（预告）春假来信问答：Sora 安息、广告与过剩、F1 反向行驶、Elon Inc.、智能手机育儿，以及更多

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, how are you doing?

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

Mid-vacation. I'm good.

Well—

It's great to be back to our roots here.

Vacation's overstating it. I've done a couple—I've done an update and an interview this week. It's a real stark contrast to you sitting on the beach a few weeks ago, completely taking the week off. Who's the grinder here? That's what I want to know.

Look, all I can say is I did it right, and you took the weirdest, most disjointed vacation I've ever seen, but I figured in honor of that, we would keep it light tonight and go all mail today.

Everyone keeps asking me how my vacation is. I'm like, “Does this really look like a vacation?”

It's not really a vacation. Let's just be honest. Let's call it what it is.

This is a very functional trip to Asia, and the days off are working around my travel. But I usually take a spring vacation every year, so I'm calling it a spring break. Yes, it's kind of ridiculous.

That's right. And next year I'll be right there on the first weekend of March Madness in your basement, preventing you—

That's right.

—from traveling anywhere next year.

We're going to do it next year.

Well—

Yes. I was annoyed that it was the same week this year, but I've already looked. My son's spring break does not line up with March Madness next year, so we'll see what we can do.

Oh, perfect.

This is what grinders have to do, Andrew. We have to plan like a year in advance, not like 3 hours in advance or whatever you do.

Exactly.

So—

A day before, I tell you, “I'm going on vacation. I'm going to be out all week. Deal with it.” But no, here we are, Ben. I gave you a couple weeks' heads-up on my vacation. At least 50% nonsense questions is the promise on this episode. We're going with a spring break mailbag today.

Before we hit mail, though, we do have to read this note at the top. A note from @soraofficialapp on Twitter: “We're saying goodbye—

Actually, I think the problem is right there. I saw one of the most devastating takedowns of a publication that I ever heard—who is it? Bob Lefsetz, the guy who writes the email newsletters about the music industry. I've referenced him before. Taylor Swift's song “Me!”—I know about him because he wrote a thing about how she can't sing.

One of the more devastating takedowns of a publication that I ever heard was when Recode went out of business. Bob Lefsetz, I think, is his name.

Yeah.

And he's like, “It was recode.net.” If you didn't have a .com—sorry, you weren't a business.

I don't take it seriously.

Now, I hesitate to say that to you because it is sharptechs.net.

Right.

Which I think is fine. I'm fine with a .net. It's a top 3 sort of domain. But if you are @soraofficialapp—

Yeah.

—as opposed to Sora, maybe you were not long for this world.

1. Sora Was Never A Business

Well, they weren't long for this world. “We're saying goodbye to the Sora app,” was the announcement earlier this week. “To everyone who created with Sora, shared it, and built community around it, thank you. What you made with Sora mattered, and we know this news is disappointing. We'll share more soon, including timelines for the app and API and details on preserving your work. The Sora team.” What you made with Sora—

I have a question for you. Did anything made with Sora matter?

Mattered. Fact-check on that particular statement.

That's exactly where I was going.

At least you can pop champagne mid-spring break over there because Vibes is still alive, you know?

Look, Vibes won. Just a lot of L's all around. I think I was proven correct, although this is probably one of those things where it's like, “Congratulations, you made the most obvious call ever.”

Sora wasn't necessarily it, although it did look sketchy for a while. It was at the top of the App Store. It was ultimately proved to be a novelty. Again, this is the guy who liked Vibes and then has never looked at Vibes in months, so I really do not get to claim any sort of W here.

You announced how much you liked Vibes about 12 hours before the entire internet was like, “This is the most antisocial product we've ever seen.” But hey, Meta has so much money that they're keeping Vibes alive, at least for the time being.

They will claim that the numbers are pretty good, which I don't believe. But yeah, Sora was kind of a—it's interesting because we talked about, or I think John and I talked about it at the time, watching this capability not in ChatGPT, but watching it as a standalone thing, with real community and social-network-type features.

Mm-hmm.

It was pretty audacious, particularly when you look back in retrospect. Stuff like this usually doesn't work. Facebook has launched a bunch of additional social networks that really haven't worked. The ones that have worked are the ones they just ended up buying because they came out of the ether.

So this is, in retrospect, pretty audacious, and it's interesting to consider whether it was a good or bad thing that they didn't just put Sora functionality into ChatGPT.

Mm-hmm.

I think the critique of why that was a mistake is they maybe never got critical mass. The good thing is this is a very bloodless killing.

Yeah.

I think the primary casualty seems to be Disney, of all entities.

That's right.

But if no one was using it—

That deal has been rolled back this week as well. It's interesting, though.

Right. If no one's using an app in the App Store and the app goes away, does it make a sound?

Well, this app scared the crap out of Hollywood. There was a whole frenzy. I'm sure everybody listening to this show remembers it, but Hollywood was all freaked out about copyright and what this could mean. Then Disney was forward-thinking for embracing it.

I do think, in retrospect, it reflects good product instincts at OpenAI. The choices they made and the way they promoted it were all good. It's just that the product itself isn't really a business for them, and they're now focusing on business.

Well, here's the thing. You could make the case that the reason Sora died is because they started cracking down on copyright. In the first few days, when it was going nuts, it was because people were doing stuff with recognizable figures.

I think we talked about this at the time, but maybe that is one of the biggest takeaways. It's a theory that I've put forward consistently. It's a reason why I feel optimistic about our fortunes in an AI world: things that people have in common—

Yeah.

—are going to matter and matter more and more. At the far extreme, there's live events. Live events are becoming more and more important, tremendously profitable, going up almost everywhere, in all cases except for NBA games for tanking teams—but that's another discussion.

It's a similar thing with these copyright bits. When you could do Pikachu doing crazy things, people found Sora very interesting.

Would we still be using it on a semi-regular basis? It probably wasn't going to be an everyday thing, but would we still resort to Sora to make videos featuring movie characters we all know? I think that's a good question.

2. Social Networks Need A Tool

The other thing that's interesting is that you go back to Instagram in the day, and Instagram started as a photo-filter app. On Instagram, when you would take a picture and apply the filter, there were buttons right there to easily share it to other social networks.

Mm-hmm.

It's become a very well-chronicled strategy. I think Chris Dixon wrote the canonical argument about this: come for the tool, stay for the network, or something along those lines. They were independently useful and then got permission over time to layer on the social-network aspects, and then also lifted the Twitter graph wholesale back in the day.

Everyone wants to bemoan the end of open APIs and open social graphs.

Well, there's a very good reason why stuff's not very open anymore: Instagram basically bootstrapped itself off Twitter's social network.

Yeah.

But for Sora, probably the most Sora videos people did see were on other social networks. And it's just really hard; maybe even having the social stuff at the beginning was a mistake. I mean, there was the Cameo bit, all those bits and pieces. I think probably the takeaway is this—

They were trying the right things.

It's hard.

Yeah.

Look, there have been so many attempts to build social networks, and the reality is the vast majority of them fail.

Mm-hmm.

And if you want to be a good prognosticator, if you want to have a good track record, just come out and say, “Every single one is going to fail,” and you're going to have a great record.

9 out of 10 times, you're going to come out—

Yep.

—looking pretty smart.

3. Compute Costs Decide Sora's Fate

The thing that was striking to me, though, is just when you hear OpenAI talk about focusing—

Mm-hmm.

—everyone's like, “Well, the Sora app's got to go.” It's not just that the Sora app needed to go. The amount of compute this appears to have been using, even now when no one's using it, feels astronomical. And that's, in the end, maybe the biggest critique of this. Which is—

Who greenlit this—

If you're constrained—

—if it was that expensive? Apparently, I saw $15 million a day thrown out. I don't know what the actual number is, but, yeah. I mean, it was obviously compute-intensive on day 1, and people had questions about how much compute was being dedicated to this project.

Yeah, I mean, this probably gets to the broader questions about where OpenAI's business is in general, but I don't want to bang on it too hard.

Yeah.

Again, you could get a very high batting average predicting everything's going to fail, and that's not a very fun world if no one tries because they assume everything's going to fail. So I will give Sora a parting salute.

Well, look, we're all living through the early innings of the AI era here, and there are going to be misses and weird experiments and all sorts of things that are charming in retrospect. Sora is already one of those things. I'm always going to remember that first week when everyone thought that Sora was taking over the world. And I do think if you were still allowed to use copyrighted movie characters and insert yourself into The Matrix, people would use it more often.

Right, or modify actual scenes. It had the potential to be the greatest meme maker of all time.

Exactly.

But that's interesting. It's like—

And that was always plausible. That was more plausible, just for the record, than it turning into its own social network that rivals Instagram or TikTok or Twitter.

Well, but if—again, to go back to the Instagram example, if you have heavy usage—

Yeah.

—then you have the ingredients to build a social network. One thing that is important to note, though, and this gets to the infrastructure-cost angle: to go back to the Instagram example, it was applying filters on your phone, using your local camera and your local compute, and then you were posting stuff to other social networks. The actual costs for the initial version of Instagram, on a marginal basis, were basically zero.

Yeah.

Everything was happening locally. You built the app. It was distributed through the App Store, so Apple's taking care of that aspect for you. It's running on people's local phones. Over time, then, you could layer on the—

Mm-hmm.

Now, the cost to run Instagram today is astronomical, right? Because it's basically a cloud. It's a front end for a cloud service, and everyone has their own customized feed, and there's obviously lots of video and all these bits and pieces. But it was a very natural progression from basically zero marginal cost to layering on. And as the value of the app itself and the network increased, the costs also increased, but it was a match.

Yeah.

And that is a fundamental problem with Sora and is going to be a fundamental problem for anyone that tries to do something similar. This just gets to why, from the very beginning, I've always been on and on about zero marginal costs—

Mm-hmm.

—and why that makes all these new things possible. AI generally—there's a whole discussion to be had about local versus in-the-cloud computation and the question of marginal costs. But for video generation in particular, we're nowhere close to a world of—

Yeah.

—zero marginal costs. But this almost makes the point that I make continuously about zero marginal costs in reverse. Not having zero marginal costs is sort of debilitating. And it ends up in an all-or-nothing scenario where either this takes off and is a huge deal, or within—

That's its own problem.

—a matter of months, it's dead.

Yeah. Right. Well, and if it takes off and it's a huge deal, we haven't really figured out how to make money from this. So how are we benefiting more than we're hurting from all the costs of all this product?

Yeah, maybe they should have had an advertising business in place.

Right.

Yeah. It really speaks to that being probably peak OpenAI, in terms of they could do no wrong. They would make a deal with a company, that company's stock would go to the moon, and it's incredible—like, that was only a few months ago.

I know.

So it's—

Most takeable company in history. It's all part of the ride here.

Yeah, it—

Well, speaking of which—

Yeah, go ahead.

4. OpenAI Chooses Enterprise

Adrian, who emailed last week saying that OpenAI pivoting to the enterprise was a mistake, followed up on our discussion and said, “Guys, how many great companies did well at consumer and enterprise in the first decade? Microsoft was able to dominate the consumer market not just because they were good at enterprise, but because they had an early enterprise monopoly and were a default. Will OpenAI be able to build a monopoly in the enterprise today against Microsoft, Google, Oracle, Anthropic, and Amazon, and leverage that to remain the default for consumers? In the consumer market, they only have to compete with Google, but keeping up with Google will take laser focus.”

Any thoughts on the follow-up from Adrian there?

Well, this gets to what is honestly going to be a theme for a huge majority of this podcast. I'm going to warn you right now: maybe this is the, quote-unquote, “Don't call it a vacation” laziness kicking in.

Okay.

But one thing—this is just a thing about AI in general. My answer is, I don't know. The degree of uncertainty is so high, and OpenAI is a great example. It does speak to this cost issue really screwing with things.

Yeah.

In a more traditional compute environment, which a lot of my analysis with OpenAI may be unwittingly incorporated, it's like, well, you have a billion users, or close to a billion users, or whatever it might be. Consumer is the great prize. You only get a big new consumer company every 15 to 20 years. There's usually one defining one—

Uber's like the last one to launch, really.

—per paradigm.

Well, I'd really go back to Facebook. Facebook is sort of the category that you—

Mm-hmm.

—you sort of have in mind for OpenAI, and that is the biggest prize in the world. And part of the whole thesis around why they need to be monetizing, why they need to have an ad model and all these pieces, is because then that gives you this revenue flywheel that no one can compete with because it's so hard to get consumers. It's even harder to build a thriving advertising business.

But they seemed to have such a lead, and they had so many consumers. It's like, surely they can figure this out, and once they figure this out, then they're getting revenue from this model that they feed back into their R&D, and they can push further ahead.

Yeah.

And it sort of becomes an insurmountable lead. I think I probably made this case most clearly around, I want to say, 20 months ago. I think it was at the MoffettNathanson conference or something along those lines, talking about this flywheel and why that would give them this big advantage. And it's so frustrating, to the point of being a takeable company. We don't know if that would have worked—

Yeah.

—because they were so late to get started. And now we're in a situation where, for sure, the best way to make money in the short to medium term, where you can just really turn the spigots on—

is in enterprise. We talked about it last week. Enterprise is willing to pay for productivity gains, and what they have with Codex is very compelling. And yes and no, I should say: it’s not a monopoly.

Yeah.

It’s tooth and nail with Anthropic. Obviously, they’re going to be fighting with Microsoft. They’re going to be fighting with all these folks, but at some point, the computation costs are what they are. You have to pay for these servers.

5. Microsoft Built Consumer Reach

When I think about Microsoft succeeding in the consumer space, a lot of that was because they did have an airtight monopoly on the enterprise space, and so everybody learned to use computers by using Microsoft at work. That’s a huge advantage as you’re trying to get a foothold in the consumer space.

OpenAI’s story is going to be more complicated than that in terms of enterprise adoption because I’m sure there are going to be a lot of different tools used at enterprises over the next couple of years here, and Anthropic does have a real share of that market. How big of a share is anyone’s guess because these are private companies.

Yeah, it is anyone’s guess. I got a lot of pushback when I shared that Ramp data on Anthropic that we talked about before.

There’s a story in The Information this week, too, that Anthropic is basically—when they’re talking about their skyrocketing annual recurring revenue numbers, that’s actually the total that people are paying. It’s like with e-commerce: what do you calculate? Gross merchandise value, I think—GMV—which is how much you pay for all your items.

Okay.

Do you calculate that as your revenue, or do you subtract away the cost of the item itself?

Hmm.

In Amazon, actually, one thing that’s challenging with Amazon’s business is that traditional Amazon is in a wholesale model. It’s like a retailer. They actually buy the items and then sell the items.

Yeah.

For that business, Amazon’s revenue is the whole item, and they have a very large cost of goods sold because they actually had to buy the item from whoever made it.

Whereas in the third-party business, Amazon’s revenue is much smaller because they’re not including the cost of the item in their calculation. It’s a third-party merchant that’s selling it.

Right.

All they’re doing is their skim on top, as far as their percentage of fees.

But they’re not assuming any of the risk in that scenario either, right?

Right. No, exactly. But as far as understanding the company, what you want to know is gross merchandise value, which is the total value of everything they’re selling. That’s a very different number from your actual meaningful revenue, which is what’s actually coming into your company.

Mm-hmm.

This report in The Information says that Anthropic’s numbers are basically the GMV version. It’s like everything.

Yeah.

Whereas OpenAI’s numbers are the percentage on top of what’s going to the clouds. Who knows? To your point, them not being public companies is frustrating.

But they’re different methodologies.

It’s easy to get caught up in these details.

They’re stated as different methodologies.

They’re different methodologies, but trajectories are real. I don’t think anyone denies that the Anthropic trajectory is more vertical at this point than maybe the OpenAI one, even if it’s possible the OpenAI lead is larger than you might think.

Yes. Well, both of them may not be private companies for much longer. We’ll see whether they—

Yeah.

—go public in the next couple—

That’d be good for us.

—quarters here.

Well, it is actually the public-company concept working as it should. It just had to get as large as these companies are. They need money.

Yeah, exactly. Uh, Richard said, "I have a take for your review." All right, and that is the end of the free preview. If you'd like to hear more from Ben and me, 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.