DeepSeek Rattles Markets + SailPoint's Going Public… again? | E2077

0:00Jason, you're talking about the amount of money in play here. Look how much things have changed just today. NVIDIA has been dethroned overnight by DeepSeek and is now the third most valuable company compared to the first. Crazy. That is such a change. Yeah. Wow. That's mind-blowing to me. The AI trade giveth the AI trade. Take us away. This Week in Startups is brought to you by Northwest Registered Agent. Starting your business should be simple. With Northwest Registered Agent, you can form your entire business identity in just 10 clicks in 10 minutes. From LLCs to trademarks, domains to custom websites, they've got you covered.

0:34Get more privacy, more options, and more done. Visit northwestregisteredagent.com slash twist today. Squarespace. Turn your idea into a new website. Go to squarespace.com slash twist for a free trial. When you're ready to launch, use offer code twist to save 10 % off your first purchase of a website or domain. And LinkedIn Jobs. A business is only as strong as its people, and every hire matters. Go to linkedin.com slash twist to post your first job for free. Terms and conditions apply. Hey, everybody. Welcome back to This Week in Startups. It's Monday, January 27th, a day that will live in infamy or intrigue, or it will be a footnote in the history of AI.

1:14But we have a lot of hand-wringing going on. We have a story here that we touched on on Friday that seems to have captured the imagination of everybody in our industry and geopolitics. And you know it's a big story when it actually hits like the Sunday shows or it starts to hit mainstream consciousness. But China has, as everybody knows who listens to this program, released a model for$6 million with old iron, maybe, sort of. That could be revolutionary or it could just be a little smoke and mirrors. So we're going to get into it here. And it could be both of those things. All of those things.

1:56and uh how are you alex i'm doing really good you know i was a little bit offline yesterday but i woke up very early this morning hyped and stoked and ready to go and the first tweet i saw was nasdaq off five percent over deep and i was like oh man all right let's go well the ai trade giveth and the ai trade taketh away the deep seek project has uh shaken the public markets specifically nvidia uh i think because what this project we'll see how true it is uh has done more than anything is made people question the need for capex spending at the apex capex spending announcements which happened uh at the end of the inauguration with uh masayoshi-san as all apex funding announcements do when masa comes in you know you've reached the peak man that's it you're at the top of our errors.

2:53That's the peak. That's the top tick. When Masa comes, he puts that little straw on the camel's back and then just snaps in half. So what are we talking about? Project Stargate, the plan that was announced very recently to invest$100 billion now and up to$500 billion, Jason, over the next couple of years. Essentially a massive data center build out and also meta last week dropped. And we talked about this, I think it was on Friday as well. A 60 to$65 billion CapEx spend for this year, a two gigawatt data center so large, it would take up a big chunk of Manhattan. And now all of that seems ever so slightly silly because DeepSeek has shown that you can make models that are very high quality and can challenge the state of the art for what is considered to be a much lower price tag.

3:37Now, you said$6 million earlier. You're referring to the white paper that DeepSeek put out. I read the same thing. There has been discussions about what's included in that price point. Should it be$6 million? I say, great, 10 exit, 60 million, still cheap compared to what we've seen out there in the market. But if you can do things more cheaply, Jason, the gist there is not really the dollar amount, but just how much compute you need to make it happen. And if you need less compute, then well, maybe your data centers are more like financial albatrosses that will depreciate versus a way forward.

4:08So we've had a flip from look at the money we're spending, therefore take us seriously to look at the money we're spending on data centers. Perhaps we took it ourselves too seriously yeah and there's a number of interesting ways in which they did this in terms of building the model and how it's executed and that's all in the paper i think how they did it is interesting of course and i think leveling up the conversation the question becomes and we actually had this discussion here which is when you have constraint when you have limited resources yeah you get clever so we call this sort of macgyver it uh if you remember the tv show for those of you there was a guy macgyver who would you know basically stop a rogue nuke with gum and a paper clip and basically proved hey uh you can get a lot accomplished with a small amount of resources if you put your mind to it and so this is what survival is like this is what happens when you ban h100s being sold into china is they take what they have and they do the best you can now there is uh a bunch of and i think uh scale ai ceo in davos said you know they have a bunch of h100s in china all these ones that go to um which is it taiwan or the philippines there's one or two markets where a lot of these get sold singapore i think is another oh singapore yeah so 20 of the H1s, 100s get sold in Singapore, but Singapore is not using 20 % of H100s.

5:39Wouldn't think so. So where do they wind up is always the question. So like many things that occur in China, it's a black box. There might be more to the story. Putting that aside, it's an accomplishment. It's a white paper. It's open sourced. So the experts seem to think this is largely legit. Yeah, that's the consensus is like, hey, this is not faked in some way. And then the next card that turned over was, well, maybe they ingested other models to make this model. To which I say, okay. So? Yeah, that sounds like they stole something, but they're all building off each other. And okay, maybe.

6:15But the output is still the same. And then I saw that the DeepSeek app was number one in the App Store ahead of some other, which by the way, you can just pay like an outside firm five grand and do that. So it's pretty easy to game the App Store. So don't read too much into that. But this is now game on, because if the Chinese want to prove a point to us about AI exceptionalism, to have a hedge fund manager with but a handful of scientists and engineers in their spare time, basically make something that races up the charts for a small amount of money. Well, what does that say about all this crazy investment we're doing?

6:54Well, it goes to show that right now, some very smart people with access to at least some GPUs in quantity, whatever level of quality they were, whatever, can do quite a lot. And to me, the biggest point that that makes is a little bit less about China versus US, but about the fact that we are nowhere near having wrung out all the possible gains, all the possible improvements. I read the white paper, Jason, and I'm not going to lie when I'm reading through about how they use pure reinforcement learning and different types of cold start data and supervised fine tuning through rejection sampling.

7:30Okay, I'm not going to lie. That's not where I spend my time. So I'm not going to pretend to know every single word here. But when we're dealing with open source stuff, it just doesn't get held in the same kind of national grip. So to me, these improvements will filter into the US market. We'll learn from them. Sure. And then the question becomes, is there any advantage to after we've ingested and digested these upgrades and improvements to having a lot of GPUs? Because if there is, then with DeepSeek's learnings, we're going to be able to accelerate very quickly. And some people are arguing that really, this will make the Project Stargate effort and therefore XAI supercomputer and Anthropic and Amazon, etc, even more valuable.

8:09Gary Tan said that the news makes Project Stargate, I think, 30x more valuable in his view. Clearly hyperbolic. but like directionally interesting. There's a new paradox, you know, everybody who's in intellectual circles, if you want to win the social media game, if you want to win the argument, introduce a new paradox, a new framework for thinking, a new cognitive biases. And Jevon's paradox seems to be the one that Satya Nadella brought up, I guess, late at night. So he was up late and thinking about this, or maybe that would be - I think everyone was up late thinking about this. Yeah. You need to know this.

8:44So when you go to your next dinner party, you can mention, yes, that's just Jevons Paradox. Absolutely. So Jevons Paradox was originally put together to discuss energy consumption. And as energy got cheaper, people used more of it. So essentially, the idea behind Jevons Paradox is that when technological progress increases the efficiency of something, i.e. a winter resource is used, people use more of it. So essentially, falling cost increases demand. So if AI, in this case, to kind of complete the point, becomes much cheaper, as DeepSeq showed with its relatively low priced, excellent model, it doesn't mean that it's going to collapse the industry.

9:19It means that we're going to use quite a lot more of it. Does that make sense, Jason? Absolutely. We have a concept called induced traffic. I've talked about it here many times. And so people will consume more of something if it's more available. So when you had to go walk to a well and get water and pump it and bring it back to your house and you had to walked to the other side of town, people did not have lawns and lawnmowers. But when they built pipes to your house with water and it was so cheap, people decided they would put sprinklers on their front lawns and make beautiful green patches of grass that were non-native and they could look at and go, wow, look at that beautiful green grass, as opposed to just seeing it seasonally.

10:04Hey, founders, you want to build the next great billion dollar business, right? Well, you're going to need two things, a killer idea and a properly set up company. I can't help you pick the idea, but I can help you get your business started the right way with Northwest Registered Agent. With Northwest Registered Agent, you can form your business for just$39 plus state fees, one of the most affordable options out there. No hidden fees, no upsells, just a simple, straightforward process. And here's how it works. In 10 clicks and 10 minutes, your LLC is officially formed and ready to operate. Northwest handles all the paperwork fast, accurate, and hassle-free so you can focus on building your business, not navigating all those crazy legal forms.

10:47Thousands of entrepreneurs trust Northwest Registered Agent because they make business formation affordable, efficient, and stress-free. And with their expert team standing by, you'll always have the support you need. So here's your call to action. Don't let paperwork hold you back from your entrepreneurial dreams. Get started today at northwestregisteredagent.com slash twist. That's northwestregisteredagent.com slash twist for just$39 plus state fees. Your business could be up and running in no time at all. Get more value, more convenience, and more peace of mind only with Northwest Registered Agent.

11:20So that's a great example of induced consumption there. Same thing with when you make a freeway, people go, hey, wow, there's a freeway. I can get from Austin to Houston, Dallas, or San Antonio in X number of hours, minutes, and maybe we'll go there for a day trip. Maybe we'll go see a concert there. Whereas if you had to take a dirt road, you'd be like, yeah, I'm not going there. I'll wait for the show to come here. I'll wait for the circus to come here. So that's essentially what people are now saying is, hey, if that iron is sitting there, we will use it for something. In our industry, I remember Dropbox and YouTube were based upon the premise that all this dark fiber and all of the storage had suddenly become really cheap and it was plummeting in cost therefore there was more available therefore the idea of hosting somebody's video for free on youtube became viable whereas previously you had to pay every time somebody watched your video you had to pay your hosting company for it yes it became free when it became free then the entire concept of putting a vlog out like we're doing right now a blog and then eventually a podcast People said, yeah, okay, we can do this.

12:27It's free. Now, if I had to pay a dollar for every time somebody, or 10 cents, or even a penny every time somebody downloaded one of these podcasts, we'd be like, whoa, we got to throttle this thing. So what happens to this extra capacity, I guess, becomes the question. Yeah. Well, so I want to bring up another point along this line before we answer that question. So talked about Jevon's paradox, increasing efficient usage of something increases the demand for it. So Pat Gelsinger, up until 48 minutes ago, Jason, the CEO of Intel, put out a very interesting tweet, and he compared this moment to the gas law.

12:59He says, computing obeys the gas law, making it dramatically cheaper will expand the market for it. So I actually am a little bit torn between pessimism and optimism here. Because on one hand, it's lovely to see such amazing technological improvements. But on the other hand, we worry about what we're going to do with all this hardware. I wonder if we're just going to end up, instead of having a decrease in demand, we've just brought the future forward a little bit. That's correct. Yeah. Okay. So in that case, then none of the capacity is wasted. I mean, it's kind of like if you were to build extra apartments, you know, in a market where you had massive constraint, would it cause damage or would it allow young people to get an apartment and be able to live alone as opposed to five people living in an apartment like they did in.

13:48you know, the Tenderloin district of San Francisco. And I love that analogy because who are the young people in this AI moment with all this capacity built on to improving AI models? Startups. Exactly. Yeah. So the, you know, this idea that startups were not going to have access to H100s, you know, there might be so many H100s available. There might be so much capacity available if this gets easier and easier and easier to, you know, build these models that, uh, yeah, startups will use one to do something that would be considered trivial yep and so when we deploy these we're gonna we're always thinking well how do we make the money back right how do we make the money back so with that 500 billion dollar number it's like well is there 500 billion dollars in software revenue from ai available to make the money back uh okay well we replace every factory worker if we replace every you know truck driver i mean you really start to have to think about these things but now hey maybe you don't have to think about it as much and putting a bunch of cpu towards you know putting a filter on top of your photos which is what happened with iphones you know the the processing power was so great we we came up with some uses for it and and that's where that's where we're at right now it's like we can do things that are seemingly uh unnecessary with this do you remember when filters were cool like i remember when instagram was really little like people would like pick that like kind of like purplish filters make things look a little bit more kind of exotic that was fun Yeah.

15:14It used to be done by a job known as a graphic designer. So if you had a photo of your family, you could have a photographer take it because you didn't have access to a camera because they were very expensive. Yeah, yeah, yeah. Then you would have a graphic designer touch it up and they would touch it up for 200 bucks. And then you'd have a Christmas card that you would go to a store and have printed. Oh. And now you just take a picture on your phone and send it to print and you're done. It's pretty great. That's a little bit different than the way things work now, in which we upload our contact book to an online database, run that through the system we're purchasing cards from, and then they come free, filled out for us with everyone's address.

15:52Fantastic. Not at all at risk, though, when it comes to security, now that I think about it, giving everyone's address over to some random company. Jason, why don't we do a little demo? Why don't we show R1 in a couple of different contexts here? So what I have done is downloaded DeepSeek's R1 7 billion parameter model, and I am hosting it locally on my MacBook Pro. If you're curious, this is running a M3 Pro chip, I think 12 or 16 gigas, something like that. So Jason, let's ask it, you know, who is the president of China? And let's just see what it does here. What I want you to watch is how it talks to itself.

16:27So what it's going to do here is it's going to start to riff. All right. So user wants to know this, looking back at our history. Let's think about this. I don't know why it's asking, it's giving me emojis. And it says the president of China is Xi Jinping. Very, very simple, pretty easy. Took into account my prior context. Nothing there that's going to blow anyone's mind. Here is the DeepSeq interface. And I'm running R1, as you can see right here. And I'm asking the same question. And in testing, we discovered a relatively different set. oh, it got cut off. So what we have seen here is power of open source.

17:04The hosted version of DeepSeek R1 has clear lines around what it can and cannot answer. Clearly, President of China, off the list. Whereas the hosted version does not. What's cool, though, is you don't have to use the web one. You can just download it. It's open source and run it yourself. And so it strips away all that cruft. And I bring all this up to say that people are saying, oh, you know, it's a censored model. I can't do anything. Look at my screenshots. Well, yeah, or you can just download it. And I think that's incredibly powerful and needs to be discussed more. At the early days of these with the image creation, a friend of mine, you know, they had these filters.

17:38You can't do celebrities. You can't do racy photos. And I had a friend who was like, hey, look, you can do anybody. You can do anybody and you could change somebody's gender. You could do this. You could do that. You can make somebody a squid person. So the censorship in the hosted versions, I think, is going to lead a lot of people to actually run this stuff locally and then privacy. I did see people were taking the terms of service. If you think TikTok is a security issue, I mean, if you're using DeepSeek with any information, you got your head examined. Everything you're sending is going directly to the CCP.

18:14So I would recommend not using this. It's probably got all kinds of the hosted version. And certainly everything is being sent directly to servers in China, directly to the CCP and being used in using AI. So they're using AI to examine every photo you share, then correlate your IP address with your deep seek, and then build a dossier on every American, and then take public sources of data, run it against geolocated data in ad networks. And they have so much information on every American. It's nuts. We need to really address all of that. But, you know, big picture, the question is going to be, will this slow down hardware consumption?

18:51And I think that that is actually possible that you might have the leaders of some company saying, hey, let's spend the next dev cycles optimizing this and just making it run better on the existing footprint of hardware we have. And then when you're a capital allocator, you're looking at a pool of capital. Is this billion dollars better spent on more H100s? Or is this billion dollars better spent on more headcount? And what people might realize here is it might be better to hire more people to do more optimization on some aspect of what you're working on than buy the incremental H100s. All right, founders, let's kick off the new year by talking about your website.

19:37If you're launching something new in 2025 or you need to refresh your brand, and I bet you do, you need to use Squarespace. Squarespace makes building a stunning, professional website ridiculously easy. Whether you're selling products, offering services, or just sharing your ideas, Squarespace gives you the tools you need to make it happen. And you know Squarespace is at the cutting edge with AI tools that make building your site easier and faster than ever. For example, its design intelligence tool. It's just like having a world-class designer at your fingertips. You ask it a few questions and their AI takes it from there, building a fully customizable website that's perfectly tailored to your brand in just minutes.

20:14We're talking professional layouts, on-brand visuals, and premium content all ready to go day one. So here's your call to action. Start the year strong. Check out squarespace.com slash twist for a free trial. And when you're ready to launch, go to squarespace.com slash twist to get 10 % off your first website or domain purchase. That's squarespace.com slash twist. that was always going to happen at some point people would say hey we're utilizing x percentage of this we're out of training data let's optimize this and then where would we deploy the capital better how could the capital be deployed for a better return and that's the key here what i've always felt would happen is i think gemini and co-pilot at microsoft is going to become a free product eventually they're just going to say you know what we can just abstract this cost into our existing business and apple if they ever get their act together which it seems like they're just going to draft off open source um if they ever get their act together or they buy claude or something like that which i think is probably what will happen eventually they'll just buy one of these language model teams so they can just start on third base with the team they will um basically kill the market for paid products here and what could you run the dev and the hosted versions down too it might just be you know so de minimis at aws at azure at google cloud uh you know at oracles cloud that people just go you know what uh there's not a business here just like it's really hard to have a business in storage or transport it's maybe this like thin layer that's uh free for everybody which is good for humanity i think that's the other thing if you're thinking from a human perspective like maybe we move to the geopolitical ramifications of this now sure um it's going to be very hard to constrain the chinese uh i mean i do think we win ultimately but it's going to be hard to constrain anybody by just limiting their access to h100s uh they'll be able to get other chips and when you train something people get creative so maybe they're making knockoff h100s um and other kind of devices other hardware platforms that will do the same thing so so you wanted this josh gusner to be pulled up because i think it does detail an interesting split uh in the world of venture uh mark andreason called um deep sea's work a gift to the world and then you know uh josh lux was pretty critical of that and then here we have josh kushner or josh kushner really i'm saying that quote pro-america technologists openly supporting a chinese model that was trained off of leading u.s frontier models with chips to likely violate export controls and according to their own terms of service take u.s data back to china and then he does the thinking emoji face and jason i want you to explain what josh is trying to say here decode the investor speak for the regulars you know how you can tell if an investor is talking their book how they're tweeting and so they're tweeting so what josh is doing here is you know he is the last investor uh at the 150 billion dollar chat gpt round around i said could possibly be the peak valuation of open AI.

23:21I don't know if you remember me making that claim on all in and I made it here as well. So, you know, I've been thinking about this for a long time when a company gets extremely overvalued and there's a lot of competition from extremely deep pocketed people and the technology has open source going on, you start to look at and go, Hmm, I wonder if that is the highest valuation this company will ever reach because of this competition. because so many people are going after the same price. And so there used to be a business in selling packaged browsers. I remember going and buying Netscape, the browser in packaged software at CompUSA on Fifth Avenue by the Public Library in Manhattan, like as a kid, just like, oh, wow, look, Netscape in a box and then go get your five and a quarter inch floppy, three, 3.5 inch floppy and install it.

24:10You can't, you don't have to buy a browser anymore, right? There's not a business in owning the browser. So I think - before you go on there's a company called island that's making an enterprise browser with extra cyber security that's growing very quickly and it's on the twist 500 but for regulars you're dead on back to you i mean so you're paying for security so exactly so you can build things around it and you know obviously people did a search engine around it and you know people will pay for other things or pay other ways so i think he's probably concerned uh about the fact that open source is just the ultimate headwind against a paid product.

24:47And this open source and this kind of progress coming from small teams, I think is probably going to make people wonder, should I invest in the next round of open AI? Should I give open AI the next 10 billion, 20 billion? And so it's quite possible this might be the peak open AI valuation, and that would be pretty nerve wracking. It's also probably pretty tweaking if people are saying like, just use this. Like, why would I pay for open air i just use this um because remember that valuation gets backed into based on revenue and the revenue picture for open air i was surprisingly good i mean i would have six billion in revenue this year uh or last year so yeah you know it's hard to know because they keep talking about the run rate um it's a little fuzzy you know it wouldn't be fuzzy if they would just go freaking public like do you know how annoyed i am with databricks and their run rate versus arr versus historical revenue numbers that i have to then tease out just yeah the other complications that open ai is you know the lawsuit switching to a public uh entity from a non-profit and if it is a non-profit the non-profit is entitled to the increase in value so i think they have to buy it for 150 billion because that's what its value was or you know somewhere close to that number before the flip happened.

26:08So maybe the investors only get from$150 billion forward. It's one of the most - Convoluted. Convoluted, yeah. You ever see a Chinese F1 filing to list in the US and then they show the ownership chart and it has like 40 ,000 boxes of like JVs and like dotted lines for control and like offshore, that's what OpenAI feels like to me. But what's interesting is that there's still some reasonable optimism out there from folks like Aaron Levy that we are seeing both high-end and low-end demand. So Jason, we talked about OpenAI and their operator product. This is essentially like computer use from Anthropic and it's how it can like go there and play with your browser for you.

26:46Very expensive right now. And then Aaron Levy points out that DeepSeek is cheap and also blowing up. And he says all this points to the race is only going to continue to accelerate. So my view is, I don't think we're going to have a data center overcapacity problem in five years, but maybe we might have our first period of time in which if all this shakes other people think we're not gpu or compute constrained that would be interesting and not a catastrophe i don't think i just wonder if you were hoping that nvidia would double again you might be a little bit sad but i i'm not at all pessimistic about any of this it just doesn't seem bad yeah i mean it's um We'll see what the truth is around the model over time.

27:32And like you said early in the conversation, whatever, because this is an open source, because there's a paper, because they're sharing their innovations, it'll get incorporated into any interesting techniques or approaches will get incorporated into our research. And we have the money to deploy the hardware in America. So why not have that as a potential runway? It's like having extra capacity. but this would change i think a lot of how people think about nvidia the stock and they were off double digits yeah 10 15 at the uh at the start of the at the start of the show this is a chart of nvidia's share price for the last week uh inclusive of where we were earlier today and it went from roughly i don't know jason 150 to about 120 and if you're a multi-trillion dollar company, each 1 % decline of 1 trillion is 10 billion.

28:24So every point of 3 trillion is 30. And if you have a 10 to 15 % drop, that's 300 to$450 billion in deleted market cap in a day, which is a lot. Almost 17 % right now. So that's a lot. That's a lot, a lot, a lot. So there's going to be some putting up, I think. NVIDIA needs to show the demand's hot. I think that the closed source model companies now need to put up some big points on the board to show that they're not essentially dinosaurs that are trying to build something closed source that's going to be open source. And you know what? That's going to be fun to do. Fun to watch. Yeah, absolutely.

28:57I think the operator product, which we've seen different versions of that taking over the browser versions, is going to be the feature of 2025. I think a lot of folks are going to look at repetitive tasks. And this was part of the thesis with our investment in Athena. Go to athenawow.com and get like a month off or something. They give a really good deal. to my followers. And so, you know, if you're watching somebody do a repetitive task in their browser over and over again, you could basically anticipate what they're doing and build it out. So when you are building out the docket, go to thisweekinstartups.com slash docket, and you can see the docket we're working from.

29:33Let's say you're talking about NVIDIA. And anytime you talk about NVIDIA, you pop in a chart. It might just fire off a tab and say, say you know in the early version of it might say insert chart from y charts or wherever you use and you say yes or and then next it would say inserting chart from y charts 10 9 8 and you can cancel it but if you don't cancel it does it and then eventually you're going to wake up in the morning and it's going to be like here are the five most important news stories of the day that we know jason's been talking about chamat's been talking about satia's been talking about this account's been talking about that account's been talking about and here are those stories.

30:14All right. Growing your business in 2025 comes down to really one thing. How well can you hire? And we all know smart hiring starts with LinkedIn. LinkedIn gives you the strongest data and insights to find the right candidates faster. From unique skills and interests to connections you have in common, LinkedIn will help you hire with confidence. And here's why it works. There's a billion members on LinkedIn now. The march to a billion has happened. These people are across 200 plus countries and LinkedIn will connect you with the best talent, period, end of story. 72 % of small businesses say LinkedIn helps them find the right hires.

30:50And 86 % of businesses get a qualified match within just 24 hours. Here's what makes LinkedIn so unique. It connects you to over 171 million professionals who aren't actively job hunting, but they're open to new opportunities. That's kind of the magic, right? They're open to the possibility of joining your team and crushing it. This means you can reach top talent who aren't actively looking. It's a pool of candidates you simply can't find anywhere else. And we have hired so many of the most talented people I have on this week in startups all in and at launch my investment firm. So knowing the right information about these candidates is going to just make you super confident.