump. I mean, remember this started under Trump, sort of, seven or eight years ago.

Trump won, yeah.

And, yeah, we're going to make stuff in the U.S. I'm like, wow, all this is going to be totally dependent on Taiwan. It's going to be a few generations behind.

Even Morris Chang was giving interviews saying, "We make less money in the U.S."

Yeah.

"I really love this."

He's like, "We're doing it because we're being forced to." Yeah.

Right.

But once you've put in the effort—and there were well-chronicled challenges and difficulties in getting this fab built in the U.S.—

Oh, yeah, all kinds of cultural challenges with TSMC, for sure.

Well, yeah, but there were cost challenges, and there's just so much to get it started. Once it's there, how do you maximize all the time and effort that you put in?

Make more.

You build more.

Produce.

It's the exact same as a fab, right?

Yeah.

Once you've built the fab, you need to maximize the number of chips that are being sold by that fab so you get a return on your investment.

In this case, whatever the motivation and whatever the intention of that original fab in Arizona, 7 or 8 years later, has resulted in an extremely aggressive, large-scale... They're talking about—I think it was—30% to 40% of their advanced chips being made in the U.S. in a number of years. That's a large number. That's much larger than I think was thought about at the beginning.

Mm-hmm.

It's going to have its own R&D center, so it can be somewhat self-sufficient if Taiwan gets taken off the map. They refer to it as a GigaFab.

Right.

What is a GigaFab? A GigaFab is a new term that's entering the lexicon, but I think it's actually really interesting.

The way TSMC used to function, 10 or 15 years ago, they would build a fab once and then just run that fab forever. That's how they made their money, because they were way behind Intel, so they couldn't charge a premium for their chips.

Mm-hmm.

But they would depreciate that fab over 5 years or 10 years or whatever it is. They'd use that fab for 30 years, and so it would just generate free cash over time because you've already paid for the equipment once and you're just selling. Even if the chips cost $1 each, it's $1 of basically pure profit. So you're making money again and again for ages.

They faced a real challenge when they became the leaders. I critiqued them heavily for N3, and I don't know—I think it's why their former chairman of the board and former CEO lost his job. From what I've heard, he didn't necessarily leave, especially the chairman of the board post, willingly. They underpriced N3.

Hmm.

They were the leaders in the market, and they priced it too low given their dominant position. The reason is that they had to have a cultural mind shift from, "We're going to make money over 30 years," to, "We need to make money immediately because this stuff's so expensive, and also we have pricing power."

The other reason is that below 28 nanometers, there was a shift to FinFET, which is around 17 or 18 nanometers, and then at 5 to 7 nanometers there was the shift to EUV. The cost became so large that for a lot of applications it didn't make sense to go lower than 20 nanometers.

It's a perfectly functional chip. We don't have any power constraints or performance constraints. It's going in a refrigerator. It's totally fine, right?

Mm-hmm.

We don't need to—

Right.

The problem is they had these massively expensive fabs that the Apples of the world would come in and use. Then, when they had the new massive fab, Apple and AMD and everyone would move forward, and suddenly, "Wait, who's going to use this fab?"

Right.

They had this gap in the middle between the trailing edge and the leading edge. So they've had to rejigger their business model to make much more money up front, but also rejigger their fab model into this GigaFab idea, which is that our 3, 5, and 7 nanometer fabs are all together.

Hmm.

We are designing them to be repurposed. We can start with 7, and we only need to change out some equipment and it can do 5. Then we only need to change out some equipment, and it can do 3. Building in this flexibility from the beginning, knowing that we're not going to have as much backfill as we're used to, means that what we need to do is reuse our equipment to the greatest extent possible on the leading edge.

The thing about this is that that was the Intel model, right? They—

Hmm.

TSMC is becoming Intel, and they're becoming Intel much faster than Intel was able to become TSMC. That's really, at a fundamental level, what has happened over the last 5 to 10 years, if you were to sum it up. Man, that's a good line. I'm going to have to write this article for sure now.

There you go.

I thought it was meaningful in that context that they're now referring to Arizona as a GigaFab.

Yeah.

A GigaFab is not one fab. It's a bunch of fabs that are worked together, a flexible sort of organism that's meant to continually be on the leading edge.

This is not a “we’re gonna throw some sort of satellite out there that’s copying one of our fabs from before. It’s just gonna sit there and run.” They have a fab in Washington because they tried the U.S. thing in 1999—

Mm-hmm.

—or something.

Or is it Oregon or Washington? Yeah.

It’s somewhere around there.

It’s the Pacific Northwest.

Yeah. That’s just stuck there, frozen in time, right?

Mm-hmm.

It operates because they’ve spent money on the equipment, but it’s still 1999 technology, right? That’s not gonna be the case in Arizona. So, number 1, as far as the U.S. position is concerned, TSMC’s presence in the U.S. is substantially different, more meaningful, more sustainable, and more independent of Taiwan than it was previously.

Mm-hmm.

And kudos to the U.S. for making that happen. So that’s number 1.

Yeah.

Number 2: Samsung. This whole fab is in Austin. Samsung is still, I think, similar: their core functions are in South Korea. Their R&D is there, and their fabs that pioneer all this stuff are in South Korea. I think it behooves the U.S. to now push Samsung to make similar investments around this Texas fab that TSMC has made in Arizona. But the likelihood that that happens, given this Tesla deal, is much higher because these—

They’re investors, yeah.

—investments have their own gravity.

They have their own gravity. If you’re gonna spend all this money to build a leading-edge 2-nanometer fab in Texas, it just starts to make more sense to keep building around that, to make it a center.

Hmm.

That would be a big focus for me. If I were in charge of this sort of chip stuff, my focus would be on continuing to push TSMC to be more and more independent in Arizona, which sounds like has happened.

Yeah.

And getting Samsung to do the same. Intel, I mean, it’s a bummer. But it’s always been a fight against gravity to a certain extent, and I think that culture is hard to overcome.

Mm-hmm.

I was just so struck by Elon Musk saying, “I’m gonna walk the line of the Samsung fab.”

Yeah.

On one hand, okay, Elon—

What are you gonna do?

—you’re an expert now—

As you walk the line.

Number 2, everyone was like, “Okay, Elon, you’re a rocket expert now,” 15 years ago, and now look, right?

Yeah.

So, hey, maybe he’s gonna actually have some input. You sort of think about it like: What if we put Elon Musk in charge of Intel? That actually might solve the problem.

Figure some things out. Right.

Yeah, I think that would—

Break a few eggs.

—cut through some red tape.

Maybe cook an omelet eventually, I would say.

Yeah. It’s exactly what Intel needs. There’s no country in the world that could probably benefit more from Elon Musk being in charge of them than Intel.

Yeah.

Just going in and breaking things. Intel needs to be broken.

Mm-hmm.

They need to be broken and reconstituted.

Well, the reason I highlight TSMC and Samsung is because I came away from your coverage earlier in the week feeling optimistic about the future of leading-edge manufacturing in the U.S., despite all of the dysfunction compounding at Intel and them potentially getting out of the leading-edge manufacturing game. Given that story, the future actually looks fairly bright for the U.S., generally speaking.

Does it matter that it’s not a “U.S. company”?

American company. Yeah.

In the modern stock market, does it really make a difference?

Does it matter to me?

I think it’s really interesting because there are lots of industries where this is a very pertinent question. Take shipbuilding, right?

Mm-hmm.

There’s a big push for the U.S. to make ships. Guess what the U.S. cannot do? Make ships.

South Korea.

The Koreans and the Japanese.

Yeah.

And should our shipbuilding endeavor, which seems very reasonable, manifest as American companies having to make ships in America, or should it be some sort of partnership with Japan and South Korea? Of course, we should make some in America because of the risk of war in Asia, but we should be leaning on these companies, their expertise, and their capabilities to a much greater extent. This is like the Nippon Steel thing.

I was just gonna bring that up. Yeah, it was so infuriating to hear people pushing back against that deal. If Nippon wants to do it, then let’s let them do it.

If they want to invest money in America to upgrade steel mills, why are we saying no, right?

Yeah.

At some point, are we going to trust the Japanese or are we not?

Right.

Right? And I think a similar thing is true with South Koreans. You could push back on that. Look at a map: South Korea is not an ideal location.

Mm-hmm.

We fought China over South Korea in the past. It’s something that has actually happened. But at the end of the day, is this gonna be a pursuit of the perfect and an abandoning of the good along the way?

How long it would take for Intel’s manufacturing to get to perfect is its own question in the midst of this.

And Intel, by the way, is not exactly scared of outsourcing. They’re the ones that started it, right? They went to Hong Kong, as did Fairchild Semiconductor in the 1960s. They were the first ones to do it. All in all, I think you’re right to feel encouraged. I think the TSMC situation—and again, I think this is a real credit going back to the first Trump administration for really pushing and putting pressure on this specific point—

Mm-hmm.

—and this has been a bipartisan effort. I think we’re in a much better spot than we were, even if Intel is arguably in a worse spot.

Yes. Shifting gears here, I look forward to your Intel article next week, although that’s not a promise.

No, I never promise articles now.

Who can say? Who can say?

I think I wrote everything I wanted to write, but it wasn’t packaged. I don’t know. We’ll see.

Yes. Time will tell. For now, last week we recorded before you’d had a chance to parse Google’s earnings. You wrote about them for Thursday, and then Bob C. sent the following question. He writes, “Your update on Google’s earnings has me thinking. Are any of the big tech incumbents going to lose in an AI world? I am now long on everyone.” What do you think, Ben?

One of my better articles—I mean “better” because it was both prescient and everyone hated it at the time, and I think it turned out to be right—

Okay.

—was “The End of the Beginning,” which I wrote in January 2020.

I analogized tech to the car industry, where the car industry started out with a gazillion car companies between the 1890s and 1910s. There were hundreds of car companies that were started, and then it rapidly congealed into 4 or 5, and then over time down to 3, such that Tesla was the first new car company in about 60 years.

I was gonna say 80 years. Yeah.

And, yeah, there was one that started in the 1940s or 1950s that got absorbed right away. So realistically, it’s really the first new car company in 100 years.

Mm-hmm.

People wonder why Elon Musk gets a lot of credit from people. That sort of says it right there, right? A seemingly impossible task.

What happened was you had this explosion in innovation in a new industry, then it just settled into economies of scale and production, and all the innovation happened on top of that. Obviously, there were foreign companies—the Japanese and the Europeans and all those sorts of things—so it’s not just a U.S. story. But the overall take is that—

People might be biased by the fact that they were around for tech’s early years, where you had a gazillion new companies—

Anything was possible for anybody.

—and everyone competing for these spaces—

Yeah.

—and XYZ. Actually, it might be settling down into a very rational structure of the big tech companies. Each of them sort of does their thing, and there’s some competition, like in the cloud layer, for example, but by and large, each dominates their space, and every time someone tries to challenge them, they just get overwhelmed with scale and distribution.

Mm-hmm.

That's just sort of what happens. I noted that this was combined with the sort of end state of computing: a phone in your hand and cloud computing on the other side, and you have continuous computing everywhere you go. Yes, we'll have devices beyond the phone, but it's really hard to beat the phone for being the sweet spot for everything you want: portability, battery life, connectivity, a screen big enough to do stuff, and the ability to fit in your pocket.

With cloud computing, you get access to computing everywhere. It's continuous, thanks to 5G in the air and the internet, and whatever it might be. Given that, it makes sense that this is the point where things are set.

That doesn't mean innovation is over. It means all innovation going forward is going to happen on top of this foundation. I think Bob's question here is, “Ben, were you totally right in January 2020?” And, Bob, it looks like I might have been. Google—AI comes along. Google was always sort of the most interesting, and not because they couldn't technologically do it. Everyone knew they were ahead, right? OpenAI famously was started because everyone was saying, “Google's going to run away with this.”

It was the business-model question: What happens when the 10 blue links question goes away? And the whole thing about Google and this update—that's still a question.

Mm-hmm.

Like, it's still there. But the thing about the cloud and GCP, and the reason why I've always been very bullish on GCP, is that I'm one of the world's biggest Thomas Kurian fans. I think what he's done for that organization in general is very impressive.

Thomas Kurian is running GCP, for anybody who's not familiar.

That's right.

Yeah.

He's a former Oracle executive. When they hired him, it was like, “Why are they bringing an Oracle guy into Google? How's that going to work?” Turns out that's exactly what Google needed. They needed some Oracle discipline to actually be a reliable organization.

Let's get Elon into Google, you know? Mix it up. Tell me your—

No, no, that's the wrong mix.

Thoughts on this.

No, we need Elon to go in and break things. Google is so loosey-goosey and flexible. They don't need someone to break things; they need someone to stiffen them up.

Some structure. Okay.

They need the exact opposite. That's right. Thomas Kurian gets GCP into a functional enterprise organization just in time for AI to show up. It was so perfect because of all these investments that Google makes. I think we mentioned on a podcast previously, people were like, “Well, you know, Google's responded to OpenAI.” The point was they needed an OpenAI to respond to.

They didn't have it in them to build an AI product. They invented the transformer but didn't build crap with it in the real world until OpenAI forced them to. That is a picture-perfect example of why an incumbent holds back innovation for business-model purposes, even if it's sort of subconsciously.

Right.

But GCP takes advantage of all of Google's strengths. It takes advantage of the technology; Google has the best cloud and all its own networking. This goes back to the dark-fiber days of the dot-com era, when Google bought up all this dark fiber. They still have huge advantages that are reaped from that.

They have everything that's great. Also, with GCP, other people are building the product. So Google's product sclerosis and inability to build anything effectively don't matter because other people are doing it for them. That's why I've been posting this graph—or this chart—of GCP revenues, margins, and growth rates, earnings after earnings after earnings, because this is the key.

It fits in perfectly with Google. It's kind of a business they can get for free as long as they can execute and as long as their AI offerings are compelling enough for third parties to actually switch to them. All evidence is that that is happening. GCP is crushing it. Their margins are going up, their share is going up, everything's going up.

I don't know if this is why Google has a new CFO. All I can say is, ever since they got a new CFO, it's been pedal to the metal, and the pedal to the metal is CapEx investment explicitly for GCP.

So who are the GCP customers, just for example?

Enterprises all over the world. Everyone. They brag that a lot of AI companies are on GCP. OpenAI is now on GCP, apparently with some workloads.

Say you're Bank of America—I don't know if Bank of America is one of them—but say Bank of America. They need cloud computing, right? The first wave was on AWS, particularly a lot of startups. A lot of companies went to Azure with Microsoft. But there are a lot of companies that are either going to the cloud now, so they go to GCP, or they might be on other clouds, but Google's AI offerings are so compelling, and the AI stuff on Google's cloud is much cheaper than it is everywhere else, in part because of Google's implicit infrastructure advantages.

So they're like, “Actually, we will keep Azure, but we'll also go to GCP.” The great thing from a GCP perspective is that once you have some workloads, they have gravity. They'll start pulling more workloads in.

The bull case for GCP was always that Google's AI offerings would be differentiated, and that would be the anchor to start pulling in other business. It looks like that's happening. The beautiful thing for Google is that this is completely independent of the search questions. The search questions are still out there.

But GCP was always compatible with the search model. The question was whether they could execute on that. It appears that they are, and they're investing for it. They did $55 billion in CapEx in 2024. It was clearly not enough.

Yeah.

Now they went up to $75 billion, which blew everyone's mind. I came out in January and said—even though I've been the Google skeptic—“Investors are totally missing the point here. This is amazing. You should want them investing more.” I wrote the same thing last week. They went from $75 billion to $85 billion.

If GCP is such an opportunity that they don't have sufficient supply, that means they screwed up previously. Again, they do have a new CFO. I don't know if that's why. All I know is that they definitely invested too little. Now they have a new CFO, they're being super aggressive, and they absolutely should be.

Yeah. Well, zooming out, as far as big tech generally, we go back to your article, “The End of the Beginning.” I don't think we were working together at that point. In fact, I know I was a lawyer at that point, but we were friends, and I thought that was a really poignant article. Was it 2021?

No, 2020. January 2020, right before—

Oh, COVID was dawning there.

COVID was billowing out from China. Yes.

The framework makes sense, and the analogy to the auto industry makes a lot of sense. You look at some of these companies now, and they have so many built-in advantages and various forms of lock-in. It's easy to see how AI will be more sustaining than disruptive.

Google has always been the unique case within that conversation, only because Google is so dominant. They have the entire market share of the greatest business in the history of businesses. So if the volume of traditional searches drops 20%, or Google's market share of searches drops 20%, Google is still ridiculously profitable in that context, but less dominant than they've been for the past 20 years.

4. OpenAI Faces Incumbent Economics

So who's going to lose in an AI world? Losing may be relative, particularly if the cloud business emerges and becomes dominant in its own right over the next however many years. It sounds like it's dominant now.

Well, it's interesting because the cloud business is a much lower-margin business. So there is a risk. Google is committing to lowering its overall margin profile. If GCP is massively successful, that means Google has worse margins as a company.

Right.

That's always a challenge for companies to accept and deal with, and kudos to Google for leaning into it. They are investing heavily to get worse margins, which is exactly what they need to be doing.

The other companies—yeah, that's why you go back to AI in the Big Five. Google has definite disruption questions. Everyone else is looking pretty great. Distribution and all those sorts of things matter. I think the real question, if it's true that AI is sustaining and not disruptive, is whether we should be a little more worried or skeptical about xAI, OpenAI, and Anthropic.

Mm-hmm.

If they're in a fight with incumbents who are funding their efforts with profits, and you're funding them with VC money and debt and whatever else—

And escalating losses—

—you can get your hands on.

Every single year.

Yeah. That's a tough fight to be in. It really is, right?

Yeah.

I mean, it depends on the chart that you look at, but with OpenAI, it has market share, it has mindshare, and it's growing every single month. The rate has been unbelievable in terms of the accelerating growth. But then you look at the projected losses. Relative to the way tech worked 20 or 25 years ago, you're supposed to lose less money as you take over the world and people are using your product constantly.

But I don't know.

I'm almost sure their unit economics are still positive. People's subscriptions to ChatGPT—if they stopped all their R&D investment and whatnot, they'd be fine.

Mm-hmm.

But does that count if you can't stop?

Right.

Because you will fall behind. And this is why Mark Zuckerberg's war for talent is really debilitating, because even if OpenAI keeps people, their costs are just rising through the stratosphere.

Mm-hmm.

Mark Zuckerberg wins either way. Either he gets the talent, or he really hurts OpenAI's profit margins.

Yeah.

And OpenAI and Meta, that's the fight. It's OpenAI versus Meta for consumer mindshare. So there's a very real factor here where Meta is waging a no-holds-barred assault on OpenAI, and this talent war is like a boxer doing body blows.

Mm-hmm.

You don't see it. It doesn't look as good on TV, but you're just sapping the energy and the vitality of—

Make it—

Yeah.

—harder and harder with each passing round, each $100 million deal that comes raining down from the sky from the superintelligence team.

And Mark Zuckerberg is paying for that with free cash flow. He doesn't have to go to SoftBank and get money.

Yeah.

He decided to pay them more.

Absolutely. I mean, it's—

So—

—again, built-in advantages. You have a money-printing machine. If you're Google with a search business, you're not underwater as you invest in the cloud business. You also have YouTube, which is just this empire sitting right there. Somehow, YouTube never comes up in the Google conversations, but it's just as dominant as any entertainment platform in the world.

A mistake I've made for years. Yeah, no, they're just taking over TV. That's one of the underappreciated things. This is the real Netflix threat. The Netflix threat was not other streaming services.

Yeah.

I think we were all over that. They're going to be fine in the long run.

HBO trying to become Netflix before Netflix becomes HBO. Yeah.

No, it's YouTube just devouring everything with a fundamentally advantageous cost structure because they get their content for free.

So can I just say, alongside this conversation, what I enjoy about the landscape for the next 5 to 10 years is that there's still a decent amount of unknown baked into all of this. You could go company by company, and the individual bull case is more compelling than the bear case for each of the big tech companies, whether it's Microsoft, whether it's Google, whether it's Meta.

But if I were to step back and ask myself, are we going to get 10 years down the line and still be talking about the same 5 companies as the most powerful consumer tech companies in the world? That seems like a bad bet, despite everything we just said about Google and despite everything we're saying about Meta. Meta has not had a hit product in a really long time. I mean, they bought Instagram and are monetizing that. They're great at monetizing, but in terms of a new thing that people care about, it's been a while.

And so the jury is still out in terms of whether they can make AI work. Apple is legitimately lost right now. And Liquid Glass—I heard you and Gruber talking about Liquid Glass. It sounds like the stupidest thing in tech history. I don't understand what they're trying to make happen there.

And so I'm closer than ever. If OpenAI released some hardware in 2 or 3 years and it was really cool, I'm open to trying it at this point. And whether Microsoft—how does ChatGPT change workloads?

This is why OpenAI is so interesting. This is why OpenAI is one of the most interesting companies. They're basically taking on everyone.

Exactly.

Right? They're trying—

You can go company by company—

—to walk the narrowest path, right? You—

—the threat is OpenAI. It's wild.

Well, you can make the case that OpenAI should never have hired Jony Ive. They should have really buddied up with Apple and cemented that relationship, and they didn't. And this is classic Altman. He wants it all. He's going for it all. And so, yeah, it's super compelling.

Trying to be Google and Apple at once.

It is funny, though.

Absolutely.

There is a bit where, tying into Stratechery and what we talked about in 10 years, there was a point in the mid-2010s where I'm one person. I can only be stretched so thin. People are like, "Why don't you cover X? Why don't you cover Y?" I'm like, because the answer to doing that—

There's only 4 days in the week.

—is having multiple authors.

Yeah.

That's right. Well, it's not just even that. I want to write authoritatively about what I write about, and I could lampoon myself as an expert on Company XYZ, but I'm making it up as I go along.

Mm-hmm.

And that ultimately is not good for you, and it's not good for me. And so I'm just looking at it like, Stratechery's really about the big tech companies.

Yeah.

That ultimately— And I will talk about other companies to the extent they intersect with that, and occasionally I'll dive into other companies as a side note when I think I have a really interesting angle. And it was interesting because that was just a functional decision that I had to make at some point in Stratechery's evolution.

Mm-hmm.

But it also ended up being, I think, strategically very good. The end of the beginning is good for Stratechery. That is like, I'm going to be Car and Driver in 1930 or whatever it might be. It's—

Yeah.

So it's funny that it just sort of has that impact. It's something that I have to think about. It was like, am I talking my book by saying it's the end of the beginning, that it's only the big tech companies, which happen to be the companies that I cover specifically?

Right.

But no, maybe it ended up working out.

Yeah. Well, and also I would say it really was fortuitous timing because AI arrived and everybody immediately treated it as this disruptive threat to several of the companies that we're talking about here. And the reality is there are so many advantages and so many different ways for these companies to integrate AI into how they make money, that there is—

I mean, I think the search question's still out there, by the way. Again—

Mm-hmm.

—I don't think that's been resolved. I think Google's responded as well as it can. Like I said, all the product shops that still exist in the company appear to exist in Search.

Yeah.

And I think they're responding as well as they can. Again, the GCP thing is great, but that was never in conflict with Search. It's still an open question if it's going to have an impact.

Well, I mean, that's what I mean. It's fun that there's still a little bit of unknown here, and all of these companies can feel confident. But the Earth could shift underneath them over the next couple of years and spice things up a little bit.

And as far as Google's concerned and Google Search, my wife was complaining about her sister last night. We were sitting together, and her sister was texting her all these different questions. She was like, "Why is she asking me these questions? She can just go to ChatGPT and get answers from ChatGPT."

Right.

So I think that's how a lot of people are living at this point. I'm not that unfriendly to various people who reach out with questions, but I don't think that bodes well for Google, obviously trying to protect 98% market share of that sort of use case. I guess we'll continue to check in, but a good note from Bob, because it does dovetail with a lot of the conclusions that you've drawn on Stratechery over the last year or 2 as people try to size up what the landscape will look like.

Do you remember—were you ever a user of LMGTFY?

Yes. But I always thought it was dickish when people would hit me with one of those links.

I did too. I never used it. I think there is a Let Me Google That for You website or something—

Yeah.

—and you can send someone a link, and then it replays you typing the question into Google and figuring it out.

Well, and look, here’s the reality, Andrew.

Where is the new Let Me ChatGPT That for You?

I hit you with questions several times per week that I could ask of ChatGPT, but I don’t because I want to talk to my friend Andrew.

Aw.

Communication is all part of the human experience. We don’t all need to just—

Man, you just bought yourself 100 new queries with that one.

That’s right.

That’s so sweet.

There you go.

5. Grok's Entertainment Bet

Keeping it moving and keeping it LLM-themed here, Daniel says, “Grok has been discussed by you both as being a potential big winner in the AI space. Presumably—”

Just a second. Just a second. When?

I don’t know. The first thing I was going to say to Daniel is that I don’t recall ever talking about Grok as a potential big winner in the AI space. They’ve been a peculiar entrant in the AI space for the last year and a half or so. Do you want me to read the rest of his question, or do you want to just reject it out of hand?

No, because the whole question is predicated on what our angle is. The theory of the case is that, at the end of the day, some of the LLMs are pretty annoying in what they will or will not answer.

Mm-hmm.

Is there actually a space? It’s not just what they won’t answer; it’s the coy nature of them. It’s very—

Yeah.

I’m very good at dismissing that and ignoring that, but if you do actually go back and read the text, it’s like, “Stop—”

Most annoying friend.

“Stop brown-nosing. This is really annoying.”

This is why I ask Ben. Exactly.

Right. And so is there room for, quote-unquote, truth-telling? Is it so much truth-telling as it is just saying it like it is, or whatever it might be? I think in theory there’s something for that. In reality, ChatGPT has won the consumer space. It’s hard to see anything getting in the way there.

To the extent I’ve talked about Grok and xAI, it is that they’re super useful for everyone else as a stalking horse. If it doesn’t work out with you, I’ll just go with xAI, right? If you’re Cursor and Anthropic has a problem, you can say, “Well, we can just plug in xAI.” It’s great to have more players. The problem is that being the designated second bride, or whatever it is, usually isn’t a very good business model.

The bridesmaid, I think, is what we’re looking for there.

Bridesmaid. Thank you. Second bride, whatever. So is xAI actually going to pay off with Elon’s own companies? Maybe. That might actually be the better way to think about it.

Yeah.

But, yeah, this is—I mean, the whole question—

Well—

Can’t do better, Daniel. Come on.

Hold on. In Daniel’s defense, this came up on a Dithering episode last week, and it wasn’t you. I think you were quoting someone else in a group chat. You were recapping group-chat back-and-forth stuff. Great midsummer content. Love it.

Oh, yeah, that’s what it was, with the three dragons and Grok being silly.

And being a potential winner in entertainment. The reason I raise that is because I do think that’s a really good question two and a half years into this new era: Who are the entertainment winners actually going to be?

I don’t know whether Grok—

Yes.

—is nonsensical enough to fill that role, but I do think it reminds me of the Nat Friedman discussion maybe a year into the AI era, where it’s like, where are all the products and where are the people who are going to capitalize on these new capabilities? I don’t know that there’s been a good entertainment solution that’s working on the strength of LLMs at this point, but maybe I’m missing it.

Right. The whole anime-girl-talking-to-you thing—everyone gets upset about it, and then you remember that some of the most valuable companies are stupid toys that everyone dismisses. Because it turns out most people just want to be entertained. They don’t actually want to get stuff done, right? And to the extent they want to get stuff done, it’s so they can get back to their entertainment more quickly.

Right.

And so, yeah, that was the angle. Fair.

Would Grok be the company that does that? I don’t know. What am I doing on my iPhone all day?

Right.

Mostly screwing around in group chats—

Right.

—and reading Twitter.

Yep.

It’s not me being productive all day, to be honest with you. Granted, I’m talking to my boss here on this podcast, but it’s an 80/20 situation in terms of iPhone usage, with 80% being entertainment. So who is going to serve that 80% market? Maybe it’s going to be ChatGPT, and ChatGPT is just going to win everything at the end of the day.

Yeah, yeah. Well, maybe. Or maybe it’s going to be Meta. Maybe Grok wins just by not caring. We’ll see.

By just being stupid and spewing out nonsense and occasionally pretty dark nationalism. I don’t know. In any event, good question. Well, complicated question from Daniel there. But we appreciate all the people who write into us.

6. AI Doom Becomes Marketing

To keep it moving, this is not an email, but it piqued my interest last week. This is Matt Levine at Bloomberg. He’s riffing on Sam Altman’s discussion of ChatGPT agent. He could be describing any number of product launches in AI. He writes:

“I have described Altman’s marketing approach as business negging. If you say, ‘Our product is good and will make the world a lot better,’ that is sort of ho-hum. Everyone says that, so your audience will discount it. But if you say, ‘Our product will probably make the world a lot better but could destroy humanity—tee hee, aren’t I a little rascal?’ That’s exciting.

“People are drawn to danger, and if you go around warning people that your product might destroy humanity, then that strongly suggests you think it’s powerful. Similarly, when you launch a product whose pitch is, ‘This product could do your whole job for you,’ that’s boring. Everyone says that about everything.

“But if you launch a product whose pitch is, ‘This product could do your whole job for you, but there’s a 10% chance it will email porn to your boss,’ more people will be excited to use it. The do-everything-for-you-and-maybe-ruin-your-life AI assistant is just obviously more psychologically appealing than an omnicompetent AI assistant that definitely won’t ruin your life. Aren’t you a little curious?”

I just wanted to read that into the record because the psychology that he’s describing there does ring true, and it will stick with me as I digest future product launches in AI. Does it spark any thoughts on your end?

I think, in general, all the AI rhetoric—I’ve always been pretty critical of the whole “it’s going to destroy the world” sort of thing. I still think Washington, D.C., is really screwed up in its evaluation and understanding of AI and chips and NVIDIA and China and all these sorts of angles.

In part, that’s because you had folks like Anthropic going there and convincing everyone this is a doomsday device, that it’s a one-time chance to control the world. I think that was wrong and false, and is actually causing us considerable problems—an ongoing sort of issue that is just out of touch with reality.

It still drives the—yeah, it drives the conversation to these bizarre places, like with the H20 being listed.

You had the Biden administration just destroying U.S. industry, as far as a lot of the chip stuff goes, and destroying the NVIDIA moat because they were convinced this was a one-time opportunity.

Right. Well, and there’s this notion that there’s going to be one superweapon that nobody else is going to be able to replicate, and if we don’t ban everything, then we could be losing everything in that scenario and ceding the future to X, Y, and Z. It’s just not really—

It’s not ceding the future. It’s that we’re going to magically control the whole world forever. I mean—

Yeah.

Like, I mean—

And it hasn’t been borne out over the last three years of AI progress and development at various places. It all just strikes me as kind of crazy, but you’re right: It has absolutely taken hold and informed how a lot of people think about these issues.

And this is a very frustrating mode of argument. The base case for AI—and where I’ve generally been—is that the impact over a smaller number of years is probably going to be smaller, while the impact over a larger number of years is going to be large.

Technology as a whole is going to be hugely impactful, massively improving productivity. It's going to eliminate a bunch of jobs. It's going to have all these sorts of effects, but in a way that we've seen technology play out before. People who are like, “I'm not saving any more money because the world's going to end in 2 years”—it's like, do you have any concept of how technology actually disseminates over time?

And again, the problem is, the people hearing this who believe that are like, “Yeah, but it's different this time.”

Mm-hmm.

And the problem is that they were saying it's different this time in 2020 and 2021. At what point do we get to decide, actually, no, it's playing out like we thought? We're going to get to 2035 or 2037, and they're going to say, “Look, the world's totally different. Told you.”

I'm like, no, I also thought the world would be different in 2025, 2030, 2037. I just think it's going to be this really steady change that we're going to acclimate ourselves to in a surprising way, just like we had this discussion on the podcast a while ago. The world's drastically different in so many ways, and we barely notice. That's sort of what happens.

Yeah.

And again, this isn't to dismiss what happens. It's just to say the method in which it happens and the way in which we experience it is different than popular conception. But people come out with these scare scenarios, and it's like, “Shouldn't we just in case?” It's like the whole COVID thing.

Mm-hmm.

It's like people came up with these spectacular means of transmission and harms—“What if it's the case?”—and then they convinced everyone to protect against this worst-case scenario that was basically a function of imagination.

Yeah.

And no one's doing an actual expected-value calculation, which is like, what's the— It's not just the potential harm of an outcome, but what's the chance that harm is actually correct and real? If it's a 0.1% chance, that harm has to be multiplied by 0.1%. And now these people say, “Look, the world's ending,” then that's more than 1. They have to think through, like, okay, what capabilities do I actually have to stop this? Is it actually even solvable or tolerable, or do you have to just accept the fact it might happen—

Mm-hmm.

—and manage all the other potential outcomes along the way? And what I feel has happened with AI, particularly in Washington, D.C., is very much similar to what happened with COVID. It's people with the largest imaginations got the most influence and control, and no one was actually considering what's the best policy for the most likely outcome.

Yeah. I mean, I can't argue with you. It does remind me that the AI executive order under the Biden administration was drafted after Joe Biden watched Mission: Impossible – Dead Reckoning at Camp David, as far as imagination is concerned. Wonderful stuff. A note that will delight me to no end until the end of time, so thank you to the Biden administration, who put that out favorably to show that Biden was taking the issue seriously.

And as far as the rhetoric from Sam Altman, it just cracks me up because it seems like every time there's a new model—

Well, Altman's bad, and Anthropic's worse. It's like they're—ugh.

Well, it's always like there's a tweet that precedes the new model by a week or 2 where Altman's like, “I don't know if we should even release this.”

Ah, it's so true.

“We're really worried.”

It's really aggravating.

“But here it goes.”

It's super annoying. Yep.

And it is very effective marketing, though, you know? It piques my interest. I can't wait to see ChatGPT-5. Let's check it out.

And it's really annoying because all these people who are doomers—it's like, to disprove them, you're disproving a negative.

Yeah.

Right? Could we get some metrics about what the actual—

Or a date. A date—

Yeah.

—where we can all measure if the world hasn't ended.

Well, it was 2025, I think. Now it's 2027.

Yeah.

Everything, 2027. What's the— It seems auspicious. That's like the date everyone's putting on for China and Taiwan, too.

Ah, yeah. Well, I think that's the Xi date.

A real big opportunity for the nothing-happens crowd to make some anti-2027 bets.

Oh, boy. Well, I'm going to enjoy the rest of my summer, and we'll cross the 2027 bridge when we get there.

7. The Tea App Breach

For now, we'll keep it moving with Joe T., who says, “It's surprising to me that Apple isn't taking any heat for the Tea app. The ethos for their App Store monopoly power is partially based on promises of protections and app standards. SaaS apps use SOC 2”— SOC 2 or SOC 2?

SOC 2. Sock two.

SOC 2. Sock two. Sock two it. “SaaS apps use SOC 2 and more to standardize trust.” Apple should implement something similar in the era of professional-looking vibe-code slop. What do you think, Ben and Andrew?

For context, for anybody who is enjoying his or her summer and is blissfully unaware of what the Tea app is, it was an app that allowed women to upload a selfie, sign up, and share their experiences with men they'd dated, men to stay away from, et cetera. The whole affair looked pretty dystopian. I didn't investigate any further than seeing some tweets.

But then in the past week, there were reports that the company inadvertently exposed the names, selfies, identity documents, and locations of thousands of women. And then on Tuesday there was a second report that the app's direct messages had also been compromised. So it's turned into quite a mess. Do you think Apple has a role, a responsibility to police something like this? How does it happen?

What, to police apps that are predicated on randomly slandering other people who can't—

Well—

—defend themselves?

I will say there was a similar website that emerged when I was in college that was quickly banned, like 15 years ago. I didn't know we were doing this again 15 years later.

You could choose any angle you want on this, and it's incredibly dystopian and—

Yeah.

—fairly horrible. And again, I'm not defending bad actors, to be clear. But the entire premise of this is not great.

Really, really crappy—

And what's the reason?

Yeah—

—and depressing, no question about it.

Right. So, yeah, I mean, it should— I think the reality is, I'm not going to get all worked up like Joe is because I never believed Apple's BS about the App Store in this regard anyway.

Uh-huh.

I've been calling it out for years and years, so I'm not going to— I'm already on my high horse. I don't have a higher horse to get on.

Well, I don't know if that's true. Every time we talk about the App Store, you say, “Well, they are very secure.” That part is true. They do work hard to ensure the security of your devices. I, and I—

No. No. No. What I say is, the fundamental architecture of iOS is secure in that apps are in their own sandbox. They can't interfere with other apps, which makes it fundamentally different than a Mac or than a PC or the way computers used to be architected. My point is, the security comes from the architecture, not from the App Store.

Not from the screeners.

I'm disputing Apple's characterization of the App Store. Yeah, I think it's BS.

Okay.

I think it's all BS.

So is it something that we should worry about, though, as AI lowers the barrier to entry for app development and we get—

We should have been worried about Apple's BS around the App Store for years. That's been my position since as long as I've been publishing online.

Okay.

So I see, this is just—

Do we multiply those worries?

—an affirmation. Yeah, exactly.

Now I sound like a doomer here, because it's—

Exactly. Yeah.

—suddenly going to be really easy to make a professional-looking app that is completely insecure.

Right. But the insecurity here is on the cloud side. It's not the— And so, do you want Apple going in and verifying someone's cloud implementation? We could go that direction, but that is what Joe's referring to. And to be fair, that's what Apple seems to suggest with their claims about the App Store. I'm just saying I've called those claims BS all along. The security comes from the way that it's architected on the system layer.

Yeah.

And in this case, no one's hacking into someone's app on an iPhone and getting this information. They got a publicly available URL to an S3 bucket in the cloud.

Yeah.

Well, fair enough. You talked about zero-trust security in the past. I have zero trust in any security. Anything I'm putting into my phone, putting into an app, is something that I'm comfortable with people hacking, because at this point I just don't really feel that secure in any of it. So, similar to you, nothing is going to change for me in the future.

Well, and it also speaks to— The other thing you can critique is the whole know-your-customer set of laws and regulations generally—

That conditions people to continually upload their identity and verify themselves and all these sorts of things. Nominally, to stop smugglers or counterfeiters or people or money launderers, whatever it might be. And the reality is, like so many things, what you're actually doing is compromising the security of everyone else who are law-abiding citizens that are forced to compromise just to live their life on a day-to-day basis.

Yeah.

KYC is, in many respects... Again, I'm not—there are, of course, legitimate reasons to know your customers, but it's like any trade-off. By focusing on one thing, you're trading off a bunch of other things along the way.

Yep. Well, we will see. Good luck to the good folks behind the Tea app. I'm sure it's been a very stressful last couple of days here.

Yeah, this whole thing feels like an overthrow, so...

Nothing but the worst of luck.

Everyone's getting what came to them. That's right.

Oh, God. James says, “Ben and Andrew, you all have had several thoughtful conversations about digital media recently, and that leads to a question. How has The Ringer lasted twice as long as the greatest sports website of all time, its predecessor Grantland?” Great question from James. I worked at Grantland, so it's near and dear to my heart. I didn't work at The Ringer. I still have a bunch of friends over there. Do you have any thoughts, Ben? Any guesses for James?

Yeah, Grantland—ESPN decided to stop funding Grantland, and then Spotify keeps funding The Ringer. I mean, like, it's the—

It's as simple as that.

Yeah. I mean, The Ringer, much more than Grantland, is very explicitly a podcast-first organization, and the website supports it. I think the website is probably there just because Bill wants it to be there.

Mm-hmm.

Bill still has enough pull at Spotify to make it happen. Spotify's there for the podcasts, which are successful. And this was, I think, a shift Bill made—

Yep.

After Grantland, where Grantland really was a website first and foremost, and one of the greatest websites of all time—

Yep.

Despite the fact that you were there, you overcame that—that sort of challenge.

Crippling handicap. Well, it's funny. At the end of Grantland, I was not all that familiar with your work at that point, but you were writing about Grantland's business and Bill's podcasting. The end of Grantland was very frustrating because I was on the inside as things were beginning to fall apart, and people were speculating about how much money it was losing, this, that, and the other thing.

It was the first time that the media was writing about something I knew a lot about, and I watched people repeatedly get the story wrong, like Deadspin, a habitual offender in that regard. But you noted that Bill should be monetizing the podcast first and foremost, and that's what ESPN just did not do very well—

Yep.

During the Grantland era. And I think the economics have become a lot more favorable at The Ringer, where five or six podcasts can pay for a lot of ambitious work elsewhere because they're actually doing a good job monetizing in that space, which monetizes much better than digital print media in the modern era. So I think that's part of the answer in terms of The Ringer's longevity versus Grantland's cursed fate at ESPN.

Yeah. Grantland is like the Kurt Cobain of websites: It will always be thought of as great and can never be diminished by virtue of going out early. It never had the chance to fade, as it were. It's a shining light forever and ever.

It also existed—

But it was really awesome for those who were there.

It existed at a time when everybody online was in a much better mood than they have been over the last 10 years, so that helps its legacy as well.

Absolutely. So it was one of my favorite articles I wrote, and it goes into some of my thinking around Stacker and Stacker+, this idea that you need to think about a media company in a multimodal sort of way.

What is text good for? Text is good for customer acquisition because it's easy to share, easy to spread. It's really bad for monetization. Podcasts are good for monetization. You have a captured listener, really hard to grow and spread. So you need to think about them together. You have people that get popular through text, and then they can be on podcasts. You need people that can fit that mold and can do that.

I wrote that about how ESPN was not thinking about Grantland the right way, how you had to have this multimodal approach. And that's actually how I met Bill. For sure, this was a core piece. I think it's fair to say this thinking went into The Ringer—

Mm-hmm.

And how The Ringer is going to be formed, with a much more explicit link between website and podcast and how they work together. Ultimately, it probably doesn't really matter to Spotify that much. I think The Ringer website is more a function of Bill wanting it to exist—

Yeah.

And so it still exists. But I think overall, the fact it lasts longer is because business models matter even more than product sometimes, or having a model that works is important to longevity, not just having a great product. Grantland was an incredible product that lost a lot of money.

Yeah. Well, and I would also say that Bill cares about the written word and cares about preserving the website, and he has a tremendous amount of leverage because Spotify has figured out how to make money on his tremendously profitable podcast and suite of podcasts at The Ringer.

Yep.

And so he's allowed to swing that weight around and keep writing alive for a couple more years here. Who knows whether anybody's going to be reading writing in 2035, but I've enjoyed The Ringer, and I'm glad it still exists. Not quite on the level of Grantland, to be honest.

No, it's not even close to Grantland.

It's not necessarily a destination site.

I'm glad it exists. It's not close to Grantland, yeah.

Yep. All right, closing out here, Andrew says, “I know Ben is reluctant to go all in on the endorsement game, but given how consistently helpful his recommendations have been, I'd love to see some sort of middle ground. The Above & Beyond Group Therapy playlist has carried me through many late nights building presentations. The Kiziks, big win with my father-in-law.

“I've worked in tech, so I'm not a total normie, but I still find myself asking the same questions I imagine many listeners do. Should I get the M4 MacBook Air now or wait for the M5? What's the best desktop setup, travel setup? And, yes, some of us are very much interested in watch talk, even if it is dangerous territory for host and listener alike.

“You've built trust by being rigorous and practical, but I think everyone would love to learn more about your product recommendations, your heuristics for buying, what's worth future-proofing on a regular basis, as well as having a place to find those recommendations.”

Thank you for the note, Andrew. As Ben's friend and co-host, I don't know what the next 20 years of your career look like. I don't know what the next 10, 15... Who knows? You could do any number of things. I do hope that your final job in life is life coaching and product recommendations, because I think that's what you're most passionate about in this world. I don't know if you have any response to Andrew, but I feel like you've been training for this for quite some time.

There was something I was talking to someone about. Oh, yeah, it was key chains.

Oh, God.

My key chain is—you want a carabiner? What is it, the clip you use for rock climbing?

Okay.

There are key-chain carabiner clips. Those suck. Don't use those. You want a real one that's actually strong.

Carabiner?

Carabiner. Is that what it is?

Yeah.

And then every individual key gets its own keychain that goes in the carabiner, so you can not only carry them all together but also hook it on your belt. I'm a big key-on-belt guy. Really great, real father-in-law territory.

Yeah, I was going to say, a real dad move there.

But also, you could take one key off super easily for something else if you need to give it to someone to use, X, Y, Z, and it goes back on. So, of course, I was in the U.S., and I got a new carabiner, an extra one. I was giving someone my other one, saying, “Oh, this is how I do my keys.” They come back and they're like, “This is such a better key system. I can't believe I never thought of this before.”

And I'm like, look, there's literally every single aspect of my life I've probably thought too much about and have my own system.

Yeah.

So all that is to say, yes, I think your recommendation is correct.

I’m happy to write your keychain article at some point. But I’ll have to stick with the big tech companies for now.

Right.

But if you want to drop the occasional question, look, here’s my key recommendation on how to carry your keys. I have—

That’s what the podcast is for.

—more where that came from.

There will be room on Sharp Tech for Ben life coaching for years to come. I assure you, Andrew, and anybody else who’s interested. I’m interested depending on the day.

No, you’re a big problem with this thing.

So the issue is, what makes you a good podcast host—

Mm.

—is sometimes even you might agree or be inclined to agree, but you’ll push back—

I’ll tease out—

—not just because—

—the devil’s advocate perspective, for sure.

Yeah, not just because it makes for good podcasting, but because you’re instinctually inclined to disagree.

Mm.

You are a contrarian. Speaking of Silicon Valley clichés, you are the biggest contrarian I know. And that can serve you well. It’s good for podcasting.

It has, I assure you.

It leads to some horrific life choices, particularly the longer you know me. You’re like, “I can’t listen to Ben for everything, so I have to disagree in particular areas to maintain my sense of self and independence.”

Yeah.

And I’m just like, do I need to watch the movie Inception and trick you? Do I have to start saying stuff I don’t believe so you start doing the right thing?

You know what? I actually do that—

It is some real friction in our relationship.

I do that to my 2½-year-old son sometimes. When I want him to do something, I’ll tell him to do the opposite of that thing, and then he’ll naturally gravitate to doing what I actually want him to do. So I’m glad to know that I’m the 2½-year-old son in this analogy. Context for this conversation is that Signal, @signal on Twitter, said, “Every millennial is now paying Apple $9.99 per month for 2 terabytes of data just to not delete their life.” I responded to this midday Friday in a group chat and said, “I actually got so upset when Apple tried to charge me $9.99 that I went out and bought an external hard drive and moved all my stuff onto an external hard drive just because I’m not going to pay Apple $10 a month for the rest of my waking life. I draw the line at a $10 tax for living and taking photos.” Then, 6 hours later, you log on.

I was in Taiwan, so it was delayed, yes.

So it’s Friday night. I’m sitting with my wife, and I’ve got Ben waking up in Taipei, just coming off the top rope about how stupid this decision is, how you could lose the external hard drive any day.

An external hard drive is not a backup. All you’re doing is transferring your data into a less safe device. It still only exists in one place, and now that one place can be—

Yeah.

—bumped or knocked or thrown away or—

It can be lost.

Yes, it’s—

It can be burned.

—and $9.99 is not a tax. It is a service. You are—

Mm.

It is insurance. There is nothing in your life that is more irreplaceable than photos of your kids.

Yeah.

And you can, by definition, never take them again. And for the low price of $9.99, you can have that—

$9.99 forever.

—taken care of immediately and backed up. It’s worth it. It’s not a tax; it is a service. And, yes, I’ve never woken up so quickly because I was so filled with disgust—

So much anger.

—and rage. First, at you for doing this, and 2, at myself, knowing that by coming from the top rope at you, I was just solidifying your obstinance and making it more likely you were not going to do it. So I was just filled with self-loathing, all at the same time. It was terrible.

Here’s the thing. I may be contrarian. I am open to changing my mind. On The Greatest of All Talk, I’m known as the greased pig, because as the facts change, I’m willing to change my mind. This one, however, you were so annoying on Friday night as I was sitting there. I will never be paying $9.99 a month for the 2 terabytes of storage—

All you’re hurting is yourself.

No, I’m not.

And someday—

No, I’m not. So this is the thing. As a millennial, there are so many freaking photos of my kids. We print photos of my kids every 3 or 4 months. One thing I can promise you is that 20 years from now, no matter what happens to my photos, the external hard drive, whatever, I shouldn’t have even bought the external hard drive. I’m not going to be suffering from any lack of photos of the early years of my children.

The problem is, fine, you make that choice for yourself. At least let your wife have the backup plan.

Well, she does.

Do you want to deal with her despondency when stuff gets lost?

Yeah. Look, my wife is nothing if not a—

And then get a family plan. The family plan’s very reasonable.

—a big tech adherent.

She is much more responsible.

She’s paying the $9.99 a month.

Yeah, I’m done. I’m going to text her right after this call.

No, I checked with her. Don’t worry, she’s got the extra 2 terabytes.

Okay, good.

Get a family plan.

I know.

It’s very reasonable.

Maybe I owe it to her to get a family plan, except—

That’s right.

—you were so annoying—

Do it for her. Do it for your children.

—over the weekend. I can’t possibly—

I know.

—cross that bridge.

Like I said, what made me most upset was knowing how annoying I was, and how this was going to be counterproductive—

Yep.

—but I couldn’t contain myself.

I’m dug in now. Sorry, Signal. I’m one millennial who does not pay $9.99 a month. But Ben, great to see you, as always. And we are coming back next week. People can continue emailing email@sharptech.fm. Again, we're going to be in the heat of August, so hit us with anything you've got, because God only knows how slow it's going to get over the next several weeks. But we'll be having fun on the podcast regardless. For now, go enjoy—

I’ll give one recommendation of podcasts. You have to come—

That’s right.

—up with your item.

Product recs, life coaching.

It’s a little risky, because who knows what we’re going to get? But that’s why the key thing was so funny, because I’m like, “Yeah—

Well, and—

—actually I do. I have a take on literally everything.”

Of course you do.

For better or for worse.

It’s unbelievable. It honestly is really charming. You enjoy helping other people improve their lives and optimize various aspects of their lives. What gives you the most charge of anything at this point in your life?

It’s so true.

So—

And it’s really rough, because you resent it. You’re not going to do it. It’s like a very dysfunctional relationship. You know what, though? Dysfunctional relationships make for a good podcast—

Good podcast.

—so I think that’s the real takeaway.

—great content. It’s all about content at the end of the day. All right.

This was why I wanted to hire you in the first place. You and Ben have a very dysfunctional relationship. Great podcasting.

Great content.

I love it.

10 years running.

That’s right.