20VC：Anthropic递交S-1，瞄准2万亿美元IPO｜Instinct传融资300亿美元｜微软裁员27,000人，Meta则通过股权市场筹资950亿美元，推进资本开支狂潮

As crazy as this may sound, I think folks are almost underestimating Anthropic because it took 5 years to IPO, versus 24 for SpaceX. It’s not just something we’ve never seen before. We’ll see hints of it with OpenAI’s IPO, right? It might be worse; it might be better. Probably slightly worse, based on what we know: less capital efficient, slightly slower.

But JFC, getting here in 5 years from a spite startup is just beyond unprecedented. Time value to liquidity still matters. Markups are fun, man. But getting out of the greatest IPO in our lifetimes in less than 6 years and getting distributions—I know Rory will roast me for it, but you gotta think: “Why would I do anything else? Why would I wait?”

To pay for all this infrastructure, all these GPUs, infrastructure, and inference, we gotta get rid of the people. I think the Chinese open-weight models are pretty much over in the US. How much money is there for the Instincts of the world? It feels infinite now. I think we should put 20% of the fund into OpenAI at the $1.25 trillion round.

We have some big news to start with: Anthropic files its S-1. It’s the moment we’ve been waiting for. $16.3 billion in first-half revenue. I just want to hand over to you, Rory, because whatever metric I give, you’re going to say it was the wrong one to cite.

No, I’m not.

So, which ones were you drawn to? And what are the takeaways for us from this?

1. Anthropic’s S-1 Stuns Investors

It was a great S-1. Frankly, all the haters and the worriers, at least for now, have to go pound sand. The burden of proof is most definitely on the side of the naysayers. In other words, the S-1 was great in almost every dimension. It’ll be wildly well-received.

There are still some open questions. They didn’t give the Q3 revenue number. The overall revenue numbers are known, and they were amazing. It was better than expected on margins, operating income, even operating income after stock-based comp. It was great on customer concentration and compute efficiency. It was stunningly great on cash and the cash balance sheet. You have to struggle to find questions, and there are questions we can come to in a second. But it was a really top-tier performance.

As crazy as this may sound, I think folks are almost underestimating Anthropic. Five years to IPO versus 24 for SpaceX—it’s not just something we’ve never seen before. We’ll see hints of it with OpenAI’s IPO, right? It might be worse, it might be better. Probably slightly worse, based on what we know: less capital efficient, slightly slower. But JFC, getting here in 5 years from a spite startup is beyond unprecedented.

Time value to liquidity still matters. Markups are fun, man. But getting out of the greatest IPO in our lifetimes in less than 6 years and getting distributions—I know Rory will roast me for it, but you gotta think: “Why would I do anything else? Why would I wait?”

No—

These C deals are like 18 years to full distributions, right? I could get full liquidity in 6 years with Anthropic. Why would I waste my time with anything that’s—

Totally.

—10 to 12 years to IPO? It’s just one deal in the fund, then another 2 years to distribute. It just works in 5 years?

I won’t roast you, Jason. I saw a tweet that basically said the same thing: every founder and every investor is looking at this and asking, “Why am I bothering with anything else?” These are such stunning numbers in terms of size, scale, and speed. You’re right, it’s going to dominate the results for a long time to come. Everything else just seems like a lot more work for a lot less return compared to it.

All credit to the founding team, and all credit to the investors who got in early. Things going up and to the right faster than anything else in human history for 5 years, adding up to a trillion dollars—that’s as good as it gets in venture.

It was funny watching Byron Deeter from Bessemer on the TV show saying, “This is the greatest investment in the history of Bessemer.” I think Bessemer was founded in the 19th century or something like that.

Yeah, I think so. In the 1800s.

I don’t know what the fund was, right?

Yeah.

And it was their greatest investment and their largest checks, right?

Yes.

So to see that all happen now after 100 years—that’s the whole story, right?

Yes, it is the whole story.

Greatest investment, biggest check, fastest time to liquidity. Why are we doing anything else?

Yes.

Rory, you said there were questions there. There are always cynics in the audience. There are always people throwing peanuts. If you were to present the alternate view, is there a striking number or two that—

2. The Six Month Model Cycle

No. The interesting one is that 80% of their revenue comes from models that are less than 6 months old. Separately, you’ll note they didn’t give you a gross margin, an actual gross margin calculation. They just didn’t lay it out that way, because you can do that in accounting. You can just put all the compute costs in one place. Then they give a footnote breakout that about 70–80% of the compute costs are for training, not inference.

In other words, the models are wildly expensive to build. Once you build them, they’re fairly wildly profitable to serve, and you only get to serve them for 6 months before you have to invent another one. It’s not like software, where you build something and maybe you can sell it for 10 or 15 years. In this case, you invent something, get most of the revenue from it in 6 to 12 months, and then you have to keep investing masses of money to build the next one.

The question is, what do you make of that? The funny thing is, those are the facts, and you could look at it one way and say, “Oh my God, that’s totally scary. If you miss even once, you’re screwed.” On the other hand, you can say to yourself, “That might actually be great for Anthropic,” because they’re best positioned to build new stuff, and they want to build new stuff. They said that very clearly: “We want to build new stuff.”

Very few people are going to be able to compete with this, right? Very few people. If the next turn of the crank on the model—if they spent $8 billion on compute year to date as of June 26, and they’re going to spend, I don’t know, $16 billion next year—there are very few people who can write a $10 or $12 billion compute check.

So they can continue to invent these new models, get masses of revenue from them, and then maybe, at the back end, other folks have it. Even though that statistic was kind of weird and scary—“Oh, my God, the shelf life of your amazing product is 6 months”—I would argue it’s actually an advantage if you’re the kind of leader who’s willing to be aggressive about it.

But it does point to the incredibly dynamic and scary nature of the game. This is a game where you have to spend billions and billions and billions of dollars just to play.

But that was Elon’s point: most of these new labs are just going to die after this, right? They just can’t—it’s not possible to raise the capital to survive. I thought there was one bad metric that wasn’t in the S-1. I thought the worst metric wasn’t in the S-1.

Go for it.

The worst metric was that Broadcom had to bail out this debt issue so Anthropic could buy its chips. The market was saying, “It doesn’t get any better than Anthropic. We just read the S-1, right? It doesn’t get any better.” But they weren’t able to get this debt for this semiconductor deal with Broadcom, and Broadcom had to come in and guarantee most of it. They had to pay even higher interest rates than expected.

It’s not terrible, but it certainly suggests that going forward, fundraising—even for Anthropic and OpenAI, and even if it’s somewhat indirect—is getting harder. Combined with the fact that it’s getting harder and harder for the banks to lay off and sell the AI lab debt, I think that’s the real, the only thing to be nervous about: that the debt is exhaustible here.

And—

—that the debt is exhaustible here.

Pushing back on that, because I do hear you. I think the amount of debt the banks are digesting has hit the limit. But I think the interesting thing about this week and the last few weeks has been that we’re equitizing the issue. In other words, it was all debt. But look at it: Anthropic is about to issue plus or minus $100 billion of stock. Meta issued $70 billion of stock. Earlier this year, Google issued about $80 billion of stock. Obviously, SpaceX did its IPO.

What you’re seeing is that there’s a class of companies that have cheap costs of capital and are willing to take equity dilution to make it all happen. My aha from that is that, at the macro level, it’s just going to keep happening. This trend is not stopping, and the people who want it to happen are willing to take equity dilution to keep going, so the spend will be there.

You’re right, Jason. Elon’s doing his thing, cutting off the oxygen from the other people. It’s like, “Oh, you people who just have debt are screwed. You’re going to die. I, on the other hand, have one of the world’s most inflated market caps, so I can issue equity and get access to all the chips I want.” So it was very much a kicking someone else while they’re down move, which is very vintage Elon, and you’ve got to respect it.

But they’ll have, like, $200 million of total cash on the balance sheet after this. Anthropic will. So that is a lot of cash.

I mean, let’s be precise. I think it is either the first- or second-largest pile of cash in all corporate America.

I read it was the second, but it doesn’t matter which one it is, right?

Yeah. This is a company going public that already has more money than every other company in America bar one, with $110 billion on the balance sheet today. They’ve done a masterful job of raising $130 billion and only spending $10 billion or $11 billion.

Yeah.

Masterful.

Do we think it’s going to have an amazing reception?

I do. I mean, I’m not sure. Hell, I don’t know. It could—

Oh, we’re in the predictions game, Rory. Go for it.

I know. I know.

Go for it.

I think it should. I mean, I think it’s an amazing company. You just can’t argue with it.

Yeah. I mean, doesn’t SpaceX almost instantly have to go into the S&P 500?

Not the S&P 500. If you’re listing on NASDAQ, which I think it is—I didn’t check—you go into the QQQ index. But the S&P index takes probably… there’s a profitability metric or something like that, so it takes some time. But yes, they would be in some index funds pretty immediately. I don’t think this is going to need that lift in the same way SpaceX arguably did. I think this is going to be just a stock people want to own.

But even just getting indexed into everything in the NASDAQ is a retail boost, right?

Totally.

It’s just hard to bet against. They’re not going to miss a quarter, whether it’s 900%—

No.

I mean, within a range. It’s not that I don’t think they could disappoint, right? But within the next 2 quarters, I think we have to be careful to use NRR as a metric here. But it’s still true, right? With 900% NRR—even if in reality it’s more like 300% to 400% because it includes pilots and trials—I just think that, plus the brand, plus being in the NASDAQ-100, it’ll probably have a good retail run, plus SpaceX. Why would you bet against it?

For Sam Altman, reading this and thinking about taking OpenAI public himself, do you think he takes anything away from it? Does he change anything as a result?

You can’t change the past. If you were the CEO of OpenAI, you’d probably look at this and regret decisions you made because you let these guys get away from you and do better. If there’s any self-reflection, there should be a ton of that going on. There’s nothing you can do about it. You just have to get your business together and converge and get out yourself. In the meantime, obviously, compete aggressively on price to try and screw up their Q3.

Never has such a successful startup named itself so poorly, that I can think of, as OpenAI. I mean, what they went through would have killed almost any other startup, right? With the CEO drama and all the other nonprofit-to-profit drama, despite Anthropic doing better than them, they basically let Anthropic control part of the market for almost—

Yeah.

—a year, which is like 5 or 10 years in AI time, right? There is a certain amount of parity with Astra and the open-weight models today, at best, but Anthropic owned the—fucking owned the playbook for the chessboard for almost 12 months. There was no point in using any other coding solution, and coding is the killer app, as Rory has said 100 times, and the math is true.

They literally let Dario walk out the door rather than build it in-house, and then fired the CEO and brought him back during the night of long knives. My God. It’s all going to be okay, but talk about own goals. Every startup’s full of them. These are some epic ones.

I had the joy of having Satya Nadella on 20VC earlier this week. Great show. It was a fantastic guest to have on. What an honor. Microsoft had some pretty big news: they were offering voluntary buyouts to 60,000 of its 223,000 employees. It’s an enormous amount of the employee base to get this offer, roughly 27%. Jason, you’ve spoken before quite vociferously about this topic. What are your thoughts?

First of all, Rory may have observed more of these than I have. Maybe he has, and I hope he hasn’t, but it’s not like all 27% are going to take it, right?

Totally.

And they’ve already been through a—I already think they went through a 10% layoff round.

Yes.

Right? So there’s some interesting game theory here. You want to offer it to everybody you don’t want to have, and then you’ve got to figure out what the yield’s going to be. But it’s brutal, because what it says, no matter what people are going to say on LinkedIn, is: these 27%, we don’t want them.

Yes. That’s exactly my—

Whether 7% take it or 2%, we don’t want them. It’s a worse scarlet letter than a layoff, than a pure layoff. At least with a layoff, you can kind of wave your hands. This is 27% of the company they’re saying, “You do not meet our minimum. We don’t want you.”

But will they really leave?

And I bet it could have been higher, but I bet they got this done in a week. It’s so damning that Microsoft doesn’t want 27% of its people. That is the age of AI. They don’t want them.

I’m not sure. Why would you do that if you’re Microsoft, though? What a culture killer. So you stay and then know that you’re not wanted? Isn’t that awful?

I want to get Rory’s take. I think you’re trying to do this in a humane way, and I think what we learned is there’s no—when layoffs started, in late ’22 or so, everyone was flamed for doing them on Slack or by email. “Oh, it’s so inhumane. Go meet with people,” right? It’s not that much more humane to meet people. Have you seen Office Space? It’s not more humane to meet with someone in HR and get fired.

And then they were like, “Don’t do it over the holidays,” so they fired people on January 1. That was no better. So this is a way to make it more humane. It’s great for folks who were thinking of leaving anyway. It’s great for them.

Yes.

Right?

They’re the winners.

It is great for the folks who are at the tail end of their career or want to check out of the workforce. That’s why this is so hard to do from a game theory perspective, because you might get all the folks who were going to leave anyway, right? But Microsoft has so much cash, who cares?

It’s an attempt to be humane. I’ve been on the other side of a layoff, though it’s been a long time. It’s hard. You try to be a good person, but it’s hard to make it humane. It sucks no matter how you do it, and I think this is the suckiest way, because 27% of the company has just been issued a scarlet letter.

In attempting to be humane, I think this is the least humane way to do it. Just fire another 9% in one day, do it on Teams, and move on. It’s rough.

It is rough. I think you’re right. I processed the same thing: you’ve been told to go, but you decide to stay. The only winners are, as you say, the people who’ve got another job lined up and now have a call option on a severance package.

But it is interesting that they think they can do without that many people. If their take rate is 10,000 or 20,000 people, it’s clear they think they can do without another 10 or 15% of their roughly 200,000 employees, which possibly is true. You wonder how much of it will be AI efficiency and how much of it is just that you never needed them. It would be fun to find out.

As the depreciation bill for the CapEx comes due at these large companies, you’re going to have to find it somewhere else, and we’ve talked about this. Frankly, earnings this year, 2026, are in that golden period where the people selling semiconductors are booking the gains, but the people buying the CapEx aren’t booking the depreciation, so everybody’s winning.

But as you start to look to ’27, that depreciation is going to start hitting at Microsoft, Google, Meta, and places like that. To keep operating income up, you’re going to be really pushing on headcount. This is when the bill for AI comes due, and this is how it’ll be paid.

It’s brutal. In a sense, it may feed the anti-AI frenzy, because it’s the direct narrative of the story, right? To pay for all this infrastructure, all these GPUs and inference, we’ve got to get rid of the people. It is true.

So you think 2027 is when we’re going to see that really crystallize even more?

Yes. Intuitively, as time goes by, everyone’s trying to call when the hyperscalers start going down the other side of this. As the depreciation bill mounts up, so far, even though free cash flow’s been going the wrong way, everyone’s operating margins and operating income have been going up. That’ll get harder now. This will be the problem. This will be the issue.

I’m just surprised the Cloudflare layoffs aren’t being discussed more, because I think it’s—

Oh, I missed that. You told me that.

—a case study 101 for—

Yeah, yeah.

—the CEO MBA, right? Matthew comes in, an epic year for Cloudflare. Every quarter is growing faster, but on May 7 they lay off 20% of the folks. Bam. We don’t need these folks. We’re too big. We’re better off with 80% of the people than we were before, and the business keeps accelerating.

Every CEO arguing about this should just look at that. You can say at Microsoft, some of it is the old company, and we’ve got to pay for all the AI, but Cloudflare is just like, “This is holding us back 20%.” It wasn’t 8%; it was just, boom, we’re doing 20%, with no downside to the business that’s apparent yet.

I think layoffs 3.0, or 2.0, are just going to happen. Part of it is to pay for the infrastructure, and part of it is just, “Enough already.” Enough of the performative AI, enough of this. Think about every board meeting you’ve been in over the last week, or every meeting you’ve been at. No matter how well you’re doing, it’s just stressful. Everything is moving even faster than at the last board meeting. There’s even more competition in every space.

We just can’t carry people any longer. We can’t, or we’re going to die. We’re going to be killed by that new company Harry just funded or Rory just funded. So there’s no stability. The tough part about Microsoft, Salesforce, SAP, and every SaaS company back in the day is that we were all predicated on getting to a level, a plane of stability at some point—$100 million, $200 million, $300 million, $400 million, $500 million of ARR—where we could breathe.

There are no planes of stability at all. Maybe north of $10 billion, $20 billion, or $40 billion, but they just don’t exist. I think layoffs are just going to accelerate because we have no choice.

Speaking of having no choice and where cash comes from, Meta goes to the equity markets. Zuck needs $95 billion, and this week, on Monday, it launched a raise of up to $95 billion: $35 billion in common and preferred stock, and a $60 billion at-the-market program. The thing to take from that, in combination with the fact that chip stocks lost $200 billion in a day, is that public markets are in a bloodbath as neoclouds shit the bed.

No. The short answer, Harry, is no. I think those are 2 separate points that you merged, and I’m dealing with each of them in succession. The Meta thing was what we talked about. This was the first quarter where their actual free cash flow dipped below their CapEx, right? So they were free cash flow negative.

Remember, their core business is one of the best businesses on the planet. It spits out money. It’s awesome. They have $33 billion a quarter of cash flow. It’s a great, great, great, great, great business. This is the first quarter they went negative, so they just did some equity. That’s that, and it’s totally unrelated.

The chip thing: chips moved last week, but let’s be frank. Since the Anthropic S-1 hit—and I don’t know where they’ll be when this show comes out—the consensus is that it looks good. One thing the S-1 was good for was chips, so those stocks are actually up.

In fact, DRAM is way up today, which is the definitive one I keep an eye on: the index of memory chips. That’s why Korea’s up 10% today. SK Hynix, right? I actually think the news is that if you’re in the semiconductor CapEx business, it’s been a good back half of the week. It was scary 2 weeks ago with CoreWeave, but CoreWeave is a $20 billion or $30 billion company, and Anthropic’s a $2 trillion company that has $200 billion when this deal closes. They’re going to spend it all on chips, so the chip companies should be happy.

The only counterpoint, not to get off track: it’s all green, green, green, but it is a different ecosystem. I know it’s apples and oranges, but the STAR 15 China, which is their version of the market, is down 40% this year. We do have interesting markets that are out of sync here for AI. China is just as gangbusters in its own way, but it’s interesting that that one’s down 40%. Different market, weird liquidity, while ours is on fire.

Speaking of China being on fire, DeepSeek raises about $15 billion, the largest-ever private Chinese round. The largest investors are battery maker CATL and Tencent. Anything of note here?

I think the Chinese open-weight models are pretty much over in the U.S. I think it’s the beginning of the end. It’s great that DeepSeek raised this, but it’s beginning to be banned across all federal agencies.

There was a horrific thing this week that was oversensationalized: this Cohere-Qwen thing. I don’t know if you guys saw it.

Yes.

It’s pretty terrible.

I’m sorry—

But it’s terrible on a number of—

Can you explain it?

But it’s not as bad as it looks.

Explain.

What happened, Jason, for people who don’t know?

Well, listen, Cohere is now a fourth-tier LLM, a fourth-tier player that has only sovereignty on its side, as near as I can tell. Harry, you’re welcome to correct me, or Rory.

Sure.

But it has sovereignty. If you talk to anyone in enterprise, sovereignty is in the top 5. They found that Alibaba’s existing Qwen model was calling Alibaba’s servers in both China and Singapore, pinging them and uploading data to them when it said it wasn’t doing that on the internet. It’s everybody’s nightmare in the U.S., that our data is being uploaded to China.

Now, I think Cohere was more careful about how they did it, even though they’re talking their own book.

You can’t really trust anybody in AI. They’re all talking their own book because this is good for them. There’s no evidence this is a backdoor, as some on social media thought. This is probably a training mistake at some level, but it doesn’t matter.

It doesn’t matter if, 1 out of 100,000 times, your confidential data is being uploaded to Alibaba servers in China or Singapore without your knowledge and without your consent, when it’s been directed not to. What CIO or CISO is going to approve this model? I can’t imagine anyone would. If this happens multiple more times and Alibaba has said nothing, you just lose your job if you use these models.

You’re just going to lose your job. You’re just going to lose your job. So it’s pretty crazy, and it could have been much, much worse. This got some press. I’m surprised it didn’t get more.

It might be subtle, but again, this is everyone’s nightmare: using these open-weight models based in China and having your data go to China. And it did.

So what happens to the stack when the Chinese open-source market shrinks? What happens to the gateway providers, the routers, your OpenRouter of the world? What happens to the inference providers that thrive in large part because of a variety of models, often Chinese? Your Modal, your Fireworks, your Baseten. What happens to the stack if that Chinese model landscape shrinks?

Well, as Dev said when he was on the show, what goes up that quickly with developers can go down that quickly. It’ll change in a week.

There are going to be some U.S. open-source options, right? The question is, given what we’ve seen about the compute expense it’s taken Anthropic to build their model, and the fact that 75% of that compute spend actually goes to building, training a model, it would be interesting to see how advanced a model a U.S. company like NVIDIA can build and make available on an open-source basis.

Someone has to do it. If someone doesn’t, then all you have are the closed-source providers, and obviously their economic leverage goes way up. You should buy even more Anthropic in the S-1, in the IPO.

I think it’s bullish for Anthropic that this happened.

Of course it is.

I agree it’s bullish, right?

Absolutely.

I think you’ll know the moment Elon’s able to get Cursor off of these models. That’s the tipping point. If he could, he would. He just can’t yet.

Why?

Because Grok isn’t good enough.

Yes.

And he already kicked OpenAI partially out, right? But he can’t yet, because this bet has to work. He said when he bought Cursor, part of the reason was that he wanted that team’s DNA to make Grok more competitive, right? So there are multiple benefits here.

But there’s some deep irony that his $60 billion Cursor acquisition is running primarily on Chinese open-source models. It’s not going to last, right? Salesforce forced Fin to rip theirs out before the acquisition closed. But Salesforce can tolerate a little bit more in Fin than Cursor can, because Cursor is in the most competitive space there is. As soon as he can rip it out, he’s going to rip it out. It can’t run on Kimi forever. This just doesn’t make any sense if you think about it. It’s temporary.

But is this not just highly illogical? If you download it and run it locally, the security risks are limited to none. I understand that CIOs are CIOs and—

How are they limited to none when it’s dialing China and uploading your data? And what if, instead of a 404 or whatever it was, there was a prompt injection sitting there? Very easy to do. People have seen this in several iterations of the model.

I think the question you wrestle with, Harry, is: what can you know? How much can you know with 100% certainty about what the model’s doing? Separate from what the model does, you can try to block it to some extent, sandbox it from executing, to prevent it from doing that. But it’s like having a recalcitrant and difficult employee who’s actively trying to subvert you.

If you had an intern and you knew there was a risk they would actively leak your guest list on your show to your dreaded podcast competitor, it would just get in your head. But it gets to this whole question of what you can know about what the model is doing. If you read the S-1 for Anthropic again, one of the risks they mention is basically that the models will lie to you. The models will tell you they’re not doing shit, and they’re doing it.

So you get into this question of what you can really know. Once you get to that point, it’s really hard. You do worry a lot about a Chinese-sourced open-source model. So there will be that pushback, I think.

Qwen lied. It says, “I’ll call no tools,” and Claude will lie, and Astra will lie too, right? It said, “I’ll call no tools,” and what did it do? It called servers in Beijing and Singapore and uploaded data. It said, “I won’t do it.” So you really have to go deep into these open weights and test them.

Depending on how you test it, what happened with Cohere only happened in 1 out of 100,000 calls. A bunch of folks tested it a few times and found this issue a couple of months ago, but you had to run it at scale. At SaaStr, we run 40,000 API calls a day. It doesn’t take very long to get to 100,000. If you’re GE or another Fortune 500 company, you’re running millions and millions a day. That could mean it’s uploading data every hour, every 10 minutes.

Again, this doesn’t appear to be malicious per se. This appears to be a training error, in a sense. But who cares? If you’re a CISO, who cares whether it’s a training error or a spy—or, even worse, whether it’s a training error that some other actor exploits? It’s the job of all our intelligence services to exploit these things, right? To have pagers blow up.

I don’t believe Alibaba itself did it as an exploit. It doesn’t look like the engineers did it, but the government could take it over and force them to. Why do you want to have this meeting? Let’s just use Haiku 5.5. It’s pretty good. And Anthropic is on the GSA list, or whatever it’s called again. It’s back in the government’s good graces.

Yep.

It’s across all the GSAs with a 40% discount. A couple of months ago they were persona non grata. Now every government agency can use it with a 40% discount.

Yeah.

And it’s already pre-listed. Why would you use Kimi? Plus, I also think—and again, these are great companies—the fact that inference providers have to pay Kimi 30% of their gross revenue to use it at scale adds a lot of complexity. It may not be worth it at some level. The juice may not be worth the squeeze for many people.

I’m just trying to understand, though. Last week we said that Modal at 15.7 was an obvious buy. And then you’re saying Chinese open source is going to decline significantly, which is great for Anthropic and OpenAI. That is inherently bad for Modal. I’m just trying to—

Yeah, it’s bad for Fireworks and Modal. I don’t know how bad it is for OpenRouter.

Yeah, so how do you—

I’m just not sure, right?

I think that—

Because it does limit use cases.

Well, you make a fair point. I didn’t participate in the investment committee, but to be fair to Jason, the assumption is that there is demand for an open-weight model, which is obviously the demand that inference providers primarily meet. You have to believe there are a lot more alternatives: there will be some U.S. alternatives, and there will also be people for whom China is okay. There will be people who want the Poolside/NVIDIA one.

Provided something exists in the U.S., that business is fine. I agree with you that it’s not as clear a run as owning your own closed-source frontier model. But there is a role for inference provision, just because there’s so much raw demand for an alternative to the frontier models.

To answer the question, I think Harry’s got a good point. A while ago, Rory brought up that there probably won’t be another deal like Manus, right? That was a direct investment.

Yeah.

I’m paraphrasing you, right? But after that one got unwound, no one wants to do that deal again: the sort of direct investment in an Asian startup.

Yep.

This isn’t the same as investing in U.S.-based folks running open-weight but not open-source models. It’s not the same as Manus.

Yeah.

But there is risk, and I think 24 months ago, most VCs would not want to take that risk because it’s a clear risk. You don’t have to fast-forward to today, with the Cohere/Qwen issue, to know it’s a risk. The growth is so phenomenal that we’re taking risks today that you wouldn’t take if we didn’t see growth that was jaw-dropping, and that’s the job of venture today.

It’s okay if it blows up. It’s okay if Fireworks or whatever goes to zero in our portfolio at IC, because we have other investments. That’s the job, and we’re taking risks. We wouldn’t take this risk if it was a triple, triple, double, double, but it’s not a triple, triple, double, double.

But that doesn’t mean the risk level didn’t go up a lot in the last week, with DeepSeek being blocked by the U.S. government, this whole Qwen thing and dialing home, and Anthropic being on the GSA schedule. These are negatives. These are negatives. I think what we may look back on—and I still believe in the IC—is, this is just my sense: I think there’s a 56.3% chance we look back on OpenRouter and this was just a moment in time.

It was a moment in time, going back to earlier in the show, when OpenAI stumbled, so Anthropic was the only alternative for state-of-the-art coding, and then the Chinese open-weight models got really good at coding. That’s what Elon said. It’s really good, but… And he even said they’ll catch up to Fable by the start of next year.

So there was a gap these models filled, and now OpenAI and others are going to fill this gap, and the low-cost models are going to fill the gap. They each now have a much cheaper model that’s much better. We may see these open-weight models go from 60-something percent of OpenRouter to 10% in 6 months, and that would not be an extraordinary amount of change based on what we’ve seen. That would not be—

It wouldn’t, but Jason, you’ve talked very well in the past about maiming markets in terms of the Fortnite-ification, the shrinking land that’s available.

Yeah.

And to Rory’s eloquent point, there will still be some companies, and there’ll be some companies that want Reflection and Poolside, but it is a maiming of an open market—or not?

I think you have to look at it the same way. I understand the point you’re making, but you’ve just got to look at it: the inference market is still so big. We added up all the competitors, and they only had $5 billion. And we just agreed that the 2 closed-source frontier models are already $100 billion.

So if 10% of the spend is willing to put up with the brain death of open source to save real money—and you do save real money—then there’s going to be demand for that. Does it get to $30 billion? I don’t know. When the overall market is so big, 10% of it is still big, and that’s the bigger pie here. This is the mother of all markets, so everything else works.

Jason, you teased Black October, but I do think it’s significant. OpenAI models and Wiz found 2 very critical PHP flaws, hidden for more than a decade, that have been exposed thanks to where we are in the cycle and the sophistication of AI and some models. What do we take away from this, Jason? How do you respond to this? What should we know that we don’t know? Help me out here.

Well, we said months ago on this—I said everyone’s going to be hacked. The biggest change is, back in the old days before the LLMs got good, I remember my CEO said the only reason we hadn’t been seriously hacked was that no one cared. Now anyone can be hacked in seconds. People attempt to hack SaaStr every day.

This PHP thing is actually worse than it looks because there are so many websites in PHP that no one’s going to update. It’s great that a lot of stuff runs on WordPress, or that Replit and Lovable and others immediately ran gateways or barriers to protect you. They’re not all running on paid plans, but they immediately ran. A lot of stuff went up, and people released patches and updates, but a lot of government and other sites, and nonprofits, just took their websites down. Who’s going to fix it?

PHP’s still, like, 40% of the internet, right? And I bet it’s more of the old stuff. I bet it’s like 80% old stuff. This is like Mr. Robot live. All these old websites are running on PHP. No one’s patched them since this happened. No one cares. The IT team was laid off 3 years ago. They’re not vibe-coding anything, and we’re going to find our way into all this. It’s just going to get worse.

Everyone, basically, including Sam, just said it’s going to get worse and worse and worse and worse, right? I think we’re still grappling with the fact that everything that is a threat vector will be exposed almost in real time.

Agreed, because if you’re not up to date with the latest stuff, you’re way behind. You’re right, Jason. There’s going to be a year or 2 of this, and it’s inevitable.

I think the only thing that’s going to save us is that there are only so many people who can pay large ransoms, and only so many systems that control dams and other natural resources. People are going to give up on putting too much energy into it if they can’t make money or debilitate something, right?

Yeah, there’s no money in it.

So it may actually decline for a while when they realize hacking the 20VC Substack only gets you so much money, right? But anything that can be hacked is going to be hacked. Ransomware is just getting more and more professionalized. It’s just the next level, right?

There were a number of large private rounds. There was Lagora raising $550 million at $8.8 billion. We’ve spoken a lot about Lagora versus Harvey in the past, and it passed $200 million in ARR. Just in the last 24 hours, there were rumors—and we obviously never comment on rumors, do we, Rory?—of Instinct raising at a $30 billion price.

Oh, I saw that. Yeah, yeah, yeah.

We can go into it. We can have some fun with that one, because a lot of people are suggesting that is peak, peak, peak. The revenue multiple for Instinct is the same as it was last round.

Which has sort of meant the valuation didn’t go up.

I mean, yeah.

Yeah, no, exactly. It’s just—

We have others like the ex-OpenAI research chief, Bob McGrew, with Arda hitting $1.5 billion, raising $232 million.

Yeah.

I’ll tell you—

New chiefs.

Can I ask one question on the Instinct thing? It’s a question to you guys, because you guys would know better than me.

Go ahead.

We can make fun of the $30 billion at token revenue, right, and say it’s a sign of the times, which perhaps it is. The question I just wonder is: when will we exhaust the equity here? We just talked about equity versus debt for the public players, right? In a sense, the $30 billion doesn’t surprise me. They need the capital to pay for the inference, right? They need the money.

What are they raising? It’s rumored. Let’s say they raise $3 billion, right? They need it. Forget about the rest. This isn’t just a vanity raise. They genuinely need the capital to compete.

Yes.

But even if you’re getting every dollar that every sovereign wealth fund has, including GIC, how much money is there for the Instincts of the world? It feels infinite now, right?

A lot more.

Yeah, but how much is it?

Yeah.

I just don’t know how much it is. When that merry-go-round ends, we just have a structural issue in this late-growth phase. We just have a structural issue, because right now it’s game on, man.

I did an Instagram on Instinct, and I had over 1,000 DMs, and over half of them were asking to invest.

I mean, yeah. Look, why are we surprised? We’ve just seen the most successful private investment in history take place. Four years ago, Spark brilliantly put money into a pre-revenue deal that’s now turned into hundreds and hundreds of times $75 million.

Everybody reads about that and says, “Oh, that’s easy.” When you have that kind of tangible evidence, everyone’s just going to look for the next ticket. Everyone’s going to conflate—everyone’s going to think, because it’s always true to say whenever you look at one of these deals, you say, “Well, it could be huge.” We all have that sense, and that’s going to be true of every one of these things.

I actually—this is why I would unwaveringly do Instinct at 30. If I was coming into the IC, I would unwaveringly do it.

Yeah.

For years, we’ve been promising a super app for the West. A WeChat of the West, which handles your finance, your identity, your security.

And finally, we have that with Instinct. I’m not saying it’s guaranteed, but there’s a small chance that it is the super app.

Yeah.

Here’s why I’m, unfortunately, Harry, I’ve got to override you on the IC. And I rarely—

Absolutely not.

I rarely do this, and I’m going to put my little black pebble in here. It’s not that I don’t believe your argument. I do believe it, Harry. At this point, I need three to four months of data. There’s not enough data at Instinct to prove to me, for our fund, given our reach—which is substantial, but not the same as our friends at Andreessen Horowitz—that this momentum is real.

I’m not saying this isn’t an exciting play. I just want three to four months to believe the momentum is real. I’m not looking for moats, but I am looking for evidence that the 900% NRR at Anthropic is slightly durable, and I’m just not seeing it at $30 billion. I’m sorry, I’ve got to veto this one.

I—

It’s not that I don’t believe in it intellectually. I’m just tired of doing growth rounds with one month of traction, Harry.

For me, they’ll have raised $4 billion. Your late pref on that means that, given the quality of the team and the strategic nature of the market, you’ll definitely get that back. Microsoft would buy this in a heartbeat, as would several other hyperscalers, including the Metas of the world. One, your downside is protected, and your upside is really, as we said, the super app of the West: a move into a subscription business, an ads business, and a transaction business across sales and e-commerce.

I think it’s one of the most no-brainer deals for us to put money into, if we’re an Andreessen of the world, Jason. Again, we’ve just raised a huge fund, and we need to get it out the door in 18 months. Amen, I want to do this deal.

I hear you.

Yeah.

And listen, I don’t want this to break our relationship. But I feel like for our growth rounds, we need 100 days of traction. I feel like we need this rule. We need 100 days of something. I’m willing to do the multiples. I’m willing to roll the dice. But I want 100 days of traction for a growth round.

We’ve had too many stumbles when we do a round with two weeks of traction. I just believe there’s too much risk.

Jason, what would you want to see? I think the cohort data would be incredible. I haven’t seen it, so I’m not commenting, but I’ve heard the usage will be incredible. It’s unlike anything any investor doing it has seen before.

What would you want to see? There’s no revenue data for you to see. Usage data? What do you want to see in usage data? DAUs, MAUs, replacement of ChatGPT? Help me.

I just think with no revenue and an unclear business model based on travel, we should unfortunately pass at $30 billion. I just want to see 100 days of something besides negative gross-margin growth of a fungible product. That’s just me on the IC. If we lose $2 billion, it will maim the fund, and I’m not convinced of the liquidation preference here. I’m not saying you’re wrong. This one might be a bridge too far for me.

Jason, you don’t get it. If we lose the $2 billion, it was Rory’s deal.

Yeah, but—

I’m comfortable with Fireworks at $40 billion. This one at $30 billion just makes me nervous. Fireworks has met my 100-day rule. It’s got 100 days of compounding, genuine growth that I don’t believe is durable, to our earlier point, but I believe is truly measurable and bankable.

Wow. So you’d rather have durable growth where you don’t foresee it getting better than not having that data, but where you do see it getting better? That’s interesting.

I still think it’s a better investment than Hopin, but I’m not comfortable with this consumer—

No, I—

—negative-gross-margin bet. I’m just not comfortable.

You ask an interesting question, Harry. Stepping back, you’re right. What you’re basically saying is: How do you compare a known market with solid growth, growing nicely—though “growing nicely” understates Fireworks—and a ripping market where it’s inherently not magic pixie dust? There are going to be 5 competitors, so your upside is real but capped at some functional level because there are 5 competitors.

And you’re comparing that to this thing where, if it becomes everybody’s thing, it’s a $200 billion company and everyone uses Instinct. It’s the replacement for iMessage plus, um, Fa—uh, yeah. It has that potential. You’re right that there’s more uncapped upside in Instinct. How do you weigh that up? You’ve got to do the 3-way payoff matrix: if it’s amazing, if it’s good, and if it doesn’t work, and look at the 2. What kind of action are you looking for on the table? I suppose that’s the question, to some extent.

Every big fund should be fighting their ass off to put as much money as possible into Instinct today.

Yeah. Okay.

So I’m voting for it. Okay, next one on the IC, Jason. You’ve got Kevin Weil leaving OpenAI Science to launch Warp Science with $215 million. It’s a very early pre-seed, so you may not be comfortable with that because there’s no data. You’ve got the OpenAI research chief, Bob McGrew, who raised a $232 million Series A. You have Lagore’s $550 million round.

Odd things to put together. I was trying to make sense of this. There’s a plethora of deals this week that I would describe as an absolutely proven executive, a big opportunity, a hairy technical challenge, but a big opportunity if it works. That’s, to some extent, the flapping-airplanes thing; to some extent, it’s Arda, which is Bob McGrew’s deal. I don’t know as much about Warp Science, but it feels like the same thing: a big idea, a talented executive, and clearly a wild appetite for that at fairly high prices with relatively little de-risking. That’s one whole group of 4 or 5 of those kinds of deals here.

The other type is the late-stage, known-good company doing really well—maybe Lagoa, maybe Bedrock in a different way—where it’s not as binary based on a person. It’s just, how do you price these late-stage companies? That’s easier. At some level, it would be easier to get your head around Legora or Bedrock than it would be to do flapping airplanes, Arda, or Warp Science, where all you know is that the executive in question is talented and the opportunity, if it works, is amazing. Welcome to the game.

I’m going to take a radical approach for the IC, Harry. I know it might make you uncomfortable, and you should veto it too, like I want to veto Instinct. But I think there are moments—and I know you don’t totally believe me, but I do believe—that even in many of the best startups, sometimes there are dislocation moments where the market briefly struggles. I think we should put 20% of the fund into OpenAI at the $1.25 trillion round.

Hmm.

I think it’s mispriced. It’s a moment when there’s a little market panic. Anthropic is on fire. SpaceX has done well. The models are wildly competitive today. OpenAI has closed all the competitive gaps. It’s a well-run machine.

It reminds me a little bit of that Anthropic round at $183 billion or so, when the valuation made no sense vis-à-vis OpenAI. It was a dislocation. A lot of people made money on that round. This is a dislocation moment. People are misunderstanding. They’re thinking OpenAI is struggling when the numbers clearly say otherwise.

Mm.

We can fart around with these little $20 billion, $30 billion, $50 billion rounds. I think if we put 20% of the fund into OpenAI at $1.25 trillion, at least we can return the fund with that, and then take our upside with some of our other exciting bets.

Why do you say that? What is the upside on OpenAI at $1.25 trillion?

Look, these are both going to be $10 trillion to $20 trillion companies. Just look at the growth and the scale. Look at Anthropic. It’s only 0.1% of Fortune 500 budgets today. Clearly, it’s going to be in the 2% to 3% range, even by the most conservative estimates that we, Coatue, and ICONIQ have done. Going from 0.1% to 2% or 3%—I need Rory’s help here—that’s significant growth ahead of the investment.

If it’s 20% of the fund, obviously it needs to 5x, dilution aside, to return our fund. There’s a clear path to a $10 trillion outcome. At $20 trillion or $30 trillion, the rest of the fund doesn’t even matter. It’s at a discount to Anthropic that I don’t think it deserves. Worst case, just like with some of your investments, you get your liquidation preference back. Worst case here, we double our money.

Worst case. It’s not that bad to double your money, worst case. That’s what you get from these dislocation moments. You get to double your money when ordinarily rounds like Inflection or Instinct are priced to perfection. OpenAI is not priced to perfection, and that’s rare. That’s why I think we do 20% of the fund, because that’s the limit under our LPAC, and we can do that without LPAC consent. We just do the maximum without consent. We just let them know on Monday.

The other thing is the quality of Codex today is so significantly better than it was in a prior generation. I know every coding team that we work with has switched from Claude Code to Codex in the last 3 to 6 months.

That’s why we should put 20% of the fund in.

People aren’t getting it. They’ve achieved parity. This is when you go all in, but the markets don’t see it. The markets are too superficial to see it. Twenty percent of the fund. Nothing smaller, or it won’t be enough. It’s got to be 20% of the fund.

Interesting. Not crazy. I must say, Jason, you’re right that Anthropic will price at $2 trillion, and maybe a pop. If you knew this was the last private round and the trajectory of OpenAI was converging, you’d want to know that they’d gotten religion and wanted to do something like what Anthropic has done. Yeah, not crazy. If they continue to say, “We’re gonna spend, we’re gonna burn forever, forever,” then you wouldn’t want to do it if there’s another private round in it. But it’s not a crazy idea because, risk-adjusted, it’s probably less risky than some.

I love that as an out-there suggestion. My only counter to that would be, could we not do Instinct and OpenAI and actually hedge? It’s not a bad risk profile, to your point.

If that’s your definition of hedging, please don’t ever run a hedge fund.

Why? My friend Leo told me that was how you did risk management.

In all seriousness—

All right.

This may be my weakness. My scar tissue is that I want 100 days. Whether it’s seed or growth, I want 100 days.

But Jason, you’re just not getting that today. I agree with you, dude, but you’re just not getting that today.

I might not be, and I don’t want to lose you from the fund, even though it’s Stebbings and company. So if we have to do it to not lose you, I don’t want to go all gotcha.

So just be company.

I don’t want to go all gotcha here. We’ll do it, but if I can use one black pebble this quarter, it’s going to be on this one, and I’ll—

Okay.

And I’ll take the hit when I’m wrong.

Any other rounds that we think are worth bringing up? As we said, Flapping Airplanes I thought was an interesting one. They’re building models that learn from farms—

How about Superdogs at $1.5 billion? That one I need a little help with—the TAM. I need Rory to give me my TAM analysis for bionic dogs at $1.5 billion. Or is this just another Elon bet, like he’s gonna make it work?

Yeah, so just for context, War Hounds is a startup co-founded by Palmer Luckey, the founder of Anduril, that is essentially—I mean, genetically modifying dogs to be next-generation conflict war machines would be an apt description.

Yeah.

They put a valuation of $1.5 billion on it because it’s Palmer Luckey. He deserves it, to be honest. Discuss.

I realized that I had zero value to add on that. I genuinely saw it and said, “Wow. Genetically interesting. How do you price that thing?” I saw some analysis of the number of dogs the military used, blah, blah, blah, but you’re all just moving paper around.

I think the investment memo probably wasn’t one, but to the extent there was one, it was, “Wow, that’s a cool idea. Let’s put in 50.” I think that was the summary investment memo: let’s genetically engineer these dogs to be savage so they can be better war dogs. What can you say?

And is Palmer lucky or, like—

Yeah, yeah.

Yeah, okay.

Yeah.

Cool.

Exactly. Go team. I won’t be doing it, but who knows? Once you’ve founded—I nearly said Anthropic—once you’ve founded Anduril, once you’re the sword of the West, you can do what you want.

There is a connection between this and his other startup he just started, this laptop one, Blue Chip, right?

Oh, yeah.

He’s solving sovereignty issues where key national security things are dependent on vendors outside the US. The dogs come from Europe, apparently. I didn’t know this.

Yeah.

I certainly know that almost all the rugged laptops, which you can probably sell for $10,000, right? Or $20,000—

Yeah, yeah, yeah.

Those components are almost entirely from outside the US. How big these models are, I don’t know, but he’s at the core of this industry, and he wants to plug the white spaces he thinks are important. Maybe the TAMs are small, but the models make sense. He’s right there, and he has the channels to sell these products. He could get 100% of all the mil-spec laptops as soon as this thing launched. It wouldn’t surprise me, right?

So he could get 100% of the bionic dog market. These are real markets where he has visibility into them that other founders can’t penetrate. Whether they’re worth it, I don’t know. But I guess if you’re Sean Maguire, your returns are so strong and consistent, you just back them.

Yeah.

People were mocked in that decision when Larry Ellison and Marc Andreessen gave Elon blank checks for Twitter with nothing. They just said, “What—”

Any man, he—

“Tell me the price. How much do you want me to send?” He was mocked when he wanted to get out of that deal, but in the end, they’re making money.

Absolutely. No. Yes.

So maybe I’ll have to merge the dog startup into Anduril or something. Maybe it’s okay.

Haven’t you seen the brilliant graph where you start here, go up here thinking you know a lot about the market, and then you come down to where you were? It’s like “founder, founder,” and it’s where you are in your career: zero years, 15 years. In the middle period, where you think you’re an intelligent investor—no, no, no—there really just is the top.

Yes, agreed.

Yeah.

Yeah, no, I’m very familiar with that, yes.

Yeah.

You just—

It’s so, so true. Any others, Jason, that you’d like to bring to IC this week? We’re a very busy committee.

Well, 20% of the fund is enough for one week. I mean, we can’t do that 50… What is 20 times 52? What percent of the fund is that? Help me.

It’s so interesting, though. I speak to so many LPs. We just had our AGM. The single most common piece of feedback that I hear from all LPs is, every mega-platform went out to raise. They raised bigger. They spent faster. They’re coming back next year.

You know, we’re Pavlovian creatures. If you put money into Anthropic 3 years ago, it feels amazing, and most of the platforms did at some level, so they feel amazing. So we’re gonna do other things over and over again, and it’s not gonna stop until something goes wrong. We are risk-on until we’re not.

The public markets, oddly enough—let’s be really technical. On a 10-year P/E, they’re at an all-time high. Objectively, they’re at an all-time high. Interestingly, on a forward P/E basis, they’re actually quite cheap because, oddly enough, earnings have been so amazing, in part driven by one-off gains from privately held investments in public companies.

On a current P/E basis, stocks like Nvidia look cheap on a forward P/E basis. Now, you’re gonna have this long discussion on cyclicality, blah, blah, blah, but the aggregate market isn’t hurting. What is true is that 450 of those stocks are hurting. It’s a really weird market right now. A better statement, Harry, is that there are 40 or 50 stocks in the S&P that are working, primarily associated with tech and AI. Everything else is quietly subpar.

Oddly, despite that, there haven’t been a ton of IPOs because everyone’s been waiting for this trillion-dollar thing, and the whole world is in suspended animation. So that’s where we’re at. To draw this to a conclusion, it’s a perfect time for these platforms to put a whole ton of money out in private companies because there’s a big demand for money.

Any other stories, boys, that I’ve missed this week?

There’s a lot. The interesting thing is, I did a pre—It was odd because I thought you’d come in with Anthropic and nothing else, but there was a whole bunch of financings we talked about.

You’ve got the continuing SpaceX–Tesla drama, which is going to run and run and entertain everybody for hours to come. And then you’re right, there were a ton of really interesting companies.

Yeah.

Oh, yeah.

3. Zuckerberg’s California Standoff

I had one that was underdiscussed. I didn’t fully understand it, and I should’ve done more research on it. There was much more drama with Zuck and the board about leaving California than was discussed. It was clearly a high-stakes drama. Zuck hired lawyers, and the lawyers negotiated with the board.

The board was not his ally. He had to threaten to sell all of his homes and leave in 1 week if they didn’t give him the equity he wanted to stay in California. There are a lot of mechanical things happening there with the billionaire tax and others, but the fact that it had to go through lawyers for months, and there were threats to sell, I get it, but it’s certainly the opposite of the era where founders are controlling everything, right?

Yes.

There’s a lot going on.

Yeah, the one thing you can control is your own comp. I get it, because even Elon has to do that. If you’re board members and you do this without CYA 100 different ways, you’re vulnerable to being very second-guessed. My guess is, even if they’re reasonably amicable, everybody should lawyer up so they have protection and a process, because process is your friend in these things.

But on top of that, it was super interesting. I don’t know the detailed dynamics on a lot of the voting, but the billionaire tax is up, and we’ll know in a week or two. People who think there’s no impact from that have never met a billionaire when you’ve tried to take 5% of their money away. It turns out they take it really badly, especially when you want to take 5% of the money away and there are voting criteria when you have more voting control than anything else, so it’s more than 5%. And especially when the legislation says, “We can do it again soon.”

I thought they’d already given it all away. Isn’t this the point of the Priscilla and Mark Foundation, that there’s a tax-efficient structure?

Yeah, it turns out people like to give it away, but they don’t like it to be taken away.

Mm-hmm.

It’s all about agency and control. You’re right, Jason. That was probably much noise under the circumstances. Whatever happens with this billionaire tax, it looks like it’s been remarkably successful in driving out enough billionaires that you’ll have an ordinary income tax hit. So, go team.

And do you know what I’m excited by with this show, boys? It’s obviously recorded on a Wednesday. On a Wednesday night.

All right.

By the way, it’s a Wednesday, Rory. So there’s not going to be 24 hours before release.

Yeah, whatever.

And so there’s no chance there’ll be some massive news announcement within a 36-hour period.

No. Actually, just to remind you, I think Microsoft and Google—or Alphabet—are reporting tonight. In theory, you could have anything, but who the hell knows? I’ll be shocked to discover that they’re going to spend more on CapEx. Are we good, Harry?

We are rocking and rolling, my friend.

Good.