Tae Kim Sounds Off, Big Companies Are Hiring Again, NVIDIA $50B Tenant | Tae Kim, Ben Zweig, Aakash Thumaty

21:40Tae Kim:Yeah, Apple running the do-nothing win strategy. Yeah. Jensen doing thousands of deals. They didn't even sign the open letter. There are three companies that still, I believe, still haven't signed the open letter. Only three companies on the entire surface of the earth. No. There's three major companies that are – how do I actually get that to go away? I don't know. You can keep looking at Alphabet. But there are three major companies that haven't signed that NVIDIA open letter about banning open source and or not banning open source. And it's Amazon, Apple, and Anthropic. Anthropic put out a post yesterday, very, very clear response, sort of outlining their view on open source, their stance.

22:29We should go through it. But the interesting thing that Ben Thompson was talking about today was the fact that Apple and Amazon haven't signed. And they both have very physical elements in the world in the sense that they're sort of unsloppable. You can't vibe code an Amazon warehouse. You can't vibe code an iPhone. There are threats to those businesses, of course. And, of course, Apple should benefit from open source, and so should Amazon, because they'll be able to serve open models across AWS. But it's just potentially interesting. I think the Apple standing back is more just like, look, we're not jumping on with this crazy open letter that everyone is signing.

23:10Like, we just have our own brand. We're thinking different. We're doing our own thing.

23:13Tae Kim:Well, and based on other Apple AI timelines, I would expect them to sign it in maybe a year or two.

23:21Potentially. if they just signed it in 2028. It's just like robots.

23:26Tae Kim:WWDC 2020. We're ready. Now we're signing the open letter. These glasses have changed you. They turned you a new beast. Let me, let me run through the, the three anthropic proposals because it's an important response. So Dario Amadei, anthropic CEO responded directly to that letter, summarizing supporting open weight models that circulated over the weekend. So to summarize his, he makes three claims just to sort of clarify things that I think are important. One, he says Anthropoc has never advocated for a ban on open weight models. Now, that's a blanket ban. There's obviously like defining what a ban is and what an open weight model, what a distilled model, what a foreign model is.

24:06These things all matter. But he has come out and said, look, we never advocated for a total ban on open weight models. Two, he says undergirding all of this is the U.S. must beat authoritarian governments in the AI race. He points to China, but he identifies any authoritarian government. If they get really powerful AI, they'll come over here and steamroll us and you won't be free to do whatever you want to do in America. Three, powerful AI models may be misused to carry out cyber attacks or biological attacks. There are risks to having really, really powerful AI open source systems just running around.

24:41So he's worried about those three things, clarifying those three points. But he makes three recommended actions. He makes three proposals. First, he says, let's continue to sanction chips. Let's not sell chips to China. He says, we should not sell powerful chips or chip making equipment to China. So this has been debated for years, like going back to the Biden chip controls. Everyone knows every different angle on this, the basics. I mean, there is a pretty good argument for chip controls, even on purely geoeconomic competitive grounds. Like even if you don't believe in the risk of authoritarian governments having powerful AI, even if you just think it's like, you know, fancy autocomplete, it's like, well, it's the engine of our economy.

25:21And if you can slow down a rival economy, that's beneficial to you, right? And so, and there also seems to be basically unlimited demand for chips in America. So by restricting sales to China, that shouldn't actually hurt American chip companies all that much, but yes.

25:37Tae Kim:Well, they just like their argument would be we fully lose the Chinese market. Yeah. Which is the second largest computing market in the world. No. Right. So. So I think. But but you're the counterpoint to that is you were going to lose it anyways. Yeah. And a lot of that stems from the fact that China has been building an indigenous chip supply for decades. We've talked about going back to a whole bunch of their state-led, state-funded chip and fab processes. They've always been a few years behind. And so maintaining that gap, all else equal, is an advantage for the United States. The second point Dario makes is he says we should crack down on industrial-scale distillation operations.

26:24This seems totally reasonable. Companies can set their terms of service. and they have a right to maintain intellectual property with proper legal consequences for violations. Anthropic's been fighting distillation attacks, but according to them, it's not that effective. Dario proposes policy interventions to deter this behavior. And this is where I'm still not clear on where that goes next. What is the correct policy intervention? Policy intervention is a very, very broad thing. It can mean anything from a tax, a tariff, a fine, a sternly worded letter, not getting invited to a golf tournament.

27:02There's so many different things that policy covers these days, right? Where does this actually go? He says he doesn't want a blanket ban on open weight models, but it does seem like one possible policy intervention would be to sort of ban, restrict, or pressure open weights models that can be reasonably shown to have been distilled. So if there's someone who's just a perfect distillation, It just doesn't quite feel right. It's hard to quantify these things. We don't have a binary where you run some sort of algorithm and you say, yes, this was distilled. Because you can distill half on Opus 5 and then throw in a little GPT 5.6 and then mix in some Mistral and just be distilling from all over the place, fine-tune stuff, change the flavor, change the RL environment.

27:48There's so many different pieces of it. And, Tyler, you were making a point about Tinker or...

27:53Tae Kim:Yeah, I mean, so the Inkling model from Thing Machines, like, it used some synthetic data that was created with, I think, Kimi K2.5. Yes. So, like, does that count as, like, distillation? Like, probably not when people usually talk about it, but, like, it definitely benefited from Chinese open source models. Yeah. So, if you're downstream of Kimi. Yeah, I wouldn't call that industrial scale distillation. But it's sort of downstream of industrial scale. There is some big gray area where, how do you actually define these things? And so defining that is going to be the conversation that plays out in DC, like behind the scenes on the basis of this.

28:33And that's where the actual negotiation is going to happen between the position of NVIDIA and everyone that signed the letter versus the position of Anthropic and everyone who didn't sign the letter. They're going to sort of decide, okay, well, if you can prove this, this, and this, and you can show us that your API was getting hit by these different things and you have a really solid report of what happened. And then the model also sort of checks these boxes quantitatively. When we do this eval, then maybe we will pressure it. And then what does that actually mean? You could go after the lab that committed the distillation attack with lawsuits, but that seems really difficult given the international nature of these attacks.

29:14So we're sort of back to where we started, where, you know, you're like, what can the government do that the lab can't? Like the lab should be looking at every customer and saying, oh, this seems like someone who's trying to distill. They keep asking for basically what looks like a lot of training data. They're not acting like a normal user just being like, build me a website. Okay. Anyway, third, he says, all sufficiently capable models, open and closed, should go through mandatory safety testing. So this was recently outlined by Demis Hassabis over at Google DeepMind as well. And it seems like the two companies are in alignment on this in particular.

29:53And it's a somewhat reasonable position, although the risk is that small companies who have safe models that aren't distilled could get tied up in a review queue for years before they can release. Like, that would be very, very annoying if you're recursive superintelligence, for example, and you don't have a Washington, D.C. office and you're like, hey, we want to release our new model. And they're like, yeah, totally. Like, you got to go through the review process, get in line. And then it's like every, you know, every trillion dollar company is there with a ton of lobbyists being like, we'll review our model first because we want to get out a week before the small startup.

30:32And that's the frustration of biotech, the FDA, anything that goes through approval. We've talked about this with the nuclear stuff. It gets very tricky. And so you want to avoid that. And you don't want to wind up slowing down innovation that's happening on small scales and decreasing competition. Dario does do a good job of acknowledging up front that he says it would protect USAI companies from competition. But that's never been my goal with anything that he's saying here. And so it's still worth working through what happens in a really adversarial situation. Like, what if a foreign lab distills a bunch of frontier models that are the most aggressive, they're just distilling everything, then they jump forward a bunch in capability, they get a bunch of smuggled chips, they take all the restrictions off of cyber, all the restrictions off of bio, and then they just drop the weights on, like, a torrent.

31:21Or they put it up on Hugging Face, and Hugging Face is like, this is really crazy, no one likes this, there's a lot of pressure to take it down, I don't know. But it's out there. Like, what does the government actually do? Like the government probably pressures or bans like hosting the weights, maybe serving the model. You maybe won't be able to run it in American data centers. You go to the neocloud and say like, hey, this thing is actually bad. And I think people are divided on this because they see the current models not as actually dangerous, which is totally reasonable to assess. Yeah, it's not that bad.

31:53But like if there was a model that was like, yeah, it's actually just like the killing machine. Like I think most people would be like, yes, I'm democratically voting to not serve that because it's just like it's an annoyance at best and like actually at it worse.

32:07Tae Kim:And the other big question is like how much compute do you actually need for it to be dangerous? Yeah, totally. Is like having some GPUs in the back shed going to be enough? Yeah. Maybe for sufficiently advanced model, yes. Yeah. Or do you need access to a ton of racks? Yeah. A ton of power? Totally. And then you do need to work with a neocloud in that case. And as soon as you're a US-based company with a real data center, with a bunch of NVL-72s in there, you probably have registration and all sorts of just like business registrations where the government can reach out to you and say, hey, we're actually really worried about this.

32:44Just like there are other things you can't host in a data center. There's all sorts of stuff that's illegal, even if it's just intellectual property.

32:51Tae Kim:Yeah, exactly. Yeah, that's it. Like you can't even just because you have a data center doesn't mean that you can like take an open source, you know. You can't as a data center. Oh, yeah. Open source Marvel. Like they'll be like, no. Or even, you know, a CRM company can't knowingly support like a organized global cartel. Yeah. That is like trafficking narcotics. Yeah. Right. You'd have to imagine like that they have. They have to vibe their own. Chats and balances. Maybe.

33:26So what's interesting is like, what is the next step of that? So if there is a bad model and everyone agrees like, okay, yeah, we got to not host this, not distribute this. Like, yeah, the weights are out there. People are trying to like sort of run it a little bit. But does it go offshore? Do we wind up in like the crypto scenario where there's like these offshore things and people are using VPNs to get access to it? Like what level of aggression do you see from the U.S. government in that scenario? It probably should be proportionate to the danger imposed by the model. If it's just a model that's annoying or slightly IP infringes, but no one's really being like, I'm canceling my Disney subscription because this new model will generate me Disney IP.

34:07That's probably not like, okay, put up a crazy firewall. But if it is the ultimate hack machine that's stealing everyone's money from the banks, then yeah, you are going to put up the firewall and be much more aggressive. So I think the response will be in reaction to whatever the power of the models are, but it'll be interesting to go back and forth. Anyway, all in all, the letter clarifies a lot about the anthropic position, so I think it's good that it came out. But it's still worth working through the game theory of what happens down the line. Policy interventions is all we got here, and I think it's still too generic at this point.

34:40I want to know what policy looks like. I want to predict that. I want to understand what's actually being proposed, what people like, what people don't like. And so I think we'll learn more about this in the coming days. Let me tell you about Figma. Agents, meet the canvas. Your AI agents can now create and modify your Figma files with design system context. We have Tay Kim in the waiting room. Let's bring him in to the TVP on UltraDome. Tay, how are you doing?

35:05Tae Kim:Hey, guys. Doing great. What's going on? So tell me, last time you were on the show, you bottom-ticked it? What's going on? uh i think i made the bullish call on cpus memory and nvidia nvidia it's up like five ten percent but nice the cpu names have still doubled even after this big drawdown and the hbm names are up 100 so i'm hoping that you know it's the same thing again i come on here it's not to go up again yeah ideally we could have like an emergency reserve of take appearances so if the market is ever down strategic reserve we call you up you jump on and then it was funny because it was literally the exact bottle and you know it went exponential after that okay take him effect so where are we right now with the level of fud the level of uh downward pressure on the ai trade broadly the chips the semi trade like reset for us on like where sentiment is and then we can work through the different pieces of counter examples?

36:06So I think sentiment's very negative. We kind of had this huge parabolic up move the last few months, and likely a lot of retail and hedge funds piled in. And we're seeing this unwind now. I think the first big part of it was the Iran war getting worse. Every time we had the first ceasefire and negotiations, stocks started taking off right after that. And then when we had the actual ceasefire, we had a follow through. And then as soon as Trump started bombing Iran again, you know, chip stocks have kind of plummeted in the last two, three weeks. And then now we're seeing just, you know, back to the old pattern of media and the viral hot takes spreading a lot of FUD.

36:52I think we saw earlier this month, I think Reuters quoted like Zuckerberg's about agentic AI. They took it out of context. And then every media person was with the hot take that this meta was seeing bad returns that they're going to cut a CapEx. And then we had leaks right after that saying that it looks like meta is going to raise CapEx. So we're seeing a lot of this hot take FUD. Yesterday, I think we had a flurry of stuff that scared people. The Wall Street Journal vendor financing article that we'll see what happens with that. We had CMXT IPO in China, and everyone freaked out over that. We had the information article on ASML.

37:34We could go through each one. And then the Kimi thing, it's obviously a big thing. Yeah, we'll definitely get there. And I want to talk about open source and NVIDIA's strategy there, obviously. Starting with the Mark Zuckerberg news in Reuters. This was July 2nd. Meta's Zuckerberg says AI agent tech progressing slower than expected. Zuckerberg added that the company's reorganization that included major job cuts was not as clean as it could have been. Zuckerberg and other meta executives have been seeking to moderate some of the organizational changes introduced this year. And they said that the trajectory of agentic development over the last four months hasn't really accelerated in the way we expected.

38:16The company's bets on new structure haven't come to fruition yet. And so people were sort of reading this as maybe Meta is going to pull back, but then it felt like the response was extremely quick with Boz going on a podcast and Alex Wang sharing a whole bunch of progress across a few different models and data points. And then Semi-Analysis wrote a whole bull case for MSL talking about how they have compute and also they have more of like the internal structural alignment to sort of properly YOLO in the AI era if I'm boiling it down as brutally as possible. Just because with Google, there's always this debate between, oh, do you sell the TPUs or do you sell the cloud?

39:02Do you have vended in the product? Whereas Mark Zuckerberg is able to sort of like go all in on this new idea. And so maybe there's more glimmers of hope there. But what else have you been tracking downstream of Meta's ambitions? Well, I mean, they've been very upfront that they're investing heavily in AI. Alexander Wang is tweeting multiple times every few weeks that they're going full force. They're going to redo open source AI models. I think he said that at the YC event over the weekend. And it's, I mean, if you actually look at, and then Reuters came out, I think, with an article saying that they're actually going to raise capex dramatically uh this year and next year so all that kind of fear that that quote about a defensive ai from the town hall um that kind of like spooked the market for a few days it kind of it was completely false um the stuff like this yeah it feels like it's a comms error because the the language that's been coming out of meta has been a little bit like ai is going to replace our employees and it feels like it'd be much better for them to come to the market with a message of, we're going on the offensive.

40:12We're a hyperscaler. To be fair, that was the internal town hall. They didn't mean to leak it, and the writers leaked that one quote and put out the headline before the article. It's interesting.

40:23Tae Kim:Did Meta basically go through an eight-year period where internal town halls didn't instantly leak? I think everything leaked always. I think everything's been... I know, but there was a period where like the the the sort of attention of the media was way way way less on like what meta was doing internally relative to the 2010s and all that attention just went to the labs right yeah yeah no that makes sense um yeah i guess the question is like the question that i keep coming back to is like where is their revenue ramp where is their ai revenue going to ramp and when right because Because as you say ads, like the ads, like the AI has.

41:06Yeah. They're accelerating.

41:08Tae Kim:That's always been my view, too. But when you're when you're continuing to ramp CapEx. Yeah. With and saying, like, we're going all in on a gentick and we're building a harness and we're also going to do open source. And it's like, well, what is the strategy? Sure. Like, yeah. What is going to take you to a billion dollars of like pure AI product revenue? Yep. Or just API revenue, and then to five and 10, and what's going to allow you to justify the spend other than, I think the market would love if they just said, yeah, we actually need all these GPUs because we can actually be 10 times, we're already good at ads, we can be 10 times better, and that's where we're going to get the ROI on all of this CapEx.

41:53Well, they're definitely getting ROI on that. The market is worried about all this extra CapEx. They're going for the Frontier AI model race again. and they had to reset. They had to, a lot of people left and now Wang hired a ton of people and we'll see what happens over there. It's going to take time. It's going to take six to 12 months before we see any more progress. But that first model that came out a few weeks ago was a lot better than people expected. It wasn't the frontier, but it was much better than what people expected. Yeah, yeah. So how have you been processing the NVIDIA letter around open source and all the back and forth, all the people jumping on, the companies that have been staying back.

42:38How do you work through that? It's been very impressive what they've been, they've basically united the entire tech industry against Anthropic in the last like three, four days. 18 trillion in market cap has signed on. Last time I checked across. I mean, Google took a little time. Amazon signed on eventually. Oh, they did? Yeah. They signed on yesterday. They tweeted out. Interesting. I think Apple is still the holdout. Apple's the one, which is kind of strange because they're the one that would most benefit from open source, open weight models being more available, I would think. I would think so, too.

43:14But I mean, they pretty much got the whole tech industry to kind of corner Anthropic in their position. Open AI signed on. What did you think of Jario's? Obviously, NVIDIA is afraid. Yeah, I don't know how if, I don't know if they're really, I don't read it as being like cornered by any means, right? Well, Jensen is on the record that he said, I think to Bloomberg, that there was a rising sentiment that something was going to happen on the regulation front at the White House or whatever. So this was last week.

43:54Tae Kim:You had at least four people in the admin say, we're not against open weights, we're against distillation. And at least I was reading into that of some type of regulatory action around open weights and then positioning it as we're targeting. this is like we're hitting yesterday yeah about you know pushback and restrictions and he's doing it under the safety umbrella but definitely Microsoft and video are worried that the White House or Congress is going to do something on this front and that's why it seems very reasonable that he would have no problem with like Gemma or Llama or any of the open source from like American hyperscalers where if you find out that they're distilling, you just walk across the street and sue them.

44:41And also these big companies have huge, huge, I mean, they have safety teams, but also just like huge incentives to not have a safety incident happen on their watch because you're trying to like catch up to the frontier. And then all of a sudden you have a safety incident. That's going to be really bad for your overall brand. And you have a different business to protect, whether it's social networking or Google search if all of a sudden the Gemma model winds up being a thorn in someone's side for a cybersecurity reason or a bio reason, that would be really, really bad. But a foreign company that is just like hurling it over here can kind of just be like, you guys deal with the consequences potentially.

45:20So I think that's what Dario is worried about. What about the overall idea of like where it feels like we're sort of replaying the deep seek moment, open source is going to reduce cost. And so that's a reason to pull back on the AI trade overall. How have you processed that? It's almost a perfect catalog. People are worried about Kimmy. But when you actually read the technical paper and their blog posts, this is not a tiny efficient model. This is 2.8 trillion parameters. It's going to require a ton of compute to serve. I mean, we saw it the first day they put it out, but their servers got slammed.

45:57And even in the blog post, they say it's best run on a server with 64 GPUs. So big, super clusters that are networked well, and that perfectly runs great on NVIDIA. And if you remember during the whole DeepSeek thing about a year or so ago, the market freaked out that DeepSeek was so efficient that it will lead to a compute glut. But DeepSeek was an example of a reasoning model that actually was the opposite. created a ton of demand. And I think the same thing has been happening with Kimmy, where when you have more capable models that come out, people find uses for them. And right now, just like last year when reasoning models took off, Agendic AI and agents are taking off right now.

46:43And the market is kind of like not realizing that because right now, just like last year when reasoning models were taking off, right now Agendic AI is taking off. And the next six, nine months are going to be bigger than anyone believes. and Sam is on the record. Sam is on the record over the weekend. At the YC thing, again, like people don't, I don't know why people don't listen to you. It's on YouTube. The next six months, it's going to be much more dramatically better for AI than the last two years in terms of advanced capabilities. And I heard you say RSI before. I think it's going to be RSI.

47:19People inside OpenAI and definitely Anthropic. Anthropic put a blog post on this. RSI, I think, is a lot closer than people think. And if RSI actually happens in the next three, six, nine months, that's going to soak up an insane amount of compute. I mean, we have this exponential ramp for reasoning, exponential ramp for agentic. And if RSI actually happens, and I think it sounds like both frontier labs think it's going to happen very soon, that's going to soak up an unbelievable amount of compute as the AI models use more compute to self-develop and improve. And I think that's one thing that people are missing, that both Anthropic and OpenAI are kind of winking that, oh, it's happening.

48:00Anytime I tweet something on RSI, all these Frontier AI researchers like my tweet.

48:05Tae Kim:So I think that's good. What is your sort of framework around compute hoarding? Because certainly it has been happening. When you look at, you know, like going back to the meta example, right? They're not selling compute yet. They're maybe curious about it or exploring some deals. They have all this compute and they're betting on their own ability to create the capability that will have enough demand to justify that. Do you just think there's so much demand overall that it just, you know, even if there's hoarding, it just will leak out and it's okay? Well, there's so much demand overall. Well, I mean, the SK Hynix executives said during their IPO run that their customers are asking five to six times more than they're able to serve.

48:59And they're going to double capacity over the next five years, they said. And their customers, and I'm going to assume it sounded like Jensen, are asking for five to six times more than they're able to build. So there's overwhelming demand. You guys were at the Advanced AI A &D event. that Lisa Su raised her agentic CPU forecast just three months ago it was$120 billion for 2030. Three months later, they raised it to $220 billion. She doesn't do that. You have that just on the

49:33Tae Kim:I've got that ready. Well, I just love this chart because he called it perfectly. He actually did. It's crazy. CEOs don't raise their tans by these multiples in a few months if they're not seeing insane demand coming in. Especially not public CEOs who are serious business leaders who've been running non-meme stocks for decades and are serious people. Everyone's freaking out that this is the dot-com bubble all over again. But what if these hyperscaler GPU cloud businesses are amazing businesses? Morgan Stanley says, if you do inference, it's 60 % to 80 % profit margins. These are amazingly profitable businesses as long as we keep growing the next few years.

50:20And again, just like last year, we're on this exponential run right now over the next two quarters. And the market isn't seeing that. Everyone's freaking out that, oh, no, we're spending too much. And even Sam Altman podcast came out today. And another podcast is out there. He said that he regretted pulling back on the compute purchases. They made a mistake by not putting the pedal to the metal because now things are taking off again. So, like, Amazon, the CEO, in April, if everyone read his annual letter Andy Jesse wrote, he talks about how free cash flow works. We're not betting$200 billion on a hunch.

51:03We see the demand. We know it's going to be insanely profitable and free cash flow positive in the medium to long term. So that's why you're investing$200 billion now. And in a year or two, we're going to see insane amounts of free cash flow, the thing that people are worried about right now. It takes time to build out these data centers and fabs. And you bet now to bring that in a couple years. Like if you see free cash flow, that assumes that like the revenues have to catch up and then the CapEx can't grow more exponentially. And so that means you have to see some sort of plateauing. Maybe it's at the end of the chart.

51:42Maybe it's this 2030 range. But there is a different world of just like continued growth forever. And then we sort of run out of money. The pushback I have there, that's a static view, right? If they don't grow revenue for the next three years, yes, you can't do that. But Azure is growing 40%. Google Cloud is going 80%. Amazon is growing double digits. So if revenue is growing 40 % to 80 % this year or next year and the year after, that's more revenue you have. That's more operating cash flow you have to invest, right? Yeah. So that's what people are missing. And if the data center that you're building now, you're spending all this now, generates unbelievable free cash flow in 12 to 18 months because this agentic AI is actually aging and re-architecting all the workflows inside companies.

52:39And you need to do the agentic AI coding agents to make your product better. because if you don't iterate 100 different iterations of your product in R &D, if you don't do AI, just like AT &T is doing at the Gen Tech Advancing AI at AMD, he's talked about they're putting 100 Gen AI models into production. They're burning a trillion tokens a month, and then that's growing double-digit. The reason why they're doing that is because by using Gen Tech AI, you're providing a better customer service, you have a better product R &D, and you're helping your companies make better products and services. And if you don't incorporate AI into your company, Verizon, your other company, is going to incorporate AI and then disrupt you, and then you lose all your revenue.

53:27So everyone's worried about ROI. ROI is important, but you also need return on revenue because if you don't use AI, your rival is going to use AI to beat you in the market. Yeah, yeah. I think the diffusion story is still, even though we got like sort of jitters by the token maxing thing just the actual usage of ai across companies is still pretty limited in terms of the amount of people that are using it the time that those people are using it like there definitely is a san francisco bubble of startups where everyone is using ai a lot but if you just walk into a normal business a lot of people are like yeah i gotta check that out which is let me give you some some context here some yeah are

54:09Tae Kim:Mark Horozian, Jared Sleeper, everyone, actually saying enterprise adoption disparity remains enormous. And he cited Ara saying usage would 100x if every company adopted AI to the degree of the most advanced companies. There's a small group of companies that are – People forget in the RAMP data, adopting AI can mean having a ChatGPT Pro account for someone, which is not exactly the same as using codecs and coding agents and stuff. It's important. I think that, you know, if I have someone on my team, I want them to be able to go and do a deep research report. But that's like table stakes. The question is, like, are you actually speeding up anything that's repetitive in your job?

54:52And that diffusion is just starting to take hold. So the total market size in terms of IT and knowledge management in corporations, it's about$6 trillion, right, a year. the two main frontier AI model companies, OpenAI and Anthropic, I'm going to say, I think this is roughly accurate, are doing$120 billion combined in ARR. Why can't that go to$200,$300,$400 billion in the next year or two? I mean, they're growing at exponential rates when we're taking off. And if the market is$6 trillion, why can't they grow to$200,$300,$400 billion in the next couple of years? I mean, it's like just do a little logic and rational deduction.

55:36This is definitely possible, and it's happening right now, and it's accelerating, and people aren't, you know, they're just taking, you know, these big headlines where we had this, you know,$50 billion for Financial Times, and we find out it's over 30 years. It's like on the homepage. Wait, wait, yeah, yeah. Okay, I wanted to ask you about this. NVIDIA revealed its tenant for$50 billion data center that will use its chips. Explain what is actually going on here. So the Financial Times put on their homepage today that NVIDIA is going to backstop a lease for a data center in Texas for$50 billion.

56:11And I saw that. I was like, oh, my gosh. Oh, that doesn't sound good. No, it literally sounds like they're buying their own chips. It sounds like the most bad thing you could do. Yes. Then they actually read the article like halfway down the article. It's like a 15-year lease, and it's only$50 billion if they renew the lease after 15 years. So it's like over 30 years if they renew it. Then if you think about that, you're like, wait a minute,$50 billion divided by 30 if they renew it. Oh, okay, yeah. NVIDIA's 15-year lease commitment for the Texas site is worth basically$20 billion, and renewal options would take the total value to$50 billion over 30 years according to HUT8.

56:53Tae Kim:Okay, but what are their plans for the site? Is this they are going to have some, like, what do you expect them? So my point is this is a billion, you know, whatever, a billion or$2 billion a year, right? It's a non-story, but it's a big headline, sensational headline on the homepage. Yeah, and also it's not like you're taking a$2 billion loss every year. you are the tenant and then you are also renting that out. So hopefully you're making profit. It's a rounding error. It's like, you know, they're doing 320 billion run rate a year now. That's going to go to 400, 500 billion next year. And we're talking about something that might be a billion, you know, like this is not a story, but this is how people run with the sensationalized headlines and people panic and freak out.

57:47I think they just wanted to say the biggest number. That's exactly the point. And we're going to see what happens with this Wall Street Journal article. Both OpenAI and NVIDIA are not commenting so far. We'll see. But take us through the rumor. We have to wait. Rumor will time. Well, it's not a rumor. It's the Wall Street Journal and other people reporting that. Yeah. NVIDIA is in talks with OpenAI. It's a backstop, soft bank, up to$250 billion. We don't know the details. And I don't want to speculate and comment. But let's actually see the details before we – I think the market had a really big negative reaction yesterday to this story.

58:24Because everyone, I mean, Jim Cramer was telling his audience, like, sell everything at the open today. Because AI and data centers, dot com, you know, it was insane. It's just, let's see the actual deal and the metrics and the numbers before we panic and freak out. Yeah, yeah, yeah, yeah, that makes sense.

58:45Tae Kim:so yeah honestly when you say freak out and sell everything sell your dollars sell your house sell your house sell your stocks then i'll freak out but until then tay i feel i feel okay i mean i just see the fundamentals i see the ceo of amd expanding her tam you know dramatically over the last three months i see rsi under horizon like every ai researcher is like oh my god this is going to happen. We have to get there sooner. And then I see the obvious use case of agentic AI, where you have to re-architect your workflows internally. Every company has to do this. So everything is taking off. You see when the president of Korea came to San Francisco area last week.

59:33They had a day in the valley. Instantly, NVIDIA CEO Jensen Huang, Broadcom CEO Hock Tan, Dario, Sam Altman are there, right? Do a little logic deduction. Why are they there like crazy? Because they need HBM memory and they're dying to have it. So if you think about that, that means there's insane demand and HBM memory is in shortage. There's tremendous demand for it, right? Talk about the NVIDIA CUDA mode. It feels like a big piece of AMD's advanced AI event was maybe the CUDA mode isn't as much of an issue anymore in the age of agentic AI. You can have an AI agent write you the software that you need to use any chip, and that creates less pricing power for NVIDIA.

1:00:29but there's another world where you're not really like nvidia doesn't necessarily need a moat because everything's just growing so fast that they're still growing but how have you interpreted the processing of like the potential death of the cuda moat so amd is on it kimmy wrote like a couple paragraphs in their blog post about how they created a gpu kernel all that so everyone you know sure get scared or whatever it's like we'll see what it's like in real life you know this is It's just, you know, AMD is incentivized to say, oh, Kuda is not a problem anymore. Kuda has been a tremendous moat. And I think it continues to be a moat.

1:01:07And the reason why is it's super reliable. All the bugs have been optimized and fixed. And that comes from hitting the software, you know, millions of times and millions of times. Right. Like you don't know if, you know, if you use ClockCode or Kimi that what they figure out using their training data is going to work in the real world. Right. they could talk about one little piece that does well. Let's see how it actually works. But NVIDIA's big moat is its scale, its co-design of actually working through the networking, the CPU, the GPU, and how everything works together. And the other big thing is their balance sheet and their ability to get supply commitments from, you know, I think I said this before, optical startups are like upset because NVIDIA secured all the supply for all the optical components.

1:02:01Same thing with TSMC wafer, same thing with HPM memory. So NVIDIA is using their size and Gorilla and being able to prepay and get components that are in shortage. So they've become the dominant, over the next year or two, you're going to see NVIDIA able to add tons of revenue because they were able to lock up all the supply components. That's another thing that people don't really talk about is their supply chain and their ability to work with partners and secure component inventory. Is there still energy FUD that we would run into an energy bottleneck before we run into a chip bottleneck? So Jensen said this last week on the Bloomberg interview that there are a lot of bottlenecks, including data center shell, power, and all those things, components, energy, whatever.

1:02:48So all those things. It sounds really bad, right? And then right after that, he said, I think the chip industry has enough supply to double their revenue every year, basically implying NVIDIA has enough supply for energy and all that stuff. No one is pricing that in. So everyone talks about bottlenecks. NVIDIA CEO just basically told you on Friday that they have enough supply chain and all the bottleneck stuff to double revenue every year. And no one, you know, NVIDIA's revenue estimates for next year are a lot lower than double. I'll tell you that.

1:03:23Tae Kim:Do you think the market prices in just how much of almost every important AI company in every category NVIDIA actually owns? Like it feels like every single, like we're constantly focused on who's going to raise CapEx next and where is this quarter coming in? And it feels like in two or three years, people will look at NVIDIA's balance sheet and be like, wait, they have what I imagine then will be, you know, could end up being a trillion dollar plus of just like ownership and all of these great companies, which, again, just goes back to the advantages of that early scale. while all these companies are trying to compete away NVIDIA's margins and all these different things, they've been able to accumulate, again, positions in all of these incredible companies.

1:04:16Tae Kim:I mean, we saw the SSI news yesterday is a great example of that. But how do you see it? So I think look at Jensen's history in investing in these companies and CoreWeed and see how much money they made. They just bought, staked in the optical companies, Lumentum and Coherent. Jensen is enabling the future because he sees this overwhelming title of demand, and he needs these companies to be able to build up their supply chain and to give supplies and chips to NVIDIA so they actually ramp very hard. Everyone's freaking out that this is vendor financing. saying, what if hyperscale GPU cloud is so profitable and these companies need capital to build up that supply so they can serve the GPU cloud services over the next year or two?

1:05:09Maybe Jensen sees that coming like he did with all these other companies like CoreWeave. And that's why he's investing in these companies to be able to expand their ability to make the components the industry needs. So I think you're exactly right. In a year, two, three years, NVIDIA is going to have all these stakes in these companies, and it's going to look like he was a good investor because he has been in the past. I mean, the man can buy a leather jacket for like five grand and sell it for a million dollars. I don't know what else you need to see. I mean, think about Melagon. Are you the secret bidder?

1:05:45Did you win that? No. We've got to get you a jacket. the real question is how long until someone distills a jacket and open sources it you can get a dupe of a Jensen jacket for two bucks that's what I want thank you so much for coming on the show Jordy you got anything else?

1:06:03Tae Kim:this was great thanks for putting up with all of our jokes hopefully this becomes the lucky charm for the markets yes I agree I agree i agree we'll talk bottom is in great to see you tay have a good rest of the week goodbye let me tell you about crowd strike your business is ai their business is securing it crowd strike secures ai and stops breaches um i wanted to run through on the power issue there's an interesting article in the journal an underground nuclear reactor is coming to this kansas town and it's dividing locals this is something that i had tweeted about years ago like why don't we just put the nuclear reactors underground, put solar panels on top, best of both worlds, optimal use of energy.

1:06:46But there's a lot of fear and uncertainty and doubt about this one. They say no one has tried operating a commercial one a mile down until now. It's great that it's here. It's kind of bad that we're the guinea pig, says the residents of Parsons, Kansas. Residents of the sleepy farming outpost agree on many things, but whether to put an experimental nuclear reactor a mile deep in the granite beneath their town isn't one of them. Elected officials and some others see a chance to create jobs and lure data centers and manufacturers to a rural patch whose economy has been flatter than the surrounding cornfields.

1:07:22Another group is effectively saying, not under my backyard. It's newbie, not nimbie, because it's not under my backyard. Noombie or something like that. I put$125 ,000 into my house and now a nuclear reactor is coming to town said gerald johnson an it professional who planned to retire in parsons i can't think of a worse idea no one has tried operating a commercial nuclear reactor deep underground until now i'm surprised no even like the soviets in like 1950 didn't try it i feel like they were trying everything i'm surprised yeah but the so-called gravity reactor is is the creation of liz mueller and her father richard mueller emortus professor of physics at the university of California, Berkeley, it's Berkeley people again, and an inventor, they founded Deep Vision, a three-year-old California startup that raised$150 million in the past year, including$40 million last month through an IPO, largely to fund the work in Parsons.

1:08:18Parsons, with a population of 9 ,600 people, sits about midway between Kansas City and Tulsa, Oklahoma. Deep Vision drilled a first test hole this spring on a 100-acre...

1:08:27Tae Kim:Office chairs like that have really fallen off, They have. Which tells you now might be the time to bring it back. No. Right? I need a new office chair. I might go for one of those. The high-back leather. It's good. Good. Deep Fission drilled a first test hole this spring on 100 acres at a mostly overgrown industrial park dotted with old munitions bunkers just outside town. On a recent day, Maurice LaFountain, division's senior engineering director, showed off a pink-flecked granite retrieved from the company's first test hole and joked that the billion-year-old rock would make a nice countertop.

1:09:03An empty steel canister sat on a cleared, drilled pad waiting to go down a second hole this year. The plan is to send another one loaded with nuclear fuel into a third hole to heat water a mile underground and generate electricity on the surface in 2027, 2028, an astonishingly short time frame by industry standards. Interesting. Anytime you're putting a nuclear reactor in a hole, it's kind of scary, he said in his office. It's great that it's here, but it's kind of bad that we're the guinea pigs. Very interesting. I'm surprised we haven't heard more about this company, this idea, everything that's actually being planned.

1:09:39And there's something a little, I understand where they're coming from. There's something a little bit nerve wracking about, like, even though you would think a mile deep, if something goes wrong, it's less of an issue. It feels like, well, people, it's harder to get to and, like, just go and solve the problem. Deal with it. As opposed to, like, oh, yeah, it's a building over there. I see people coming in and out all the time. The experts are in control. I don't know. What do you think? Are you pro nuclear underground, a mile underground? Could be the future.

1:10:08Tae Kim:Could they not find maybe a place to do that that wasn't right under a town? It's not right out of town. It's outside of a town. You need some infrastructure. Yeah, yeah, yeah. I bet what they would say is like, look, it's a 10 ,000-person town. We went miles away. We're on 100 acres of land. We are outside of the town. But, yeah, there aren't that many places that are truly uninhabited for hundreds and hundreds of miles just because of the nature of America. There's towns all over the place, every street. And you need roads to be able to deliver equipment and whatnot. Anyway, let me tell you about Codex.

1:10:45Codex is a powerful workspace for getting work done with AI agents. Whether you're writing code, analyzing data, creating content, or automating business workflows, Codex helps you move projects forward from start to finish. What is this? How did we get here? Anyway, we have Ben Zwieg from Reveglio Labs coming on the show. How are you doing, Ben? Good, good. I love that intro.

1:11:07Tae Kim:That was just for you. We're testing that out for the first time. Nice and matches the vibe. The vibe of the labor market. Take us through a little bit on your background, how you work, and then some of what you're tracking in the labor market and how it ties to your actual business. Yeah, for sure. So I'm a labor economist, been tracking labor market data for a long time and started Rebellio Labs. So RevealioLabs is a workforce data company. We're collecting, curating, synthesizing all labor market related data that's out there in the world. And of course, you know, a big question is how is AI affecting the labor market?

1:11:43Of course. So, you know, we're uniquely positioned to answer that question and it's on everyone's mind. So we started putting out this labor market. This is kind of AI labor market tracker, which is really about answering like how is AI affecting the labor market today? So not really getting into the speculation of what might happen. Yeah, just today. But really, what do we know? Really quickly, what is your business model? Who gets value out of this data? And then I also would love to know, how do you go about getting more accurate data? Because I see, obviously, the Census Bureau, the government has access to do polling.

1:12:19ADP is a very logical place to get data because they run payroll so they can see the data. But what's been your strategy there? And then who's the customer? Yeah, yeah. So I'll start with the customer. So a lot of it is hedge funds. So they're speculating on the performance of companies. Yeah, nice. Thank you for your service. It's here for hedge funds. Yeah, they don't get a lot of love these days. But yeah, they are speculating on the performance of a company that they have no affiliation to. So you have to understand what's happening in the company, the workforce dynamics, HR departments for benchmarking also.

1:12:52So strategic workforce planning, people analytics, talent intelligence. These are all kind of segments of analytical HR. and academic research. So, you know, they, of course, want to know what's going on. So these are, so basically we get the data not through surveys, not through payroll, but really from the internet. So, you know, LinkedIn profiles, job postings, Glassdoor reviews, layoff notices, immigration filings, freelance platforms, like anything and everything that is in the public domain. And that information has to be, you know, enriched and synthesized in a smart way. Like there's all sorts of sampling biases, There's lags in reporting.

1:13:30There's raw text. So we have to classify that to occupations, to skills, seniority levels, and importantly, work activities, which is more of a recent thing for us, but important these days. And then I imagine you can sort of backtest against the historical actuals to see that the model is working and then you can be more up to date. So we don't backtest against financials because if we did, no one would trust it anyway. I mean about like if you ran your model on like what was the employment rate in 2021, you could look at the actual employment rate to sort of calibrate that your system is predicting employments correctly.

1:14:04Is that roughly correct? Yes and no. I mean for some models, we can see what was retroactively revealed. So when someone changes their job, they don't necessarily update that right away. And we can see that every timestamp has like more information than they said before. So that's like a solvable problem. But in terms of saying what's happening in the labor market at large, we can kind of use BLS data. So BLS is a Bureau of Labor Statistics. We can use that data to kind of like proxy for it. But that's got issues itself. So I don't know if we want to use that as ground truth. So I think BLS has a view on what's going on from survey data.

1:14:40ADP has from payroll data. And we have from Internet data. And they're all kind of independent in their own way, kind of uncorrelated errors. Okay, so I want to understand how you look at your data in the context of AI diffusion, right?

1:14:57Tae Kim:So a company, an individual company or an industry might have fluctuating labor data, right? Maybe they're adding a lot of people. But then individually, if you look at those companies, maybe some companies are adopting AI quickly. some companies in that sector aren't really adopting AI at all, or they're doing it in a very minimal way. Let's say they just have a basic, you know, ChatGPT$20 a month subscription. So like, how are you, I was talking to John maybe, was it six months ago? I was saying like, I really want there to be a firm that is just studying AI diffusion in specific industries and getting into the nitty-gritty, probably doing surveys to actually understand how, because every company says they're adopting AI, but we all know that there's such a broad spectrum.

1:15:48Tae Kim:And then, of course, some people are saying that just because they want to be and feel like they're a part of the club. Yeah, yeah, I think it's probably mostly those that want to be part of the club. But I agree. I mean, so there's a few ways to get at adoption data. So I think adoption is the hardest part of all of this because that's really a firm level piece of information, whereas AI exposure is like more of a person level piece of information. So I'll tell you the way – we do it in a couple of ways. So one is that we had a partnership with – we still have a partnership with Ramp. So I know, friend of the pod.

1:16:22Let's go. So they can track adoption just using like AI spend. Yeah. So they can see like dollars spent on tokens, et cetera. So that's like a pretty good way to get adoption. The problem with that is that, first of all, it's like a self-selected sample. You know, ramp skews toward more like tech, which is fine. Like that's overcomable. The other issue is that it's anonymized. So they can't release information at the firm level. So, you know, when we collaborate with them, like we have, you know, the labor market data and they have the adoption data. So, you know, it's like complicated. You know, we have to send the data.

1:17:00They have to like run something. We have to do some matching. So it's like a little bit, it's got some friction. The other way to do it is through, we use this measure, which is used in a paper, a recent paper that measures adoption by like sort of hiring AI integration teams. So the thought is that, you know, if someone's like hiring AI integrators, you know, beyond some threshold, that they're like taking it seriously. And they're embedding it into their business processes. And by that metric, we see about 9 % of firms getting very serious about AI. It's a very conservative way to measure AI adoption, but it seems to be pretty good.

1:17:40It's correlated with all sorts of other things. Yeah. I sort of hate that idea as a metric, but it probably makes so much sense in larger organizations that that is a great signal. But it just feels like completely the wrong way to go about actually changing a business. Like, I feel like adoption should be so much more ground up than like, oh, we're hiring a special team to do this. But that's the way businesses work. And I think you're correct to identify that. It probably is very indicative of a change in the stance of the business.

1:18:09Tae Kim:Where is an area that AI is really good and you're seeing job loss? Hmm. Good question. AI is pretty good at software engineering now or generating code and, you know, the companies that are adopting it the most are hiring a lot of engineers. Whereas I've heard in L.A. specifically, apparently the models that do product photography. So men and women that, you know, wear a bunch of clothes for like an old Navy when they're releasing a new collection. like that work has been very impacted because that talent they don't have a brand yet right and so maybe certain companies will just say like yeah let's just take this shirt that we have and just generate it on 20 different ai models and you're uh and we're good to go right it just doesn't doesn't really matter that much if they're using real talent or not and so they choose the easier cheaper route yeah i think that's a great example i mean for the most part you know across the board adoption is generally correlated with growth.

1:19:15But where we're seeing reductions, I mean, I think the creative fields are a great example. So, you know, if you need like video B-roll or just like, you know, stock images or just, you know, podcast intro music, you know, that is like very easy to get from these kind of AI generated, you know, creative elements.

1:19:34Tae Kim:sure yeah so it's so fascinating because how many people yeah it's like it's just quite interesting because when you when some of these things how many people were actually in those roles like would it actually does it show up in labor data at a large scale at all right people that are just doing you know stock photography and making their living that way or Yeah, and a lot of these people might have sort of sloshed around. Like, I mean, I see Instagram reels from people who years ago were posting like After Effects tutorials, Premiere Pro, DaVinci Resolve, like little video editing tutorials.

1:20:15And now they're posting like AI-enabled workflows. And instead of showing you how to deal with a green screen the old-fashioned way, they're just doing it the new way. And they're probably still doing it for clients. and the client spec is just like, I need ads that convert and they're just doing more of the work, but then there's other stuff that's bleeding out, all sorts of different stuff. Yeah, I mean, one kind of framing I would put this in is that we're seeing a lot of automation of things that are very task-based, things that are really micro jobs that aren't full jobs at all. So we're seeing declines in freelancing across the board.

1:20:54So freelancing has hit pretty hard. but that's really an environment where people transact and tasks. They're not, yeah.

1:21:00Tae Kim:Yeah, this is why we were just talking about this earlier. Historically, if you needed a really specialized website, it's not your main site, but let's say in our case we're doing a drop. Three years ago, we would have gone and maybe gone to Upwork and said, hey, I need a simple website made and just find somebody to do that one-off. And now AI is just so good. Or like a basic logo for a first draft. would be like a 99 designs. Before you bring in like a real branding firm, you might just get a freelancer to mock something up for you. Now image generation can do that for sure. What do you make of the computer science shifting?

1:21:38Because there are so many opportunities for entrepreneurs, startups are growing. There's some tech layoffs, but at the same time, it feels like just in general, if you have a computer science degree, you're probably gonna be a bit better at using AI broadly. And so there's lots of opportunity And yet the number you have here is computer science enrollment is down 28 % from its 2022 peak. Yeah. Yeah. So I have mixed feelings on it. First of all, it's very dramatic. Yeah. So one thing that kind of one optimistic take is that the supply side of labor markets is actually quite responsive to changes in technology.

1:22:17And that wasn't obvious before. And, you know, if people can reorient themselves flexibly, that's great. That means we can be adaptive, we can have more of a dynamic economy and worry less. So I'm encouraged by that responsiveness. I think it's an overreaction for two reasons. One is that we are not seeing declines in employment based on the firms that are adopting a lot. And that's true in engineering, it's true in tech. We're not seeing mass layoffs despite the narrative. So I think it's premature for that reason. Another reason is that I think even just a couple years ago, maybe even less, I mean, time is like elusive to me, but I think, you know, not so long ago, you know, we thought of AI as chatbots and code assistants.

1:23:04And now it's more agentic tools. So it used to be such a low barrier to entry type of technology where, you know, anyone's grandma can use it. And coders, engineers, were really just replacing their work at high rates. Now we're seeing complicated tools. Agentec systems are hard to use. They kind of favor the digitally native and people who have experience with engineering. And even when they're easy to use, there are a whole bunch of, from a business, from an enterprise perspective, cost tradeoffs, privacy, security, how deep is this system? Even just firing up a coding agent today, you're hit with prompts like, do you want this to have access to your documents folder?

1:23:58And that's a question. And a lot of consumers are like, I don't know. And a lot of businesses are like, I don't know. So there is some sort of capability overhang. Yeah, yeah. And I think it's a different job than it was before. Engineers are spending less time doing the front-end engineering for a website, but they're doing more of kind of that DevOps. So I think it's premature, and I think we'll – I mean, I suspect we might have a shortage of engineers in the way that now we have a shortage of radiologists. Everyone was nervous that radiologists were going to be a thing of the past, and now there's a shortage, and wages are super high.

1:24:40It's like the final boss of AI automation. Every AI researcher is like, one day I'm coming for you, radiologist. You imagine that it all started with a radiologist just bullying an AI researcher and being like, what you're doing is so useless, and the AI researcher is like, I'll show you, radiologist. I'm going to put you out of a job and the radiologist is like, I'd like to see you try. And then years and years go by. Talk to me about hires to posting ratio. It's down 38.6 % since late 2022. I can imagine that there's a lot of slop posts. We were debating this before, but how do you tease that out?

1:25:22What do you make of the hires to posting ratio dropping? So this is the thing that I get the most nervous about. So, you know, we're seeing some slop posts, some slop job postings. But we're also seeing a lot of slop applications. When a job goes up, you know, you get, I don't know if you guys have posted a job recently, but I just did last week. And I got, you know, a thousand applications in the first like five minutes. There are all these like job boards that are kind of helping people auto apply. Even Indeed is doing this, which I think is a bad move for the record. but they'll do what they want.

1:25:57So basically employers are getting completely signal jammed. They're getting overrun with these applications that look strong, but they really have no way of verifying. So the utility of each job posting is going down. It's not as good of a way to find candidates anymore. So employers are relying on networks. It's getting harder to hire. And in the economy at large, we have this kind of low hire, low fire environment where there's just not a lot of movement in the economy. And I think that is the result of, you know, AI usage in the search and match process. Yeah, you would think that, like, I've been surprised that social media has not been that overrun with slop.

1:26:41Like, there's definitely some slop problems here and there, but in general, the algorithmic feeds have been sort of set up to deal with this where the bad slop gets filtered out pretty quickly. Except for LinkedIn. But yeah. Sure. But I've been surprised that there hasn't been as much of an intermediary where you put up a job post. Yeah, you get hammered with a thousand applications, but the filtering is really, really good so that you're really only looking at the top 10. Maybe you dip into the top 100, but you're not at all annoyed by the bottom 900. Because I guarantee you that there are millions and millions of sloppy Instagram videos out there that would annoy me if I saw them, but the algorithm will just never show them to me.

1:27:28And then maybe there's one that uses AI, but it's good and it will show it to me because I still enjoy it. So it feels like, hopefully there's people working on this. I'm sure that people are, but that feels like the next iteration to unclog this? Because that seems like a major problem. You need the matching in the US economy to be really, really strong. Yeah. I mean, there's been some regulatory challenges there too. So a few years ago, it became illegal for employers to sift through candidates using AI. And I don't know how enforced that is, but it's a liability for employers and not a liability for candidates.

1:28:04So there's some asymmetry in who can use AI. That's very interesting. I had no idea. When did that yeah i remember i remember that that you can't use ai to to to to filter out candidates i think of it as like so i i i understand where that came from on like bias based into models and like very preliminary uh deep barely deep learning algorithms to sort of like look at the person's name and look at the graduation date and like try and filter for that like i'm just thinking about like, is the resume complete slop? You know, like a complete, like a pangram level that doesn't seem to impose like bias in the same ways that they were trying to avoid.

1:28:45So we're in this weird like knock-on effect world. But that's the way these things go. Jordy, anything else?

1:28:50Tae Kim:No. Come back on as there's more. Come back on as there's more data that's notable. You can tease the hedge funds a little bit. Give them a taste. Yeah, for sure. And congrats on the progress. Thanks so much for coming on. Yeah, great to meet you, Ben. You can talk to me soon. Cheers. Have a good one. Let me tell you about MongoDB. What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI. Own the data platform that powers it. And let me also tell you about Cisco. Critical infrastructure for the AI era. Unlock seamless real-time experiences of new value with Cisco.

1:29:23Up next, we have Akash from Takeoff. He's the founder and CEO. He's been on the show before, but this is the first time.

1:29:29Tae Kim:Here he is. Find a new flag. How's it going? Give us the news. It's going well. It's great to see you guys. The news is that Sierra just bought us. We announced it last Thursday. She already completely missed. That helped me. Let's go. That's the first. First miss. First miss. Brutal. Sorry to do that to you. Congratulations. I got to go back to practice. I'll work on this. I'm excited. I'm excited. Thank you, guys. So tell us the story of the company. I mean, we got enough time for, I think, for you to tell the entire story from start to finish because all of this sort of happened pretty quickly.

1:30:11Where were you before you started the company? When did you start the company? What was the growth like? Take us through the journey. Yeah, absolutely. So we started the company a little over a year ago. We started to build basically agents that would be slightly more capable than what we're seeing today. Sure. We fundamentally founded the company on this hypothesis that there were two different kinds of agents. There's human-in-the-loop agents, and then there's truly autonomous agents. Human-in-the-loop agents are the agents that we all love to talk about. Like we're talking about cloud code, codex, things that you prompt.

1:30:40They do things. They can do them for a very long time. It could be minutes, dozens of minutes, hours, even in some cases days. But fundamentally, you are the person that kicks them off and evaluates their work. And then, like you were talking about in the previous interview, clicks accept viewing my downloads folder. Autonomous agents are not that. Autonomous agents, when you think about it from the perspective of a buyer, it should feel like you are multiplying your labor force. And when I say feel like, I mean it should be a one-to-one translation. I should feel like when I buy Takeoff, I'm buying 100 ,000 agents that can do what I might have a team of disparate humans, software, et cetera, are doing today, but at a much more massive scale.

1:31:16That was like the foundational thesis of like how do we actually build agents that can do this? We call them long horizon agents. We call them fully autonomous agents. and then we'd go after explicitly revenue-aligned use cases. And the reason we went after revenue-aligned use cases is, well, there's a lot of different reasons. But the most obvious reason is like, why are you going to buy mission-critical AI software from a kid with crazy hair? Like the only thing that's going to get you to do that is if I can prove to you I'm going to make you more money. And the way I get to prove to you I'm going to make you more money is I make it very zero risk for you.

1:31:44I'm like, give me your lowest quality leads if I'm talking to a lending company. Give me your patients that are going to like churn if I'm talking to a healthcare company. And I'm like, let's see what I can do with the agents that I build for your company. Let's see if I can recuperate that lost revenue. Let's see if I can increase your top line. And if we are all successful here at the end of the day, I'm going to be in your board deck at the end of the year because you bought a piece of software and revenue is up double digits. That was the foundational pitch thesis, the whole idea of what the company was going to be.

1:32:09We tried building agents in different ways. We started actually with browser agents because we figured if we can use software, we can do what humans do. But we thought that that would actually – we realized, not thought, we realized because we get eaten up, chewed up, spit out by the market over and over again. We realized that that's not actually a thing. Over and over again for like six months. For six months. For six months. That's fair. It's not like you were like, yeah, we were going to chewed up.

1:32:31Tae Kim:Then you've got one simple trade. For the viewers, what Jordi and John are referring to, if you haven't read our literature, which I don't expect you to, is that we entered this calendar year at effectively$0 in committed revenue. And by the time we got acquired by Brett and Sierra, we were at near eight figures in revenue. So what they're referring to is that very short and vertical kind of, no pun intended, takeoff in revenue ramp. Let's go. More people should name their company Takeoff. Great nominating term. That's a lesson. It's amazing. Yeah. I mean, it's got its own SEO things. Former member of Migos, rest in peace.

1:33:05We're honored to carry the name with a positive light. that being said like you're only going to make money if your agents are trying to sell or like you know trying to be sold upon the value of adding revenue if you actually add revenue and so we de-risk it because you know i'm not brett taylor i can't walk into a room or at least i couldn't walk into a room previously and get someone to pay for something that's not already driving results so we go in we do pilots they're not necessarily free but they're paid on outcome and so if i drive this outcome that we're talking about usually directly revenue or something tied to revenue, for example, for a lending company, loans funded, loans originated.

1:33:40You're going to pay takeoff the way you pay a human being on commission and some sort of base units based on the amount of tokens, voice, SMS that are used. So the customer thinks, I'm only paying when I create X thousands of dollars in margin. I'm paying hundreds of dollars of cost of goods sold. They love that tradeoff. And they're like, if it fails, it fails. And if it succeeds, we're making more money at the end of the year. That's how a crazy hair walks into a room and ends up selling multimillion dollar contracts over and over and over again. Because like once it starts actually working, even what I did not expect really is like the compounding nature of exponential growth.

1:34:12I love it. Walk me through how deeply you're integrating or you were integrating with some of those first customers because there's a world where you're just like, give me the stale leads. I will go off and I will do the email. I will do the SMS. I'll do the whatever happens and I'll sort of like either build those systems or maybe you'll set up your own MailChimp account or whatever you want. Or there's another version where you're like, I want to live within your CRM, within all your tools. I will do API integrations into whatever legacy systems you have to actually collect all the knowledge to make the correct move and actually drive revenue.

1:34:50No, it's a fantastic question. And it's categorically the latter. So there's this lecture, for lack of better words, I give for every potential candidate and existing employee of Takeoff, which is we are given the privilege, right? Not the right, but the sheer privilege of sitting between our customer and their revenue. I could explain this in a million ways why it's so important, but the most important thing to explain is that our buyer was always the CEO or a C-suite member. It wasn't some VP of something that reports into something that reports into the CEO. When you are selling revenue, you are selling to the CEO.

1:35:22That's what he or she is getting greater on at the end of the year, whether they're a public company of which some of our customers are, whether they're a massively multi-billion dollar private company. they're getting graded on revenue and then it's fired at the chief revenue officers in the audience but no no no i know i know um it's it's it's one it's one of the things brett pointed out like it's kind of amazing that like every one of your customers your contact like the person who's in my iMessage top five iMessage is the ceo um and so to answer your question i would give this lecture that when we are referenced by our customers they have to think of us as their best employee when i say us, I mean myself, like Akash, Spencer, Shred, my teammates' names.

1:36:01They have to think of us as their best employee. And the way we get there is we have to understand their business as well as any individual that works for them. They could be a 10 ,000 employee company. They should be able to ask us about anything that is even remotely related to the line of work that our agents are doing for their business. And we should be able to answer it. I'm talking about gross margins. I'm talking about conversion rates. I'm talking about time to fund. I'm talking about time between first contact to revenue generated, literally everything. And we know we've succeeded when the CEO starts asking us questions about their business.

1:36:30That's when you're in like, you know, the promised land. That's when you are literally their friend, when they're texting you 530 in the morning, 10 o 'clock at night. And like none of your family or friends are in your top five iMessage anymore. It's just your customer CEOs. And so it's very much understanding like the intricacies of that business in order to build an agent that could actually do what's going to drive that company's revenue. And so we have to understand every piece of software that they're using, every single thing that somebody might do because we are trying to genuinely scale the workforce and you can only scale the workforce if you can do it end to end.

1:37:00And that's like a really important thing that I think most agent companies don't get. If you're going to sell an agent into some work stream, but you're only going to take like a horizontal slice, it's virtually useless because then you have to educate everything below and above it, how to drive the end to end result. So if you want to actually drive the business outcomes, own the whole thing. If you want to own the whole thing, you have to be capable of owning the whole thing, which means you have to understand the business well enough to build the agent to do so. You guys had Marky, who's been a friend and an incredibly founder, who I've learned a lot from, on the show, I think a month or two ago.

1:37:28And Marky talked about how her entire company and product is rooted in this foundational philosophy that we have to translate whatever language the company is speaking to what the agent is going to do. We think very, very similarly. Our culture is entirely predicated on that assumption that if we don't understand the business better than our customer or as well as our customer, our agents aren't going to do it as well as we need them to. So what does that translation look like for you? Is it a bunch of markdown files and then your agent can interpret those? Because you could go all the way to like we pre-trained a model just for you.

1:38:01And then we could be like, we fine-tuned a model for you, sort of the thinking machines model. And then you could be like, well, we're using the frontier models, but we have a custom harness for you, or we customize our harness for you, or we write a special integration, or it's just or the agent just shows up and it figures it out. I love that you're giving me multiple choice because if you didn't, I would just ramble. It is the second half of answers that you just said. Another foundational philosophy that we kind of built this company on is that the inference API is a commodity. That's a sound bite.

1:38:29You can clip me. That'd be great. That's a crazy thing to say. It's a crazy thing to say that an inference API is a commodity because when you think about Claude and Anthropic and OpenAI at tens of billions in revenue, people are going to be like, that's all inference. I would disagree. I would say it's a function of the things built on top of inference. We're talking about JTBT, Claude, Codex, cloud code. And what we need to be able to do is, you can call these, most people would call these harnesses, like harnesses with a great GUI with a great command line interface. But it's functionally the thing that's delivering end-to-end value.

1:39:01And coding was a great first, coding and chat agents were a great first product because that was the end-to-end value. Again, it's a human-in-the-loop type agent. So they're giving the value to the person that's using the product. The subsequent type of wins in enterprise AI, And when I say wins, I don't mean like hundreds of millions in revenue or even billions. I'm talking about like the next wave of tens of billions of revenue. It's going to come from the fact that your harness, which is just a fancy way of saying agents that can do multiple things and operate across multiple different surfaces as opposed to a single call and response API, should be as capable as someone that is on a job listing or something that you are hiring to drive an outcome or a result for the business.

1:39:40And so it's a combination of harnesses. And specifically at Takeoff, what we built was what we call it as a DSL, a domain-specific language. So you should be able to educate, direct, and build the agents on Takeoff using our domain-specific language that is built around the idea of trying to handle something end-to-end. The other kind of unique thing about this is your agents have to be answerable to the outside world. You can't just say, go do a thing. The thing that our customers are calling APIs for is go fund this loan or an API call to go onboard this patient or go get this patient's prior authorization.

1:40:13That requires multiple actions by the agent that then in turn require input from the outside world. Let's use the borrower example, the loan borrower. If you're getting an API call to your agent that says go fund this loan for this borrower, this lead, you have to call that lead, contact them, help them with that initial rate quoting, understanding what options they have, whether it's a HELOC, a refi, a home equity loan. Then you have to have a second call after whatever happened between the first call and the second call. You have to go reach out to third parties, the e-notaries, the underwriters, everything else involved, document collection.

1:40:40Then you have to have a third call saying, hey, I noticed you got stuck here because you have this weird Iowa borrower question about co-borrower co-signing. Then you have a fourth call that's getting it over the line for funded. This is something that for an agent to actually handle end-to-end, your harness is transcending just like tool calls. It's an always-on harness with a heartbeat that's answering anything that could happen by, on behalf of, or in relation to this central entity in this example of the borrower. Tell me a little bit about post-merger integration. I could see Sierra having a product called Takeoff.

1:41:14I could see Sierra just being a service or company that you work with for a bunch of different things. And I don't even know if I need different products because AI is so broad that everything sort of merges together into like one product that can do multiple things. And I just flip on a switch and say, OK, I want you to handle this. I want you to handle this. But how are you thinking about integration? I know it's I know it's like really, really early. But I imagine that this was what you were talking to Brad about was like with a vision of what these two companies can do together. So like take us through a little bit of it.

1:41:45Well, I mean, that's actually great. I'm going to go in reverse order of your questions here. Like when Brett and I first chatted, we basically realized that we have a similar vision of what the world was headed towards. And what we realized was that by virtue of just, again, being a kid with crazy hair, like there's no chance that I was going to like compete and win in customer support. There's a dozen companies, three of which that are like.

1:42:07Tae Kim:Selling yourself short. You seem, to me, I'm getting young Brett Taylor. Yeah, let's pull up a picture of Brad Taylor's hair, please, and see if you can tell Brad his hair. I think you got a shot, but yes, okay. The point being that we had to come from a different angle. We had to sell a thing that only the early adopters were ready for. And it's not like we – when I say we had a similar vision of the direction the world was headed in, we were further along on that timeline. And we had done this thing that I don't think most of the world realized was possible yet. And it wasn't until we proved it was, right?

1:42:42Like, that's what was really exciting to, I think, Brett and the company is like, hey, we'd love to get to where you are, but we're realizing that you're already there. And why not get there together and then scale it times a million? Sure. Right. And so that's kind of how the original conversation started. We then kind of came to this, like, realization that, and I think what was actually really interesting for you guys to understand, or for anyone who's listening and watching, is that we realized we were onto something when our first, like, three, seven-figure customers were, like, they already had customer support vendors, right?

1:43:12They had a Sierra or a Decadon or something else there. They were spending, on average, between a few hundred grand to maybe a million dollars with them. With us, they were spending at least three times more. They had an AI support vendor, and they also had takeoff. They were spending three times, in one case, eight times more than they were spending with their support vendor. And that makes sense, because you're driving revenue. And you're talking to the CEO. And if you're driving revenue, there's three kinds of software. There's revenue driving software, there's functional software, and then there's must-have software.

1:43:40And if you're the first category, Google ads, Facebook ads,$1 in equals more than$1 out, I will keep spending until I flat that line. And that's what we were going for. It's exactly how we want to be thought of by our customers. And so we realized, again, all the things that Brent and I were talking about that we were excited about, a lot of the same shared ideas around where the world was headed. We're like, hey, together this can be one plus one equals a thousand. And so I don't know if you guys saw, probably not because you have a lot going in your mind. We as in Sierra and Sierra together launched this product called Horizon, which is this new thing.

1:44:10And the reason it has to be this net new thing is because we want people to realize this is a step function jump in capability. Yeah. A step function jump in capability, which is going to drive revenue for your business. It's not just agent and like one second cost savings, but it's agents that your CEO is buying. And that's really freaking exciting. Yeah. I don't know if I can swear. I'm sorry. But that's really exciting. That's awesome. Yeah. No, it makes it makes so much sense.

1:44:34Tae Kim:You're great. You're great at naming. Yeah. These are all every every name is great. Like these are all good. Yeah, I love them. Well, thank you so much. It's great. Great to meet you. I found that I found the whole pitch very compelling. I was just imagining myself as a CEO or enterprise buyer being like, I'm sold. Just send the contract. I will say, like, I've been on a few sales calls with Brett now and Brett. And it's very flattering to hear this from Brett Taylor, right? Like one of the best salesmen probably ever lived in software history. We've had a few sales calls together. It is magic in that room.

1:45:06People get really freaking excited when we show them Horizon. And it's like people start imagining what they're going to be doing for their business, all the awards they're going to get, the fact that in the board deck, it's going to be plus double-digit percentages at the end of the year. And that is very exciting for us as a company. It's very exciting. Amazing. Thank you so much. I can see why you guys did the deal.

1:45:24Tae Kim:Great to hang, dude. We'll talk to you soon. Let me tell you about public investing for those who take it seriously. We've got stocks, options, bonds, crypto, treasuries, and more with great customer service. Mark Zuckerberg is in the Wall Street Journal opinion section with a new piece, The AI Future is for Everyone. He says, the history of democracy and economics has proved that centralized power stifles human potential. And it's quite long. I'm going to go let you guys read it, but let's head into the comment section. Let's get a quick reaction. This is the Wall Street Journal. I think it'll be pretty.

1:46:03Tae Kim:No, it looks relatively tame. But yeah, making a clear effort to position to be the overtly – there was a white space for a guy investing hundreds of billions of dollars a year in AI that says, hey, this is going to be really great for everyone. Yeah. And I'm going to help us get there. It is interesting. Facebook does have some monopolies, but the competition for attention is constant and there are always sources outside. They've never had a full monopoly on social media even with TikTok and Snapchat and LinkedIn and Twitch and YouTube and Netflix and the podcast feed and SMS and iMessage. There are so many other platforms for disseminating information.

1:47:01I don't know. It's hard to jump straight to a critique here. But the key quote that Andrew Curran pulled out was that he said, in most cases, like cybersecurity, the history of open source software has shown that giving everyone full access to powerful systems will be the best way to protect safety and security over time. So he's firmly on the side of democratizing powerful AI. And he is yet another one. I imagine that they signed the letter. I've lost track at this point, but you can imagine that he did. Anyway, thank you so much for tuning in. The other piece of news is that Apple is launching Apple Upgrade next week, an iPhone, iPad, Mac, and Apple Watch leasing slash subscription program.

1:47:51They said you will own nothing and you will be happy.

1:47:56Tae Kim:We're launching our new program, you will own nothing and be happy. it's partnering with Klarna to launch in the United States at online and retail stores it's now official leasing prices start as low as$20 or$17.99 per month for iPhone $11.99 for Apple Watch$24.99 for Mac and$11.99 for iPad so interesting I mean a lot of people are saying this is a direct reaction to increased prices for memory increased prices for products There was a time when an iPhone was a couple hundred dollars and there were incentives to jump on a Verizon plan and you sort of amortize the cost over that. Those days are gone.

1:48:38Like we're in the world of like a$2 ,000 iPhone. It's a significant delay. Same thing with our gongs. For people.

1:48:42Tae Kim:Honestly. You want a subscription gong? No, I'm just saying there was a time when a TVPN gong was$200. Yeah. Now it's in the tens of thousands of dollars. It's actually so expensive. A friend of mine texted me and was like, where do we get the gongs? I need a gong. And I was like, I think you should start small. And this is not like you can't handle the big gong. I was more saying that there is a joy to being on the hedonic treadmill of larger gongs. You don't want to jump straight to the biggest gong. You want to start with a small gong. And work your way up. Work your way up because every gong that we've added has been so electric when we get the bigger gong.

1:49:18I think it's time for a new one. You want an even bigger gong?

1:49:20Tae Kim:Yeah, I want one that's hanging from the rafters. You want a giant gong? Maybe. and also every gong has a different flavor different sound different amount you gotta warm them up all sorts of things we always warm up the gong anyways folks that's our show for today enjoy the rest of your July 28th leave us 5 stars on Apple Podcasts and Spotify money never sleeps you shouldn't either call me back sign up for a newsletter at tbpn.com And we will see you tomorrow at 11 a.m. Pacific. Goodbye. Cheers. Morning flashbang.