E165: Vision Pro: use or lose? Meta vs Snap, SaaS recovery, AI investing, rolling real estate crisis

25:14So Snap is more egregious. The difference is that Zuckerberg is listening and Spegel is not. The reason why Snap is doing poorly is not because its revenue has deteriorated. So I looked up, let's put this back, I asked chat GPT for their key metrics. So assuming GPT is not hallucinating, if you compare 2021 to 2023, their total revenue went up from 4 .1 to 4 .5 billion and gross profit went from, call it 2 .4 to 2 .5 billion. So not a huge increase, but revenue and gross profit were slightly up. But if you look at operating expenses, they went from $3 billion to $4 billion a year. And that is why their operating income or operating loss went from a $700 million loss to $1 .4 billion loss in two years.

26:10So that's the source of the problem is that they increase their operating expense by a billion dollars a year from 2021 to 2023. It's pretty simple. So they seem like they're the last ones to get the memo. Yeah, they were the last ones get the memo and just just finished the point. So you saw that a few days ahead of this quarterly announcement where their stock got crushed, they put out a press release saying they're going to cut their head count 10%. It's too little too late. Yeah, they knew, right? They knew that a problem. So they released the press release saying, oh, we're going to cut. Well, you should have done what Zuckerberg did.

26:47You know, Zuckerberg did a 20 % cut last year. He got serious. So you got lean and fit. And instead, these guys held out, did nothing. Then when they know that the markets can accruciate, they put out this lame announcement 10%, no, not 10%. Really, if you just wanna get back to where you were two years ago, in terms of operating expense, you need a 25 % reduction. Yeah, but it's more than that. If you look at the numbers, let's use operating cash flow. It's 165 million for SNAP for the quarter. So their operations generated $165 million a profit, but for the entire year, because they lost money in the quarter as prior, they generated free cash flow of only $35 million.

27:30So the business net produced $35 million of incremental cash. You know how stock based comp accounting works? The charge happens when it vests. So this is what employees are vesting. During the year of 2023, employees vested $1 .3 billion of stock based comp. So that means new shares or options were issued that on an accounting basis the options are valued using black sholes and the shares are valued based on the share price So they issued 1 .3 billion a stock based comp so they generated 35 million of free cash and They used 1 .3 billion dollars to compensate employees beyond their op -ex So that means that they paid employees 40 times the free cash flow that was generated for shareholders during the year which is also equivalent to 10 % of the enterprise market value of this company.

28:18So the enterprise value of the company is $15 billion. 10 % of that was issued to employees to compensate them. Now, let me give you the story of another city, Meta. And by the way, Snapshare Count, because they issued all the stock, the number of shares outstanding increased by 4 % during the year. During the year Meta's number of shares outstanding decreased by half a percent, because they used cash to go and buy back stock. So they were able to reduce the shares outstanding. Now, as you guys talk about Metacut employee count by 22%, and SnapCut employee had counts by 3 % during the year. But here's the crazy difference in performance.

28:55The stock -based comp expense for Meta during that year was about $14 billion that vested that year. That company generated 71 billion of operating cash flow. So while SNAP gave employees 40 times the free cash flow, Meta gave employees about 20 % of the free cash flow. Then Meta went around and they used some of that extra cash to buy back $20 billion a stock. They bought back more shares than what the employees were issued back that year were. It showed such a difference in looking out for shareholders. So if I'm an investor, and by the way, Meta is creating a 25 times free cash flow, which all the new businesses that they have in Lama -2 and the progression to cloud and other things that they might do.

29:41If I'm looking at those two businesses as a shareholder, you got this guy that controls the whole stock. He's giving employees a billion three of shares a year when he's only making $30 million of free cash flow a year. And then the other guy is issuing $14 billion of shares, buying them all back, and he's making $70 billion of free cash flow a year. I don't know. It's very hard to decide which one to go ask. He's being brought it up in an interview I saw in a lot of the layoffs were top heavy. So he got rid of a lot of the top people who had these huge comp packages. And then what I'm hearing from a lot of executives is cutting these highly stock comped executives who are also have big cash comp, cutting them, putting lieutenants in charge and then moving more jobs to other locations where people don't expect stock based comp.

30:29If you're in India, or you're in South America, whatever, stock -based company is not like the obsession it is here. As everybody optimizes these businesses, Facebook, even if you didn't. Why did they eat 5 ,000 employees? They announced roughly 500 job cuts out of what, 5 ,500 employees. That's crazy. I mean, should that company be operating with 2 ,000 employees? It's a good question. I've actually long cut the number of Twitter employees from 8 ,000 to 1500. When you look at the number of apps that they're running and the number of products that they're running compared to Meta, Meta has far more apps, far more infrastructure.

31:06Meta is serving 3 .2 billion daily active users' snap is about 400 million. So Meta is 8x the users with many more applications and much more infrastructure. So I think it's another great kind of ratio to look at the performance of these two. 12 times. I think you're exactly right. Exactly, yeah. The other advantage that Meta has is because they're so profitable, they have the resources to go big and AI. Big time. Which is very expensive. So yeah, so they are the leader. You get all this option value at Meta, which you don't get at Snap. There's all this infrastructure that they can leverage, much like Amazon did with AWS, into things like cloud AI tools for third party developers, third party applications.

31:49And then obviously the, you know, So Meta is the biggest advertising platform next to Google in the world now. And there's much more that they can start to do to extend further into the platforms. They did get them not some say remember Apple screwed them and was like, you can track devices now. And like that just took a massive hit in the ad network and was all those headwinds. They were like, okay, we're just going to use AI to optimize ads. And supposedly the AI optimization of ads I was talking to somebody on the inside. They said like, yeah, we got it all back. We gained it back. We've got massive AI advertising optimization going on.

32:25So yeah, that's great that Tim Cook You know kicked us in the nuts, but we don't care. By the way, that's a great point, Jay -Tal It really says a lot about how meta was able to respond to that change Which a lot of people speculated would destroy the advertising business and the fact that they were able to engineer solutions to drive advertising revenue up to $40 billion. It's just mind blowing It's a really kind of impressive outcome for the team and I think it speaks a lot to the quality of the engineers there. I think it's a great point. SACs you tweeted that you're seeing a little SaaS bounce back all of a sudden.

32:58That's interesting. I am seeing something similar. Last year or last two years, you had a ton of people cutting their SaaS spend maybe removing the number of SaaS vendors they had, consolidating vendors. You tweeted many public and private software companies are experiencing accelerating growth after six to seven quarters of deceleration. session appears to be over according to the SAS master, David Sachs. Do you want to unpack this for us? What do you see? Well, it's still pretty early because not everyone's reported. But if you looked at the big tech cloud performance in Q4, you could see that there's a bounce back in here.

33:34This is NetNewARR added for AWS, Azure, and Google Cloud. So you see here in Q4 that there's a huge increase in NetNewARR for the big cloud computing platforms. And then I think another bell weather is that last year. So we're still waiting to hear from HubSpot Sales for Zoom, Adobe, companies like that. They haven't reported yet, but if you look at it last year. And last year, Nate, Jira amongst other products, they're based in Australia, the major. Yeah, exactly. Collection of SaaS companies, right? It's a collection of SaaS products. Yeah, so Net and UARR would be the amount of growth in that quarter.

34:13And this is on a year over your basis. So you can kind of see, Q4 of 21 was the absolute peak and then plummeted and then it actually went negative for about a year. That's tough to be in a company with a new ARR going negative. Yeah, that doesn't mean by the way the company shrinking. It just means that the amount of net new ARR, which is the amount of growth is actually smaller than that same quarter of year before. Yeah. And then in Q4, you could see there's some acceleration here that they're starting to add more, they added more net new ARR. I guess 33 % more in Q4 than they did over the previous year.

34:55And part of that, SACs is because the comps are lower and they kind of bottomed out. Yeah. They bottomed out and now they're re -accelerating. So we're starting to see this in some of my board meetings as well, where in 2022, everybody was missing their numbers and re -forecasting down. and then they would miss the reforecast. So by 2023, the forecast were very, very conservative. And I would say, I'm now seeing companies beat the sort of the lower forecast in Q4. This wasn't happening earlier in the year, but finally I think people are starting to beat their sort of their lower forecast for Q4.

35:33That's the question that I was curious about. What do you actually think is happening? Is that we've rebased line these businesses. So now what would have looked like just a massive miss over the last two years now look like a beat because we've just completely reset expectations. Is it that or is it that the economy is actually expanding and we can count on some reasonable growth rates? Is it a combo with the two? What do you think it actually is? Yeah, I mean, it's definitely a new baseline in the sense that, and you can go back to 2020 or 2021, we considered good growth to be, you know, two to three X year over year.

36:13And now if it's going from 60 to 80 percent growth year over year, you're happy. So there's definitely been a lowering of expectations. That being said, you still see in these numbers, there has been a bottoming out and we're We're starting to now grow from this new baseline. So for example, I think with Atlassian here, we are seeing an increase in spent basically in growth, right? So the way our recession is typically defined is two quarters of negative growth, right? We had six to seven quarters of decelerating or negative growth. In SaaS, in tech. In SaaS, which is why I called it the SaaS recession or B, it's actually kind of a depression right.

36:56But now we're seeing quarter over quarter growth. So growth is re -accelerating. Growth is higher than it was. So is it going to get to where it was? That probably will take some time. But it feels like the problems in the ecosystem work themselves out and now we're back to growth again. Yeah, I can add psychologically because I'm on a couple of SaaS boards as well. And psychologically, it felt like you tell me if I'm right, SaaS, SaaS, or you saw the same thing. There were two years of calling up customers and they were like, we're consolidating vendors, and by the way, we did a RIF, and so we need 20 % less seats.

37:28So we're gonna have 20 % less SaaS companies that we're buying from, and we're gonna have 20 % less seats. So you start putting that all together. Man, everybody was just in psychological triage mode. We cannot spend money, I don't wanna lose my job. So if you're a procurement person, you're the CTO, you don't wanna lose your job, you don't wanna have more cuts. So you're like, well, I can cut some software costs. Do I get points for that? And the points you would score for the last two years were cutting costs. With the market ripping and you now got a really efficient company, you're like, hey, can we spend a little bit on SaaS to make the remaining employees even more productive?

38:07Okay, maybe that's a reasonable discussion. And then people are playing ball in terms of negotiating prices. So that's the other thing I see is people are like, we'll take your software, but here's what we want to pay. And then they're coming to the board and saying, can we do this deal? Well, it would have been a million dollar deal, but it's a $200 ,000. It was a get, take the money, take the money. Let's bear how to that customer. The market is generally an escalator on the way up and elevator on the way down. So the recovery is going to take a long time, but at least we've bottomed out and we're in recovery as opposed to continuing declines.

38:38By the same token, if you're a startup and you're not seeing improvement in your Q4 sales, then you no longer have a macro excuse for why you're not doing well. interesting. And then freeberg, you added, you know, you're like, I'll make my own software. You said, you know, some software is too expensive. I'll put a developer on it. And so, how's that working out for you? Are you still in that mindset of like, yeah, maybe we just build our own software? Yeah, I mean, it's not just us. I think we're seeing a lot of companies pursuing this path. A couple engineers can rebuild the functionality of core applications, particularly because I think If you think about the business model that makes SaaS so great is they could value share.

39:20Rather than charge the cost of an engineer plus some margin, the great business model, the equity value that comes in software, if you can build something once that creates $100 of value, you can probably charge your customer $30, $40 for that product because it's saving them $60, $70, and they'll make that switch to software. So, you know, the ROI -driven value share model in SaaS has made it incredibly valuable. The problem now is that an engineer can be hired to build the replacement and so it creates price compression. So the SaaS company can no longer capture that much value because the savings is actually less than that because the enterprise might say, hey, I'm going to hire someone and instead of spending 60 grand a year on your software, I'm going to allocate a quarter of an engineer's time to build that software and it's going to replace that cost.

40:13So I think that that's still the case. So while there might be bookings, there's still, which are driven largely by a search for efficiency gains, a search for more profitability, for more productivity within an enterprise. There are other options for that enterprise to realize that productivity gain today. And that's what's going to cause perhaps price compression and more competition than has been the case. But I don't think that the adoption of software is going to slow down. it certainly seems to be accelerating, which is great. More competitive, right? We're moving into a hyper -competitive market, right?

40:45Especially with AI. It's a mix of internal software. It's a mix of internal software and tools. As you guys know, there are very few traditional non -tech enterprises now that don't have a software team that can write code. So now that so many companies have software teams at write code, they're all going to be asking the question, should we be buying the software, or should we be building something internal? Yep. It's a classic buyer -build situation. All right. Let's talk a little bit about VCs and how they're investing in AI. There seems to be three camps shaping up here, Chema. One group is like being compensated for going to win.

41:15Microsoft Google Amazon, everybody, they're going to win the day. So they're going to wait and see. Then there's another group who's sitting at App because they're like, hey, open source is going to win. Metas committed to open source and collaborative platforms I've been playing with hugging face with Sandeep as well as you Jamoth and it's pretty amazing what's happening over there. And then a bunch are obviously placing bets right now. The valuations are absurd. Founders fund and in Dresden Horowitz, two notable firms are approaching it differently. Founders fund bought into OpenAI at a $29 billion valuation.

41:51But aside from that investment, they're generally avoiding the AI deals. My other hand in Dresden is betting heavily character AI, Replet, 11 Labs, You're also in Riplett, Sachs. So, what do you think? Is open source going to win the day? You've been fixing channels the whole way. You've been talking about compression. Maybe this isn't actually a good market. What you're thinking as a capital allocator, Juma? I think foundational models will have no economic value. I think that they will be an incredibly powerful part of the substrate, and they will be broadly available and entirely free. Wow.

42:28So if you think about that, any closed model, especially a closed model that operates on the open internet is not very valuable. And any open source model that operates on the, that trains on the open internet will make that so. So in that world, things like Mistral and Lama will essentially decay the market to zero. So if you, if you're looking at any economic value that has been captured up until today, if it has been captured by having a proprietary closed model trained on open data, that economic value will go away. And I think Google and Microsoft and Facebook and Amazon and all these startups have a deep economic incentive actually to make that so.

43:19So now you can evaluate what that means. So if you get an open model from hugging face that's just kick ass. Where do you spend money? Well, you're going to have to spend money to actually train it, to fine tune it, maybe to have some pretty zippy inference. And all of that means that there's a new kind of substrate that has to be built, which is all around the way that the tokens per second are provisioned to the apps that sit on top of the model. What that means is you need to go back to 2006 and 2007 and say, Okay, when we first created the cloud, who made money? And fast forward 18 years later, it's the same people that are still making money.

44:02So the people that made money in 2006 and 2007 were Amazon, principally, because of EC2 and S3. The perfect analogy of EC2 and S3 in 2024 is the token per second provider. Now, there you have to double click and say, Okay, well, what does a token per second provider need to do to make a lot of money? And I think the ultimate answer is you need your own proprietary hardware. So who is in a position to do that? Amazon has announced that they have an inference and training solution. For training, Syribus has announced a pretty compelling solution. Google obviously has TPU. Then there's a handful of startups, including one that I helped get off the ground in 2016 that I funded called GROC.

44:46All of those companies are in a position to build a tokens per second service. Then you have companies like Together AI, which basically just go and take venture money and wrap Nvidia GPUs. And you can debate what the advantage will be there. One could say, well, it's not really a huge advantage over time. So my refined thoughts today are sort of what my initial guess was when we started talking anybody I a year ago, which is the picks and shovels providers can make a ton of money. And the people that own proprietary data can make a ton of money. But I think open source models will basically crush the value of models to zero economically, even though the utility will go to infinity, the economic value will go to zero.

45:34Did any of you guys see Chamoff's interview with Jonathan Ross? Nope. Nope. You put it out, right, Chamoff? You made it public? You know, I did it just from my subscribers, but Jonathan is the founder. and CEO of GROC, the company that I just mentioned, and the quick version of that story is I would pour over the Google earnings results in the mid teens of 2000 because I was pretty actively investing in a bunch of different public equities. And Sundar said in a press release, he mentioned that they had rolled their own silicon for machine learning called TPU. And I was like, what is going on that Google thinks that they can actually roll their own silicon.

46:13What must they know that the rest of us don't know? And so it took me about six or nine months, but through sunny, I got introduced to Jonathan. And then we were able to get Jonathan to leave Google and he started and he, Jonathan was the founder of TPU at Google. And then he started rock, which I was able to lead that funding round in 2016. So eight years ago. Anyways, I did a space with Jonathan talking about the entire AI landscape and AI acceleration to my subscribers, but it was so good. I got to say he was so impressive that we kind of like figured out a way to just play the space and tape it and then we published it to everybody.

46:58So it's on my Twitter for anybody that wants to listen to it. It is. It is amazing. He is really impressive. I was sitting on the 17 going to Santa Cruz not moving for hour and a half and I've listened to it So I kept me alive, but I thought it was really great. Yeah, he's great. No, he's great I have some great insights and I think he's very compelling in arguing why some of the big cloud providers today that are offering infrastructure for AI Model training and inference are going to be challenged if someone can build full stack and do it successfully. So it was a really good interview. I actually think it's really worth listening to, but I enjoyed it.

47:45Yeah, thanks for putting it out there. I was like literally just sitting in the car, browsing the computer, and I saw your thing and I clicked on it and then I just said, it was a little hard actually. When you do a space for your subs, you can't actually just flip a switch and then release he said to all of your followers. So we actually had to like literally play it and then just capture the audio out and then republish it. But anyways, despite that inconvenience, if anybody's interested in learning about AI hardware, he is very compelling and he's very educational. So Saks your thoughts on just how you're approaching investing in AI, if you're specifically investing in the underpinnings of AI, picks and shovels, you have to add it, or if you're just looking on the application level and it's, That kind of approach.

48:30Well, we divided it into three categories. One is the models themselves, the foundation models, which can be either open source or closed source. There's infrastructure. So like Timoth's saying, it could be like model training. It could be vector databases. Tools that developers use to create the AI stack, typically inside their enterprise. and then the third would be applications, which can be things like co -pilots or it could be a pre -AI app that's using AI to kind of turbocharge its capabilities. Yeah, most SaaS would be in the Application bucket and so that's principally where we're focused.

49:10Although we do look at Infrastructure plays and models. However, I do think there is an argument for I mean really with the question of Commodization will like all the model companies just get totally commoditized We're really we're talking about open AI right because we're the leader so the question is can they maintain their lead? I do think there is an argument that open AI will stay in the lead and actually do quite well and I think there's a few points there one is That if you're a consumer you just want to use the best GPT you want to use Google It's just like search, right? If Google is a little better or the perception is, it's a little better than being or the other search engines, you don't win a plurality of search traffic.

49:57You actually end up winning it all because consumers just want the very best one. So, most of the tests show that OpenAI is still ahead of the open source models and I think even people in the open source movement will tell you that OpenAI is called six months ahead. They have no doubt that open source will get to where open AI is now in six months. Nonetheless, if open AI just maintains a little bit of a lead over open source, then it could compound. Yeah, it could basically win the vast vast majority of the call of consumer search or consumer GPT market. So, that's point number one. Point number two is now that open AI has these hundreds of millions of consumers using it.

50:41That's a pretty attractive audience for developers to want to reach. OpenAI has done a really good job creating a platform for developers to create what are called custom GPs. Most developers don't want to go through the hassle of training a model, fine tuning a model, doing all of that work that you have to do in the open source ecosystem.

51:10They of data or documents, information, have it learn what it needs to learn, fine tune it in that way, maybe add some lightweight functionality using OpenAI's platform to create a custom GPT. That's what I think most developers want is they just want a simple stack to work with. And they're gonna prize, again, simplicity and the power of the developer tools over the theoretical control they get by rolling their own models, training and fine tuning their own models and open source. And so I think what you're seeing now is, I mean, how many custom GPs have already been created? I mean, that's it.

51:46It might be tens of vast. I mean, there's so many millions. Maybe a million. Yeah, it's so easy to create them, man. So you have a classic developer network effect where you've got OpenAI aggregating hundreds of millions of consumers because they perceive that chat GPs is the best. Then you've got developers wanting to reach that audience. So they build custom GPs on the OpenAI platform. that actually gives chat GBT more capability. And that's something that open source can't easily catch up with. So actually, just first the point. So it is a flywheel where a classic operating system developer network effect where you want to use the operating system that is the most programs written for it.

52:27Yeah, and interestingly, Hugging Face has realized this and Hugging Face released this week their own version of Gpt's, which is really interesting. and you can pick SACs, which opens your project you want to use to make it. So unlike GPDs on chat, GPD, we have to pick theirs. On the hugging phase one, you could pick Lama or whichever one you want. There's an account called artificial analysis that you can follow. The thing to keep in mind SACs is that for any of this to be true, these APIs need to be usable, right? I mean, I don't know if you remember, but when we were building apps even as back as the late 2000s and early 2010s, one of the things was there was a pretty important paper that was published by Google about attention span.

53:11And it would look at page load times in a cold cash environment, right? And it basically said you have to be at like 150 milliseconds, right? That's like best in class performance or faster. And I remember when we read that at Facebook, we went crazy so much so that at one point, A small team and I kind of actually launched a stripped down version of Facebook to compete with Facebook. If there's a Nick, you can probably find this article on TechCrunch and we did it without telling you it was called like Facebook zero. Anyways, the point is speed matters because in the absence of having very snappy response, you could have the best model in the world.

53:47But if it takes 10, 20, 30 seconds to basically initiate and get back data from a fetch request, it's an impossible thing to do. So I think one of the things that you have to keep in mind is that there are these two things that need to move at the same time. One is the quality of how the model is, but two is the speed and its responsiveness, which is a function of, again, hardware and your ability to basically tokenize tokens per second very, very quickly. So that developers are incentivized to not just play around in a sandbox, but to actually build production code. And I don't think we've seen that second thing happen because nobody is delivering it.

54:22And that's the big thing that nobody talks about. For example, like AWS, if you look inside of how expensive it is to build an app there, I've tried, even when they give you credits, the credits they give you aren't sufficient enough to even pay for half the power. And then the way that they schedule and the way that they try to orchestrate you to use hardware makes building production apps unless you are willing to spend millions and millions of dollars for a very slow app, unfeasible. And so if you go back to a startup economy raising money here, the venture investor should start asking the question, well, what is the speed and usability of these services that I'm funding?

55:01And the reason is because you could build the best experience in the world that runs on local host. But if all of a sudden you actually try to launch it as an app and the thing takes 35 and 40 seconds to generate something, it's DOA. And I don't think enough people ask those questions or understand that that's true. So this is why I think you have to sort of be looking at both of these two things at the same time, but this account is Interesting because it kind of just strips things down to the bare facts and they start to allow you as a third party to understand what you can do Yeah, speed is just such a critical Componer this and what Google found was as you know free burgers you were there Every time they lowered a certain number of milliseconds, the usage went up, right?

55:46People did more searches, which makes sense if you get your results back faster. Yeah, it was a key metric from day one at Google, Marissomair. She ran all the consumer -facing products at Google during this, you know, earlier era. She was like, beat it into the team. I mean, if you guys remember, one of the first, the first kind of early feature of the Google results page was the amount of time it took to load the results. They'd show you how many milliseconds it took. Yeah, they'd show you that. Yeah, they literally put your North Star metric exposed to the consumer, which that must a little fire under the asses of all the deflated person server people.

56:19Yeah. Well, I mean, they were kind of showing off the quality of the infrastructure and the way they did indexing and everything. But the result really played out in usage, the faster the results, the more frequently you would use the search engine and the more likely you were to come back. And it's amazing how much consumer behavior drifts based on milliseconds. You have a few milliseconds of the way. I mean, if you look at the ever -shedded movie, the founder, where they explain the McDonald's process, they learned it too. Guys, look at this. This is really interesting on this analysis. I mean, Tremoth, are you saying that you don't think OpenAI can achieve the necessary levels of performance?

56:54No, I'm saying two things. OpenAI is three different businesses. OpenAI has a closed model that's trained on the Open Internet. I think economically, it's going to be very hard to sustain that unless they start buying all number of apps So that they can get some fine tunes that they control that are proprietary to them. So for example, if OpenAI were to buy all of Reddit, that would be a really interesting development that would improve the quality of OpenAI in a unique and differentiated way, relative to where things like Lama and Mistra will get to at the same time, as well as X's GROC. I think they're all going to converge to the same quality in the next probably 12 to 18 months.

57:34That's point number one. your belief there is there's enough data in those pools that everybody reaches parity. No, did you guys okay Nick did you so I published this primer on day primer? Yeah, there is a slide in there Nick that you can pull out but it just shows you that there is a converging in the quality of the results as the number of the parameters of the model gets higher and higher and what it effectively shows you is that we are already in the land of diminishing returns when models are trained on the same underlying data. So if you are using the open internet, Lama, Mistral, OpenAI, they're all getting to the same quality code point and they will be there within the next six to nine months.

58:17So that's business number one on OpenAI. Business number two is a consumer facing app called ChatGPT. That has a lot of legs because I think people are, you know, develop habits. It'll be very sticky and I think it'll get better and better. And then the third business that they're in is selling enterprise services to large fortune 500s. In fact, if you look at their open AI days, what they talk about is they sell, they've sold already to like 94 % of the fortune 500. What does that mean? I think what that actually means is they've sold a lot of test environments in sandboxing, but again, in order to translate that into functional production code that's used by Bank of America, right?

58:57Or Boeing in production, you have to have zippy, zippy fast SLAs and a level of performance that no cloud provider yet has delivered, none, nobody. So Nick, if you just go to that, please, the thing, I just wanted to show you this because it's really interesting. Sure, this is not mine. This is theirs. If you look at quality versus price It starts to show you like where do you want to be? You want to be in the upper left quadrant in their analysis, right? And so the point is what you can see is that a ton of different models are getting to this same place. And so obviously you'd want to use the model that's the cheapest or most convenient.

59:40Well, who's gonna pay for that? If you and your LPs want to pay for that, that the person that figures out the way that it's the cheapest to give you the same answer will actually end up winning because you will run out of money and they will not. I don't know. I mean, I think that there's a lot of business problems inside companies where people just want to very quickly set up their own, again, custom GPT without having to go through the time, the cost, the hassle of trying to do model training or fine tuning. So let's just back up. Here's the path that OpenAI is on. So step one, get hundreds of millions of consumers using it and getting them to view OpenAI or ChatGPT as the Google in this area.

1:00:25Right? Strong presumption. This is just the one you go to when you have a question. Step two, these same people, these same consumers now want to use ChatGPT at work because there's some research they want to do. So OpenAI has just rolled out both enterprise licenses and team work spaces. So you can work collaboratively on the same queries in a team work space. Step three is a rolling out a very easy -use dev platform that allows developers to again create custom GPs by just pointing open AI at repositories. Okay, and so let's say that you are the customer support team and you want to create a GPT to help customer support answer cases.

1:01:13You could basically then train Chad GPT on let's say every customer support ticket and email that the company is ever produced, right? Now, you could wait for the company's IT department to get us hacked together and figure out how to train an open source model on the same thing, but do you really want to wait for that or do you just want to get going? And now OpenAI has given you the enterprise license that you need to pacify the concerns about security and privacy and all that kind of things. To some degree, there's always going to be those superparanoid like Fortune 500 companies that will insist on doing, on owning everything and doing it, doing it open source.

1:02:01Let me build on your example. So I run a small software company during the day called Hustle and we saw a lot of tickets related to this specific legislation that exists whenever you're texting or you're doing auto dialing stuff called 10 DLC. And so we wanted to eliminate those tickets, right? So I actually went and I built a GPT, which was called the privacy policy generator, because a lot of these trouble tickets were because the privacy policies were bad. And we trained them using a handful of ones that were good and a handful of ones that are bad with a bunch of rules, and I trained them on.

1:02:42And it's wonderful, except I can't run it in production, because it's not the kind of thing that is usable in that way right now. It's still very difficult. And so all I'm saying is I'm happy to keep spending a few hundred dollars a month, a few thousand bucks a month, whatever it is that I'm spending I don't quite exactly know. And I agree with you, it was very easy. I think OpenAI does an excellent job of getting off the ground. But what I'm also saying is that when you actually translate that into a mainline use case, right where I want to now give it to my support team and say this is now a tool you can rely on, it's integrated into your workflow, into your other tools, it's integrated into how you pipe out data into Salesforce or what have you, it's just very hard.

1:03:30And I'm not saying it's not going to get fixed, I'm saying we're just not there yet. And one of the ways in which it's not there is that there is no place I can go, including OpenAI, that actually makes it fast enough to be usable in production. You wrote this on OpenAI stack. You wrote a custom GPT. Yeah, built myself. Yeah. And you could do the monoghugging face now. It's going to be a lot of options. In terms of integrating into your workflows, I think it's a really interesting point, because I saw a demo somewhere where now, actually, I think OpenAI announced this, that you can at mention a custom GPT.

1:04:05Yeah. Sunny showed me that this week on the pond. Yeah. In chat GPT, you can now at mention a custom GPT to kind of invoke it. Yeah, so how it works is you'd say, Hey, I'm heading to New York, what flights can I get at Expedia, at Chi -AC, whatever, and then it gives you the results here and you're kind of pulling that up. Just to the point about where data advantages lie and that's ultimately going to drive value, I've tried to think a lot about this. I cannot think about a better data advantage that is orders of magnitude better than anything else. Say YouTube. Say YouTube. Yeah. YouTube. It is.

1:04:47So I. So here's the numbers. I pulled this up. You guys know like GPT3 and 3 1 1 2 half were trained with a heavy waiting on Common Crawl, which is this open source. Yeah, we talked about this before, Gil L. that runs it. Open source crawling of the web. The total amount of data in Common Crawl, which I think accounted, and I could be off on this something like 40 to 60%. of the waiting in GPT 3 or 3 .5. I'm off on this probably. So the total amount of data in that common crawl data set is about 10 petabytes, okay? Based on YouTube's public statement recently, they're seeing about 500 hours a minute of video uploaded or 720 ,000 hours a day.

1:05:29And if you assume somewhere between, you know, just under 1080p on that video, we're talking about probably one to two petabytes of data being uploaded to YouTube per day. So if you assume like over time the definition of the videos gotten better and the amount of uploads gotten up, you could probably assume that there's roughly, I'm guessing, there's probably somewhere between 2000 and 3000 petabytes of data in YouTube growing by one to two petabytes per day, which makes YouTube data repository 300 times larger than CommonCrawl, which makes it bigger than anything else that anyone else has. And here's the amazing thing about it.

1:06:12It has video, it has image, it has audio, it has text, it has everything. And it is growing. So if you were to take a bet or build a thesis around this point that the data advantage is going to drive value creation, if Google gets its act together and leverages the data repository at YouTube, it is an insurmountable moat that will only continue to extend because the quality of the YouTube experience and the network effects continue to accumulate for them. I think it's the most valuable asset in the world today based on the species that AI value is going to accrue to the data owner. I think you're making such an important point.

1:06:49This is why the counterfactual is true and it's actually showing up in the data and Nick will show you this slide again from the AI primer. but that is why we're seeing these diminishing returns for your burger and all of these third -party bench parts of these models. Yeah, obviously the same data sets. It's all using the same data sets. So what we are proving is not that the underlying hardware can't scale nor that transformers are only efficient to a point. That's not what all of this convergence is showing. It's that in the absence of proprietary data, you're just going to get to the same model quality.

1:07:19And we're seeing a bunch of different models get to a very early finish line, which, again, And if people like Facebook are doing for free, that's much easier to underwrite because you don't have to underwrite it being a differentiator in five years. But if you have a startup with equity value tied to a model, I think it's much more of a tenuous place to be in the absence of proprietary data. And everyone in the world, as a camera and a microphone in their pocket, and high -speed internet now, from the phone in their pocket, it, and more and more people are uploading that content, that data that's being generated.

1:07:59YouTube's got this free data vacuum, and it's just out in the world. And most of it's getting up to the issue. Well, it is public facing though. So it's not just true for text, it's also true for all of the image generation. So like if you're looking, they can train more than just an LLM on it, right? They can build all sorts of... Yeah, go ahead. No, no, no, no, I was just going to say like the version of Common Crawl for training these image models also exist. And so to your point, it's like, we are all operating from the same brittle, very fixed, small quantum of training information. And so that is why I think like Facebook and Google are doing a really important job by deciding that these models should be free.

1:08:42Right. And then being able to the so then the question that just accentuates their data advantage. It does. And I think that it allows them to decide how much to leak out. So for example, whenever like if you were using a lot of Google services like GFS, Bigtable, Big Query, you know, TensorFlow, the versions that you had access to via GCP was always one or two generations behind what the Google employees got to use, right? But it was still so much better than anything else that we could get anywhere else that you would still build to those endpoints. And I think there's a similar version of this where Facebook and Google probably realize, like, look, we'll have version five running internally to optimize ads and all of this other stuff that makes our business that much better.

1:09:30And we'll expose version three to the public. But version three is still trained on so much proprietary data that it's so much better than version 10 and anything else is just operating on the open internet. and you know, to your point, Freiburg, that's the outward facing stuff. YouTube is a collection of things people want to share. What Google also has is Google Docs and Gmail, things that people say privately. So they have another data resource there that they can tap, you know, and there'll be regulations and privacy around that. But maybe there's a difference there, but I honestly can't think of the quantum coming close to YouTube, not even close.

1:10:06Well, the thing to Jason's point, which is really interesting is like, you know, there's a modality in AI called rag where you can actually just Augment with very specific training on a very specific subset of documents to improve. It's like a it's like a hacked version of a fine tune But the beautiful thing about that is like if you have a Google workspace my entire company runs on on Google workspace in fact most of my companies do at this point to click a button where all of a sudden now All of that stuff and all of my G drives, all of a sudden is trainable. So that the N plus first employee comes in and has an agent that's tuned on every deck, every model, that's every document, that's a huge edge.

1:10:50Huge edge. By the way, and as a CEO, if you gave me that choice, I don't think anybody underneath that reports to me has any right to make that decision. But as a CEO, I would click that button instantly and I have that right as a CEO. And so that's the CEO pitch. It's like, look, I can just give you these agents that are like the next version of a knowledge base that we've always wanted inside of a company. Notion has this, you know, they basically, you can start asking your entire notion instance questions about notion, which is incredible. And yeah, you can just, and as a CEO, you can see across everything, Chema, because as you know with Google Docs, if you're on a compliance -based industry, like finance, you can see everything, every message, every email, every document, and you can start the security model and the data model becomes very complicated and all of that stuff.

1:11:42Like, for example, how do you know that this spreadsheet is actually, you should learn on it, but who gets to actually then have that added to the subset of answers, right? All of a sudden, like salaries, right? The HR information, information gets put into the training model, very dangerous. or subset A of companies working on a proprietary chip design, but they actually like the way that Apple runs highly highly segregated teams where nobody else can know. So there's all kinds of complicated security and data model and usage questions there, but Yeah, for every world. So there's been a lot of discussion real estate.

1:12:17You shared a video with us Why don't you kick it off for us here? Freiburg what's going on in commercial real estate and sacks you've got holdings and a lot of that's as well So let's kick up the commercial real estate challenges of the moment. Well, I mean, I think we're teeing off of Barry's comments at this event last week. He and I met backstage because I spoke right before him and then he gave this talk, which is available on YouTube where he talked about the state of the commercial real estate market. And particularly he talked about the office market just to take a step back to talk about the scale of commercial real estate as an asset class in the US.

1:12:53and if you'll pull up this chart, the total estimated market value of commercial real estate in the US across different categories is about $20 trillion. With about $3 trillion being in the office market, which is specifically what he was talking about and he was saying that in the US we're seeing people not coming back to work and all these offices are empty and we've talked a lot about these offices being written down. So how significant of a problem is this? So $20 trillion asset class, obviously the multifamily market is probably not as bad as office and retail, which are the most heavily affected, each of which are about $3 trillion a piece.

1:13:29The rest of these categories seem relatively unscathed in comparison, industrial, hospitality, healthcare. You know, those real estate sectors are probably pretty strong. Data centers obviously growing like crazy, self -storage, the great market. If you pull up the next image, so it turns out that of the $20 trillion of market value, There's about $6 trillion of debt. So you can kind of think about that $20 trillion being $6 trillion owned by the debt holders and $14 trillion by the equity holders. And the debt is owned roughly 50 % by banks and thrifts. And this was this concern that we've been talking about with higher rates is the debt on office actually going to be able to pay, the debt on retail going to be able to pay.

1:14:12One half of that debt is held by banks and thrifts that as we talked about have such a close ratio to deposits that you can actually see many banks become technically insolvent if the debt starts to default. Barry's point that he made was if you look at the office market which you know is marked on everyone's books as $3 trillion of market value he thinks it's probably worth closer to $1 .8 trillion. So there's $1 .2 trillion of loss in the office category. And if you assume 40 % of that $3 trillion is held as debt. You're talking about $1 .2 trillion of office debt. A reduction from $3 trillion to $1 .8 trillion means that the equity value has gone down from $1 .8 trillion to $600 billion.

1:15:03So they've lost equity holders in office real estate have probably lost two -thirds of their value, two -thirds of their investment. And who owns all of that? Most of that, 60 -plus percent call it two thirds of that is likely owned by private equity funds and other institutions where the end beneficiary is actually pension funds and retirement funds. And so if two thirds of the value has to be written off in these books and it hasn't happened yet, what's going to happen to all these retirement funds? And this is where going back to my speculation a couple months ago, kind of gets revisited. If you're actually talking about a two third right down on the value in these funds, most of that being pension funds, you're not going to see governments let that happen.

1:15:45You're going to see the federal government. There's going to be some action at some point. And it's unlikely the office market is going to suddenly rebound overnight. If this stays the way it is, who's going to fill that hole for retirees and pensioners? Because we're not going to let that all get written down. Someone is going to step in and say we've got to do something about this. And there's going to need to be some sort of structured solution to support retirees and pensioners because that's ultimately who ends up holding the bag in this massive right town. He didn't go all the way there in his statements.

1:16:15He was talking more about his estimate of three trillion to one point eight trillion And then I tried to connect the dots and what that actually means and ultimately there's gonna be some pain felt by retirement funds That's gonna need to be dealt with somehow. So I I don't know if that if that sits right with you I mean, I think the big picture is right. I think you're applying a lot of averages. Right. I think in the office market in particular The typical office deal is more like one third equity and two thirds debt. There's a lot more leverage. Right So that'd be point number one, which makes the situation worse.

1:16:45Even worse, yeah. So I would say that there's a huge amount of equity that's been written off, but in addition to that, there's a lot of debt holders who are in trouble too. And that debt is held by regional banks. So these commercial loan portfolios are significantly impaired. That's what we saw with community bank in New York is that they're stock cratered when and they reported higher than expected losses than their commercial real estate portfolio. So, Friberg, I think the point is just, the pain from this is not just gonna be on the equity holders, but also on these banks, which can't afford to lose or not to - It's not evenly distributed, yeah, right?

1:17:29Yeah. Right. And we saw this in San Francisco where some of these buildings have 70 % debt equity ratios and the value puts them in the hole and that could be wiped out completely. And the deadholders have to take a head. And normally, you know, that debt is not really written off very often. Well, this is why the deadholders, the deadholders don't want it for close. They don't want to get these buildings back because when they do, they're going to have to write down the loan. As long as the loan is still outstanding in the MF for close, they can pretend that the value of the building is not impaired.

1:18:00Kick the cand out of the road is the best strategy for them. So it's called pretend and extend. So what I'm doing is they'll work out a deal with the land with the equity holder that the equity holder say, listen, I can't pay the interest. So they'll just tack on the interest basically as principal at the end of the loan and they'll extend out the term of the loan. Which would wipe out the equity at a certain point, yeah. Well, what it does it allows the equity holders to stay in control on the building, right? Because Yeah, the equity holder can't make their debt payments today, but they're going to postpone those debt payments until the end of the loan.

1:18:39And again, in the meantime, just kind of hope that the market will come. It could mention debt at some points, if they have so little equity in these buildings, typically just exceed the value of the property. And it's like, I'm just working for the bank now. And why am I even putting this work in? Because everyone kind of hopes that the market will recover. or the value of their equity will go up and they'll be able to make their debt payments again. So if you're the equity holder, you'd rather hold on and have a chance of your equity being worth something in recovery, then definitely lose the building.

1:19:09And if you're a regional bank, you'd rather blend and extend, or pretend and extend, as opposed to having to realize the loss right now and showing the market that your solvency may not be as good as you thought. The same thing happened with government bonds. Remember that with SEPB and these other banks, they had these huge held to maturity bond portfolios? These are mostly just T -bills that were worth, I don't know, 60 cents on the dollar when interest rates spiked from zero to 5%. But they didn't have to recognize that loss as long as they weren't planning to sell them. Right. And then when they had the bank run, they had to sell.

1:19:50Well, yeah, that's right. So when depositors left because they needed their money or because there was a run or because they could get higher rates in a money market fund, all of a sudden these banks had to sell their health maturity portfolios and they had to recognize that loss. And that's when everyone relies, oh wait a second, they're not actually solving. Okay, so Jamal, supply demand matters in real estate. We have a tale of two cities here on one side in real estate for commercial real estate, no demand for office space, which is in way too much supply. Paradoxically, on the other side, we have this incredible market for developers, which is, gosh, there's not enough homes, I think we need seven million more homes, and the demand is off the charts for homes, yeah.

1:20:33Yeah, I mean, I think you're basically right. It's not, I keep trying to explain residential, it's not a great market either, because it interest rates are spiked up. So there's not a vacancy problem. Multi -family developers are still able to lease the units. They're still able to rent. The problem is their financing costs have shot through the roof. So, again, let's say you were a developer who built multi -family in the last few years. You took out a construction loan. That construction loan might have been at 3, 4%. Now, you want to put long -term financing on it. But if you can even find debt right now because there's a credit crunch going on, you might have to pay 8, 9, 10%.

1:21:10Yeah, but at least you can find a renter. You can find a renter. That's true, but only at a certain price and let's see you under wrote that property To I don't know like a five cap like a certain yield. Yeah, but now your fancy costs are much higher than you thought You might be under water Yeah, that situation isn't as bad as what's happening in Why I think it's I think it's worse in some ways if you're fully rented and And your building is underwater because now your debt payments are much higher than you expected. Then there's no business model. At least seeing that, are we seeing tons of multi -fabric under?

1:21:46Can I make two points? I think David is right, which is that I don't know this market very well, but just as a buy -stander, here's what I observed. It seems that the residential market has a feature, and I don't know whether it's good or bad, but that feature is that you reprise to market. demand every year. So to the extent that supply demand is changing and default rates are up or whatever, that's reflected in rents. And you see that because rents change very quickly and most human beings are signing six months to one year leases. So that reset happens very quickly so it can more dynamically adapt.

1:22:25So to the extent that a market segment is impaired, you see the impairment quickly. on the, on the office side, what I see is that there's been a structural behavior change in COVID that has reset in every other part of the world except for the United States, where there are these, frankly, typically younger, typically more junior employees that have held many of these companies hostage in a bid to return back to office space. And so we know that there is this vacancy cliff that's going to hit commercial real estate. We just don't know when, because there are, they're in long -term leases, they're canceling these leases over long periods of time.

1:23:04So the reset cycle was longer. That's just my observation as an outsider. I don't know what that means for prices or anything else, but it just seems at least the residential market can find a bottoming sooner because you can reset prices every year, but commercial just seems like a melting ice direction. I correct you, you sax. That assessment. Commercial has both the demand and a financing problem. Multi -family just says a financing problem, but it's important to understand. We're talking about office. Because it's retail and then there's office and then there's other industrial. Do you guys see a China?

1:23:37China has 50 million homes ahead of schedule. 50 million additional supply that can house 150 million people. So as acute as our issues are, the China issue might be much much, yeah, seismic. Can we just give you an example on the multi -family site? Okay, let's say that you buy a building. Okay, let's say you bought a building in 2021 the absolute peak of the market and you could get debt at say 4 % okay and you penciled out let's call it a 6 % yield that with the debt you're getting. So let's say you did 2 thirds debt at 4%, you could now lever up that 6 % yield to 10%. Okay, that's like sort of the math right?

1:24:19Now all of a sudden and to get there you'd have to do some value added work on the property, you have to spruce it up. Now it's a few years later and your short term financing is running out and you need to refile and you've done your value added work. But here's the problem. The overall valuations in the market have come way down. So before the bank was willing to give you two thirds loan to value, now the values come way down, you may not even be able to get two thirds loan to value. So you're going to have to do what's called an equity in refinancing. You're going to have to produce more equity.

1:24:53You're going to have to pony up more money since instead of taking equity out like when the deal goes well, you're going to have to put equity in. You may not have that equity if you're the developer. The other thing is that your financing cost now might be 10%. So now you've got negative leverage. You're generating a 6 % yield, but you're borrowing a 10 % to generate that 6 % yield. So the debt no longer makes sense. Again, you're not positively leveraged. You're negatively leveraged. So you're not gonna wanna take out that debt. And if you do take out that debt, the buildings will be under water.

1:25:24It's not gonna be generating, not operating income, it's gonna be generating losses. So that's why even categories like multifamily, where you don't have a vacancy problem, there's strong demand. Those properties still don't make sense. If you had long -term debt on your multifamily, if you were able to lock in that 4 % loan for 10 years, You're fine, but for all the people who are refinancing now who are coming up this year, last year, next year, they're in deep trouble. And that's why there's a rolling crisis in real estate is because the debt rolls over time. It's not like everybody hits the wall and has to refinance at the same time.

1:26:05Well, that God, right? I mean, this would be a cataclysmic if it was. If ever, and can you imagine if Silicon Valley and San Francisco had to say, here's actually the reality. Anybody want to actually pay for this office? all on the same year. Right. That would be insane. But the crisis is growing as as the leases roll and those old reds that were higher than market roll off and now you have to take on new leases if you can even get them. It's going to be bad. And a much lower rate. And as the old loans roll that were at a much lower interest rate, you have to get financing even if you get it at a much higher interest rate, that's when all of the sudden these buildings go from being basically solve it to insolvent.

1:26:45Yeah, I mean Janet Yellen's just gonna bail these folks out. I mean she won't bail out the banks themselves, but she'll bail out the creditors obviously of the people holding the bag. They'll get bailed, yeah. That's everybody agrees? Janet Yellen? Yellen, our treasury secretary. I don't know if she's gonna be the one to do. I think there's gonna be congressional action on this stuff. Yeah, I mean they tend to lead it so. Alright, for the Sultan of Science, David Freiburg and David Sacks and Shema Paul Haapatia, the Chairman Dictator, I am the World's Greatest Monitorator. We'll see you next time in the Nolan Pond.

1:27:19Bye -bye. Bye -bye. We'll let your winners ride. Brain man David Sackett. We open source it to the fans and they've just gone crazy with it. I'm the queen of King Wild. I'm going on a win. What? Your winner's ride. Besties are gone. Go through it. That's my dog kicking it. I wish you drive away. Sit down. We're at all. Oh man, my ham is the actual meat. We should all just get a room and just have one big hug or two because they're all just just like this like sexual tension that we just need to release them out. What your beef? What your beer of beef? Beef of beef? What's good for you? We need to get mercy.

1:28:07I'm doing all of it. to