This Week in Startups
00:41:17 · D.A. Wallach, Thomas Tunguz, Jason Calacanis
[00:41:17] S1: Yeah. When you say the demand is growing so much faster, I mean, this is what all the companies in the space say too. Right? Yeah. And and my question is, well, then why don't the prices just go up? I mean, they could raise the price to the point where there is only the amount of demand that they can meet. [00:41:34] S3: Well, they're already at 85 percent margin. They're 85% You can look at the Anthropic S1, approximately 81%, 82% gross margin compared to a publicly traded software company of about 71%. And we've seen a price war. So you look at what OpenAI did with Luna. They cut prices 50%, then they cut prices 80%. You have open source models that are significantly less expensive, 80 to 85% less expensive than closed source. You have a lot of deflationary pressures that are going here. The problem is, think about it like a pharmaceutical company. If it takes me $500,000,000 or a billion dollars to make a drug, if I'm a pharma company, have seventeen years to commercialize that with a patent. If I'm a state of the art AI lab, I have forty one days. Because in forty one days, someone else is coming out with another soda model and then all of a sudden share starts to shift. And so I need to generate a pretty significant and healthy margin in those forty one days to be able to finance the development of the next model. [00:42:24] S0: Capitalism is so brutal, D. A. It's so brutal.
00:42:24 · Jason Calacanis, D.A. Wallach
[00:42:24] S0: Capitalism is so brutal, D. A. It's so brutal. [00:42:28] S1: Speculate? I I I would love to get your speculation, Thomas, about what the steady state margin ought to be for these companies. What you just said sounds like an argument for Frontier Labs having no sustainable margin in steady state. Tell me one, is that a correct representation? But two, if you start to think through these other layers, are the memory guys gonna be doing 80% gross margins four years from now, five years from now? I mean, Micron's trading at a forward P of four and a half, Hynix trading at a forward P of 2.8. I mean, if you believe what you're saying about the scale of this, how do you not put all of your retirement funds into Micron? [00:43:08] S0: I'll be right back. Jay trading like a lunatic. Well, as the market came down, I just bought a million dollars at Uber and Tesla, and I totally hit the bottom. But, Thomas, answer the question here about this. This steady state because we're kind of evoking Jevon's paradox here, I think.
00:43:25 · Thomas Tunguz
[00:43:25] S3: Yeah. Look, think the problem... Okay, so the problem with the hardware companies and why I wouldn't load up on those stocks is primarily because there are GPUs lying fallow in warehouses. We don't have enough electricity to light them up. So at some point, NVIDIA bookings plateau just because we can't deploy them within The US, and then you have a seismic shock to those stocks. I think on the other question, D. A, what is the terminal margin for some of these businesses? You look at AWS, right? AWS, massive, the largest hyperscaler. A lot of the business before the AI wave was selling commodity compute. If you rent a server on AWS, it's a server just like there is on Amazon... Sorry, on Azure, just the way there is on GCP. The operating margins not the gross margins, but the operating margins of that business in other words, the profitability of that particular business unit is 35% to 36 net. So even if they're selling a commodity good, even if you're an oil company and you're selling crude or diesel, you can still have a really nice healthy business with competition. Coca Cola's gross margins and net margins are not zero because they're competing with Pepsi. They just continuously try to develop new products and innovate. So I think having even for the broader macro economy geopolitically, having these companies generate 70 to 80% gross margins is very healthy, sustainable, and it's probably the right thing because we want to encourage them to continue to develop and innovate.
00:44:44 · Jason Calacanis, Elana Gold, D.A. Wallach
[00:44:44] S0: Yeah, at the $50 to one ratio I put my spend, Elana, at $50 in value for every dollar. Now, I might not have built that stuff, so maybe it's really 10 to one or something. But I see the revenue going up in the projects I'm doing. I'm seeing the ambition, the dollars invested. Everything's going up into the right because when we deploy the software, we kind of get more efficient. That argues to me that they're undercharging me. And I think everybody's getting undercharged here massively. And plus, I just ordered a 12,000 Mac studio to run local models, which is exactly what I paid for my PC Junior, PC XT, PC AT in the eighties. They were $4,000 computers, dollars 5,000 computers, maybe you can get down as low as I think $2,000 back then. In today's dollars, that would be three or four X. So here we are, it used to cost the equivalent of $15,000 to buy a desktop computer. You just go back to $15,000 desktop computers and you got unlimited unmetered, well not unlimited, but unmetered local compute. [00:45:49] S2: And then just to add to that, you're still hiring at your podcast. So it's not like you're doing everything, it's all automated and you're completely done. Like I saw on your Twitter, you're hiring for 10 jobs, 14. [00:46:01] S1: Yes. [00:46:02] S2: I had to check the numbers and it's like, okay, sure, you're able to automate a lot of the manual tasks, maybe some of the editing that's made for your podcast, some of the preparation, but there still needs to be hands on. And I think some people forget that, especially as we continue to grow.
00:46:02 · Elana Gold, Thomas Tunguz, Jason Calacanis
[00:46:02] S2: I had to check the numbers and it's like, okay, sure, you're able to automate a lot of the manual tasks, maybe some of the editing that's made for your podcast, some of the preparation, but there still needs to be hands on. And I think some people forget that, especially as we continue to grow. [00:46:15] S3: But on this local inference point, Jason, I think you're right. In an era or a year in 'twenty seven when GPUs are limited, local inference will become really big. We're investors in a company called Ollama that has about 10,000,000 people using local inference. Of course. Stanford released a study that 90% of the work that non coders do with AI can be executed on a local model. So I think this is a really big wave, and you'll start to do stuff locally, and then when you need a cloud model, you'll spike, and this is one of the ways that we'll maintain cloud. [00:46:40] S0: I would have no problem, D. A, buying everybody in the company, a Mac studio or NVIDIA, Windows, Spark enabled computer. I would spend 20 ks per person, Because if I'm spending on average right now, 2 ks per person per year, it's going to kind of net out eventually, know, if I sell them in five years for half the price.
00:46:40 · Jason Calacanis, D.A. Wallach
[00:46:40] S0: I would have no problem, D. A, buying everybody in the company, a Mac studio or NVIDIA, Windows, Spark enabled computer. I would spend 20 ks per person, Because if I'm spending on average right now, 2 ks per person per year, it's going to kind of net out eventually, know, if I sell them in five years for half the price. [00:47:05] S1: I agree. And I think it's like, as they used to say in the Internet 1.0 era, information wants to be free. The Internet wants to be free. I think intelligence wants to be free. Yes. And hardware can't be free. And so I don't know how long this takes, but ultimately, you're going to have to, to Thomas's point, you're going to have to have hardware companies that earn a big enough profit to attract the people who know how to do these crazy complicated things. Right now, margins are maybe people are over earning. You've got a situation right now where Elon is financially incented to put literal rocket engineers on general contracting type jobs. Right? Because it's just so valuable to build these data centers as quickly as possible. That isn't steady state. Mean, it's not what we want those rocket scientists doing over the long run, but still someone's going to have to be assembling those ASML machines. Someone's going to be having to produce all the stuff that we need to make this work over the long run. And that I think is the most durable business in a way.