Grok 3, AI Memory & Voice, China, DOGE, Public Market Pull Back

00:00:00.000 | I witness almost daily people that are either in government or even friends of ours who

00:00:07.280 | say we have to win the AI war with China.

00:00:11.000 | And I don't know what that means.

00:00:12.560 | I can't imagine an end state where we control all the AI and they don't have any.

00:00:18.520 | It's already too late.

00:00:19.520 | It's too late.

00:00:20.520 | And they're smart.

00:00:21.520 | And the reality is that we just need to focus on running our fastest race.

00:00:25.800 | We need the Teslas.

00:00:26.800 | We need the open AIs.

00:00:27.960 | We need rockets that land themselves.

00:00:30.280 | We need all of this.

00:00:31.360 | But to think that they're not going to have BYD building great cars, or they're not going

00:00:34.960 | to have DeepSeek building great models, or they're not going to have rocket companies

00:00:39.160 | that copy us and can land themselves like that, that would be naive.

00:00:43.360 | It's remarkably naive.

00:00:49.320 | Bill, it's good to be with you.

00:00:58.240 | Good to see you.

00:00:59.240 | I mean, we're in this-- wait, should we tell them that Steve Ballmer gave us this man cave?

00:01:03.480 | It's a private cave in the interior.

00:01:06.320 | I mean, the truth of the matter is, the hardest thing about this pod-- I love this pod-- but

00:01:12.440 | you and I getting our schedules to match and actually getting together.

00:01:15.360 | And so I've gotten a ton of feedback.

00:01:17.080 | I've seen on Twitter people like, when are you guys going to record the pod first?

00:01:20.720 | Thank you for the audience encouraging us to do this, because I love doing it.

00:01:25.200 | Know that we would love to do it more.

00:01:27.240 | It's just a little challenging to get together and to do it.

00:01:29.860 | We have an ongoing dialogue, I think pretty much 24 by 7, about this stuff going on in

00:01:35.440 | the world.

00:01:36.440 | And then occasionally, we get to get together and share it with you all.

00:01:39.600 | So I don't know.

00:01:40.600 | I thought maybe today, Bill, we'd kick it off with Grok 3.

00:01:44.720 | So we're now like 10 days out since Elon and his team unveiled, in pretty record time,

00:01:52.360 | an unbelievable model.

00:01:53.920 | And so maybe you can just help us zero-base what you thought when the model came out,

00:02:01.360 | where it stands in the rankings, then we can have a conversation about the impact and what

00:02:06.720 | it means.

00:02:07.720 | Yeah.

00:02:08.720 | So we've talked about this in the past, but everyone in the ecosystem was super impressed

00:02:14.040 | with how quickly they built the Memphis facility, and how big it was, and it's the largest contiguous

00:02:20.760 | cluster in the world.

00:02:23.240 | And there was a lot of chatter about that ahead of time.

00:02:26.040 | And I can remember some of the investors there saying, this will prove that pre-training

00:02:32.240 | still has headroom, because this will be the biggest cluster ever trained on.

00:02:37.220 | And you can decide what your expectation was after that kind of line in the sand.

00:02:44.020 | The things that happened, I mean, I think the generic way of saying it is it went right

00:02:50.060 | up near the top of all the benchmarks.

00:02:52.820 | Right.

00:02:53.820 | Ahead on some, not on others.

00:02:55.860 | Some people argued about whether they-- The reasoning component, the beta reasoning.

00:03:00.300 | Or did they cheat or over-tune to a benchmark?

00:03:03.200 | But I don't think it matters.

00:03:05.420 | I think the biggest positive takeaway is there's a new player in the model market.

00:03:11.500 | And we had often, a lot of people said, as a sport of Kings, there's only going to be

00:03:14.900 | so many players.

00:03:15.900 | Correct.

00:03:16.900 | There's a new one in the market that invested what they needed to do, that has access to

00:03:20.060 | capital, has a data asset that they argue is important and special, and was able to

00:03:26.380 | get at the front of the race, let's just call it that.

00:03:29.980 | Right.

00:03:30.980 | And you're looking at, we're looking at this artificial analysis that just shows, I mean,

00:03:35.140 | there's this clustering here in the upper right.

00:03:38.180 | Deep Seek kind of got up there a couple of weeks before.

00:03:40.660 | You had Grok.

00:03:41.660 | You know, what's interesting is they all seem to be coalescing in an impressive way around

00:03:46.660 | the top of these benchmarks.

00:03:48.140 | And when we say they all, I mean, we're really only talking about five or six players who

00:03:52.660 | have a chance to be in this game at this point.

00:03:55.180 | Correct.

00:03:56.180 | Correct.

00:03:57.180 | And I saw people who interpreted Grok's fast rise as proof that pre-trainings still got

00:04:03.300 | legs, and to me, I kind of had the opposite reaction, which is I felt like they just slammed

00:04:09.060 | up against the ceiling that's holding everyone in.

00:04:12.860 | Right.

00:04:13.860 | Although, again, an incredibly capable, right, level.

00:04:18.140 | So...

00:04:19.140 | Oh, no doubt.

00:04:20.140 | I just, you know, I've said this for a while, I've been concerned that the way an LLM works

00:04:25.680 | and the way it's optimized, that building bigger clusters and more parameters won't

00:04:30.100 | buy you much, and whether I said it or not, Ilya said it, Andreessen said it, other people

00:04:36.140 | said the same thing.

00:04:37.500 | And so, to me, this reinforced that point.

00:04:40.780 | I was expecting, if there were pre-training headroom, I was expecting this to go through.

00:04:46.700 | Now, I will qualify.

00:04:48.860 | This was their first run.

00:04:50.380 | Right.

00:04:51.380 | Maybe there were some tricks they didn't know.

00:04:52.660 | Right.

00:04:53.660 | They could very well back up and do another run on that same large cluster and maybe shoot

00:04:57.660 | past these people.

00:04:58.660 | Yeah.

00:04:59.660 | So, benchmarks aren't the exact right thing to be looking at.

00:05:02.100 | Right.

00:05:03.100 | Well, I would say a couple of other things.

00:05:04.100 | Number one, it's not just a pre-trained model.

00:05:06.300 | They also have an inference time reasoning component to the model that's incredibly capable.

00:05:11.460 | We have this benchmark chart, right, that I tweeted the other day, and I compared it

00:05:15.020 | to kind of the search index benchmarks that we all used to track.

00:05:18.860 | And, you know, the benchmarks are one thing, but the reality is, how do we feel when we're

00:05:23.820 | using the product?

00:05:24.820 | And what I will say is Grok 3 rocketed to the top of all app downloads, you know, on

00:05:33.220 | the iPhone charts.

00:05:34.220 | You know, at least my Twitter thread was full of people having great experiences, showing

00:05:40.740 | those great experiences on Grok 3.

00:05:43.340 | So it had a personality and an interaction with people that I think people were enjoying.

00:05:48.220 | So number one, it just has to be capable enough.

00:05:50.980 | Right.

00:05:51.980 | And they clearly crossed the threshold of being capable enough.

00:05:55.740 | Now the real question shifts to, can they leverage the X platform, right, which reaches

00:06:01.820 | a massive and important audience to really drive that.

00:06:05.340 | And what I would say, the early indications to me, when you compare it, for example, to

00:06:10.860 | how Meta has used Meta AI, like, as incredible as I think Zuckerberg and Meta are and the

00:06:17.580 | advancements they've made, I have not particularly been impressed by the productization of Meta

00:06:22.340 | AI.

00:06:23.340 | Right.

00:06:24.340 | It's basically just a search box stuck at the top of Instagram or stuck in my WhatsApp

00:06:27.820 | thread.

00:06:28.820 | And when I'm on it, I never intend to be there, whereas on X, they figured out, you know,

00:06:35.060 | the first thing they did is they put that button at the bottom of the app that clearly

00:06:39.540 | distinguishes it as its own standalone application.

00:06:43.060 | They launched a standalone application.

00:06:45.100 | They're using X to drive those app downloads.

00:06:48.340 | And now I just opened up my, you know, my X app today, and it said, "Hey, go out and

00:06:54.020 | try the new voice for Grok 3."

00:06:56.580 | So to me, the execution on the product side to drive consumer use has been pretty damn

00:07:02.140 | impressive and took them to the top of the charts.

00:07:04.940 | Yeah, and only DeepSeek and Grok, of all the others, have shown the ability to break into

00:07:12.460 | the top 10 on the app store download.

00:07:15.900 | Exactly.

00:07:16.900 | And so I think that, you know, while we all have a fascination where, where did they get

00:07:20.340 | to on the benchmarks?

00:07:22.260 | My own sense at this point in time is, you know, this is going to be one of these battles,

00:07:27.820 | kind of like search was at this point in time, where, you know, the five or six players,

00:07:32.380 | and it's going to be, you know, just out today, literally, as we're about ready to go on,

00:07:36.620 | you know, OpenAI has released ChatGPT 4.5.

00:07:41.580 | And they've kind of hinted in, I guess, this presentation as to ChatGPT 5 or 6, I guess,

00:07:47.500 | that was shown on a screen.

00:07:49.100 | And if you look at 4.5, one of the important distinguishing elements that they're pitching

00:07:53.580 | is it's more human-like.

00:07:54.580 | Yeah.

00:07:55.580 | It gives better answers, more concise answers, etc.

00:07:58.460 | Not a big breakthrough on the evals, although there are some improvements in the early looks

00:08:02.820 | against the evals.

00:08:04.980 | But I think ultimately, we're going to measure the success of these things by how many people

00:08:09.700 | are using them.

00:08:10.700 | Well, so I think one thing to be good for the audience, I know you've said it in the

00:08:14.740 | past, but you're an investor in OpenAI, and I think you have a theory about their prowess

00:08:20.580 | in the consumer market and their lead in the consumer market.

00:08:23.380 | So why don't you reiterate that?

00:08:24.940 | Yeah.

00:08:25.940 | Well, I mean, you know, I've showed, I've showed this chart before, right, that in the

00:08:31.060 | search wars, we had Google and Yahoo and AltaVista and Lycos and Ashgeese and Excite and Infoseek.

00:08:36.500 | And by the way, they all did pretty damn good on the benchmarks, right?

00:08:40.800 | But the reality is that didn't get them to any value creation because ultimately, all

00:08:45.580 | the consumers aggregated around Google.

00:08:48.900 | So the real question is, does that same pattern play out of winner-take-most in consumer around

00:08:55.500 | AI, right?

00:08:56.860 | It did in search, and it did in social.

00:09:00.100 | But it's not necessarily, you know, follow on that it will in AI because, you know, I'll

00:09:04.900 | stipulate X has an incredible installed base that they can market into, Meta has an incredible

00:09:10.260 | installed base, Google has it.

00:09:11.780 | And it's existential for those companies in order to, you know, market to those consumers.

00:09:16.840 | So I don't think it's going to be winner-take-as-much.

00:09:19.980 | I don't think we're going to see a 99% monopoly here.

00:09:23.260 | But I do expect that, you know, we're going to see 70% or 80% share go to the winner.

00:09:28.460 | Now if we look at the numbers, yeah, if we look at the numbers today, I think last week

00:09:32.740 | Sarah reported that OpenAI has crossed 400 million weekly average users.

00:09:39.000 | That's a user number, not a paid user.

00:09:41.580 | That's a user number, right?

00:09:43.020 | The number of paid users is a fraction of that.

00:09:45.700 | I think they also reported last week something like $11 or $12 billion in expected revenue

00:09:50.700 | this year.

00:09:51.700 | So you can reverse engineer your way into kind of what percentage are paying for that.

00:09:57.920 | But more importantly, I think the number of monthly average users must be somewhere on

00:10:02.820 | the order of magnitude of 700 to 800 million monthly average users.

00:10:07.780 | And there, you know, you and I followed consumer for a long time.

00:10:11.080 | There's this magic number around a billion that, I mean, like I already think they're

00:10:14.420 | nearer at escape velocity.

00:10:16.480 | But at a billion monthlies, you can funnel all of those folks into weeklies.

00:10:21.740 | And then you funnel the weeklies into paying subscribers or people who are consuming advertising.

00:10:26.540 | So what I have seen is everybody else catch up on the benchmarks.

00:10:31.180 | What I have not seen is people catch up on the consumer velocity.

00:10:35.980 | Let's handicap some of the other players a bit.

00:10:39.780 | Who do you think is closest from a user standpoint?

00:10:43.540 | Is it probably, you'd have to count the Gemini searches in the Google search, right?

00:10:48.580 | To get to a number that's close to OpenAI.

00:10:50.540 | Yeah.

00:10:51.540 | I mean, listen, I think, you know, let's just start with Google, OK?

00:10:55.940 | So there's been a lot of reports out over the course of the last couple of weeks.

00:10:59.060 | We have public companies now reporting that are reporting their Google organic clicks

00:11:04.780 | are down 20 to 40% year to date.

00:11:08.120 | So the question is, why?

00:11:09.120 | Like, why is Google's clicks down so much?

00:11:11.620 | Because if I do a Google search today on my phone, half of the page is taken up with an

00:11:16.140 | AI answer to whatever my Google query is, and the rest are all paid links, right?

00:11:21.980 | So I think Google, I think that's the right decision for them to make, right?

00:11:26.180 | If you want to compete, you ultimately have to be willing to, you know, take the innovator's

00:11:31.460 | dilemma head on and really just cannibalize your product with AI.

00:11:35.140 | Now if they do that-

00:11:36.140 | And if you're one of these humans that thinks SEO wasn't dead already, which I would have

00:11:40.380 | declared it dead a while ago, it's really f*cking dead, yeah.

00:11:44.660 | SEO is dead.

00:11:46.840 | But just, you know, those are the free links that was the core product that used to attract

00:11:52.180 | everybody to Google.

00:11:53.820 | And the idea that SEO is basically now gone is pretty absurd.

00:11:57.460 | You know, I think, and this is just an aside, but I've been remarkably frustrated with Google's

00:12:02.380 | organic links for the past five years because you go in and search for your favorite team's

00:12:08.180 | schedule and all the ticket guys are up front.

00:12:11.460 | Now, like the link you're looking for, you have to hunt for it.

00:12:14.540 | Okay, so let's talk about that for a second.

00:12:16.580 | I mean, now the obscure link or the obscure information that you and I may be looking

00:12:21.820 | for may be on page three, four, or five.

00:12:24.420 | You and I are never going to get to page three, four, or five.

00:12:27.700 | What's so interesting, for example, about OpenAI's deep research, now if I launch a

00:12:33.560 | query using deep research, it will go to page four or five or 10 or 100 and find those obscure

00:12:41.420 | pieces of information.

00:12:42.760 | So I think the evolution of Google actually provides acceleration to the deep research

00:12:49.160 | project because I don't want to go do that deep research.

00:12:55.640 | So Google, I think you just can't discount their installed base.

00:13:00.000 | The number of people going there who will, inertia will continue to carry them there.

00:13:05.360 | But I would say this, and you can go search for this on Twitter or anywhere else.

00:13:11.320 | And I know certainly with my own behavior, the amount of activity that I used to do on

00:13:16.640 | Google has been 80% cannibalized by ChatGPT because there's search embedded within ChatGPT.

00:13:23.480 | And so I'm getting all of that information, all of those answers.

00:13:27.080 | So I think that they're going to be formidable.

00:13:30.760 | I think they're being bolder than they've been, but I think they'll have to continue

00:13:35.040 | to do that.

00:13:36.040 | I think that, and we've talked about this, but I think some of their assets are remarkable.

00:13:39.680 | I mean, you've got the YouTube data set and all the search queries over all the years,

00:13:46.400 | their understanding of structure, of structured data around a lot of the consumer verticals.

00:13:51.800 | I mean, they built that out in airlines and things.

00:13:55.040 | They should be able to do those agent type queries better, faster, should.

00:14:00.720 | Their velocity on product has not been impressive.

00:14:03.160 | Wait, let me finish.

00:14:04.160 | Their velocity on consumer has not been impressive.

00:14:05.960 | So they have them.

00:14:07.200 | They've had them for a long time, Bill.

00:14:08.960 | They had ChatGPT before ChatGPT.

00:14:11.520 | They also have Android, which is a massive asset.

00:14:15.640 | And they also have browser, their own browser, which both Perplexity and OpenAI have started

00:14:23.200 | toying with the idea of either having a browser or in the operator case of using a browser

00:14:29.040 | in the cloud to go do this work.

00:14:31.440 | Anyway, they have so much.

00:14:34.200 | I still think they have a bit of the innovators dilemma in that they can't, they still have

00:14:38.940 | to try and maintain those paid links on that page.

00:14:42.960 | This chart here, the black line, is Google's paid click growth plotted against the weekly

00:14:48.840 | average user growth at OpenAI.

00:14:51.040 | It's not going in the right direction.

00:14:52.640 | Well, and I think it benefits from the fact that informational searches are what ChatGPT

00:15:00.080 | cannibalized first, not the commerce searches, which is where most of the money is on the

00:15:04.440 | paid link.

00:15:05.440 | Now, again, and we're going to see this out of X, we're going to see it out of everybody.

00:15:10.620 | Everything, the entire domain of the internet is the domain of agents.

00:15:15.780 | So if you think about Operator as one of the first agents rolled out by OpenAI, what does

00:15:21.400 | Operator do?

00:15:22.400 | It goes and it mimics me as a human going out and researching a hotel and booking a

00:15:26.780 | hotel or whatever on the internet.

00:15:28.620 | And we're in a very embryonic state.

00:15:30.380 | I agree with you, we're not there yet, but it's very clear what the roadmap is going

00:15:34.780 | to be.

00:15:35.780 | It's going to want your credentials.

00:15:36.780 | And whether you give it your credentials or not is going to matter because it's searching

00:15:40.180 | against it.

00:15:41.300 | So let's talk about Meta.

00:15:42.700 | Actually, one last thing on Google, there was a point where Meta went public at 40.

00:15:49.340 | We had some.

00:15:50.340 | Backed up the truck.

00:15:51.340 | We had some, so I was paying attention.

00:15:53.860 | And Zuck, as he has many times, got woken up on mobile.

00:15:59.180 | 100%.

00:16:00.180 | Everyone thought he was dead because Apple was messing with him.

00:16:03.620 | And he had also built an HTML5.

00:16:05.660 | He didn't believe in native app.

00:16:07.780 | There's a whole thing that they weren't going to be able to monetize mobile.

00:16:11.860 | It was on the cover of Barron's magazine, the weekend magazine.

00:16:16.740 | It was like Meta's dead or Facebook's dead.

00:16:19.900 | But he woke up and fixed it.

00:16:22.020 | Can Google do that here?

00:16:23.700 | Is that possible?

00:16:24.780 | Can they have a similar, and what would it look like?

00:16:28.180 | And where would it take?

00:16:29.180 | I mean, listen, I've said publicly that Google's moat was not a technological moat with search.

00:16:37.620 | Their moat was a distribution moat.

00:16:39.540 | Their moat was a mind share moat.

00:16:41.580 | We Googled everything when we wanted to know anything.

00:16:45.540 | And the only thing that could attack Google was never anything head on.

00:16:49.420 | It had to be an orthogonal attack from something that was 10x better, 100x better, because

00:16:54.340 | it gave us answers instead of blue links, right?

00:16:57.600 | That's why it was such a mortal sin for them to ever, ever allow anybody else to go first.

00:17:02.780 | Because the only thing that could give you a trillion dollars worth of free mindshare

00:17:07.500 | is going first with something that was 100x better.

00:17:10.260 | And that's exactly what ChatGPT did at the end of 2020.

00:17:13.540 | All right.

00:17:14.540 | Go to Meta.

00:17:15.540 | So, I mean, you know, again, Meta, if you had to handicap the big guys, 3 billion users

00:17:22.040 | of their product, I think they have products that are tailor-made for chat-oriented AI,

00:17:29.100 | whether it's Instagram and having shopping agents and, you know, co-shopping agents or

00:17:33.420 | whether it's WhatsApp and just having a bunch of agents live within my WhatsApp channel.

00:17:38.620 | It feels natively much better positioned for AI.

00:17:42.720 | And we know that Zuckerberg is, you know, in complete beast mode.

00:17:48.700 | But I am surprised, I have to say, that we're now kind of 18 months into kind of the llama

00:17:54.420 | thing and it feels like the manifestation of it into the product was slower than I expected

00:18:00.460 | in 2024.

00:18:01.460 | Back to your product point.

00:18:02.460 | Right.

00:18:03.460 | Exactly.

00:18:04.460 | Right.

00:18:05.460 | Like Meta hasn't.

00:18:06.460 | And I will say, I will say even, you know, we know he was ripped about DeepSeek, right?

00:18:12.880 | Kind of blindsiding llama in the release of R1.

00:18:17.920 | And so I would say it's not just product for them.

00:18:20.340 | I think they have, you know, I heard from several inference players that you and I are

00:18:25.440 | friends with that all of a sudden DeepSeek rather than llama is the enterprise open source

00:18:31.920 | model of choice that everybody's experimenting with and playing with.

00:18:35.700 | And so that becomes a real problem for them as well.

00:18:38.640 | So I think 2025 is a critical year.

00:18:42.020 | I think they will come through.

00:18:43.300 | And remember, when it comes to almost all product stuff, stories, copy, you know, catching

00:18:48.800 | up with with Snapchat or or whether it's Reels catching up with TikTok, they've always showed

00:18:54.840 | up to the party late, but they are grinders and they always deliver the product.

00:18:59.520 | It would be interesting to see what they do.

00:19:01.400 | Who else would be in the list?

00:19:03.200 | I mean, Anthropic has really not been.

00:19:05.680 | No, they've pretty much seeded the game on consumer.

00:19:08.840 | You know, there was a product announcement yesterday about they're going to be powering

00:19:11.840 | Alexa.

00:19:13.560 | But you know, now we're stretching.

00:19:15.800 | Right.

00:19:16.800 | And Amazon did do a big Alexa launch yesterday.

00:19:19.820 | Correct.

00:19:20.820 | Pretty late in this game.

00:19:22.240 | Yeah.

00:19:23.240 | And by the way, Alexa is not really in again, it's it occupies a different space in most

00:19:28.480 | consumers minds.

00:19:30.200 | It is not what chat GPT does.

00:19:32.400 | And so I think to dislodge something that has the momentum chat GPT does, you have to

00:19:37.320 | go at them and do better than what they do at the thing that they do.

00:19:41.920 | And this is why I think X, you know, if I go through the whole list here, X to me is

00:19:46.560 | so interesting because they have a platform that is the number one news platform in every

00:19:54.240 | country on the planet.

00:19:55.760 | The people who are most actively engaged are using this platform and they go there for

00:20:00.000 | information.

00:20:01.000 | Right.

00:20:02.000 | They go there for answers.

00:20:03.000 | They go there to engage.

00:20:04.580 | So I think it's an audience that's very well suited for AI.

00:20:09.560 | I think the integration they've done is as good.

00:20:11.960 | By the way, they've done this in a very short period of time, you know, and I mean, I'm

00:20:16.240 | talking everything from the logo.

00:20:18.400 | Some of the tweets will have the logo pop up and then it'll summarize or do more.

00:20:24.080 | So I'm I'm really, really impressed at the velocity of not only catching up on the benchmark,

00:20:30.680 | but catching up on the consumer product side.

00:20:33.080 | And so, you know, listen, they're number one on the App Store and that stands for something,

00:20:38.500 | you know, coming out of nowhere.

00:20:39.500 | And people said that Elon couldn't, you know, couldn't do this.

00:20:42.680 | I never doubted that they would catch up on the benchmarks if they got a big enough cluster

00:20:47.700 | because Elon set a mission that people become messianic about, you know, and his engineering

00:20:53.060 | capability to build out the cluster and do all those things.

00:20:55.340 | So that was never the question for me.

00:20:57.100 | The real question was, can anybody, can he close the gap on the consumer race, OK?

00:21:04.580 | And there I think, you know, the odds on favorite there has to be OpenAI.

00:21:11.260 | I think they continue to widen their gap, by the way, Bill.

00:21:14.220 | I think they're accelerating at scale, but that's where the race is.

00:21:17.820 | And it may very well be that coming in second place with 20 percent share is a pretty good

00:21:23.740 | place to be.

00:21:24.740 | I want to mention one more company and then I'm going to make a guess at four ways someone

00:21:30.900 | could try and win this game.

00:21:32.860 | But the one I want to mention before I do that is just Perflexity, real briefly.

00:21:37.880 | I will give them credit for being product centric, to your point, and innovative in

00:21:45.260 | ways that the others haven't.

00:21:47.860 | And kind of on their own terms, they don't have near the usage of OpenAI.

00:21:52.940 | So it is a question that kind of looks like an acquisition candidate to me.

00:21:57.660 | I don't know if anyone can agree on price, but for one of these other players that hasn't

00:22:01.860 | been as successful from a product standpoint.

00:22:03.980 | I mean, you could, you and I just had this conversation.

00:22:06.500 | I mean, you could imagine, for example, a world in which Microsoft were to buy Perflexity

00:22:10.820 | and now they have a consumer brand to go battle it out.

00:22:13.080 | We know how much Satya wants to win in consumer.

00:22:16.140 | Now, he owns a bunch of OpenAI and so he's got some potential channel conflict there.

00:22:22.140 | But I think the bigger issue at this point now, you know, we know with Leanicon out,

00:22:26.260 | you're probably more likely to be able to do a deal like that.

00:22:29.820 | But when founders are raising $8, $9 billion, you know, it becomes...

00:22:35.140 | You take it off the table.

00:22:36.140 | It becomes a much, much more difficult decision for a company like Microsoft.

00:22:40.500 | Not saying that it couldn't happen, and I will say this, that when it comes to, you

00:22:45.000 | know, punching up, being innovative, being scrappy, product velocity, the founder there,

00:22:53.380 | Arvind Syk and the team, it's been super impressive to watch.

00:22:56.960 | I think they've made other people better.

00:22:59.340 | But the numbers, I think, as we look at them today, you know, they're really powerful,

00:23:04.800 | but much, much smaller player.

00:23:06.360 | So here we go.

00:23:07.360 | Here we go.

00:23:08.360 | I have four things I'm watching out for that could potentially lead to either further lock-in

00:23:13.240 | by OpenAI or a window for someone to do something else.

00:23:17.520 | And some of them I mentioned before, but I think, you know, memory is still this thing

00:23:22.600 | that could just tie you to something.

00:23:25.480 | And OpenAI has probably done more with memory than anyone else, but no one's really got

00:23:31.400 | to the place where I'm telling it to remember things, to store things, to create lists,

00:23:37.760 | like where it starts to become like an executive assistant for you.

00:23:43.100 | And I haven't seen that yet.

00:23:44.400 | I still think that's a dimension that could really be important.

00:23:49.760 | Voice, we've talked about, and they're all playing with it.

00:23:54.480 | I think voice also ties in with device type.

00:23:58.840 | And this is where Alexa may have some, you know, some assets, but like, you know, if

00:24:06.360 | the voice were spectacular, they might not have to carry the phone around as much.

00:24:10.560 | And by the way, by the way-

00:24:11.560 | But I need it.

00:24:12.560 | I need it.

00:24:13.560 | Yeah, you need an earbud.

00:24:14.560 | You need an earbud.

00:24:15.560 | I will say, advanced voice mode on ChatGPT is excellent, Grok 3's new voice, excellent,

00:24:21.440 | and they're getting better and accelerating rate.

00:24:23.760 | You know, we're an investor in this company, LiveKit, that's powering a lot of this voice.

00:24:28.120 | And I will tell you what I see in the product pipeline is super impressive as to what's

00:24:32.260 | coming with voice.

00:24:35.320 | The third one's nebulous, but just someone could focus on a feature that no one has to

00:24:41.840 | date.

00:24:42.840 | And right now, the game looks so much like with the benchmarks and voice, like everyone's

00:24:48.000 | running at the same place.

00:24:50.160 | So I don't, that's an easy thing to say, but it'd have to be really out of the box.

00:24:55.200 | And then fourth, I just have been thinking about this, no one's really thought about

00:25:00.680 | a network effect.

00:25:02.400 | And I wonder how you could make the quality of the AI experience a function of your user

00:25:09.600 | base.

00:25:10.600 | Let me give you an example of a network effect I think that is happening.

00:25:15.040 | I think around model improvement, if you have seven or 800 million monthly average users,

00:25:22.160 | your diversity of information and questions and answers and follow-ons, et cetera, is

00:25:26.000 | much, much higher.

00:25:27.560 | Those questions and data, that's now being fed back into the models to improve the model.

00:25:31.040 | And some users may have seen, I know I have, you get two answers and the OpenAI asks you

00:25:37.720 | to.

00:25:38.720 | Right, right.

00:25:39.720 | So I think that's an example of OpenAI very actively building, attempting to build network

00:25:45.640 | effects in terms of the quality of the model, the quality of the answers going back to the

00:25:49.360 | user.

00:25:50.360 | But there could be a more intense form of network effect if you found a way to leverage

00:25:56.200 | the user base as part of the value property.

00:25:59.000 | Let me go back to your first one, memory.

00:26:01.320 | Because you and I have talked about this a lot, right?

00:26:03.320 | If you get memory, the switching costs explode, right?

00:26:06.960 | And I would argue not only is switching costs explode, the conversion rate from free to

00:26:12.240 | paid probably also goes up, right?

00:26:14.160 | I think you're right.

00:26:15.160 | Just because the value delivered.

00:26:16.160 | I think you're right.

00:26:17.160 | And so I was with my 89-year-old mother last Sunday.

00:26:21.320 | And my mom has wanted to write a story of her life for a long time.

00:26:26.400 | And the reality is she's never going to sit down and write the story of her life.

00:26:30.840 | And yet when I'm with her, podcast style, I'll ask her questions.

00:26:35.480 | And I'll just record it on my phone, right?

00:26:38.380 | So that I have it and I could perhaps go back later.

00:26:41.200 | And then I started thinking about it.

00:26:43.880 | And I said, I don't need to be the interviewer, right?

00:26:47.280 | Advanced voice mode could be the interviewer.

00:26:49.200 | And so I was sitting there with her last weekend.

00:26:52.080 | Here's the prompt that I gave to advanced voice mode.

00:26:55.600 | I said, I'm sitting with my 89-year-old mother tonight who wants to write her life story.

00:27:00.920 | I want you to interview her about her life, to ask questions about her childhood stories,

00:27:06.360 | having kids, working, growing up in the depression, her love of computers and travel, to remember

00:27:12.160 | everything you talk about, and then compose a story of her life that her grandchildren

00:27:17.320 | would like to read, OK?

00:27:19.840 | And then advanced voice mode just started asking her questions.

00:27:22.840 | How long did it go on?

00:27:23.840 | I mean, my mom was really nervous at the start, but then like a little tear wells in my mom's

00:27:30.000 | eye, you know, because she realizes all of a sudden that, oh my God, this could be a

00:27:35.280 | massive unlock.

00:27:36.320 | So here's the thing, Bill.

00:27:38.560 | Advanced voice mode in ChatGBT already has memory.

00:27:41.660 | You can already do these things.

00:27:43.060 | The problem is the nature of the product, you don't know that it can do those things.

00:27:48.740 | So part of the challenge about designing a product where prompt is your way in is you've

00:27:55.500 | got to help people imagine, like you and I could have imagined in the age of internet,

00:28:00.020 | somebody building an internet website that just did that thing, OK?

00:28:04.420 | So I think that's one of the challenges all these companies faces, and the innovation

00:28:08.820 | around that top end of the funnel in the prompt that can help people better get into it.

00:28:14.220 | I'll give you another example, Deep Reasoning, right, which is really fascinating.

00:28:21.100 | They basically took the O3 series of models and fine-tuned it end-to-end based upon all

00:28:26.820 | these browser interactions, right?

00:28:30.260 | But you need to, the more specific the prompt, the better the deep research report is going

00:28:35.420 | to be.

00:28:36.420 | So a lot of people are using O1 to help them build sophisticated prompts that they then

00:28:41.540 | feed in to Deep Reasoning.

00:28:44.480 | And so I think that there's something in there where we're effectively using AI to get us

00:28:49.300 | to the point where we're better prompting, and one of the ways will be very simple, right?

00:28:54.860 | Once I have this assistant and I'm having an interaction, I just say to my assistant,

00:28:58.540 | "Hey, my mom wants to tell her life story.

00:29:00.380 | I'm not sure how to go about doing that.

00:29:02.300 | You have any ideas?"

00:29:03.300 | And she would say, "Hey, yeah, just use this prompt."

00:29:06.540 | And I also think that there are other assets that could play a role like a contact database,

00:29:13.300 | an email.

00:29:14.300 | >> Yes.

00:29:15.300 | >> I mean, I can't even imagine moving my email to one that's integrated inside just

00:29:20.260 | because-

00:29:21.260 | >> Yes, yes.

00:29:22.260 | Contacts is a great one.

00:29:23.260 | If they just cleaned up your contacts.

00:29:24.260 | >> Well, knowing my contacts.

00:29:25.260 | >> Right.

00:29:26.260 | And know your contacts.

00:29:27.260 | >> Yeah, yeah, yeah.

00:29:28.260 | But send an email, send a text.

00:29:29.260 | I mean, there's a lot big.

00:29:31.300 | And then for anyone that works with content, there's some app, you know.

00:29:36.100 | All my writing and everything I've done for the past 10 years has been in Quip, but some

00:29:40.220 | people use Notion.

00:29:41.220 | >> Yes.

00:29:42.220 | >> And like that type of repository and how these things can interact, there's just a

00:29:48.060 | lot of surface area to figure out.

00:29:50.340 | >> Yeah.

00:29:51.340 | Well, and this-

00:29:52.340 | >> Like where does that story land?

00:29:53.340 | Where is it stored-

00:29:54.340 | >> Right.

00:29:55.340 | >> Once you did it?

00:29:56.340 | >> Exactly.

00:29:57.340 | >> Or do you have to take it out of OpenAI?

00:29:58.340 | Like, you know, you'd rather just have a place.

00:30:00.740 | >> Correct.

00:30:01.740 | And now you have projects in OpenAI and other things.

00:30:03.540 | >> Right.

00:30:04.540 | >> And it brings me to a point.

00:30:06.740 | When you think about these research labs and you look at the number of people who work

00:30:10.900 | there, right?

00:30:12.340 | Just the fact we even call them research labs, you and I haven't, you know, nobody called

00:30:16.440 | Google a research lab.

00:30:17.980 | >> Yeah.

00:30:18.980 | >> Right?

00:30:19.980 | It was a company.

00:30:20.980 | It had product teams, it had marketing teams, it had finance teams, et cetera.

00:30:22.940 | But I think because a lot of these people came out of research, that you look at them

00:30:27.740 | and they're still very small teams, very heavily tilted toward building toward the benchmark.

00:30:33.460 | OpenAI now has thousands of people.

00:30:35.300 | I know, you know, Kevin Wiley who runs the product team over there.

00:30:38.980 | So all these companies, if you're going to win this race, you got to do all the things,

00:30:43.380 | great things that great product teams do.

00:30:45.620 | And it's building all the shit that you're talking about.

00:30:47.740 | And that's hard.

00:30:48.740 | And you got to be thoughtful and you got to growth hack and you got to, you know, get

00:30:51.580 | those customers to use the product more and more.

00:30:53.900 | >> One thing came out this week, which I don't know if it was intentional or not.

00:30:58.260 | You may know more than me.

00:30:59.620 | There was someone published the internal forecast of OpenAI, which included, I think in '25

00:31:07.020 | and '26, losing $20 billion each year.

00:31:10.140 | And someone said to me, why would they publish that?

00:31:12.260 | Why would you lose that?

00:31:14.020 | And to me, I think there is a information war out there trying to scare capital in or

00:31:21.900 | out and to lay a statement that if you want to be in this game, and keep in mind, there's

00:31:28.980 | a variable cost every time you serve a deep researcher.

00:31:32.580 | And so, if this is one of these situations where people think it's winner take all, just

00:31:39.740 | like we had with Uber Lyft, and they're going to go hard at trying to win, you probably

00:31:47.160 | need to be willing to lose $20 billion a year to step into this game.

00:31:51.220 | And X looks like they have the potential to raise that kind of money.

00:31:56.340 | I don't know if a Microsoft or an Amazon are prepared to lose incrementally that amount

00:32:03.440 | of money.

00:32:04.440 | Well, I mean, this is just such a fascinating segue into--

00:32:07.820 | By the way, there's one company we didn't even mention.

00:32:10.840 | Like when we went through all this, we didn't mention Apple, like at all.

00:32:14.420 | Pretty shocking.

00:32:15.420 | Right.

00:32:16.420 | Pretty shocking.

00:32:17.420 | Why didn't we mention them?

00:32:18.420 | Well, I mean, so Apple has self-selected out of the race, right?

00:32:23.280 | They're not building a big model.

00:32:25.100 | They've been very public about, they think that they can be kind of late mover here.

00:32:29.940 | They did this Apple intelligence integration with chat GPT, and now they're going to do

00:32:34.760 | a Gemini.

00:32:35.760 | And the reality is, as I shared with an Apple executive the other day, I said, here's the

00:32:40.900 | only integration that matters, my chat GPT app on the front page of my Apple phone.

00:32:45.300 | Right?

00:32:46.300 | Wow, that's mean.

00:32:47.300 | Well, I mean, it's like the truth.

00:32:48.420 | It's the truth.

00:32:49.420 | Like I just don't use any of the integrated features on the phone, which I think creates

00:32:53.060 | vulnerabilities for Apple.

00:32:55.460 | But I think, you know, listen, this is the first time that they've been faced, I think,

00:32:58.620 | with this level of product risk.

00:33:01.180 | But the reality is, they have such lock in on this device.

00:33:04.500 | For somebody else to build a device, maybe Huawei around the world, they're going to

00:33:08.100 | be able to ship, you know, perhaps better AI phones around the world.

00:33:12.060 | They're not going to be able to ship them into the United States.

00:33:14.460 | We'll see what this Google ruling is at the end of the year, if Google is no longer allowed

00:33:18.980 | to be the default search app because of this consent decree.

00:33:23.260 | You know, do they really turn Android into the thing that it potentially could be?

00:33:28.180 | So I think there's a bunch of potential risks.

00:33:30.180 | So we didn't talk about Apple.

00:33:31.180 | I'd say the other company we didn't really talk about, because we're so focused on the

00:33:34.420 | United States, is when you look outside the United States, you really have to look to

00:33:38.620 | China.

00:33:40.060 | Because I would say the acceleration and velocity of AI in China is off the charts.

00:33:47.740 | You know, we've talked a lot over the last few weeks about DeepSeek clearly coming out

00:33:51.780 | of left field, very efficiently building a frontier quality open source model.

00:33:58.740 | But most people have kind of quietly ignored probably the company that's the leader in

00:34:04.020 | AI in China, and that's ByteDance.

00:34:07.220 | Their AI, you know, their chat GPT equivalent is number one in China, right?

00:34:12.920 | And they've been using AI to drive TikTok globally for a very long period of time.

00:34:18.260 | So I know you have strong opinions on this.

00:34:21.120 | It seems to me the US has underestimated China at AI.

00:34:25.680 | And now we're at this inflection point where I think there are a lot of people who say,

00:34:30.000 | well, they must be smuggling GPUs into China or this or that.

00:34:33.340 | But the reality is China is going to have frontier AI.

00:34:37.700 | And almost all of the things we do to try to slow them down and stop them are backfiring

00:34:43.140 | on the United States.

00:34:44.140 | I couldn't agree more.

00:34:46.500 | I witness almost daily people that are either in government or even friends of ours who

00:34:55.540 | say we have to win the AI war with China.

00:34:59.140 | And I don't know what that means.

00:35:01.140 | Like, I can't imagine an end state where we control all the AI and they don't have any.

00:35:07.660 | It's already too late.

00:35:08.660 | It's too late.

00:35:09.660 | And they're smart as possibly can be.

00:35:14.140 | And they're innovating.

00:35:15.140 | And you look at all the other products that they're crushing it in.

00:35:19.500 | Yeah, I just don't understand.

00:35:20.860 | And the reality is that we just need to focus on running our fastest race.

00:35:25.420 | We need the Teslas.

00:35:26.420 | We need the open AIs.

00:35:27.580 | We need rockets that land themselves.

00:35:29.900 | We need all of this.

00:35:31.000 | But to think that they're not going to have BYD building great cars, or they're not going

00:35:34.580 | to have DeepSeek building great models, or they're not going to have rocket companies

00:35:39.220 | that copy us and can land themselves, that would be naive.

00:35:43.420 | It's remarkably naive.

00:35:45.340 | Yeah.

00:35:46.340 | And it's going to lead to people making decisions, like you said, that either slow us down ourselves--

00:35:53.580 | a lot of the AI regulation would definitely do that-- or just provoke them in ways that

00:36:01.380 | isn't helpful.

00:36:03.460 | And it's not going to slow them down.

00:36:04.780 | Well, let me give you one example of this.

00:36:06.920 | And then I want to move on talking about the arms race, if you will.

00:36:10.380 | But during the Biden administration, they passed something called a diffusion rule out

00:36:14.540 | of the Commerce Department, which we've mentioned on this pod before, which created this convoluted

00:36:19.720 | set of rules by which US semiconductor companies could export outside the United States.

00:36:26.340 | Now, this wasn't exporting to China.

00:36:28.100 | We already have export restrictions with respect to China.

00:36:31.540 | But it basically made all these tiers and classifications on how much you could distribute.

00:36:36.580 | Did you have to distribute it through a hyperscaler or not?

00:36:39.180 | And the whole idea was to somehow prevent these chips from getting to China.

00:36:44.220 | But what it really does is it causes us to have to compete globally with Huawei with

00:36:50.140 | one hand tied behind our back.

00:36:51.900 | And it almost guarantees a Huawei-level belt and road initiative around the world.

00:36:57.540 | And the world is going to run on Huawei AI chips, which gives them then the demand that

00:37:02.900 | they need to build a frontier AI chip.

00:37:05.420 | And so, again, well-intentioned, perhaps, by the Biden administration, but totally backfires.

00:37:12.200 | And hopefully, Howard Lutnick and this administration will throw that out.

00:37:17.660 | I think there are a remarkable amount of people in Washington on both sides of the aisle that

00:37:23.220 | have a perspective about China that they use words, "enemy," "threat," "have to win the

00:37:30.840 | AI war."

00:37:32.660 | And those terms are so loaded.

00:37:36.780 | But I think they think they can achieve something.

00:37:39.820 | And if I owned NVIDIA, my number one concern would be excessive regulation coming out of

00:37:46.780 | Washington.

00:37:47.780 | My number one concern.

00:37:49.220 | Let's shift gears here for a second.

00:37:50.460 | You talked about OpenAI losing this report that they were losing $20 billion a year.

00:37:56.380 | One thing I would just say, I'm not going to share anything that I shouldn't share.

00:38:01.740 | However, I think one always has to keep in mind, what is operating expense and what is

00:38:07.220 | capital expense?

00:38:09.340 | And there's a variable cost of serving a chat GPT query.

00:38:14.460 | And I would posit that those variable expenses are not very high, like at maturity.

00:38:21.580 | Although a deep, like a O1 Pro search or deep research could cost $20, $40, $50x the other.

00:38:29.340 | Correct.

00:38:30.340 | Correct.

00:38:31.340 | But I would just posit for you that you'll be able to come up with a variable expense

00:38:34.700 | structure using the right mix of models that will be a great margin.

00:38:38.820 | May not be as high as retrieval was for Google, but still a great margin.

00:38:42.660 | I think what people are conflating, Bill, is when you decide to spend $20 billion a

00:38:47.860 | year to build out Stargate, to build out clusters, to do all these things.

00:38:51.860 | Now, as you know, a component of that is the CapEx needed to serve the inference and a

00:38:57.540 | component of that is CapEx to build future products, right?

00:39:01.780 | And so, for example, if we're looking at Facebook or we're looking at Google or we're looking

00:39:06.980 | at Microsoft, Microsoft, I think, is spending 80% of their free cash flow, right, on CapEx.

00:39:12.700 | Now, we don't quote that as their profitability.

00:39:15.740 | They have their net income and then they have their net income less CapEx.

00:39:20.580 | I would keep that in mind.

00:39:21.620 | But what I would say is these folks are very committed to continue to invest aggressively

00:39:29.060 | in a future that they see as big.

00:39:31.420 | What we heard from Satya on the Dworkish podcast, right, what many are characterizing as a pushback

00:39:39.660 | against these high levels of spending.

00:39:44.100 | I think of my fleet, even, as a ratio of the AI accelerator to storage to compute.

00:39:50.420 | And at scale, you've got to grow it.

00:39:53.900 | And so, that infrastructure need for the world is just going to be exponentially growing,

00:40:02.060 | right?

00:40:03.060 | So, in fact, it's mana from heaven to have these AI workloads because, guess what, they're

00:40:08.940 | more hungry for more compute, right?

00:40:11.500 | Not just for training, but we now know for test time.

00:40:13.900 | And as I said, test time.

00:40:14.900 | Like, here's an interesting thing.

00:40:15.900 | When you think of an AI agent, it turns out the AI agents is going to exponentially increase

00:40:21.060 | compute usage because you're now not even bound by just one human invoking a program.

00:40:26.820 | It's one human invoking programs that invoke lots more programs.

00:40:30.980 | And so, that's going to create massive, massive demand and scale for compute infrastructure.

00:40:36.140 | So, our hyperscale business, Azure business, I think that's like, and other hyperscalers,

00:40:40.020 | I think that's a big thing.

00:40:41.740 | And I think on the pod, he reiterated, "We're going to spend $80 billion this year.

00:40:45.300 | We'll spend more next year.

00:40:46.980 | So, there's not a world in which we're just going to have unlimited, unconstrained spending."

00:40:51.180 | Now, this week, we also saw rumored that Meta is out shopping for a campus, a data center

00:40:58.340 | campus.

00:40:59.420 | The rumored amount is $200 billion, capable of building six to eight gigawatts.

00:41:04.780 | Now, that sounds a lot like Stargate, which is kind of in that six to eight gigawatts.

00:41:09.260 | Microsoft, I think, has five gigs installed, probably is going to build a few.

00:41:13.700 | Worldwide.

00:41:14.700 | What?

00:41:15.700 | Five gig worldwide.

00:41:16.700 | Is that what you mean?

00:41:17.700 | Correct.

00:41:18.700 | Yeah.

00:41:19.700 | And going to build more.

00:41:20.700 | So, again, it seems to me that if you want to be in the group of five or six, that's

00:41:24.820 | kind of the calling card you have to have.

00:41:27.180 | You have to either have a business or the ability to raise capital, such that you can

00:41:31.580 | deploy a sufficient amount to build out that level of compute.

00:41:35.140 | Now, in the case of OpenAI, enter Masa, back to Lyft, Uber.

00:41:42.020 | And Masa's rumored to be leading a very big round, $40 billion round, with a lot, which

00:41:48.740 | we saw they announced at the White House.

00:41:50.820 | It is important.

00:41:52.060 | Many people interpreted Satya's comments as a tapping of the brakes.

00:41:57.420 | Yes.

00:41:58.420 | So tell me how you interpret it.

00:41:59.420 | Because he said, "I'm happy that some of these are leases," which I don't know any other

00:42:05.540 | way to interpret that.

00:42:06.780 | Well, there's two ways you can interpret it.

00:42:08.300 | One is he's telling you, like, "I'm hedged against this being overbuilt," or, "I'm better

00:42:15.940 | off canceling a lease than sitting on infrastructure."

00:42:20.500 | I would say it even a little bit more.

00:42:22.340 | Let's be honest.

00:42:23.340 | Satya said last June, we talked about on this pod, that it was very likely that at some

00:42:28.460 | point there would be a supply and a demand mismatch.

00:42:31.820 | And you had to build a resilient company that could go through a zone of disillusionment.

00:42:36.660 | So he basically said the reckoning is coming at some point in time.

00:42:41.060 | And so now he goes on Dworkish.

00:42:42.700 | He kind of sounds like he's tapping the brakes a little bit, you know.

00:42:47.540 | And so I think that the interpretations of that should not be that he doesn't believe

00:42:53.980 | in AI.

00:42:54.980 | I think he very much believes in AI, but he's running a public company.

00:42:58.860 | And I think that he's made commitments to his shareholders, and he's saying, "Listen,

00:43:02.460 | I need to see a certain amount of inference revenue in real time to justify that level

00:43:07.820 | of capex."

00:43:08.820 | Yep.

00:43:09.820 | Look, I mean, I think everyone believes in AI.

00:43:12.220 | This amount of spend is something we've never seen before.

00:43:19.060 | That's why I've said, you know, that it's like better than watching "Secession."

00:43:22.860 | This is just like, it's a massive sport of kings.

00:43:27.940 | And I think some of the things, whether it's the 20 billion losses, or Satya's saying he's

00:43:33.500 | glad he's got leases.

00:43:34.740 | Some of these might be part of an information war with other players trying to talk capital

00:43:42.500 | in or out.

00:43:43.500 | For sure.

00:43:44.500 | And it's a high stakes game.

00:43:46.340 | It's fun to watch.

00:43:47.340 | Well, I think resiliency, business model resiliency is going to be critical here.

00:43:53.020 | And what do I mean by that?

00:43:54.340 | It means liquidity, because we know in the internet there was a zone of disillusionment.

00:44:02.460 | We know in social, there was a zone of disillusionment.

00:44:04.920 | We know in cloud, there was a zone of disillusionment, right?

00:44:07.700 | A period, what do I mean by that?

00:44:09.700 | A period where the prices and the spend got ahead of the revenue, right?

00:44:15.020 | And given the level of competition, some people describe as a prisoner's dilemma, right?

00:44:21.920 | In the case of Google and Meta, they literally have a printing press in the back room spitting

00:44:26.800 | out billion dollar bills, right?

00:44:28.880 | So they are resilient.

00:44:30.320 | Microsoft, resilient, right?

00:44:32.080 | In the case of OpenAI, they have to raise money, right?

00:44:36.340 | So you need to have a big stack behind you.

00:44:39.280 | In the case of X, right, they need to be able to raise capital.

00:44:43.120 | I think there were some numbers out there last week.

00:44:46.000 | Obviously, Elon is the wealthiest person on the planet.

00:44:49.400 | He can sell shares in some things.

00:44:51.340 | But I think the most powerful thing Elon has is a global belief in him as an entrepreneur,

00:44:57.880 | which gives him an opportunity to raise capital from sovereigns around the world.

00:45:02.900 | And so if you said, is this still an open sport, I'd say, no way, right?

00:45:06.840 | I don't know anybody else other than Elon and Sam at this point.

00:45:09.620 | Although DeepSeek surprised everybody.

00:45:11.120 | Well, I'm saying if you're going to play that game, right?

00:45:14.520 | To remind you, DeepSeek spent more than the amount reported in their last training run.

00:45:22.100 | But even more importantly, to serve an explosive amount of inference, they would have to spend

00:45:27.020 | a lot of money to build--

00:45:28.020 | I want to make a point that we'll probably come back to much later.

00:45:32.460 | But when you have a scenario that has this much ambition, and this much competition,

00:45:43.500 | and this much CapEx as part of the game, it's easy to lose sight of the microeconomics.

00:45:52.100 | It's easy to lose sight of the unit economics.

00:45:56.220 | So if you're an anthropic, and you've got training credits over here, and you've got

00:46:02.980 | CapEx, and you do your-- am I thinking about depreciation or that when I say, oh, this

00:46:09.340 | is profitable, or when I price my API product?

00:46:13.300 | And you've got this razor edge pricing thing that I've never seen before.

00:46:19.700 | Explain what you mean by that.

00:46:20.700 | I mean, the price difference between today's model and yesterday's model is 20x.

00:46:28.880 | So it's a fast depreciating asset the second you're off the frontier.

00:46:33.940 | Yes.

00:46:34.940 | Yes.

00:46:35.940 | And so it's just-- it's a dangerous-- these are all traps.

00:46:39.300 | Yeah.

00:46:40.300 | And it makes it, once again, fascinating.

00:46:44.340 | Maybe we can transition to the public markets a bit, but there's a lot of talk that we're

00:46:48.820 | going to see a CoreWeave filing.

00:46:50.860 | And I'm just excited to see the numbers and to piece together more of the information.

00:46:55.540 | Yeah.

00:46:56.540 | There's a rumor out there that the CoreWeave is going to file an IPO, and so you can see

00:46:59.100 | the numbers.

00:47:00.100 | Well, I just want to underscore this point that you just made, though.

00:47:02.900 | Because I mean, and there is some rhyming to Masa coming back into the scene here, right?

00:47:12.660 | Masa is one of the greats of this industry of the last 25 years.

00:47:18.440 | But I think people would also describe him as somebody who's a bit of a gambler and places

00:47:24.300 | gigantic bets, right?

00:47:26.860 | And some people would say that he's a total visionary, and other people would say he's

00:47:31.500 | just not price discriminating, right?

00:47:34.220 | But clearly, he's shoving all in with open-- I don't think he knows any-- I don't think

00:47:40.140 | he has any other way of operating.

00:47:41.980 | Right.

00:47:42.980 | And so I think the point being that we're at this moment in time where the danger for

00:47:52.060 | the company-- I just want to underscore what you said.

00:47:54.720 | The danger for the company of getting this volume of capital is that it's hard to focus

00:48:00.600 | on really building the muscle and the grit and the ingenuity on how to drive unit economics.

00:48:07.420 | Think about what Elon had to do at Tesla, right?

00:48:10.260 | Because capital was hard to come by.

00:48:12.400 | So he had to figure out how to make money on every damn car.

00:48:15.740 | How do I take costs out of the manufacturing at every single stage of production?

00:48:21.100 | And when you have excess capital, right, you lose that discipline.

00:48:24.700 | You don't build that muscle.

00:48:25.980 | And so I think it's an important admonition for the board and leadership at OpenAI and

00:48:32.380 | all these companies to hear that, sure, it's one thing to invest aggressively in the future,

00:48:37.340 | but you better make sure that along the way, your unit economics work.

00:48:40.860 | No doubt.

00:48:41.860 | I know you've been thinking a lot-- let's switch gears.

00:48:44.060 | You've been thinking a lot about Doge and if it happens, what it means for the capital

00:48:49.260 | markets.

00:48:50.260 | Yeah.

00:48:51.260 | And it's interesting to even say if it happens, because as I watch the press every day, there's

00:48:55.500 | an equal number of people that say, oh, this is going to take out all these costs.

00:48:58.900 | And there's other people that say, oh, they're just saying things, but they're not actually

00:49:03.720 | going to happen.

00:49:04.720 | Yeah.

00:49:05.720 | Yeah.

00:49:06.720 | You and I said some-- so I think on our pod on like February 6 or something, when you

00:49:11.540 | asked me about the markets, I said, hey, we have peak political uncertainty, right,

00:49:15.940 | because we have a lot of things changing.

00:49:17.740 | We have peak economic uncertainty.

00:49:19.500 | And that's not just Doge, because-- Right, because we have tariffs and other things.

00:49:23.460 | And I said that we have peak technology uncertainty, i.e., it's hard to predict the future, what

00:49:28.900 | software company is going to be worth what in five years.

00:49:31.660 | And that causes discount rates to go up.

00:49:34.580 | It causes multiples to come down.

00:49:36.860 | And I said I was surprised how resilient the market was in the face of all this uncertainty.

00:49:41.900 | Well, now I would argue we're starting to see a few cracks in that.

00:49:47.420 | And so if you look at this chart, Bill, it's really the NASDAQ since the election.

00:49:54.060 | And we ran way up.

00:49:55.260 | The NASDAQ was up as high as 10% post-election.

00:49:58.660 | And now we've come off four or five points from that high.

00:50:03.260 | But we're still four or five points higher than we were on the night of the election.

00:50:08.300 | And so one thing I just have been thinking a lot about, and I've been talking a lot about,

00:50:13.260 | is this difference between stimulus and austerity.

00:50:18.740 | Over the last three or four years, we had massive stimulus into the economy.

00:50:23.540 | Now, you and I both supported it in March and April of 2020.

00:50:29.540 | Right when we were in the depths of COVID, you had to prevent the economy from coming

00:50:34.500 | to a screeching halt.

00:50:36.180 | And so the Fed went all in and Congress went all in in order to save the economy.

00:50:41.360 | But then we also were very critical that the Fed moved way too slow.

00:50:46.460 | The second stimulus package was way too large.

00:50:49.420 | And it led to this runaway inflation.

00:50:51.340 | We saw inflation hit 9%.

00:50:54.100 | But the one thing that all of that monetary liquidity did to the system is it caused risk

00:51:00.260 | assets to go up in value.

00:51:03.740 | And now we're in this period where we're talking about not adding a trillion and a half of

00:51:07.680 | liquidity to the system, we're talking about pulling a trillion and a half out.

00:51:12.220 | Now, what do I mean by that?

00:51:13.740 | Okay, so last year, we had $56 billion of tariffs imposed on other countries.

00:51:19.780 | That's the amount of revenue we collected from tariffs.

00:51:22.260 | We're talking about that going to $500 billion.

00:51:25.420 | So 10x in the amount of tariffs.

00:51:27.420 | Well, we know that some of those will be eaten by producers, right?

00:51:31.100 | The company that's producing something in China will just take a lower margin.

00:51:35.100 | But we know a lot of those will be felt by U.S. consumers who just end up paying higher

00:51:39.540 | prices for their Dell computer because Dell passes along the price increase of the computer

00:51:44.340 | made in Mexico, as an example.

00:51:47.340 | So that's $500 billion.

00:51:48.440 | On the other hand, I think Doge, there's no doubt in my mind at this point in time, and

00:51:53.300 | we'll show this chart of the likely spending cuts, they're not only making big cuts, and

00:51:59.780 | the president has now just last week said he wants Elon to be more aggressive, right?

00:52:05.820 | They sent this email out to every employee that said, you know, respond back to us or

00:52:11.020 | you'll be deemed, you know, to have resigned.

00:52:13.740 | Now they're giving them more, you know, shots on goal.

00:52:15.900 | But the message is very clear that I think there's going to be a downsizing of the federal

00:52:20.140 | government to the tune of, let's call it 40 or 50%.

00:52:24.180 | Now, a lot of people have been giving a lot of grief to Doge.

00:52:27.660 | But I remind you, and I tweeted this the other day, that Bill Clinton, right, did Doge in

00:52:33.180 | the late '90s.

00:52:34.180 | Yeah.

00:52:35.180 | I don't know the exact percentage of federal employees they let go.

00:52:37.820 | It was like between 10 and 20%.

00:52:40.020 | But we had a balanced budget, you know, in three fiscal years, we had a $230 billion

00:52:45.340 | surplus.

00:52:46.340 | Now, it was helped by the internet, but now we're going to be helped by AI.

00:52:50.220 | So like, I think that you can see some replay of that.

00:52:54.420 | But it does mean that we're probably going to take $500 billion to $1 trillion out of

00:52:59.020 | federal spending over the course of the next couple of years.

00:53:01.800 | And all I'm suggesting is that austerity has the reversed impact of liquidity from government

00:53:08.480 | into the system.

00:53:09.480 | So if you think about, go back to our GDP calculation, right, in macroeconomics, C plus

00:53:14.940 | I plus G, where G is the amount of money the government's spending.

00:53:19.060 | Well, the amount of money the government's spending is going down.

00:53:22.180 | So tariffs is a headwind to the economy, and this austerity out of the government.

00:53:26.060 | Now, I am 100% in agreement.

00:53:28.860 | This is the short-term shock therapy we need in order to get our fiscal house in order,

00:53:33.500 | right?

00:53:34.500 | But you've got to think about this as, you know, somebody says, hey, you're out of shape,

00:53:37.700 | you're going to have a heart attack.

00:53:39.060 | You've got to, you know, you've got to take this medicine, this short-term pain, you've

00:53:42.180 | got to work out every day, you've got to get fit, right, in order to avoid the heart attack.

00:53:47.540 | You would do it every day of the week.

00:53:48.980 | We need to get fit in order to avoid bankruptcy.

00:53:51.620 | And all I'm suggesting is-- It might affect markets.

00:53:55.040 | That it might affect markets.

00:53:56.700 | So markets may, in fact, right, my risk profile is lower than our standard risk profile.

00:54:02.920 | What do I mean by that?

00:54:03.920 | Very simply, you know, I own half as much as I would normally own at a point in time.

00:54:09.700 | Now, do I think that's because the future, you know, is bleak?

00:54:13.900 | No, I believe aggressively in the future.

00:54:17.080 | But I think we're going to have to take a little bit of short-term pain, which means

00:54:20.660 | we could see just a random run of the mill, 10% to 15% drawdown, right, in the markets

00:54:26.400 | while the market gets its head around the fact that the economy is going to grow a little

00:54:30.180 | slower.

00:54:31.180 | When the economy grows a little slower, that means companies grow a little bit slower.

00:54:35.340 | When they grow slower, you know, the earnings goes down and the multiple goes down.

00:54:39.420 | Let me ask you this question.

00:54:41.280 | When Elon went into Twitter, one of the stories that came out was that they found there were

00:54:47.340 | software licenses for a whole bunch of people that they weren't using and that they cut

00:54:52.440 | that dramatically.

00:54:53.980 | Do you anticipate that one of the outcomes of Doge will be a, you know, obviously a headwind

00:55:02.660 | for a bunch of companies that have sold software and/or services into the market?

00:55:08.980 | 100%.

00:55:09.980 | I mean, like, there's just no way around it.

00:55:11.260 | You know, if you go from 3 million federal employees to a million and a half federal

00:55:14.620 | employees, then you don't need as many licenses, you don't have as much cloud consumption,

00:55:18.820 | right?

00:55:19.820 | And so if you think about the multiplier, right, you take the federal government or

00:55:23.860 | federal person's salary, employee's salary, now you have all the healthcare and benefits

00:55:28.240 | and pension and all the other stuff, and then you have all the ancillary spend.

00:55:31.940 | So I won't say the exact company, but I talked to an airline the other day, and at this airline,

00:55:38.220 | their number of government tickets sold year to date is down 50%.

00:55:43.180 | Already impacted.

00:55:44.900 | 50%.

00:55:45.900 | Yeah, because they said, we don't want, you know, you traveling, we want you in the office

00:55:51.100 | every day and all this other stuff.

00:55:53.220 | And so this airline has already been impacted.

00:55:55.380 | So I think everybody in the ecosystem, if you have revenue line items, if you're a business,

00:56:00.780 | if you're a public company, you have revenue line items from the federal government.

00:56:04.540 | It's not just that the rate of growth is going to slow.

00:56:07.980 | It's that they're actually going to be negative on a year on year basis.

00:56:11.020 | Now again, I happen to think this is a generally a good sacrifice for us to make.

00:56:16.160 | Those are our tax dollars.

00:56:17.380 | There is no government money.

00:56:18.960 | This is our money that's being, you know, being consumed.

00:56:22.700 | But I don't think the public markets or investors generally, and certainly not Silicon Valley,

00:56:27.580 | has kind of got in their head around what this means.

00:56:29.820 | Now what's the-

00:56:30.820 | Well, and in fact, I would say, like ironically, this happens quite a bit in our world, but

00:56:35.660 | the Silicon Valley and the venture capitalists have just gotten comfortable with backing

00:56:41.940 | companies that sell to government.

00:56:43.580 | Exactly.

00:56:44.580 | We see a lot of that.

00:56:45.580 | It's an interesting timing.

00:56:46.880 | People get excited about it.

00:56:47.880 | Well, you saw what happened to Palantir stock the other day when, you know, the president

00:56:52.620 | directed his cabinet members, Secretary of Defense, to find 8% cuts in the Department

00:56:59.300 | of Defense every year, right?

00:57:01.980 | And so this austerity, again, is real.

00:57:05.340 | Now, that probably means we're going to have a rotation of money out of, like, the less

00:57:10.860 | technologically innovative folks into the more technologically innovative folks.

00:57:15.220 | But you know-

00:57:16.220 | He went further.

00:57:17.220 | Trump went further.

00:57:18.380 | He suggested in one thing, which I was really blown away by this.

00:57:22.180 | I actually thought it was kind of the most interesting thing he's done.

00:57:25.300 | He suggested to Xi that China and America should both cut their military budgets in

00:57:31.020 | half.

00:57:32.020 | Yes.

00:57:33.020 | Now, maybe that's provocative.

00:57:34.020 | Maybe it's-

00:57:35.020 | By the way-

00:57:36.020 | That'd be an amazing idea.

00:57:37.020 | I thought that was extraordinary.

00:57:38.020 | Back to the, you know, this idea, you know, we've been-

00:57:40.500 | We both blow up the world many times over, so-

00:57:42.860 | Right.

00:57:43.860 | There's a certain camp of folks, right?

00:57:46.860 | And I think Mearsheimer is in this camp, right?

00:57:50.220 | Which is great power politics and, like, you just got to build, build, and build, and,

00:57:54.740 | you know, eventually you're going to have a war or something like this.

00:57:57.460 | Or maybe the fact that you have these stockpiles deters, you know, the ultimate war.

00:58:02.840 | One thing that is just fascinating, I've never heard an American president in my lifetime

00:58:10.060 | suggest that he wanted to sit down at a table with China and Russia and talk about they

00:58:14.900 | could cut their- they could collectively cut their military spending in half.

00:58:17.900 | Right.

00:58:18.900 | So, from an entrepreneur perspective, like, isn't it- it caused me to stop in my tracks

00:58:23.100 | and be like, "Hmm, that's an interesting idea."

00:58:25.340 | I thought it was the coolest thing he's thought of.

00:58:28.860 | That's an interesting idea.

00:58:29.860 | Well, I will tell you back on the public markets, the other interesting thing here, Warren Buffett,

00:58:35.940 | you know, just put out his annual letter, he's going to have his annual meeting coming

00:58:38.780 | up here, has a $400 billion cash stockpile, has been liquidating stocks-

00:58:44.140 | Yeah.

00:58:45.140 | Right?

00:58:46.380 | -for a while, in ages.

00:58:48.860 | Stan Druckenmiller, Howard Marks, Stevie Cohen came out over the weekend and said, "I'm nervous

00:58:54.680 | about the markets," for the same reason that we were talking about a month ago.

00:58:58.580 | So I think there is a growing chorus of players.

00:59:02.140 | Now, what's the flip side to this?

00:59:04.820 | Well, since Trump's been elected, the cost of a mortgage or credit card or etc. is starting

00:59:10.500 | to come down.

00:59:11.500 | Why is that?

00:59:12.500 | Right?

00:59:13.500 | There are two reasons.

00:59:14.500 | The first reason, I think, is because we're saying, okay, the economy is going to slow

00:59:19.220 | a little bit.

00:59:20.220 | And if the economy slows, equities, as an investment, are a little less positive relative

00:59:26.940 | to a bond.

00:59:27.940 | So you rotate that into cash.

00:59:29.780 | And when the cash is sitting on the sideline, it's invested in a U.S. treasury.

00:59:33.620 | Just to put it in perspective, and the only anecdote I really ever hear about this is,

00:59:38.500 | well, China doesn't want to own our treasuries anymore.

00:59:41.700 | China buys 3% of our treasuries annually.

00:59:45.660 | It's tiny.

00:59:46.660 | They used to buy, 10 years ago, they bought 12% of our treasuries, and everybody panicked

00:59:51.260 | that they were too big a buyer.

00:59:53.020 | So what I see is just the opposite.

00:59:55.460 | Every sovereign around the world and every domestic investor who's starting to put more

01:00:00.340 | money into cash, who's hedging a little bit, all of that's going into U.S. treasuries.

01:00:04.680 | So I just think that one should brace over the next three months.

01:00:09.340 | I think these tariffs are very real, they're structural, and the president is committed

01:00:13.900 | to them.

01:00:14.900 | I think, number two, the reconciliation package is now rolling.

01:00:18.860 | And I think they are very committed to balancing the budget within this president's term.

01:00:24.180 | And the only way you balance the budget is a trillion dollars has to come out of spending.

01:00:29.300 | Remember, 2019 baseline, we were spending about $5 trillion.

01:00:34.220 | The COVID high, $7 trillion.

01:00:36.980 | We got to get that back down to at least $6 trillion, probably to $5.75 if you're going

01:00:42.400 | to balance the budget.

01:00:43.740 | That means a trillion out in a year, that's austerity, and that's going to be a headwind

01:00:47.340 | to the economy.

01:00:48.340 | But it's the right thing to do.

01:00:49.340 | OK.

01:00:50.340 | That's a tough note to end on, so I'll switch to something more positive.

01:00:56.100 | I got invited to the Golden State Warrior game on Tuesday night.

01:01:00.760 | The butler trade looks like it's working.

01:01:02.980 | It's incredible.

01:01:03.980 | Six and one, I think, since the trade.

01:01:06.540 | It's incredible.

01:01:07.540 | I was related.

01:01:10.160 | I happened to go get an invite to the banner ceremony and dinner afterward for good friend

01:01:15.940 | Andre Iguodala.

01:01:18.020 | And Steph gave an incredible speech.

01:01:22.380 | And I had Andre speak at our investor day maybe two years ago.

01:01:27.060 | And two things Steph said that really stood out to me about Andre.

01:01:34.940 | Number one, he said, there is no this without Andre, right?

01:01:41.100 | And by this, and he explained it to me, he said, he came at a moment in time, even his

01:01:47.380 | decision to come to the Warriors made us believe in ourselves.

01:01:52.300 | And then he came here and he did whatever it took.

01:01:56.420 | And the second thing he said is Andre Iguodala always put excellence over ego.

01:02:04.260 | The guy would be the first to never pound it on the bench.

01:02:08.380 | When he came off the floor, he was the first to get guys fired up.

01:02:11.340 | And Steph talked about game six in Boston.

01:02:13.740 | I remember that game.

01:02:14.860 | I was at that game.

01:02:16.740 | And I remember Andre.

01:02:20.860 | He must have played five minutes in that game.

01:02:23.340 | And he was so fired up and really willed all the players to up their game.

01:02:28.220 | And so I was so happy for him.

01:02:29.780 | Yes, and you know, me and our good friend Jason Chang, I never bet on sports.

01:02:36.660 | I never bet on sports.

01:02:38.100 | And he talks me in.

01:02:39.100 | We're at a Warriors game during the losing streak.

01:02:41.060 | And the odds are so great that they're not going to win at all.

01:02:45.060 | He talks me into placing a bet on them winning at all.

01:02:48.060 | And at the time it was like 40 to one, right, against them.

01:02:51.740 | And all of a sudden they're on this six game winning streak.

01:02:53.660 | They trade for Jimmy Butler and they may win this whole thing.

01:02:57.060 | So, you know, fingers crossed.

01:02:59.620 | It now has me with a focused mind.

01:03:03.740 | With a focused mind.

01:03:04.740 | It's great to see you.

01:03:05.740 | Great to be with you.

01:03:06.740 | Take care.

01:03:07.740 | As a reminder to everybody, just our opinions, not investment advice.

Grok 3, AI Memory & Voice, China, DOGE, Public Market Pull Back | BG2 w/ Bill Gurley & Brad Gerstner

Chapters