It's Tuesday, November 10th. Welcome to Deirdre Bosa Live, presented by Factory. On the stream today, we've got exclusive new financials from OpenAI. Plus, Anthropic on the road ahead of its IPO next week. What are people willing to pay? Maybe not as much as first expected. Markets, they closed just moments ago. Stocks in the red after a hot CPI report this morning sparked new inflation concerns. Nvidia and Microsoft, today's biggest laggards. We'll get into that. And as for sentiment on the AI trade itself, well, we are expecting CoreWeave results any moment, which will be good insight into the state of NeoClouds itself and all to do with AI financing. We've got a big hour ahead for you. Vinod Khosla is here. OpenAI's Sarah Friar is here. We will get into all of that in a moment, but first let me lay out where we are.
Anthropic has really been the big momentum story in AI all year, but I just got a hold of some new numbers from OpenAI. They suggest the story is starting to change. For the first time this year, OpenAI is growing at roughly the same pace as Anthropic quarter over quarter. And that is a big deal, especially because Anthropic is about to go public. Let me explain a little more. For most of this year, it felt like Anthropic was just running away with this race. You can see this chart here from its S-1 prospectus, but really the narrative has been Claude. It became the default for a lot of developers and plenty of non-developers too, myself included. Claude Code took off. People were vibe coding companies. They were building on Anthropic's models in a very big way. And really, the numbers backed all of that up. Anthropic passed OpenAI in reported revenue last spring. It was more than doubling its revenue every quarter. And meanwhile, OpenAI was still growing, just nowhere near as fast. You might remember in the second quarter, Anthropic grew 144%. OpenAI, just 18%. And that is a pretty stark difference.
I'll be honest, when I started digging into these third quarter numbers, I expected more of the same, or at least not this big of a jump. That's not what happened. OpenAI did $10 billion in revenue in the third quarter, up from about $6.7 billion. So that is nearly 50% growth, almost triple its growth rate from the quarter before. Now take a look at what Anthropic did. $17 billion in the third quarter. That's 47% growth quarter over quarter. So just think about what happened. OpenAI went from 18% growth to nearly 50. Anthropic went from 144% down to 47%. So for the first time this year, their quarterly growth rates are basically neck and neck. Now to be clear, Anthropic still reports a lot more revenue on that absolute basis. And there are some differences in how these companies account for cloud sales. So the exact size of that lead, it's hard to pin down precisely, but the change in growth is what caught my attention.
Anthropic slowing down, probably inevitable at some point. You can't just double your revenue every three months forever, even in AI. But what I expected less was OpenAI speeding up this much, this quickly. The timing, of course, is fascinating because Anthropic is just about to ask public investors to buy into this growth story, and it is still an incredible one, to be clear. But now, for the first time in a while, OpenAI has a much stronger quarter to put up against it. Which brings me to Vinod Khosla. He was one of OpenAI's earliest backers, and unlike a lot of investors in the space, he did not bet on both horses. He picked OpenAI and he stuck with it even as Anthropic seemed to be pulling ahead. Vinod, welcome to the show. Thanks for being here.
But at any point this year, did you lose your nerve at all and wish that you had invested in Anthropic as well? I know, like in past eras, it's unusual to bet on two competitors, but with companies this big and this important, a lot of investors have done that. Why sit Anthropic out?
Well, we invested in OpenAI very, very early. In fact, it was at the $1 billion valuation. So it's a very attractive hold for us. It's worked out extremely well. But what's important in the AI race is not what's happening quarter over quarter. What's the trajectory over the next two, three, five years, and what will be the situation in 2030 and beyond that. So when I look at that, it's important for me to get deep insights into what is happening in the space. And if you're an investor in both, neither company will share that with you. If you're an investor in one, you get much deeper insight, and that's what we chose to do. I think it affects everything we do in AI, and I'm very happy with that.
Okay. So if you don't want to look at them quarter over quarter, which is what I just did, but I also looked at sort of this big picture narrative. And even when you look at it on an absolute basis, I mean, OpenAI started the whole generative AI cycle in the lead, but now you look at Anthropic's $17 billion in revenue versus $10 in a quarter, it's taken the lead. So that is not a quarterly thing, that happened over a few years.
Yes. Well, that happened this year. What I would say to you is it's very hard to use that as a basis to decide who's in the lead in a year or two or three. And so I'm very happy with the kind of approach that OpenAI is taking. They have a large investment in R&D, probably the most research-oriented model company there is. They also have a large investment in compute, which gives them a fundamental cost advantage. So to me, revenue is not as important as measuring tokens. So who's supplying more tokens? OpenAI has chosen to price its tokens lower, and because of that, is growing share of tokens, which is the intelligence and the service both of these companies are providing.
You know, people always ask me what constitutes a bubble or not. It's not prices, it's actual usage. If you look at the dot-com bubble, yes, prices went up and down and then back up again during the 1998 bubble, but when you look at internet traffic, you can't tell where the bubble was. So I would urge people to look at the reality of actual usage, which is the number of tokens, not Wall Street's perception of bubble prices. That is much more indicative of what happens in the long term than over a short period of time.
Okay. So let's look at token usage then, because we've seen something really interesting this year. Everyone always says there's so much demand. That's right. But that is being spread out, right? And certainly if you look over the last few years, this year in particular, we've seen the explosion of open-source models. And I know you're talking about both the frontier labs sort of bringing down the cost of their tokens to compete, but you can't deny when you look at an OpenRouter or another third-party website that looks at developer usage, these open-weight models are certainly encroaching on usage, right, Vinod? That's what you're talking about is usage. And that picture is not totally clean.
That is true. Open source is getting token share. Here's what I would say to you. There's a misperception that open source is cheaper to run. It's not. You know, because you're paying for the 5x markup Nvidia does on its silicon.
As more and more of the Jalapeno silicon comes online and becomes most of the inferencing at OpenAI over the next few years, you're going to see a much lower cost of running the model. You've seen this with Google and the TPUs, and now they've started to compete with Nvidia on TPUs. But what is your cost per million tokens that is lower if you have your own silicon and don't pay the Nvidia premium of literally 5x the cost? Also, if you have your own silicon, you can tune your silicon to the model you're running, the technical approach you have, which gives another multiple factors the advantage. So I suspect the cost of running a certain size model at OpenAI is one-tenth of running the same size, same class of open-weight model run on Nvidia GPUs.
And that cost advantage is what wins, and what you want to use to gain market share in tokens.
So I don't disagree with you in that it is maybe more complicated this year to use open-weight models, right? It's not given to you sort of from the top of the stack all the way down like the frontier labs give you in terms of that compute and token usage. But that's getting easier, right? You have companies like Fireworks and Baseten and others that are actually helping to serve these very efficiently. So do you think that's still the case going forward? And you mentioned TPUs, right, and Jalapeno. Those started as chips for, you know, Google to serve its own AI, for OpenAI to serve its own AI, but increasingly that's moving to their own customers. Doesn't that bring the cost of compute down for everyone?
It does not. Why? Because Nvidia is not pricing the same way Jalapeno costs OpenAI. So whether you look at the neoclouds, the Fireworks, Baseten, they are buying Nvidia GPUs at premium prices. And they have to sell it at high prices going forward. Their cost per hour of a GPU has actually gone up over the last couple of years when the cost of serving inference at OpenAI has gone down. In the long term, the best cost advantage wins, and that's why OpenAI made the right bet with their own silicon. They will use other people's silicon too. There is going to be a silicon shortage for a couple of years, which means Nvidia will keep pricing at a premium. Google TPUs will be priced at a premium. And having in-house silicon is a huge advantage. Having in-house silicon, in-house models, and in-house applications, having the full stack is the real advantage.
You want to get, you know, everyone sort of looking, everyone who is a hyperscaler or has a lot of compute is developing their own silicon. Microsoft with its Maia chips, even Anthropic, Vinod, has been on this and, you know, recently said that they've made some breakthroughs in terms of chip design. And so I guess I go back to my question to you. You're so unusual in Silicon Valley right now in that you're not invested in both of these companies. And I just wonder, I'm still curious as to why.
Well, very simply because insights about where AI is going, the deepest insights are what's most important. I regularly have dinners with Sam, and I get pretty deep insight into where the world of AI is going, not only next quarter but in a year, in two or three, what's important, and I think that really helps us. So that's the tradeoff I prefer.
If you could have dinner with Dario on a regular basis, would you be in Anthropic?
But he wouldn't share as much information knowing I may talk to Sam also. So being in one camp is an advantage when it comes to understanding what's going on in the field of AI. If there's investors at Anthropic that are exclusive to Anthropic, then those investors will get more insight than somebody who's spreading their bets.
I see. I see. So Sam is willing to share more with you because you're only in OpenAI. I understand.
But have you sold any of your OpenAI stakes since you invested when it was only worth a billion?
We never comment. We never comment on what we do or when we do distributions. We generally don't sell. We do distributions. But that's proprietary to us. I don't think we are in a rush for liquidity at all. What I would say to you is, and the thought I was going to complete, was the bulk of our investing is outside of the models. If I look at new dollars being deployed, they're outside of the models, and because of that, insight into what's going on with the models is very, very important, and where they might want to emphasize for '27 or '28 or '29 and beyond is very, very important.
So seeing the OpenAI roadmap helps you think about new investments. So where are those going? What are you looking at these days?
Well, everything. Look, there's so many areas to invest in. The application areas are very, very broad. Whether you're talking about, if you take every traditional industry, take a massive business like Ansys, which mechanical engineers or chip designers use for building packaging and things. You can absolutely replace that if you understand where the models are going. I just looked at a company working with one of our companies in mechanical design, mechanical engineering. Whether you're building boats or building spaceships, you're going to use these tools. Those are all application areas. All of medicine is a massive application area. We've just done a startup that can use AI to design cancer treatments for one patient at a time. We hope not to go through a drug approval process to do treatments because the patient population is one person. I can look in other areas. Cognition's been hugely successful in software coding. Wonderful area, one of the big growth areas. In fact, where Anthropic did pretty well earlier this year, but Cognition is on a complete tear.
Cognition and Factory, right? So you know what you didn't do with OpenAI and Anthropic, you have done in the AI coding space, which is invest in two rivals, right? Why is that decision different?
Look, I mostly only work with Cognition, to be very, very clear. So we have a complex process for how we make investment decisions. My focus is in Cognition. They're a much bigger, larger, more powerful play, and I think in a very good position to both leverage whatever OpenAI, Codex, Anthropic do as well as do things on their own to add additional value. I think they have the scale and velocity to be a dominant player in a business that's clearly going to be worth hundreds of billions of dollars, if not trillions of dollars.
Okay. Vinod, while I have you, I want to get your thoughts on a major story that has everyone's attention this week. You mentioned Nvidia earlier, and it's sort of at this crisis point with what's come out with Jensen Huang and his affair and the allegations that the mistress may be a Chinese spy. You've been on many boards, Vinod. How do you think this plays out? Is the board capable of doing a proper investigation? What does Nvidia and the AI trade look like if there isn't a Jensen at the head of Nvidia?
I suspect the board is pretty well-versed on how to handle a situation. It always starts with an independent investigation. It's not a surprise to me that the Chinese are targeting US executives with this kind of, I would say, infiltration. So we clearly have a problem where the Chinese are trying to steal all our technologies, and they will not stop at anything to go after that. And boards and executives need to be super careful. You know, Jensen has apologized for it, and the board, I have no insight into how the board might handle it, but starting with an independent investigation is exactly the right thing to do.
Yeah, that's similar to what our guest Isaac Stone Fish said yesterday. He thinks that there's a lot more companies and executives that could be vulnerable to such kind of espionage attacks or honeypots, if you want to call them. Vinod, last question. I'm going to do a hard turn because I think that my 11-year-old would kill me if I did not ask you. He claims to be the biggest Seahawks fan in the world. Go Hawks.
I guess Go Hawks. He'll appreciate that.
I guess I'm curious, too. What's been the most surprising thing about being an NFL owner so far?
Well, the most surprising thing is actually how involved the whole family, even my kids, I have four kids who weren't involved in football, have become involved. We were at the game yesterday, a great game, but one daughter flew up from LA, another daughter flew down from New York into Seattle, and we had a blast. So it's been great for the family to get involved together. I'm surprised. Not one people would expect, but you know, it's a new area. I'm learning a lot and treading cautiously. I think it's very rare you get to inherit a Super Bowl-winning team that's functioning well, that not much needs to be fixed day one or season one. And so it's great to watch, enjoy the ups and downs. There's plenty of those. I call it cardiac city.
I certainly hear the ups and downs. It's been a lot of fun, a lot of learning, and hopefully because we have such a great team in place. Coach Macdonald's maybe one of the best coaches, John Schneider's one of the best GMs. I'm very excited for supporting them.
Vinod, thank you so much for your insights. Thank you so much for being here. Appreciate it. We'll talk to you again soon, I hope.
Thank you.
Let's head to the factory floor now, presented by Factory. And now we'll look at Anthropic's new releases today: Whiteboard, Sheets, and Data. Microsoft shares, they took a hit today on AI's answer to the Office suite. This on the heels of some new models being released yesterday. Plus Muse and Ember from Meta. Ejaaz Ahamadeen, host of In the Loop, joins me. Ejaaz, you are a live streamer, so I know you understand me right now. I have lost my earpiece. Give me one second. I've got the amazing Ben here connecting me.
I've got you. Okay. Right. This is, it's a live stream. It's the internet. We can do these things. I've got you now. Okay. I was actually really excited when you said that you wanted to talk about Ember, because you know, using these AI assistants, I'm sure you do too, it just takes an enormous amount of compute, right? I was talking to Sarah Friar, the audience is going to hear from her in a moment, but like she was talking about the decision to serve this to enterprise first because it just takes so much compute. So what sort of got your attention from Ember? For me, it was the cost.
So yeah, so in order to answer this question, I think it's important to kind of look at Meta's overall strategy, who was the creator of this model, Ember. So their entire strategy for years now has been, let's try and make the models as cheap as we can, because if we make them as cheap as we can, we can get it into as many hands of people to test out and use all these different AI models. So Ember is really the instantiation of that. They tried to do it with open source. They weren't quite able to figure that out. And now they've been trying to create models that are both at the highest level of intelligence that can compete with the likes of Fable 5.2 or GPT-6 Astro, as well as be cheap enough to do, say, 90% of the tasks that some of the most intelligent models can do. And that's really what Ember is. They compete directly with models such as Llama from ChatGPT or Haiku from Claude. So not the smartest models, but still quite cheap, and they only come in at a third of the cost, sometimes even less, and they work super quickly. Now, it's important to understand that you're only going to be using models like Ember when it comes to tasks that you kind of know how to do it. You don't need the answer, but you want something else to do it for you, and that can just work 24/7, and you don't have to worry about it actually completing it correctly. That's what Ember is.
Right. You know, are you using Muse yourself, Ejaaz?
Yes, I am. Every single day. I mean, so I have used Instinct and Muse and Dots, and I was telling Sarah Friar, not because she was here, genuinely I enjoy Dots more because it has more context, right? This is like the key of all of this, the more that it knows about you. And you would think if you're active on Facebook or Instagram or WhatsApp, it would know more, but that was not my experience. And what I found with Dots, it just was, you know, a lot more intuitive and better. And I do wonder if something happened at the model layer. But have you noticed a change? Ember just came out. Do you use Dots as well? Have you noticed a difference between the different assistants?
Yeah, I concur with you. Dots is just a much better experience. Really, wherever you spend the most time doing a lot of your AI queries or AI work is typically where you end up getting a better experience when it comes to these personal agents. That being said, and I'll be, you know, honesty hour over here, safe space, I haven't really enjoyed using most of Meta's earlier AI products, and I think that's perfectly a fair take when you compare it to like the frontier models back then. But Muse was the first real release that I've had from them where I open it every single day. I have multiple different tasks where it kind of hits me up in the morning about all the latest that's happened, and then it helps me kind of work through understanding some of the things that it kind of already anticipates that I don't already understand, as well as deal with like all the other tasks that, you know, I couldn't even think of doing all that. It reads my emails and does all of that. Dots does that on a completely other level because it has all that context around you. Now, I wouldn't rule Meta out necessarily because they still are one of the largest owners of data moats when it comes to your personal social life. I think Dots operates in the work side of things. So it tries to really kind of tap into that enterprise market because that's where it really is well placed to beat everyone.
Right. And sort of as Muse has this head start with consumers, they're collecting more context. Actually, I asked Sarah Friar about that. I said, how can you just, like, why are you worried about giving it up, about people having a habit and not wanting to switch from Muse to Dots when you do make it available to the consumer? But her point was that, as we talked about, like, chatbots are collecting a lot more data. People are asking it about health care, not just for work anymore, and it did start as a consumer company. So I do wonder, do you think like people will make the switch, or they're just so embedded? And we're talking a small group of people. I know this, you and I are both early adopters. But do you think people will actually make the switch, or do you think that Meta will take it away because they've been ready earlier and now they're going into the Apple Watch and they have that cute little dongle?
If I'm being very honest with you, if we're talking about the small group of people that are going to switch, I think they'll be interested, and I think some of them, let's call it 10 to 20% of them, will download the app and play around with it because they want to see what AI can do from a leisurely perspective in terms of freeing up their time and doing things that they didn't want to do. But when it comes to productivity, I don't think they'll necessarily switch. Now, on your point around most people not using it, I think it's worth calling out a report that I don't know came out a few months ago where we're looking at like less than 5% of people are using these AI models beyond just a random Google search query. So this is still a huge open field.
