DeepSeek以500亿美元估值融资 | 开源崛起对决OpenAI与Anthropic | OpenAI自研芯片

Guest

Open source is a bit of a fake because China is paying for all the training. Okay, it's not open source like a generation ago. The reason is that there's so much innovation and cost savings.

So, what do we discuss today? Google loses 2 generational scientists in 48 hours. That's a tough time. DeepSeek closes a $7.4 billion Series A. Is the Series A at a $50 billion price, but only China gets voting rights? Interesting. And then, finally, the $725 billion question Wall Street is finally asking: who's actually going to pay for AI?

Guest

There's only one thing worse than a seat-based model, and that's a model that's based on bodies. You don't get to make $10 million for working 18 hours a week. You get a watch. You get an Omega. You want an Omega or you want to be rich? Make your choice, boys.

The whole reason the OpenAI and Anthropic models work is because other idiots have spent the $300 billion on their behalf.

1. Google Loses Two Generational Scientists in 48 Hours to Anthropic

Ready to go? Boys, it is so good to be back. Jason, you are back. It is so good to have you back from China. We're going to start with the news that I put at the top of the list, which was DeepMind losing 2 generational scientists in 48 hours. We have, first, Noam Shazeer, who was at Character.AI, and then we have John Jumper, Nobel Prize winner and co-creator of AlphaFold, also leaving to join Anthropic. How significant are these moves, and what should we read from this?

Guest

Listen, it's easy to pick a turnover in any organization, right? There's so much turnover in any organization. On the other hand, when you talk to some of the smartest engineers and developers in AI, they're really looking to be in a very specific environment where they get to pursue exactly what they want to do, especially on the research side. They want to work on what they want to work on.

The best of the best—and I was thinking back in the day, I went to pitch Google for my last startup, and Vint Cerf was there, one of the creators of the internet, and popped into the meeting. I didn't get it at the time, but Google, back in the day, pre-AI, had created this environment where the best researchers in the world wanted to be there. I think that's how they lured the DeepMind guys in, right? That was the story: this persistence to create this environment. “Look, you're going to get to stay in London. You guys are going to get to build your own thing.”

I think this is probably a sign of the cracks and the realities of having to try to be number 1 in AI, and forcing an environmental change that your competitors can welcome. OpenAI can say, “Just come over here and work on whatever you want to work on for $500 million, $2 billion.” When I talk to folks at the bleeding edge of AI, that's just so appealing—to only work on what they want to work on in AI.

It's funny, because that feels like a 1-dimensional answer, but there might be 2 dimensions to this. On the one hand, you have researchers who just want to go do what they want. On the other hand, when you listen to a lot of people who left Google, there's a bit of frustration about not being able to ship. There's a lot of frustration that they had a ChatGPT alternative, and then the bureaucracy just smothered the product, whereas OpenAI just jammed it out the door and, as a result, took a lead on them.

When you have a historical, existing business, you're damned if you do and damned if you don't. Sometimes people want to just do research; other times, they want to actually get shit done and ship, and you get in the way of that. My sense is, first of all, these 2 people, in terms of their research pursuits, are somewhat different.

Noam obviously was at Google, did the original attention paper, left, did Character.AI, and got bought back to Google in large part because it was a very clever acquisition: the brains behind restarting their AI effort after OpenAI stole a march on them. I think it was a couple of billion dollars, so a good slug of that was going for him. Assuming 4-year vesting, he's probably left half of whatever he was offered on the table, right? So that's a lot.

I can imagine that it's a combination of some version of, “Hey, as you say, more research, more ability to do things,” plus a more certain ability to ship and make shit happen. I think even though all of last year we did the “Yay, Google is amazing” because, unlike the other 3 hyperscalers, they have their shit together and they have an AI story, the stock has reflected that in the last year. It's one of only 2 of the Mag 7 that's up over the last 18 months.

At the same time, when you look at things like having a viable coding model and having that next level up from just shipping a model to shipping interesting products, the truth is Google hasn't done an amazing job, and OpenAI and Anthropic have. So if you're into product shipping, which I think Noam might well be, I can see why going to one of those 2 makes sense.

Totally different on Jumper. This is someone whose pure research science—undergraduate and postgraduate degree, I think at the University of Chicago—was all focused on protein folding. Remember, he got a Nobel Prize. DeepMind has been the only company he's ever worked at since his graduation—I don't know if that's the right word after your postdoc—but since academia.

Basically, it's been a whole bunch of academia, a whole bunch of time at DeepMind, and picking up a Nobel Prize. You've got to believe you're going to Anthropic because their story there is about being able to do more research, which is almost on high-end science. Anthropic has announced an initiative on that, so it's somewhat different.

Stepping back, I think what it speaks to is that when you're top of the heap, you can promise everyone everything in a way that you're not when you're an incumbent. The truth is, being top of the heap right now means you're one of the new companies. You're not constrained by history, you're not constrained by the install base—like that old proverb or joke about, “Hell is the install base”—and you've got a stock and a currency that's huge. No one's giving you shit about stock-based compensation.

If you're Anthropic and OpenAI, you can buy whatever you want, including people, and you can let them do whatever they want, including whatever it is they've been promised to do, in a way where you have much fewer constraints than the incumbents.

You know me: I would never deal in rumors. I don't do rumors. But the rumor mill that I heard was—

Guest

Are you a rumor guy?

Come on. Bring it on, baby.

Candidly, is it that Anthropic has clearly had a breakthrough that a small number of people know about, and that's why John Jumper went there?

Guest

The reason I'm skeptical of that, Harry, is the lag time between having an amazing invention and revenue in the core LLM space is 1, 2, or 3 years. The lag time, if you have a medical invention or an idea and are turning that into real, meaningful economic value, is 10 or 15 years.

The amazing thing is that the protein-folding advance has collected its Nobel Prize and, as yet, has not had meaningful commercial success or a drug in production. So I doubt it's like, “Oh my God, we've discovered something new and this is magic, and if you don't join in the next 3 months, you won't be part of this thing.”

My guess is the initiative—whatever initiative Anthropic has around next-generation science—is a multiyear thing. It's just a question of where you want to spend the next 5 years of your life doing research.

It is a vibe, for what it's worth. Listen, I really hate to be one of those VC dads who tells stories about their kids or what happens in their spoiled kids' school as their thesis for investing. But my son is very good at a certain type of math. I don't understand Jensen, Feynman, string theory, and all this stuff, but as a college student, he's one of those top 10 in the country.

The labs find him, right? He doesn't have to apply, per se. He got an internship offer without looking from Anthropic for an amount of money that, when I was in college or grad school, would have been unimaginable. The opportunity when you graduate, even today, is incalculable, and so he instantly turned it down, with no job.

He's just like, “I can't do enough research. I want to do my type of research, for my type of math, and my type of AI.” Everything that we do on this show, or the other stuff, he kind of makes fun of me because he's so far ahead of how inference works and how open source works. Really, you should replace me with him.

My point in telling these personal dad-VC stories, which I hate, is that he instantly said they couldn't create the right learning environment for him. He's not worried about money. He's not worried about any of this. If they can't create the right learning environment for him, he just won't do the job.

I don't mean to map my family to 2 of the top researchers in the universe, but I feel the same vibe. In today's world, we're on a bull run like we've never seen in our entire lifetimes. There's nothing like this, where a so-called startup can pay billions of dollars to acqui-hire you and then you leave with 50% of your billions unvested, too.

Guest

We've never been like this. It just creates an environment for the best of the best where you will only do what you want to do. You just won't do the job that isn't 100% what you want to do.

Remember when Harry and I were in London and we were with Maggie from OpenAI? I think she said she was in sales leadership. She said she'd never been in the researcher building except for 1 or 2 meetings. They didn't allow sales in the whole building with the top researchers. “Leave, leave. We don't want these pesky go-to-market professionals bothering our researchers.”

Anyway, I hate to be that guy, but it really resonated. You just have to—it's really hard if you're not Anthropic. How do you retain this talent? How do you let people work on what they want to work on when you need the chatbot fixed so you can compete with Sierra? How do you contain them? I think it's very hard.

Guest 2

First of all, you're exactly right. These are 2 of the most talented people on the planet. One of them wrote the Attention Is All You Need paper and has made plus or minus $1 billion, and the other has a Nobel Prize for medicine and is still relatively young. These are not top 1% or top 0.1% people; they're the top 0.001% of people. Probably one of the very best researchers out there on the planet.

The funny thing for Google, and the hard thing, is that they're struggling to keep people like that while, at the same time, they're not getting the tactical [__] done and making the stuff happen to continue to make progress. The amazing thing about someone like Anthropic is that they're able to do both. They're able to create these wonderful research environments for the “leave me alone, let me do research” people while, at the same time, they're executing tactically brilliantly in terms of product and making stuff happen.

That's it. It's momentum. When you start winning, everything starts going your way. The people start going your way, the breaks start going your way. Winners win, and they compound until something happens to break that chain. This is just that, magnified.

Listen, Google overall, I think, actually has been on fire. But as I continue to learn—and I'm only so smart about the role of open source versus closed source—the most vulnerable is going to be number 3.

For talent, for people, for revenue, and for your ability to do things that maybe aren't core, if you're number 3 as the closed-source LLM, that's where you're going to hit the most pressure from open source. Sometimes, when you're in that position—and I'm not saying this is correct—you feel like you don't have the luxury of letting folks do what they want to do because you're under such intense pressure. Maybe Anthropic feels, as competitive as it is, that it has a luxury its competitors don't have.

Guest 2

I think that's true. I want to come back to “Google is executing amazingly.” I'm not commenting here; just look objectively. Over the last 18 months, since 2025, only 2 of the Mag 7 have outperformed the S&P: Nvidia and Google.

That's another way of saying that people like Facebook, Amazon, and Microsoft are doing [__]; they're not relevant here. Google has done an amazing job of being relevant here, and that's a true statement. But the other statement that's equally true is that you don't wake up every morning and say, “Let's try the new Google model. Let's try the new Google harness. Let's try the new Google coding tool.” You do try Claude Code. You do try Grok. You do try OpenAI.

The fact is, they are relevant and in the frame, but they are definitely number 3 in terms of innovation, which is a whole lot better than being Microsoft or Meta and having to say, “We don't have something yet, but we spent $70 billion. We might get something next year.”

Jason, just so I understand, why is number 3 the worst position and the perilous one, in a way that it's not for cloud?

2. Why Being #3 in AI Is the Most Dangerous Position

Guest

There are 2 threads happening at the same time. On the one hand, clearly, the price of tokens, token pricing, and the budget issues are real. The amount of folks routing models and running multimodal is exploding, right? At a small level, OpenRouter and open source— that stuff's all a big deal.

Everyone is realizing they have to get smarter and smarter, much more quickly, on routing workflows to different models. That is clearly very true today. Maybe 90 days ago on the pod, it wasn't clear how big a deal that was. Everything except the smallest startups are routing.

So, what do you route to? If you have 3 vendors—and Rory is the professor here, especially with his background—number 2 was often simpler. Number 3 was usually cheaper, right? Some variant of that.

Frankly, GCP, old Google Cloud, blew up because it was cheapest. A generation ago, Google Cloud struggled in the beginning, and then it was just cheapest and simple, so people would move certain workloads to Google Cloud. Now you're trying to do the same thing with your massive AI spend.

The question is, open source is really complicated. Open source in inference and training is not free, unlike Linux. It's not free. There are substantial costs, but it is materially cheaper, which we could talk about.

In theory, with OpenRouter and others, you could route to 10,000 models. In practice, is the number 3 thing you're going to figure out, “What's the best open-source model I can use for my application?” Or is it number 3, closed source or open source? There's so much going on in open source. It's fueling these crazy Baseten and Fireworks and all the other companies.

There is so much innovation in open source that number 3 might just get swamped by all the subsidies of the Chinese government, everything else subsidizing open source. Open source is a bit of a fake because China is paying for all the training. It's not open source like a generation ago, but there is so much innovation and cost savings.

That's my thesis, and that's what I'm seeing, too. It's a big deal. Literally this morning, I got an email from Anthropic. This is their sort of AI shot across the bow for open source, saying, “Your prompt cache hit rate is low.” Out of the blue—I don't know if you guys got this email this morning—but what Anthropic is doing aggressively here is fighting back at open source and trying to get you to cache your prompts, which are very expensive.

They offer such a massive discount on cached prompts, if they work for you, that it actually can be cheaper than open source. This didn't say, you know, “Fighting Gemini.” This is an email they sent to maybe their entire base saying, “Cache your prompts so that it's cheaper than open source.”

Got it.

Guest

So that's why I think it's just hard. Number 3, you can't be cheaper, you can't keep the researchers, and the projects are less interesting.

Guest 2

The short answer is, I agree—you're right, Jason. Typically, tech markets tend to—I mean, look, you don't end up with perfectly competitive tech markets in the economic sense of millions of players. You end up with small numbers in a tight oligopoly, where there's a leader, a number 2, and then, depending on the size of the market and the competitive dynamic, maybe there's a 3 and a 4. The vast majority of the revenue goes to 1 and 2, right?

That's just the structure in the cloud market, where AWS was first, Microsoft second, and Google Cloud third. The interesting thing about Google being third is that, unlike the typical third, if they were a standalone company and didn't have the Google balance sheet behind them, I think it would be incredibly tough. Implicit in that statement is that it's almost impossible for a number 4 with the same business model—a closed-source number 4—to emerge and catch up at this point.

I mean, a16z had a nice piece on that 6 to 12 months ago. The market is set. The game is clear. The only reason Google can keep punching is because they have a whole balance sheet behind them.

3. DeepSeek's $7.4B Series A at a $50B Valuation: Is China Winning?

The other part of it is, if there is a compelling alternative to this entire set of competitors that's 5 times cheaper, which is what open source is, then it's going to grind everyone down. When an industry gets ground down, what tends to happen is the number 1 guy makes a little less money, the number 2 guy makes quite a lot less money, and the number 3 guy goes bust, right?

In this case, obviously not bust because they've got Google behind them. But you're right: it's a real, powerful downward pressure on the profitability across all of these. That, I think, is the big story now from open source, with lots of caveats.

Rory, how do you think about sovereignty, then, and sovereign models and Mistral and the rest of the world, if there's no room for number 4?

Guest 2

There are 2 ways to answer that. We'll answer sovereignty first and then the impact of China on sovereignty.

Look, Europe is effectively saying, if someone says, “We need a sovereign model,” rephrasing that, what you’re saying is: We are no longer part of the market over there in America, where our product is third. We are, in fact, first instead of being fourth in the worldwide market for closed-source, state-of-the-art foundation models. We are first in the European market for state-of-the-art foundation models, right? So you’re effectively saying, “I couldn’t compete as the fourth player worldwide, but I can compete profitably for me, but probably less efficiently for the system as a whole, as the dominant player in Europe.”

A million years of economic theory explains why that’s a dumb idea from an efficiency perspective. By definition, everyone in Europe is getting the less good model at a higher price. But someone’s deciding that there are political or national security reasons to pay that tax. And that’s a political decision. It’s beyond the economic analysis. It’s just something that government may choose to do or not, right?

But the other part of it—and I’m going to segue to Jason’s trip there—is the fascinating thing about the sovereignty question: all the competitive models to the foundation models, all the open-source models—bar not quite every, not quite all, but most of them—are Chinese-based. And, Jason, you’re just back from China. In the context of sovereignty and security, it is amazing that the entire open-source initiative is running on 4 or 5 models, all of which are built in China. So what was your takeaway from there?

Guest

Well, it was interesting. Folks probably already know that when you go to China—and, interestingly, even Hong Kong, where Hong Kong has its own sovereignty and can do whatever it wants—Anthropic and OpenAI don’t serve there intentionally, for security reasons. They don’t allow you to access Anthropic or OpenAI. You can’t access them effectively. It’s hard.

You can do it over a VPN. You can hack it on your phone, but it’s tough. They intentionally try to block it, with IP-address blocking or whatever. It’s not China; it’s them blocking it, right? But Gemini doesn’t. So when I’m in China for 2 weeks, it’s DeepSeek and Gemini. Those are my friends. It’s a parallel universe.

DeepSeek is intentionally crippled in China. It’s not as good as it is here. It can’t search the web, and it is trained on different data, as far as I can tell. But you get it. I didn’t think I got the sovereignty argument until I was in China, right? This is China basically saying, “We do not want to be subservient to the United States.”

Whether this was the original goal of DeepSeek and others, and Alibaba and others, I’m not sure. I’m not a professor—I don’t have my professorial background on my YouTube—but it clearly is where it is today, right? Massive government subsidies, right? DeepSeek’s raising this round at $80 billion. The government’s the only one getting voting shares. It says all you need to know, right?

It is an existential sovereignty thing. The Chinese government does not want to be reliant on Anthropic and OpenAI to run the next-generation economy. It’s very smart. Whatever investment the government has quietly made to subsidize all of these providers, it’s a drop in the bucket, isn’t it? At the sovereign level, what, $10 billion, $20 billion, even $50 billion? I mean, this is nothing compared to—what does an aircraft carrier cost? A lot.

It’s easier just to subsidize an open-source model and pretend the training costs are $10 million. What did DeepSeek claim their training cost was back when it launched? $15 million to train an OpenAI competitor. Of course it wasn’t true, right? The round itself is pretty wild in a couple of different ways, in that it’s $7.4 billion at a $50 billion, give-or-take, price.

The thing that’s crazy is the founder is committing 20 billion yuan himself, which is like 40% of the round. I mean, it’s like $3 billion the founder is committing himself. Wait for it: There are fewer than 10 investors, including JD.com. I saw this and was like, “What?” And none of them are getting any rights. The only people getting rights are the Chinese state, which retains governance control. It’s wild.

Rory O’Driscoll

Well, it’s just a different world, right? These are not companies that—it goes back to the old Huawei drama. They’re not independent of the government in any way, shape, or form, right? And so you either do it the right way, or you end up thrown out of your own company, or having to recycle and give back your $2 billion that Meta took from you. I mean, you’ve got to play by their rules, right? And so this is the rule for DeepSeek to not be taken from you.

A couple of comments here. First of all, your comment on, “Oh, only the Chinese government is getting voting rights.” I felt like saying, so the only people getting voting rights are the only people who don’t need them, because the Chinese government doesn’t need voting rights, right? They have sovereignty, right? They have an army, right? As we’ve seen with Manus, they can make you disgorge your money right after you’ve got it just by leaning heavily on you.

So it’s not like—I mean, in one sense, the round’s not surprising at all. Zooming out 2 levels, 2 other things on that: 1 is, yeah, it’s a $50 billion, but we have 2 companies, the leading 2 American companies with closed-source models, which, by definition, yield more and have more economic upside, trading at plus or minus $1 trillion. So $50 billion for the open-source competitor feels roughly right, right? You know, 1/20th of the price.

And Zhipu AI—Z.AI, which is the Americanized version of it—is actually public now in China, and that’s trading at $100 billion or something like that, 1,000 times revenues. So again, plus or minus, if you think of these as the sovereign alternative and the open-source alternative, they’re trading at numbers that aren’t crazy at all compared to the U.S.

The other thing is just to say it: We’re doing a little bit of the, “Oh, you know, you have—and you do—you have a lot of state interference in China, and you’re seeing it over and over again, and your property rights are weak.” We saw it with Alibaba back in the day, with Jack Ma. We’re seeing it here and now. But at some point, we’ll have to talk about the fact that you’re seeing more of that in the U.S. as well, right?

We’re seeing Anthropic unable to ship the most recent model until they satisfy the concerns of the U.S. government. And I’ve got to give Leopold Aschenbrenner, in his 2024 *Situational Awareness*, credit—he called it. About now, national governments are getting involved, and we can argue that’s a little overwrought. But you can’t give China grief for this kind of thing when we’re doing the same thing ourselves.

Both governments are feeling this technology is pretty existential, pretty impactful, and trying to figure out: Do you regulate it? Do you take over it? Do you stop it from being used by other people? There’s a whole bunch of policy questions going well beyond economics that are being raised here. China’s solving it their way, we’re solving it ours, and Europe is doing theirs.

Rory, you mentioned Zhipu—or I hope I pronounced it right. Z.AI’s GLM-5.2 beats GPT-5.5 on coding benchmarks. How significant is this? This was deemed one of the most consequential open-source AI releases.

Rory O’Driscoll

I mean, look, to me, all it says is these guys are cranking. There are probably 6 total open-source Chinese models. 3 of them top out at or close to U.S. performance, with 3 more just behind it. The aha for me is—it’s, compared to Jason’s point, a compelling competitive alternative.

Many U.S. companies are starting to build their own model based on the open-source alternative. So it’s providing some kind of—it’s providing a ceiling on profitability for some of the U.S. closed-source vendors. It’s really economically significant. And I think the most compelling fact is that there are 6 of these companies pounding it out there. There’s just a lot of competition on the open-source, state-of-the-art alternative.

We can talk later about distillation and how much of what they do is a function of the ability to learn from what the frontier models are doing, but they’re there. They’re relevant, and they’re providing a competitive—“umbrella” is not the right word—a competitive drag on what Anthropic and OpenAI can charge.

Behind all of these, there is actually the infrastructure, and a core part of that infrastructure is seeing increases in price. Memory is one of the biggest. The price of memory has gone up 4 to 5x in certain cases. Tim Cook told *The Wall Street Journal* that Apple faces a 100-year flood in memory costs driven by AI infrastructure demand. DRAM contract prices rose 90% to 95% in Q1 alone. How significant are these cost increases? Who feels them? How will we see this? What should we take from this?

Rory O’Driscoll

I think, other than a profound regret at not buying SanDisk and Micron Technology a year ago and making 20x, look, I think it’s funny. I was thinking about this and going back to the discussion on the researchers at the start. Very different topics, but actually they’re all about the same thing: The investment in AI is commanding resources, right? And via the price mechanism, everyone else is getting impacted by that.

The impact of this is going to manifest itself, as you say, in DRAM pricing, right? It’s going to manifest itself in the price of your iPhone, the price of your electricity, and the price of your house in San Francisco. It’s going to manifest itself in terms of 20% of you losing your jobs if you’re one of the companies that wants to go all-in on capex for AI, like Oracle.

4. Wall St’s $725B AI Question: Who's Actually Going to Pay for AI?

So what you’re seeing is the positive side. And again, this is not an “AI is bad” argument so much as economics just sends its signal. It doesn’t have morality.

Guest 3

It just says, “Oh, you want to devote more resources and more DRAM to a data center in Tennessee or Mississippi.” That means you’re going to have less DRAM for Rory’s iPhone. The only way to make that happen is to raise the price of the iPhone, right? That’s just what’s going to happen. This is price as a filter.

The amazing thing about the AI capex explosion is that it’s so huge that it’s literally impacting everything, right? This is just one example. It’s sucking in the money and the resources. Specifically, what it’s going to mean is Apple has clearly decided, wisely, not to swallow the loss and lower its margins. They’re just going to raise prices. They’ll sell a few fewer iPhones because of price—there’s some price elasticity—and everyone will have to pay more for their iPhone.

Goldman Sachs projected there’d be $7.6 trillion in cumulative AI capex from 2026 to 2031. It was fascinating, and it brought up the $725 billion question. I remember David Cahn from Sequoia did the $600 billion question. I think it started—

Rory O’Driscoll

$600 billion, I remember well. It was a good piece.

Yeah. Are we going to have a trillion-dollar question next year that we’re going to be discussing? I mean, what you’re basically saying is—and it’s worth getting outside the inside baseball part of it—what Goldman Sachs is saying is that the hyperscalers are now spending $700 billion in capex a year, which times 5 or 6 years is $35 trillion, some astronomical sum. You only do that if you expect revenue to be greater than expenses, because that’s how American capitalism is meant to work.

So that implies that at some point someone has to spend $700 billion in revenue for you to make a buck, right? Right now, they’re nowhere near that. We’re probably well under $100 billion. So AI as a whole is getting $100 billion in revenue and spending $700 billion a year. That’s not a great business. What you’re basically saying is, how does that end? With the increase in capex, doesn’t that increase the revenue requirements for—

Guest 3

Absolutely not. People are going the other way, and that’s the funny thing to say. I would have thought that David Cahn’s piece a year ago was really good, and it made me think. I was wrestling with the same things, and then you step back and say, since then, all that’s happened is people have doubled their conviction and doubled their willingness to spend.

At that point, it was, “Oh my God, capex is like 60% of free cash flow for the Magnificent 7.” Now it’s 120% of free cash flow for the Magnificent 7, and they’re borrowing to fund it. Even though a year ago—and this is the classic thing about bull markets—you can be intellectually right, but the narrative can keep going for a long time. People’s willingness to invest—

It’s very hard to know how that ends—not how it ends, but when it ends. I’m only—maybe Rory can again share some history where there’s an analog. I just think, in my lifetime, in tech, being aware of things, I can’t think of another time where demand was infinite at this level. So how we deal with it and how we deploy capital—

But if you’re sitting on the other side of infinite demand, you can choose not to embrace it, right? That’s like shorting everything, but where does it get you? You’ve got to—the demand for AI is an order of magnitude more than it is today if it could be served cost-effectively. We would all be consuming tokens 24/7 if we could, so the demand is so much higher than what’s delivered today.

Do you play the game, or are you a Debbie Downer? I don’t know what the right answer is, but certainly in venture, you don’t make money by not deploying your fund. It’s hard to—I mean, fees are nice and all, but you’ve got to deploy into demand cycles and just hope to God you get liquidity before it crashes down on you.

Diving one level deeper, because you’re right, Jason, you can’t just say, “Oh, I think this is irrational,” and go home, right? First of all, that would clearly have been the wrong decision for the last 2 years. And B, it’s not the business we’re all in. We’re in the business of investing in the future.

But you do have to think about it, and the comment you said was interesting: you’ve never seen this level of infinite demand. Then you caveated it in a very important way: price, because you said price-adjusted.

I think that’s the difference between this and prior generations of software technology. With something like Salesforce, if you were buying a SaaS product, there were really only 2 states. If you didn’t need it, you didn’t buy it, right? If you needed it and were a big enough organization, you paid the $100,000 and said, “It’s just a cost of doing business,” because it was kind of a fixed price.

What’s interesting now about tokens and intelligence is there’s an infinite demand for intelligence if it’s free, right? But it’s not free, and now everyone’s just wrestling with some version of, how do we allocate that, given that it’s not free? No one said, “I’m going to buy 100 seats of Salesforce if it’s $1,000 a year a seat, but if it was only $500 a seat, I’d buy 2,000 seats.” I bought seats for my people if I needed them, and then I didn’t, right?

Whereas intelligence is this thing, as you say, where the real question is not, “Do you want it?” The real question is how you allocate it, because you now have to have this new skill that didn’t have to exist in the prior world. If you’re a CIO, how do you decide how much to spend and where the cutoff is, right?

I think—and I’ve always felt—that the slowdown here doesn’t come from some technical barrier around, “Oh, the models…” The people who articulate that the models can do everything are correct. The models are going to keep on doing everything. They’re going to get better, blah, blah, blah, right? The real question is price. Price is going to be the arbiter of how much you can do.

And I think you’re right on that. To make it practical, even in the last month and a half, there was this sudden recognition that maybe token maxing wasn’t a good idea, and now you’re seeing people try to be more efficient. The question is, do you think that shows up in actual slower revenue growth for Anthropic and OpenAI? Because that’s the rubber-hits-the-road question. If they’re sending you an email, Jason, trying to effectively cut your bill by saying, “If you’re more efficient, you won’t be spending as much with us,” do you think that reduces the growth from 10x to 5x? How does all that shape out in terms of revenue growth? Do you have an opinion?

Guest 3

I don’t know. It’s interesting. I got that email. You guys got that email, but I’m on the Max plan, where for $200 I get $10,000 of inference a month if I can use it properly, right? So for me, they may have more of an incentive to send me that email than someone on an enterprise plan where, for $10,000, I spend $10,000 a month, right?

Anthropic and OpenAI are running an A/B test where they have a free segment of their base, which they are subsidizing, right? For sure. It’s just bigger at OpenAI. They have an enterprise base, which has between 40% and 70% gross margin just on inference, and then there’s this prosumer one—the Max guys, right? Some of those they’re making a profit on, but some of them are massively subsidized.

The prosumer—the kid vibe-coding, $10,000 of tokens a month, paying $100 or $200 for a Max plan—is massively subsidized. That’s not a joke, right? That’s the whole OpenClaw drama. It’s a real issue.

So we have this weird spectrum of free, massively subsidized, and actually, if you just look at inference and not training, fairly profitable enterprise customers, right? Fairly profitable. I don’t know what the term they use is, right? It’s not—it’s an unblended gross margin, but the inference margins are attractive, right? And that’s what open source is attacking: that enterprise customer, the lucrative customer.

Rory O’Driscoll

Yes, because it’s lucratively profitable for the vendor, right? And the question is, is it profitable for the customer?

In other words, have people said, assuming you’re not on some kind of cap plan—and I think those are increasingly harder to get for the enterprise—assuming you’re on the API, can you really afford to give everyone free, untapped intelligence all the time, and does the math of that work?

As I say, the question is, what’s happening? I’d love to know what’s happening in real time now on usage and revenue across those companies, because that’s the be-all and end-all question.

Guest 3

Yeah, I think the story—it’s interesting—token maxing isn’t really the story. I think, to what Rory said, the big story of 2027 in AI in the enterprise, and all the margins in the enterprise, is, “Show me the ROI next year.”

Right now, it’s token maxing, because 2025 into early 2026 is, “Guys, just go do it. I don’t want to be behind,” right? We can make fun of token maxing, but it was the best way to get teams AI-fluent. “Just go build it, guys. Here’s $100 million, $50 million, $10 million, $5 million, or whatever. Guys, just go build it.”

We’ve even seen it in our portfolio companies. Then the reaction was, “Guys, it’s actually not a joke. You spent too much. You blew through the IT budget.” It turns out the IT budget is bounded, right?

Then the review—you’re already starting to see it—but I think going into 2027, CIOs and others are going to say, “Okay, just show me the fucking ROI. We’re not just going to ration tokens based on who we like the most in the company and who makes the best PowerPoint pitch.”

Show me the ROI. If you have ROI, if you were able to lay off 20% of your department, or you have the highest-growing division in our company, we will give you more tokens. And this group that can’t ship, can’t get anything done, or is in decline, we’re just not going to give you the tokens.

For the first time, ROI is really going to have to connect to a lot of this spend.

I agree. You’re right. Sometimes I think saying the obvious is really helpful. I think you’re exactly right, Jason. If you’re running a big company and you say, “Okay, I wrote off the first half of 2026. We spent way more than we thought, but at least my people now know how this shit works,” write it off as a one-time thing. You’re exactly right. The question now is ROI.

Bringing it back to you, the $725 billion question, the Goldman capex question: I did this math a while back on the fly on the podcast, and I’ll revisit it again. Let’s round up to $1 trillion, because if you’re spending $750 billion on capex, you’ve got to pay for electricity, too. You need $1 trillion of revenue, right? And if companies are going to give you $1 trillion, they’ve got to be getting more value than that from the spend, right?

So, $1.5 trillion, plus or minus. The total spend on labor in the U.S. across everything is—you know, GDP is $30 trillion; it’s sub-$20 trillion. You’re really talking about 7% or 8% of the labor force being replaced by tokens for the math to work. When you look at that, you go, “It’s a dauntingly high bar,” right?

There’s a little part of me that says, “I think this is amazing,” but I don’t know if that last dollar in capex is going to earn a return. If it does, it’s going to be because there’s a huge amount of labor-force disruption. Let’s call labor-force disruption the negative spin; the positive spin is, of course, productivity improvements.

If the $1 trillion is going to earn a buck, you’re going to have to have huge productivity improvements from AI. Huge productivity improvements result in labor displacement and, hopefully, eventually, new jobs for the labor that’s displaced. But if one is going to happen, the other’s got to happen.

We may reach a situation in 2027 where we just cannot prove any of these productivity gains. Going to your math, everyone may just need to lay off 10% of their company in addition to all the other layoffs, right? Maybe the Oracle layoffs and others are just early views of where it’s going, and Robinhood and everyone else are just going to have to say, “Listen, we have to fund this. We have no choice.”

We can talk about ROI and productivity, but if we don’t do it, we’re going to lose to our competition. Our team has to be 15% leaner next year, guys. It’s just this simple. It has to be 15% leaner.

Guest 3

I’m going to do the professor thing for a second: that is productivity. You’re right. You can have an improvement in productivity—and a massive improvement in productivity in an industry—by virtue of the advent of an enabling technology like AI, and no improvement in profitability for that industry if everyone adopts the technology.

If every bank adopted ATMs at roughly the same time, everyone could, in theory, save on tellers. In fact, they ended up taking more tellers because the business expanded, but we’ll skip that for a second. The problem didn’t change. The same could happen here: everyone adopts AI, and everyone’s cost function reduces by the same amount. You’re right, but the people who don’t adopt it are dead. That’s your point, right? Which is why you have to lean in.

I think that’s probably likely in a bunch of these industries. If consulting, or any of these companies, or any of these white-collar jobs are 20% more efficient with AI, then everyone should adopt it. The cost of those white-collar jobs, like audit, should then reduce by 20%, all other things being equal.

My question is: do we see the same acceleration in model progression and capability that we saw in coding, in legal, and in accounting? If we do—whoa. Andrej Karpathy says he moved from 20% to 80% in a 6-month period. If you’re able to see that same advancement in legal, whoa. Obviously, coding is uniquely well-suited, but we just built an AI VP of finance while we were in China, and it’s already better than any of the humans on our team.

It does it. It creates the quote. It builds the contract. It ships the contract. It gets it signed. It updates the opportunity in Salesforce. It logs into Bill.com, where Aurora is an investor. It sends the invoice, follows up on the invoice, gets it paid, interacts with Brex, and closes out the entire transaction.

Then it logs into QuickBooks and makes sure, for the first time in 10 years, our books are accurate. They’ve never been accurate before, and the revenue is properly characterized. The agent does all that. This is what the models today, which are not tuned for this workflow, can do for a customer.

It’s so good. We just had a human at SaaStr AI Annual this year, where Rory was, forget to invoice $80,000 of revenue. We had to write it off. Now, listen, I can afford it, but it was so frustrating that this person just didn’t do it and didn’t even explain any reason why. They just didn’t do it.

Amelia was like, “I’m going to fucking build an agent to do this,” in China when we were there for SaaStr, and then, boom. It’s just better. It’s better than the humans.

Guest 3

Did you have a full human being doing all that? Did you contract labor as a result of this, or is that person doing different stuff?

We use contractors, and for us, it’s more that we’ll use fewer hours of the contractor by accident. But this is an issue, too, right? Variable-cost human labor does exist. You may even use less of it incidentally due to agents, without intentionally trying to save money.

This is not replacing 4 humans on the team. For us, this is the agentic story: we replace things that humans are unwilling to do on our team. They’re unwilling to follow up. They’re unwilling to get the update—the sales job. They’re unwilling to send a proper invoice. They’re unwilling to do these things.

Rather than fight the fact that, as humans, we really only want to do 5% or 10% of our jobs—including investing; I only want to do 5% or 10% of that job, too—we just have agents do the other parts. If it’s variable-price, you could see what you pay these humans fall by 50% to 60% without even intentionally trying to save money.

Brandon from Record[?] posted 9 hours ago—which I sent to the whole team here—“Training agents will be the largest job category in 5 years’ time.” Largest job category.

Guest 3

Yeah. Training agents will be the largest job category in 5 years. Well, look, I’m sure he’s right. He’s much smarter than me. But didn’t we say this about prompt engineers when the show started?

Yeah. How many prompt engineers have you hired on the 20VC team? The fact that we were able to build a director—we’re going to call him a VP, a director of finance—remotely in China that is better than any human on the team is kind of the point, right?

Rory asked, “Why isn’t everybody doing it?” My response is, they can. They just don’t have the mindset to do it today. They haven’t spent a year vibe-coding, so they don’t know what’s possible. They haven’t scaled a certain amount of human learning to have the comfort.

It’s like doing your first venture investment. It’s a little scary, right? But when you’ve been doing it for a while, the next one doesn’t seem so complicated, even if it is. You know how to put the pieces together, so it’s kind of like that.

I do think it’s true. I think the number-one skill is being a master of agents. But I think what that means is going to get redefined each year, right?

Again, when we started the show, people were still hiring prompt engineers because prompts were complicated to craft. They were really hard to get right. If you didn’t craft the right prompt, the software that came out was unusable.

Today, I can go in and vibe-code something. I can say, “Build me an AI VP of finance. Connect it to Bill.com, QuickBooks, Salesforce, and our other app. Reuse what we have and just automate all of our billing process.” That can be my prompt. I think most people can write that prompt, can’t they?

So, you don’t need a prompt engineer, but you do need a master of agents to understand what you’re going to get out of that, where the limitations are, where it’s going to break, what it’s not going to see, and where it’s going to get lazy. Our AI VP of finance admitted last night that it didn’t fully read a contract. We asked it why, and it said, “I don’t have a good answer for you.”

Guest 3

Wow.

Okay. Being a master of an agent is not throwing your monitor out the window when you hear that or calling it [an idiot]. It’s, “Okay, I get what happened.” This is running on Claude Sonnet.

Sonnet rapidly goal-seeks. It tries not to finish complicated behaviors. I need to work with the agent to change how we do it and make clear that all contracts must be read from beginning to end. Then I have to be comfortable that it still won't do it sometimes.

I know I'm rambling, but in a year maybe the models and the harnesses will get so good you won't have to do that anymore. So what that means is going to change.

Even this whole idea of folks talking about loops and agents looping is an early view of where everything's going to go. If your agent is constantly looping and improving itself in the background, which is already happening, it fundamentally changes the way we build agents. They're not static; they're looping. They're constantly improving themselves in the background.

Again, when we started this show, you needed to be a prompt engineer. They would joke that the highest-paying job out of college was prompt engineering; prompt engineers were making $150,000 out of college because they knew how to write a prompt. That skill's worthless today. So I think Brandon is right, for sure. It will be really interesting to see what skills it takes to be an agentic expert in 3 or 4 years. The rate of change here is just so crazy that, with a very mediocre prompt, we could build an AI VP of Finance from China.

5. Gross Margin Is Now the New Growth

We have a lot of early-stage founders that listen. I saw this brilliant tweet that I did actually want to discuss with you guys, and it was from Nicolas Dessaigne at Y Combinator, formerly the founder of Algolia. He said, “The most common reason I see good companies fail to raise their Series A/B right now isn't growth; it's margin. I keep meeting founders doing real revenue and growing fast, but once you remove delivery costs, there's almost nothing left. Investors don't fund revenue; they fund the margin on it. If that's you, fix the unit economics first. Fast growth on revenue you don't keep is a trap.”

That was the tweet. I like that. We've looked at Algolia back there. I think it's awesome.

Guest 2

I disagree with him on this. I'll go further than that. The objective reality is that's not what was happening. In other words, companies with tough gross-margin profiles and hypergrowth have been getting funded and, frankly, have been able to improve their gross margins and pull it off. You can see the pattern: it describes the foundation models, the inference providers, and the coding agents.

Historically, what he's saying is not correct. Now, he may be picking on something in real time, which is that the ability to build a company with a tough gross-margin profile and then fix it over time probably makes the most sense in that big-bang stage of AI, which I think was the last 3 years, where you went from nothing to something and had 10x growth. It paid you to grab the ground; it paid Cursor to grab the ground.

It may be that investors are now saying, “Hmm, as a first-generation coding environment, you can be Cursor, you can have negative gross margins, and you can still be worth $60 billion because you just grabbed the space.” It may be that, in the next generation, they're seeing—because they see a lot, obviously, given the volume—a little more focus on gross margins now, which is plausible. But there's no doubt that, to date, the bet that my gross margins are shit but my growth will cover it and I will figure it out, even though it sounds stupid when you say it, has in fact been 100% true.

Guest 3

I think he was synthesizing all the learnings across the Y Combinator portfolio and this change, right? Is it okay for startups to have negative gross margins or no clear path to positive margins because they'll figure it out? That's Rory's point, and I think you could argue both sides to Rory's point, but maybe that age is ending partially, right? Maybe it should end, right?

We're sitting here, and Menlo just raised a $3 billion fund after 50 years on the back of a crazy bet on Anthropic that went very well, right, when the margins were crazy. But maybe that era doesn't last forever. I'm sure all 3 of us are sitting on a portfolio company investment we made in the last 12 to 15 months where inference was the marketing strategy and the gross margins were negative. We're sitting here today and we're like, “The company's doing okay, but I'm not sure we're going to get right-side-up on that investment.” We're all sitting with a couple of investments like that, and then we see a few others that are wildly efficient. They took advantage of AI in other ways, and they're wildly efficient.

We're wondering, in that 2-by-2, do we really want to be in the bottom left? Is it the bottom left—highly inefficient and not better growth, not top 1% of growth? I don't think we want to be in that 2-by-2, do we?

Guest 2

Everyone will have some deals where you go, “We thought we'd earn our way out of the gross-margin problem,” and, for whatever reason, we couldn't—either because we didn't grow quickly enough to get to scale, or because the foundation-model company started grinding us, or competition bundled it in—and now you're just in a shitty gross-margin-profile business, which will go bankrupt.

Because when things slow down, no one says, “Hey, let's do an acquisition of an adjacent gross-margin-negative thing, and it's going to be fun,” right?

Guest 2

Yeah, that's all these startups will just fail.

Yeah. Right. If they don't achieve massive growth, they will all fail. I think that's the good simplification, right? And so, for an early-stage investor, that outcome is not that fun.

Yes.

Jason, you mentioned Menlo. Obviously, one of the biggest winners from Anthropic raised $3 billion. I'm sure we've all got friends at Menlo. Awesome, good people. Happy for them. They're also in Legora. They're in Lovable. They've been some of the best AI ambassadors of this wave.

6. Menlo Ventures Raises $3B: Why Not More After the Anthropic Win?

I think, categorically, the interesting thing, honestly, for me—and you're going to kill me for this, guys, child of the boom, whatever, whatever—is: why is it not more? They legitimately are one of the best. They'll deliver billions and billions of dollars back in the wake of Thrive, GC, and Lightspeed. This $3 billion is pretty conservative. I thought that.

Guest 3

I actually don't have an answer. The only answer I could come up with, right? I do think, look, not only is Menlo wildly successful, but they're omnivorous. They'll invest at every stage. They will take lead positions, but they will also do smaller checks and small positions. Even if they lose the round, they'll do 2% into the round, or maybe they'll invest more later. Maybe they won't. If it's a great one, they'll just do it, right? Kind of like how Felicis got going. It's any size, like 20% or 1%.

My first answer was that that's all they could raise. Poor guys could only raise $3 billion. Okay, that was my first read. Rory's shaking his head, the professor. Listen, I don't know. I shouldn't have gone first with the answer.

My secondary read is, look, they'll just raise another $2.7 billion in other vehicles like they did for Anthropic. This is just 1 or 2 funds, and they'll raise another $10 billion or $20 billion in SPVs, sidecars, or other funds. The headline number of the fund size is not always correlated to the amount they'll end up deploying over the life cycle of those investments.

I think they were smart to raise that amount because, given their amazing performance, they'll always have access to more. By not starting the clock on a much larger growth fund, with the fees and the drag that entails, you just set yourself up for success. Remember, all other things being equal, provided you've got a big enough check size to play in any round, as a GP you actually have higher risk-adjusted return if you have smaller funds because then you have less cross-deal aggregation.

In other words, 2 separate $1 billion funds versus a single $2 billion fund: all other things being equal, you have more probability of winning on at least 1 of the 2 funds than on the single $2 billion fund. That's not the reason, by the way, they did it, but just to point that out.

I think the zoom-out question is this: $2 billion is a lot of money. How many deals are there going to be like Anthropic, OpenAI, or Anduril where you can put $1 billion or $2 billion to work? Do you create your main vehicle for that, or do you accept that there are anomalies and you have access to that capital via an SPV?

You run your business on what you think you can deploy across normal cycles on normal deals, including wildly successful $10 billion or $20 billion outcomes, and then you accept that 2 or 3 times a decade there's going to be a trillion-dollar outcome. You want to have access to that capital, but you can get that via an SPV. I think it's a reasonably rational structure.

Your fund size dictates your strategy. Once you raise $10 billion in a fund, you're signing up—there are relatively few places you can put it. You're signing up to put a whole bunch of money in those deals, and that's the only place you can put it. One of the statistics in Q1 was that 70% of the venture dollars went to 4 or 5 deals. You're really signing up to put a whole bunch of money in those deals, and that's the only place you can put it.

I kind of like it in the way that you're like, “Hey, we're conservatively sized, and we can just take advantage of deal-by-deal carry on SPVs if anything does pop, and we'll have great fund returns if not.”

Guest 3

Yes.

Guest 2

I think as long as you can spin up the SPVs on demand, it's a great—it's the better model.

Oh, dude. They'll be able to spin them up like never before.

Guest 2

When I started investing, I had 2 hedge funds as LPs, and they said, “Listen, we'll each give you a blank-check SPV, and your winners.” I'm like, “Okay, that sounds like a great LP.” I took it and called them up for a winner. Both of them said, “Well, we need to meet the founders. We need to do some diligence.” I'm like, “This is the worst deal I ever got.”

You need that for the SPV to work. It needs to be 1 WhatsApp message. And I get that. I get the $200 million to invest, then I'm all in. Then it's the best venture model there is, right? I just need $100 million. I know you haven't heard it—it's a good one. Did you read my last investor update? I need it by 5.

Guest 3

I think the impressive fact that I haven't internalized about that announcement, because obviously I've known folks at Menlo since the mid-'90s, is that they're 50 years old as a firm.

Jason, 30 years younger than you—is Menlo still with the firm? Is Mr. Menlo still with the firm? Do we know?

Guest 3

Not with the firm.

Does he still come into the office a couple of times a week? Does he have an office in the back, or is he in the front of the office?

Guest 3

Stop. Stop, guys. Keep it sane here. My point is this: they're pursuing a strategy that I think is designed to survive and be successful across the cycles. My worry would be that if you were reaching and doing a $10 billion fund, you've got to be really sure you can put that somewhere without changing your strategy dramatically.

It's not impossible. I think Founders Fund has done it very successfully, but, again, to repeat the number, the number of companies that can ingest $1 billion and $2 billion checks is pretty limited, even on a decade-by-decade basis. So I give them huge credit. I think they've done an amazing job with Anthropic, and they've done an amazing job lasting 50 years.

But speaking of people crushing it, Kalshi passes a $2 billion run rate and starts prep for an IPO. How did we think about this news?

Guest 3

Americans like to gamble, and we weren't allowed to gamble for years and years and years and years. Then the Supreme Court said, “Screw that.” So Americans started to gamble, but the states regulated it. That's why FanDuel and DraftKings did well, but there were always regulations.

Kalshi found a way to pretend that it's a prediction market, found a way to get U.S. jurisdiction from the CFTC, and has convinced everyone that it's a prediction, which is different from betting. Eighty to 90% of what they do is sports betting. They're on a roll. They found a regulatory arbitrage to a wildly popular pursuit—says the person who's betting as we speak on the World Cup.

So, it's 80 to 90% sports betting. Sports betting is really popular in America. It was illegal because of the Puritans for the longest time, and these guys are riding a wave. End of complex analysis.

I do think if Meta really can copy it, for real—if they're comfortable going as far as Kalshi and Polymarket have gone—I do think there's a chance they'll clean up. It would be so convenient to go into Facebook, right, in my feed, or whatever, and just instantly bet on anything my friends are betting on. The social aspect for Facebook could be so powerful. I just don't know if they'll cut the same corners or not. I'm ignorant, but I actually think this is one where Meta could win big.

Guest 2

Because, look, I'm a little bit slow to catch on, but if I logged into Facebook and saw that Harry was doing something for $5 or $10, I might just do it with him on Facebook, right? Grandma and grandpa might like it. Kalshi is scary. What does it mean? Is it safe? The whole world—Facebook's still doing pretty good, right? I think it's a great place to bet.

To remind folks, Harry started by asking the question, “Why is Kalshi doing well?” And Jason, when you jumped in, you covered something that we hadn't explicitly said. Just to say it, Kalshi is doing amazing and is talking about going public. It's doing $200 million in revenues. Facebook—or Meta—just said they might offer a competing product, which right now doesn't exist, but you're speculating on whether it will take the business from Kalshi. If that's Kalshi's only problem, I think they'll be just fine.

Yeah, no, I'm not saying it'll hurt. I'm just saying I could imagine it being wildly, instantly successful if they're comfortable legalizing betting on their own platform and making it as elegant and as fun and as social as these platforms are. I might do that rather than mess around with these Kalshis.

Rory O’Driscoll

I don't know. I think the demographic for sports betting is young and male, and I don't think that's the Facebook demographic anymore within a million years, right? To be honest, I think it's actually smart of Facebook to think about it because the core Facebook demographic is aging fast, and this is where, as I say, especially young males are playing.

I don't know if it'll turn out to be good business for them, but I don't think they launch it and get a whole bunch of traction. My son's an avid bettor. I worry about that sometimes. He's been on Facebook for 5 years.

Let's just be direct and chronologically correct. Kalshi: what price does it go out at when it does go out?

Rory O’Driscoll

Who the hell knows? I mean, you know, they're talking about 10 times revenue.

Where's the prediction marketplace, Rory? So, that's PredictIt.

Rory O’Driscoll

That's actually a very fair response, Harry. You win on that. And that would be good because it would be one of the 10% of bets that isn't about sports betting.

Look, the growth is so amazing that if they're talking about 10 times revenue, if it's at $200 million today, by the time it could go public, it could be only 7 or 8 times. So it could be a really big win.

I looked at it before I came on, but I can't remember. FanDuel and DraftKings—one obvious question is: to what extent do they get revenue from FanDuel or DraftKings, which are 100% betting? Those guys have some kind of geographic issues. They have to have state-by-state licensing. Kalshi, because it's regulated as a prediction market, has been able to avoid all that. So they might have an edge on that.

Obviously, it's not quite the same thing. To be clear, the structure of the business is not quite the same thing as FanDuel and DraftKings. They are a classic betting house, and in the case of Kalshi, they're just a clearinghouse and they match buyers and sellers. So it's not quite the same thing, but to a rounding error, the experience is much of a muchness. So I don't know. I mean, right now it feels like a really good play.

Guest 3

Sorry, is there a chance that in a non-Trump administration, the party stops? Yes, is the quick answer. You are vulnerable to regulation. There are currently some states suing Kalshi, saying, “You really are doing betting, so we should regulate you.”

Right now, the administration's perspective has been to swat that down because they want uniform federal jurisdiction. And, as a reminder, it was the Supreme Court that basically legalized sports betting. So, yes, there is regulatory risk consistently in these businesses, but that's just part of the bet.

No. Well, Jason, it's to you, actually, based on what you were saying. In 2 years' time, will prediction marketplaces be an ongoing activity within Meta? Will it be a part of their product that you can use in 2 years' time?

Guest 3

I just think Facebook is one of these things where we underestimate the scale and the reach of it. It's pretty powerful, and it keeps growing. I think it's the one we underestimate. Once it natively overlapped and became part of Instagram, it got much better, and I just think it's super powerful.

Are there any that we've missed that you want to touch on before I do a rage bait? But, really.

Rory O’Driscoll

Well, we didn't do it. I don't know. Is Accenture worth the effort?

7. Accenture Plummets 19%: Why AI Is Destroying the Consulting Business

We can discuss it. I don't understand it personally. I just think it's interesting that 30 days ago it was an AI beneficiary, and now it's not. The context is that Accenture plummets 19%. That's on top of already being down, I think, about 20%, so it's down about 40% for the year. To Jason's exact point, I thought these were meant to be services that benefited from AI. Why is it down 40% year-on-year?

Rory O’Driscoll

I think 2 separate things are happening at the same time for Accenture. One is that the business of helping companies adopt GenAI is probably exploding because all the companies need help. If that was zero as a part of their business a year ago, 2 years ago, or 3 years ago, it's probably exploding and doing really nicely. Every company in America needs some help on GenAI, right?

The core business they're in—90% of their revenue is other consulting—and there are few markets more prime for disruption by AI than consulting in general because it's white-collar work. It's very—it's already been outsourced. That's why Accenture has it, right? And we're seeing a whole series of companies. The mental model for us is “AI for SI,” so the whole systems integrator market is prone to disruption.

If you look at some of the core business that Accenture used to do—SI consulting for an SAP deployment—there are companies like Tercera and Cognizant that are doing that.

Guest 3

AI for SI—systems integrators for Salesforce integration. There are companies like Swan AI and others doing that. In all these markets, Accenture was probably billing maybe $20–40 million for an SAP implementation. Today, they might only get $20 million for that because $20 million of it can be done using LLMs, right?

I think what happened is their core business came under pressure. Even though the new business of helping other companies adopt AI is exploding and doing nicely, ironically, the core business—defending yourself against AI—is in a pretty tough place, right? It's not the end of the world, but it totally makes sense, right? I actually think one of the investment themes we'd love to find an interesting bet on, and have talked to some, is around this AI-for-SI space, right?

Because all those consulting dollars are massively vulnerable to compression from AI. If you look at the kind of tasks they're doing, it's gathering requirements, building statements of work, and then writing fairly simplistic code to say, “How am I going to deploy Salesforce? How am I going to deploy SAP?” Those are precisely the kind of roles that AI will replace.

When I was a VP at Adobe, there was an entire floor of Accenture there for 5 years deploying Salesforce. An entire floor of people, right? I don't know what they made off that deal, but it was $26 million a year for Salesforce. So I'm guessing they charged at least $26 million a year for 5 years to get Salesforce up and running, right? That business has to be partially disrupted in general. Also, seat compression hurts them, right?

The one thing I would add to the list—we talked about it right before we started—I didn't know D-Wave was on such a roll. D-Wave is up 30% this year, one of the big winners. We talked about it: the IPO crashed, and it's up 3× from the bottom. When I did this the other day, I was just trying to do my own little analysis to oversimplify the public markets: everyone selling seats, for the most part, is getting crushed; everyone selling variably, one way or the other, is winning. Not all of them, but pretty damn close, right?

Yeah. I think you're right, Jason. The seat element is part of it. The long-term trend isn't great, but I think the short-term crushing is less because the number of Salesforce seats at Adobe went from 1,000 to 900. That sucks. But to your point, the real point is you had literally a floor of consultants sitting there, getting $20 million a year for 3 years. When you look objectively at the work they're doing, it's probably some of the easiest work for LLMs to do. It's automating.

Guest 3

It was migrating from Siebel and your own databases to Salesforce. It took 5 years to lift data, right? And now Databricks, which is on fire—we won't hit it this week—claims they can do that lift in 30 days for their customers. An entire lift to Databricks for anybody using LLMs. It's actually an under-discussed story that Databricks promised—what are they growing, 80% at $6 billion or something like that, accelerating? I didn't really get this until one of the founders came to me and I researched it, but they tell you, in 30 days, we will do an LLM lift of all of your data, whatever the amount is, right?

I'm sure there are exceptions and edge cases, but compare that to a 5-year lift with Accenture to go to Salesforce. We just did one. We've been trying to get off Marketo for 5 years—it's our worst software—and then, believe it or not, Salesforce used their LLM thing and moved us to their product in a couple of weeks. They just lifted it with no humans. It's a big deal, this lift. It's an under-discussed story and a moat destroyer.

It is a moat destroyer.

Guest 3

It's a moat destroyer when LLMs will lift you from one vendor to the other. I don't even like—I literally was doing a pitch this week, and the founder was going on and on about their moats, and I immediately didn't want to invest. Enough. Your moat can be LLM-lifted away.

The interesting thing, staying with the Accenture problem, is that one of the things we had done about 2 years ago is just say to yourself, when you think about what work AI will replace—white-collar work—you can say BPO is a good proxy for that. Anything a company is willing to outsource to India, they're probably willing to outsource to AI. In fact, you can literally just take the BPO spend and look at the BPO, start saying, “Okay, there's a whole bunch of places where AI will win,” right?

When you do that, Accenture's top of the heap, right? They have all that business. What's challenging for those companies is that you would think they would want to adopt this technology, but their business model is—there's only one thing worse than a seat-based model, Jason, and that's a model that's based on bodies.

If your business as Accenture, or any of these SI companies, is, “I bill out 100 people, I pay them $200,000 a year, I bill them out at $500,000,” and that's how I run my world, if I don't need 100 people—if I only need 40 people and some AI—that makes my head hurt. I've got to get rid of these 60 people or find another project for them. I've got to figure out how to build that. My whole margin structure collapses, and my take, being on top of the heap, goes down.

What we're seeing is that some of the biggest consulting companies are adopting AI technology, but they're really struggling to pass on huge price increases. That leaves the room open for newer companies to come in and say, “Look, we are AI-first. We are using these tools. You got a bid for $80 million from Accenture? We'll do it for $15 million.” That kind of thing gets the attention of the CIO. We think that's a super interesting place to invest. So, yeah, Accenture has a lot of structural questions. This is going to be a constant theme there for the next couple of years—the pressure on those kinds of businesses.

Welcome to law's billable hours.

8. Work From Home Is White Collar Fraud

Guest 3

It's a little like the law comment, but even worse, I would argue. I think law, at some level—yeah, you want the work done, but you also want the wise guidance. If you're doing an SAP migration, you just want the damn thing migrated to the new version of SAP, and you don't want to talk to these people ever again.

Okay, we're going to do a rage-bait one, but real. We had a little clip that went slightly rage-baity this week: 5.5 million views. Ryan Petersen at Flexport said, kind of jokingly, obviously, “Work from home is white-collar fraud.”

I have 2 kids. When they come home at 3 from school, of course it interrupts my work. Yes, I have a home office, but of course it does. It's not the same as being in person. Rage bait or real? I saw it. I thought it's just dated, was my read, for what it's worth.

Guest 3

Listen, I get in a little trouble on the show if I agree with him. I think a lot of work from home in that era was working 15–20 hours a week. It also involved a lot—I saw it on my own team—of distractions from home, right? On the other hand, plenty of folks do make it work, right?

The reason I say it's dated is that the companies we want to invest in—and this is a very narrow set of the universe—aren't hiring folks who want to work 20 hours a week from home. We're just not hiring them. Our portfolio companies, at least our newer ones, aren't. Our older ones still are, but our new ones just aren't. The whole way you build a startup to your first 100 or 200 employees, I think, is radically changed and is under-discussed.

When we started this podcast 60 weeks ago, it was toxic to be running Cognition and telling folks they had to work 7 days a week, and laying off half of Windsurf when you acquired them because they weren't willing to work hard enough. That was utterly a toxic thing for the founder to say. Today, it is how you build a winner. You can't win in your marketplace if people are working 20 hours a week. You can't win.

The only thing I thought is that Ryan's running an old company, right? It's old. Flexport is old. I think he's struggling with trying to modernize his team and be competitive with the way startups are today. He's struggling with the fact that you can't change out your entire team.

What do you do with the folks who are unwilling to change at these big companies? What do you do with our 10-year-old portfolio companies where 90% of the folks are unwilling to change? That's what he's railing against—the folks who really just want to work 20 hours a week. But I think it's dated, right? I literally hope that no new investment I make is structured this way. I feel like the majority of my older ones are structured this way.

Can I be direct?

Guest 3

Yeah. I want small, high-paid teams that work in the office over 6 days a week. I'm not interested in investing in anything else. I'm just not interested. And it's not because I don't have empathy. It's because they're going to fail.

Jason, empathy is the word I think about when I think—

Guest 3

I am an empath.

It's okay. Let me just ask a question here. Let's do a simple 2-by-2, right? I'm pretending I'm a consultant—poor guys. There are 2 dimensions: do you want to work hard or not? In other words, yeah.

Hardworking, not hardworking, and then effective, not effective. Right? Do you think the problem with working from home is that people don't want to work, or do you think it's not effective?

Guest

Yes. But I'm just curious: we need to hire teams that are half the size they used to be, that are paid top of market, that get double the equity because the teams are smaller, and we come into the office, freaking the Cognition guy. What is his name? What's the insurance guy you had on the show that people made fun of?

Yeah. With his café.

Guest

I want to invest in startups with a café that runs 24/7 because you can't win when Cognition is running 24/7. You're not going to win.

Here's the problem, and I'm struggling with this intellectually. It is not a sprint. It is a marathon. But today it actually is a series of endless sprints. This is so hard because it's a sprint: you get about 5 minutes to relax, and then today OpenAI released the Jalapeño chip.

We didn't even know it was coming out today. Now they have their own inference chip. Maybe we don't need Cerebras anymore. Maybe the whole market's going to be disrupted in 60 days because OpenAI is going to run its own inference. Do we get to breathe, guys? You don't get to breathe anymore.

Sorry, guys. We no longer get to breathe if you want to make any money. Do you want to make money from your equity, or do you want to make $180,000 a year? This is going to be your choice in tech. Do you want to make money from your equity, or do you want to be well, or do you want to watch? Honestly, this is the choice.

You want to watch, or you want to make $10 million or $100 million. There's nothing in the middle. You don't get to make $10 million for working 18 hours a week. You get a watch. You get an Omega. You want an Omega, or you want to be rich? Make your choice, boys, and pick your path.

I say to people, pick your path. Go work for that old software company growing 8% a year. Go work there and make your $180,000 or $220,000, and wear a really nice— they all have nice watches at these companies now. Or go work in the office 6 and a half days a week at the Corgi Cafe with a chance to make 8 figures. That's the choice today. This is a different world.

I don't want to invest in anyone in the middle. And I can't afford to invest in the folks that want watches. Enjoy your $12,000 Rolex or your $8,000 Omega. I hope it's meaningful to you.

9. OpenAI Launches the Jalapeño Chip

Okay, I'm letting it go. Boys, do you want to have one final addition in the 5 minutes remaining? OpenAI has announced that it is doing a custom chip. If Google has TPUs and Amazon has Trainium, OpenAI now has Jalapeño, co-developed with Broadcom. OpenAI says this chip beats current state-of-the-art GPUs on performance per watt. Broadcom's CEO is on record saying it cuts costs by 50% versus a typical GPU.

Inference is 50–60% of revenue, plus or minus, and we know, when you peer through capex—when you look at capex—GPUs are well over half of that. So that's 30%. Maybe if you could replace all of them, and NVIDIA makes 70% margins, you can make some kind of intellectual case that you can save some money with this.

But honestly, my real response to it is: you've got plenty to be doing elsewhere. OpenAI is winning on the top-line side. I'm not sure I'd spend all my time optimizing. I'm not—I hope this isn't distracting you from the things that matter, right? So, yeah, I'm trying to give a shit.

10. Open Source Is Hollowing Out the Middle of the AI Market

Guest

I'll throw out one thought, and maybe we could talk about it next week, for what it's worth, because one topic I thought we were going to do more this week, which we didn't, is talk about open source even more, right? I do think that open source, as token maxing becomes a bigger deal, is more and more important. And I think the question is: how do you win for certain workloads if you're OpenAI or Anthropic?

Well, you cut your inference costs to be cheaper—actually cheaper than commodity open-source providers can provide it—because they still have to buy the GPUs. Open source is not a moat. Again, it's not free, right?

Listen, if I can run my own inference farm one way or another, buying somewhat expensive NVIDIA chips, which is how I have to do it, right? But the advantage is I get to keep the margin, right? Or I get to keep most of the margin, or OpenAI can provide me with a cost-competitive product because its inference costs are half of what it costs me to do it on open source.

This is not just about cutting costs. This could be about shoring yourself up against open source, which has the potential to kind of destroy the middle of their market. People are always going to use the best models for frontier applications, right? And actually, at the bottom of the market, these guys are pretty competitive, but the middle is open to massive disruption.

But if you cut your inference costs in half, the truth is open source is probably only twice as cheap for a lot of use cases. And if you cut your inference in half, your model still makes sense. That's why I think, in theory, it's a big deal. It's not just about capacity and driving down costs. It's about this flabby middle. The flabby middle in AI is at risk.

Okay, I couldn't articulate it.

Guest

Next week you're going to agree with me.

No. I'll disagree right now. I understand the bet. OpenAI and Anthropic have discovered the single best tech market in terms of consumer demand in the last 20 years, and they should put all their effort into meeting that demand.

Vertically integrating backward down the stack—2 levels down—not just vertically integrating to own a data center, but vertically integrating to own a chip that goes into a data center, doesn't strike me as the highest and best use of resources. Why do I say that? You've got 3 or 4 cloud providers who are dying to do business with you. You've got Oracle, Google, Microsoft, and CoreWeave. You've got a whole bunch of vendors 1 level down from you.

Some of those vendors themselves have chips. Google has a chip. Amazon now has a chip, right? There's a whole ecosystem of people breaking their backs to provide you with cheap compute and taking on all the capital risk of that. God bless their poor little selves, right? Oracle will rue that day sometime, as will Microsoft, and you're sitting here and should be putting all your effort right now into grabbing those end customers in enterprise.

If you've got additional time and resources and you can also take on building a chip, have a go and knock yourself out, right? And I get it intellectually: you're right, Jason. At some point, to the extent it's become a cost business, optimizing the whole vertically integrated stack might make sense.

But the whole point—I mean, if you look at it, you say that, but right now, if OpenAI and Anthropic had to do vertical integration today, they'd be fucked, because the whole reason they work is because they've outsourced $300 billion in capex to other poor fools. And the definition of vertical integration is taking in-house things that were done outside, right?

The whole reason the OpenAI and Anthropic models work is because other idiots have spent the $300 billion on their behalf. At some point, maybe you have to vertically integrate, but it just doesn't strike me.

Guest

Well, look, it's not that I argue with you. First of all, the fact they launched this as we're recording this is obviously a decision that was made in a different era.

Yeah.

Guest

Right. It was even before the TBPN era, right? It was an era of abundance. So if this decision was made today, it would be a very different conversation as to whether it makes sense today. So, all that caveat aside, we're actually looking at some sort of early-2025 decision, when the world is radically different.

Having said that, maybe we can pick it up next week. I just think that OpenAI and Anthropic have massive existential risk that didn't exist at the start of the year, which is—we glossed over this GLM-5.2 open-source whatever. It's not just performance; it's the fact that, as we come under cost pressures, the middle of their market is under massive threat.

What you would do if you had a high-margin product like software is just have a cheaper offering; you would subsidize the middle of the market. But what actually happens is they really only have the high end and the low end today. They have Sonnet, Opus, and friends, which is great, and then they have Haiku, which are these super-small models.

They don't have this middle product. The problem with the middle product is it's too expensive to provide. So if you can cut your inference costs to below open-source inference costs, then you can have this middle market where you can serve every single workflow rather than have your flabby middle hollowed out for you.

I think the flabby middle is high risk for these guys. Just when they're all ready to go IPO, they have a brand-new existential risk. The flabby middle—just when things are getting good.

Agreed on the cost side, and I totally agree. I'm just questioning whether, in order to minimize your cost, you have to vertically integrate backward 2 steps. Not only do you own the data center, but I query whether you really need to vertically integrate back through the hyperscaler and also integrate at the chip level.

It just feels like a lot of backward vertical integration when, in a competitive market, there are 3 or 4 chip providers and 3, 5, or 6 hyperscaler providers. Surely you can just beat the crap out of them on cost, because you are the largest buyer on the planet of this shit, right?

And remember, I mean—how—okay, let's actually check. Let's check if the market cares.

Why do I say that? Well, we should actually see what happened to Cerebras’ stock today. If you remember, Cerebras went public, had a decent quarter, and was going really well. Its big traction is a $20 billion chip order from OpenAI. Did the stock go down a lot? Because presumably that is now—I mean, yeah, it did. It went down 16%. So there you are. The market said, “Poor old Cerebras. OpenAI is going to build that chip instead.”

Now the question is, would they have done just as well by saying to Cerebras, “You know, we’re playing you off against TPUs and NVIDIA. We need a 20% discount”? Did they need to build their own chip to do that? I don’t know. Right.

Guest

Well, it doesn’t matter because it’s so many generations ago. The decision that was made is irrelevant today. The world changed so much, right?

That’s totally fair, which, by the way, gets to how these cycles end, people. I’ll really believe what you’re saying if the next stage is when someone comes in and says, “What we need to do is build a foundry and make our own DRAM.” At that point, you’ll know the cycle’s about to go really badly down. If you vertically integrate back into memory, then you know it’s over. Just you and the Koreans.

Look, I’ll just say one thing, maybe as we can find out next week and as the months go on. I firmly believe, especially if we just want to tie this to B2B software and stuff, for some of the heart of the show, that we need these middle-level models from the closed-source providers for the business to work. Not everyone can afford to run every workflow on Opus, right? And Haiku and Mini are too small, and this is where open source is going to disrupt the market. This is more important than it looks if it can work, because you’ve got to head off this middle disruption, this middle layer.

Guest

Yeah, first of all, I want to say I totally believe it. I think the number one question is: How will software companies access mid-priced, high-quality intelligence that hasn’t got frontier pricing? We’re seeing people start to gag on pricing. I totally agree with you on that.

In my comment, I don’t think it’s a question of optimizing the stack that gets you there. It’s, what’s the business model for these guys to provide it? Maybe with only 2 frontier providers, there’s not enough competition. If I was someone like Google, I’d be looking at—that’s why I go back to, “Come on, guys, wake up.” I’d be looking at this market and saying, “How do I put a lot of pricing pressure on Sonnet, on Opus, and on GPT-5.5 by providing a just-almost-as-good, US-based, competitively priced product?” I mean—

Yeah, they just haven’t made it happen. Okay, boys. It’s so good to have you back from China, Jason. We missed you.

Guest

Well, I thought the guy from Benchmark was pretty good. I don’t mind being replaced.