Latest / In the Loop AI / Google's New AI Chip Could Actually Dethrone Nvidia (Ironwood v7 TPU)

Transcript

Google's New AI Chip Could Actually Dethrone Nvidia (Ironwood v7 TPU) · In the Loop AI

0:00
Josh: One of the most famous investors in the world, Peter Thiel, he wrote a book called Zero to One.
0:04
Josh: Some of you might have heard it, some of you may have not, but it's all about
0:07
Josh: monopolies and how much of an advantage having a monopoly has in the world.
0:11
Josh: Now, what we've seen recently is a company named NVIDIA reaching a $5 trillion
0:17
Josh: monopoly. It's the biggest monopoly to ever exist in business, and it's huge.
0:22
Josh: It's made tons of people, infinite amounts of money, but something is happening.
0:26
Josh: Something is wrong. It appears as if that monopoly is starting to slip out of their hands.
0:31
Josh: And in this episode, we're going to talk about why and what that means for the entire market.
0:35
Josh: I mean, when you're a $5 trillion asset, the size of your company makes a meaningful impact on the market.
0:41
Josh: So when a company like NVIDIA loses hundreds of billions of dollars over a short
0:46
Josh: period of time like the last month, we got to start paying attention to this
0:49
Josh: because it's dragging everything else down with it.
0:50
Josh: So up here on the screen, we have some charts that I want us to walk through
0:53
Josh: EJS. So if you could just kind of show us the difference between NVIDIA and
0:57
Josh: who we believe to be the competition that is responsible for knocking these
1:01
Josh: hundreds of billions of dollars off their market cap.
1:03
Ejaaz: Yeah, so this really scary looking red chart that you have in front of you is
1:07
Ejaaz: the last month's performance for NVIDIA stock. And it is down a staggering rate.
1:11
Ejaaz: 13, almost 14%, which equates to over $500 billion of market cap loss.
1:18
Ejaaz: That's a billion with a B, which is just an insane amount of money for the largest
1:24
Ejaaz: company or stock company in the world to lose.
1:27
Ejaaz: So the obvious question that comes in is why and where's that market,
1:31
Ejaaz: where's that money kind of flowing towards?
1:33
Ejaaz: Well, I want to show you another chart, Josh, which is Google's chart.
1:37
Ejaaz: And do you notice a similarity?
1:39
Ejaaz: Over the last month, it is up almost the same amount in percentage market cap
1:44
Ejaaz: for a very peculiar reason, or maybe it's not so peculiar.
1:49
Ejaaz: Josh, have you heard of these? You've heard of these things called TPUs, right?
1:52
Josh: We're talking about this. Oh, there's no things called TPUs.
1:54
Ejaaz: Yeah. There's no things called TPUs. You know, Josh and I like to kind of go
1:57
Ejaaz: back and forth and discuss this a lot.
1:58
Ejaaz: In fact, we actually put out a bull episode on Google, which a bunch of you
2:02
Ejaaz: watched a few weeks back. and I don't want to be running victory laps here,
2:06
Ejaaz: but it turned out that Josh and I might have been onto something.
2:09
Ejaaz: But I want to dumb down what's going on here, showed by this really hilarious
2:14
Ejaaz: graphic or comic from the semi-analysis team.
2:17
Ejaaz: And it basically goes, Google came out with this new rock, new shiny rock called TPU version seven.
2:24
Ejaaz: It's basically their version of NVIDIA's GPUs, but it's built by themselves in-house.
2:29
Ejaaz: And it's actually really, really good. It gives you an average of 30% to 50%
2:34
Ejaaz: cost savings for the exact same performance or equivalent of an NVIDIA GPU.
2:38
Ejaaz: And it, in some cases, performs even better, up to one and a half to two times better, right?
2:43
Ejaaz: And so you've got NVIDIA in the leather jacket here on the right,
2:47
Ejaaz: which says, actually, my ROC, my GPU is faster, right? And Google's like, is that true?
2:52
Ejaaz: And then everyone ends up using Google's TPUs. And the point being made here
2:56
Ejaaz: is for the longest time, Josh, NVIDIA held the monopoly on the AI training and
3:03
Ejaaz: inference market via their GPUs.
3:05
Ejaaz: It's all anyone and everyone could use to train their models.
3:08
Ejaaz: It was the only option that they had. And now Google presents a real threat
3:12
Ejaaz: to NVIDIA's market dominance by presenting these TPUs.
3:16
Ejaaz: Now, initially, they use these TPUs to train their own model in-house.
3:20
Ejaaz: In fact, Google's never purchased NVIDIA GPUs to train their own models,
3:24
Ejaaz: and yet they have the best models, which tells us that the TPUs is something
3:27
Ejaaz: to really contend with NVIDIA's GPUs.
3:29
Ejaaz: But most recently, Josh, they've started selling these to other companies,
3:33
Ejaaz: supposedly, to train their own models. And so we're reaching a point now where
3:38
Ejaaz: Google and NVIDIA is a direct comparison.
3:40
Ejaaz: And we're seeing that in the market share dynamics that are happening now.
3:43
Ejaaz: You know, you've got NVIDIA losing up to $500 billion and Google gaining the
3:48
Ejaaz: same amount over the same month. It's just pretty insane to see.
3:51
Josh: Yeah, I can't stress this enough, how insane that delta between the two stock charges.
3:56
Josh: That's 26% combined in one month. So the market is really pricing in the fact
4:00
Josh: that this monopoly is starting to crumble.
4:03
Josh: Now, I think we have reasoning why that's not necessarily the case.
4:05
Josh: That will come later in the episode.
4:06
Josh: But for now, there is some real forces at play. I mean, EJs,
4:09
Josh: you were talking about them selling TPUs.
4:11
Josh: This morning, I saw Morgan Stanley make this announcement. They said about every
4:15
Josh: half a million TPUs Google sells can add about 13 billion in revenue.
4:20
Josh: And Google is planning to sell 12 million of those over the next two years.
4:23
Josh: So it's a significant amount of revenue that Google can expect to come down the pipe.
4:27
Josh: And it's the first time that we're really starting to see a legitimate competitor
4:31
Josh: to the NVIDIA GPU cluster. Now, that's not to say the GPU is done for.
4:37
Josh: There's a lot of competitive advantages to a GPU.
4:39
Josh: I suspect they're not going anywhere, but there is now...
4:42
Josh: Another market force at play. And when we see a market force kind of cutting
4:46
Josh: in, it starts to price cascade and the monopoly slowly starts to fade.
4:50
Josh: I do want to give a brief history lesson, E.J. on the history of Google and
4:55
Josh: their AI program, because what a lot of people don't understand is Google really
4:59
Josh: is the godfather of AI from the beginning of time to now.
5:03
Josh: And they've just had this problem where they haven't been able to actually build
5:06
Josh: products that scale or sell products to users.
5:09
Josh: But they've been doing this since all the way back in 2011.
5:12
Josh: And what we're seeing here is the original paper that a lot of people would
5:15
Josh: conceive to be the first time that a neural network proved that it was work.
5:20
Josh: And they trained a massive unsupervised model on 16,000 CPU cores.
5:25
Josh: This was before GPUs existed on random YouTube frames.
5:29
Josh: They didn't use any labels. They use no supervision. And then one neuron spontaneously
5:33
Josh: learned the concept of a cat.
5:35
Josh: So this seems so stupid. It's like, oh my God, it can recognize a cat.
5:40
Josh: But this was the first time in history a machine was able to identify something
5:44
Josh: without explicitly written instructions.
5:46
Josh: And that moment inside Google set off a light bulb that eventually led to them
5:49
Josh: creating Google Brain, which was enough of a breakthrough for them to start
5:53
Josh: creating AI inside of their in-house system.
5:55
Josh: So EJS, if you remember Google Translate, which has been around seemingly forever,
5:58
Josh: Google Translate is a result of AI.
6:01
Josh: That was an early test implementation of Google Translate. And what that actually
6:05
Josh: enabled is they could suggest to sponsors...
6:08
Josh: Or advertisers on the platform, which companies are more likely to click,
6:12
Josh: which users are more likely to click. And that created the whole AdSense model.
6:15
Josh: It created, oftentimes when you type into Google search, it'll autocomplete for you.
6:18
Josh: These were all very early versions of AI before we even realized what AI was,
6:22
Josh: which led to the invention of the TPU almost nine years ago.
6:25
Josh: And the TPU is this vertically integrated chip that we see today taking over
6:28
Josh: basically the entire world, one company at a time now.
6:32
Josh: So they've been doing this. I mean, a lot of people don't realize the TPU has
6:34
Josh: been around for nine years now that they've been iterating.
6:36
Josh: We're currently up to version seven, which is the Ironwood TPU.
6:40
Josh: And it's just this incredible testament to the fact that Google actually has
6:43
Josh: been doing this for over a decade now, almost 15 years.
6:46
Josh: And we're finally starting to see the fruits of their labor grow and be exposed
6:50
Josh: in public markets. And my God, it's explosive.
6:52
Ejaaz: It is insane that, you know, they've been working on this for over a decade, right?
6:57
Ejaaz: And like that compounded value is really starting to show now because like,
7:00
Ejaaz: I'm guessing like everyone back in the day was just kind of like,
7:02
Ejaaz: what is this machine learning thing?
7:04
Ejaaz: Like, I can't imagine any kind of like a chatbot being beneficial to us
7:07
Ejaaz: and then fast forward to 2022 chat gpc goes viral
7:10
Ejaaz: and suddenly everyone's kind of raving about gpus and google's kind of
7:13
Ejaaz: like quietly smirking and smiling not buying any of nvidia's gpus being like
7:17
Ejaaz: hey we invested in this a decade ago and it's finally paying off which is just
7:20
Ejaaz: kind of uh insane to to think about and like josh like they came up with the
7:24
Ejaaz: transformer as well right which
7:26
Ejaaz: is like 2017 architecture exactly for this alarm so super cool to see.
7:31
Ejaaz: Now, I want to kind of like step aside and kind of frame the narrative for what
7:36
Ejaaz: we're about to discuss right now.
7:38
Ejaaz: We're talking about Google versus NVIDIA, and there's many different ways that
7:41
Ejaaz: we can kind of compare the two, right?
7:43
Ejaaz: The most obvious one is through TPUs versus GPUs that you mentioned.
7:48
Ejaaz: And one of the biggest questions that I think listeners have on their mind,
7:51
Ejaaz: Josh, is like, okay, well, if Google's going to compete with NVIDIA, where's the proof?
7:56
Ejaaz: Like, who are they selling this stuff to?
7:58
Ejaaz: Like, surely they can't be selling to any major competitors,
8:01
Ejaaz: right? Surely they can't be selling to major companies.
8:02
Ejaaz: So can they actually really compete? And I just have one article I want to share
8:07
Ejaaz: with you, one little tweet.
8:08
Ejaaz: You might have heard of Meta or Zuckerberg, who is rumored to be spending tens
8:13
Ejaaz: of billions of dollars in 2026 on Google's CPUs to train their llama or big
8:20
Ejaaz: llama models coming up in the future.
8:22
Ejaaz: Now, of course, you're not too unfamiliar with Meta's kind of progress recently,
8:27
Ejaaz: or rather their spending budget.
8:28
Ejaaz: They've spent, I think, to the effect of $25 billion this year just to hire talent,
8:34
Ejaaz: to train the model we haven't even seen the model uh to
8:36
Ejaaz: begin with so the fact that they're planning on using google's infrastructure
8:40
Ejaaz: uh supposedly to train their models is is
8:43
Ejaaz: no easy feat and so you might be wondering well like why why would they choose
8:46
Ejaaz: tpus over nvidia's gpus that's kind of like the standard kind of framework to
8:50
Ejaaz: to go down well i just want to show you the scaling chart which basically shows
8:54
Ejaaz: that uh the tpu which is google's um infrastructure their gpu versus the GB300,
9:01
Ejaaz: which is NVIDIA's latest GPU,
9:04
Ejaaz: There's a significant cost difference, right?
9:06
Ejaaz: If you were to use Google's TPUs, the IronWord TPU v7, you would save 30 to
9:12
Ejaaz: 50%, depending on the amount of TPUs that you would use,
9:16
Ejaaz: in training your model. Now, when you consider that Meta's CapEx spend for the
9:21
Ejaaz: next year, I believe is going to be something along the lines of $67 to $85 billion.
9:27
Ejaaz: That is a lot of cost saving if you are able to use Google's TPU.
9:33
Ejaaz: So from an economical sense, it makes a hell of a load of sense, right?
9:38
Ejaaz: And then the other thing I was thinking about, Josh, is why would Meta kind
9:41
Ejaaz: of use Google's TPUs for their own systems, right? Aren't they directly competing?
9:45
Ejaaz: Well, there's a secret kind of detail that I learned about this.
9:49
Ejaaz: The way that Google's TPUs are designed makes it really performant for something
9:55
Ejaaz: called recommender systems.
9:57
Ejaaz: Now, a recommender system is the system that is used behind ads,
10:01
Ejaaz: behind social media algorithms.
10:04
Ejaaz: Meta is arguably one of the biggest companies which uses these things.
10:07
Ejaaz: So if they can have a hyper-performant TPU, train the AI model to use on their
10:12
Ejaaz: own social media platform that will make it inevitably better
10:16
Ejaaz: 30 to 50 percent less cost, it seems like a complete no-brainer.
10:21
Ejaaz: And if you add this deal to the Anthropic deal that are purchasing 1 million
10:25
Ejaaz: TPUs from Google, as well as another deal that we're about to talk about,
10:29
Ejaaz: that's just insane. Don't you think that's pretty insane, right?
10:32
Josh: Yeah. And going back to that chart that you just brought up earlier,
10:35
Josh: which shows the cost difference, like if you're a big AI company spending billions on training models,
10:39
Josh: Google is now offering a system that can cost only 25 to 50
10:42
Josh: cents for every dollar you'd spend on NVIDIA's best hardware
10:45
Josh: and that is a huge deal because i mean
10:48
Josh: when it comes down to it cost per compute cost per unit of compute
10:51
Josh: and training is so large when you're at the scale of these companies
10:54
Josh: spending tens of hundreds of billions of dollars that 25 50 of the cost is massive
11:00
Josh: and granted like you said they're not good for everything but if you're a company
11:03
Josh: like meta who's building suggestion algorithms that's particularly good for
11:06
Josh: this is a no-brainer and it seems to me like now the only threshold will be
11:10
Josh: how quick can google actually create these manufacture them spin them up,
11:14
Josh: put them in server racks and get them online so people can start using them.
11:17
Josh: Because this unlocks a whole new use case for AI that we haven't seen in the
11:20
Josh: past that we'll see now because of
11:22
Josh: the lower cost and also the increased efficiency of these Ironwood TPUs.
11:26
Josh: And Google's innovating quick, man. I mean, each one of these TPUs is coming
11:29
Josh: out every single year and each one is significantly better than the last.
11:33
Ejaaz: Every 500k TPUs that Google sell adds 10% to their 2027 Google Cloud Rev and
11:40
Ejaaz: 3% to their at 2027 earnings per share.
11:43
Ejaaz: That is insane. So they don't need to sell. Yeah, $13 billion.
11:47
Ejaaz: They don't need to sell near as much GPUs as NVIDIA sells. They just need to sell a couple hundred K.
11:54
Ejaaz: And if this is just one deal that they're cementing with Meta,
11:57
Ejaaz: can you imagine how much revenue they're just going to churn from this?
12:01
Ejaaz: It was rumored that the Anthropic deal, where they're selling around,
12:05
Ejaaz: I think it's a couple hundred TPUs to them, is going to earn them $50 billion.
12:11
Ejaaz: Next year just to train Anthropics kind of like next forward model.
12:15
Ejaaz: So just kind of insane to see.
12:17
Ejaaz: There is a counter thesis to this deal, which is, you know, I'm going to put
12:22
Ejaaz: my tinfoil hat on here, Josh, which is Meta's kind of doing this so that they
12:27
Ejaaz: can negotiate better terms with NVIDIA
12:29
Ejaaz: or AMD to kind of get like better chip deals saying, hey, look,
12:33
Ejaaz: we'll go with Google unless you guys give us a cheaper kind of route.
12:37
Ejaaz: I think this is kind of like conspiracy theory.
12:39
Ejaaz: I mean, the metrics around Google's GPUs kind of prove themselves.
12:42
Ejaaz: But it's just something to keep in mind. I don't want to get too much into my bold thesis here.
12:46
Josh: I was going through a lot of the deals that we're surfacing today.
12:49
Josh: One of them being with Foxconn and Google, where now Foxconn is responsible
12:52
Josh: for building 1,000 server racks a week.
12:54
Josh: Next year, they're doing 2,000 server racks a week.
12:57
Josh: There was an announcement earlier today where Google is now partnering with
12:59
Josh: AWS, the cloud server, to provide more infrastructure.
13:02
Josh: So we're starting to see, again, more of these deals that are happening around
13:06
Josh: the Google TPU world, which is super fascinating.
13:10
Josh: And then this leads to the final point, which is the head of AI infrastructure
13:13
Josh: in a meeting from a few months ago saying that,
13:16
Josh: Google must double AI compute every six months to meet its demand.
13:19
Josh: So there is no shortage of demand. There is no shortage of infrastructure.
13:23
Josh: There is no shortage of support to get these TPUs out to the world.
13:26
Josh: And what we're going to start to see is how big of an impact this really does
13:29
Josh: have on a company like NVIDIA now that there is someone else in the market.
13:32
Josh: There is a second seller for a company like Meta who wants to build massive AI systems.
13:37
Ejaaz: You could argue one of the most obvious bull signals to purchasing or investing
13:42
Ejaaz: in Google stock was the fact that Berkshire Hathaway bought a three and a half
13:47
Ejaaz: billion dollar stake in Google literally a couple of weeks ago.
13:51
Ejaaz: And then, funnily enough, the leaked information around them selling TPUs to
13:56
Ejaaz: Meta and them striking this deal with NATO kind of like surfaced, right?
14:00
Ejaaz: So there's a lot of momentum behind Google right now.
14:03
Ejaaz: There's a lot of big, valuable investors and kind of infrastructure providers
14:07
Ejaaz: getting behind the Google train right The momentum is palpable, to say the least, right?
14:12
Ejaaz: And this NATO deal is another example of it, right? Like, we're going from the
14:16
Ejaaz: hyperscaler kind of consumer level to the government level as well.
14:19
Ejaaz: So all types of organizations are treating this with a very high importance
14:23
Ejaaz: that Google is going to play an inevitably big role here.
14:26
Ejaaz: So then that begs the question, well, what's NVIDIA going to do about this?
14:30
Ejaaz: Are they just going to continue losing hundreds of billions of dollars in their
14:33
Ejaaz: market cap? Or are they going to strike back?
14:36
Ejaaz: And there's two frames of thought about this, Josh. Number one is,
14:40
Ejaaz: so NVIDIA's next generation of GPUs is going to be around the Rubin architecture.
14:46
Ejaaz: It's called Rubin, right?
14:48
Ejaaz: They introduced a new spec, Josh, after Google's TPUs, their latest TPUs got
14:54
Ejaaz: released, which upped a lot of the watts or compute performance for the Rubin architecture.
15:00
Ejaaz: Now, some might say this is just coincidental, but some might say this is a
15:04
Ejaaz: general reaction to the fact that Google just has a higher performance TPU versus theirs.
15:10
Ejaaz: And so they needed to kind of like up the metrics of their next generation if
15:13
Ejaaz: they wanted to compete and appear attractive to their competitors or to their customers themselves.
15:19
Ejaaz: But then there's also the argument where it's just kind of like NVIDIA and Google
15:22
Ejaaz: are kind of playing in kind of like different ballpark and they already know
15:26
Ejaaz: this. They're playing different games.
15:28
Ejaaz: The argument here in this tweet being that Google's TPUs are great,
15:32
Ejaaz: but they're only for very specific niche use cases. If you have, you mentioned earlier,
15:37
Ejaaz: the recommendation or search algorithm, then, you know, these ASICs are going to be really good.
15:41
Ejaaz: The benefit of NVIDIA's GPUs is that they're highly generalizable.
15:45
Ejaaz: So if you wanted to train a model in a different way or test out a new method
15:49
Ejaaz: to kind of like inference or train your model, GPUs are by far the best architecture
15:54
Ejaaz: or the best infrastructure to use.
15:56
Ejaaz: So you could argue that Google, that, sorry, NVIDIA is sitting pretty comfy.
16:00
Ejaaz: And Jensen went on a show or an interview this week, basically saying that he's
16:05
Ejaaz: not worried about Google. Obviously, you expect him to say that,
16:08
Ejaaz: but mainly for the fact that this is a positive sum game.
16:12
Ejaaz: You know, Jevon's paradox, if you create more GPUs, it's not going to be a fact
16:16
Ejaaz: that you have oversupply. There's just going to be increased demand for compute.
16:19
Ejaaz: And I think Jensen knows this, and that's why he's just kind of running forward.
16:22
Ejaaz: The fact of the matter is, there isn't enough NVIDIA GPUs to supply the customers,
16:27
Ejaaz: even if you wanted to, right?
16:28
Ejaaz: He needs to ramp up infrastructure production. That's why he's been visiting
16:32
Ejaaz: TSMC for the last couple of weeks. So I think he knows this and I don't think
16:36
Ejaaz: he's too worried, but he's definitely sweating a little. I don't know if you think the same.
16:40
Josh: Yeah, no, I mean, I'm sure it sucks. It's like you were just running the show
16:44
Josh: and now suddenly there's someone else who has a good product.
16:47
Josh: It's not to say that it's going to harm the company too much.
16:50
Josh: And I think for anyone who's listening, if you take away one thing from this
16:53
Josh: episode, it's that both of these companies are going to succeed wildly.
16:56
Josh: And there is going to be a shortage of supply for compute for a very long time.
17:01
Josh: If you believe that AI is as impressive and as important as a technology as
17:05
Josh: it really is, then you also have to believe that all of the compute around us
17:09
Josh: must be replaced by it and must have it embedded inside of it.
17:12
Josh: In order to do so, you need to shift the entire technological infrastructure
17:15
Josh: of all the hardware that exists over to infrastructure that supports AI.
17:19
Josh: Everything. And we are just a fraction of a percentage through that transition.
17:24
Josh: So as a result, there could be many more Googles, many more NVIDIAs,
17:28
Josh: and there would still be a shortage.
17:29
Josh: Now, the question becomes, is there a short-term bubble? Are we overspending?
17:33
Josh: That is to be determined, but this is a good type of bubble.
17:37
Josh: This is one that even if it does explode, we are left with unbelievable technology
17:41
Josh: across the board and a scaling infrastructure that will continue to be able
17:44
Josh: to support this new type of technology that's permeating throughout society.
17:47
Josh: So is this a good thing yes is nvidia going
17:50
Josh: to suffer maybe sure the headlines suck it's like okay
17:53
Josh: we're not the coolest person in the world now there's someone else who's like also a
17:56
Josh: cool kid but they're still going to continue to produce
17:59
Josh: the best products in the world and i just to the point that you made
18:01
Josh: earlier they're just different types of chips like um a
18:05
Josh: gpu is very different than a tpu and a
18:08
Josh: lot of people also need to understand that the whole
18:11
Josh: world isn't actually training ai there's still a lot of other things that are
18:13
Josh: happening like graphics or simulation or financial technology
18:17
Josh: scientific research tpus just don't do
18:20
Josh: that and gpus do so there's there's a lot more
18:22
Josh: going on to the story there's a lot more gpus being sold there's a lot of tpus
18:25
Josh: there's enough for everybody and then there's still not enough for everybody
18:29
Josh: so i think in the long run like this is just great for both companies this is
18:33
Josh: positive some uh there's a lot of excitement around this rightfully so because
18:37
Josh: i think it's it's great there's another person stepping in but it's not the
18:39
Josh: end of of anyone it's just the beginning for for so many of these companies still i.
18:43
Ejaaz: Mean like don't take your and my opinions either, right? Why don't you just
18:47
Ejaaz: listen to one of the smartest men or smartest businessmen? Oh, we got Elon.
18:51
Ejaaz: Yeah, we got Elon. And he was asked this question in this interview clip that I'm about to show.
18:56
Ejaaz: This was released yesterday where he was asked, Elon, if you had to invest in
19:00
Ejaaz: any AI companies today and hold it for a decade, what would you buy?
19:05
Ejaaz: And you think, you know, Elon would shill his own companies.
19:07
Ejaaz: He didn't. He shilled two companies, Josh, Google and NVIDIA. Let me show you a clip.
19:12
Ejaaz: I think, you know, Google is going are gonna be pretty valuable in the future.
19:16
Ejaaz: They've laid the groundwork for,
19:20
Ejaaz: an immense amount of value creation from an AI standpoint.
19:27
Ejaaz: NVIDIA is obvious at this point.
19:31
Ejaaz: I mean, there's an argument that companies that do AI and robotics and maybe
19:37
Ejaaz: space flight are gonna be overwhelmingly all the value, almost all the value.
19:46
Ejaaz: So the output of goods and services from AI and robotics is so high that it
19:51
Ejaaz: will dwarf everything else.
19:52
Ejaaz: And so, you know, you hear it there from Selfway is basically describing NVIDIA
19:58
Ejaaz: as a sort of toll collector because you kind of like need to basically pay the
20:03
Ejaaz: toll man for his GPUs to get access to the intelligence that you're trying to build.
20:07
Ejaaz: And then Google's mode is kind of similar but quite different in the sense that
20:11
Ejaaz: they create the GPUs, their own TPUs, but they also like kind of own dominance
20:18
Ejaaz: across the entire AI stack, right, Josh?
20:21
Ejaaz: And just to kind of maybe like round things up, I was looking at these crazy
20:25
Ejaaz: charts from the Financial Times this week,
20:28
Ejaaz: which basically showed that Google's Gemini model has now almost caught up in
20:34
Ejaaz: the number of users or monthly downloads that ChatGPT has,
20:38
Ejaaz: which is just insane. That's the chart that we see here on the left.
20:40
Ejaaz: And then on the right, which I found the most interesting, is the amount of
20:44
Ejaaz: time that each Gemini user is spending on the app using the Gemini model has now beaten ChatGPT.
20:51
Ejaaz: This kind of blew my mind because I was like, surely everyone's still using
20:54
Ejaaz: ChatGPT because people tend to use ChatGPT for their own personalized things.
20:58
Ejaaz: They kind of like confer it, therapy, ask about personal stuff.
21:01
Ejaaz: That probably spends a lot more time, but it seems like the productivity aspect
21:05
Ejaaz: that people are getting from task orientation-based AI stuff using Gemini seems to be extending.
21:12
Ejaaz: And that just kind of like shows I've heard anecdotes from friends you
21:14
Ejaaz: and I were talking to a team member just before recording this and he
21:17
Ejaaz: was like yeah I was talking to a bunch of my friends and they've fully switched
21:20
Ejaaz: to the Gemini app so I think we're going to continue seeing this trend of Google
21:24
Ejaaz: gaining the advantage not because of their infrastructure mode but because they
21:28
Ejaaz: like own all the popular apps that anyone and everyone wants to use and so all
21:32
Ejaaz: they have to do is plug in their AI model with whatever app Gmail,
21:36
Ejaaz: Maps whatever you might kind of think of and suddenly you have a really productive
21:40
Ejaaz: useful app that you and I want to use every day.
21:42
Ejaaz: Josh, you mentioned that you want to use down to banana or you're using down to banana, right?
21:46
Josh: Yeah, there's a there's an important shift that I found that has happened recently
21:51
Josh: that hasn't happened before, which is I have Gemini on my home screen on my phone.
21:54
Josh: And that to me is very high signal because I've been resistant of it because
21:58
Josh: it just hasn't been good.
22:00
Josh: And while it's still not great, I think the ChatGPT app is engineered far better than Gemini.
22:05
Josh: It is good enough to make me want to use it. So I went from not being a user
22:09
Josh: like I would really I'd use Gemini 3 Pro on desktop whenever I had a hard question,
22:13
Josh: but I wasn't reaching for it.
22:15
Josh: And Nano Banana Pro, it really was the killer use case that had me like,
22:20
Josh: oh my God, I need this quickly accessible and in my pocket at all times because
22:24
Josh: it is so far superior to any other product in the space like that.
22:28
Josh: And I think as Google starts to roll out these products, as this is something
22:32
Josh: that we talk about a lot with OpenAI, where they're really good at creating
22:35
Josh: innovation and wrapping it in a product and selling it.
22:38
Josh: As Google gets better at doing that, I really, I strongly suspect Gemini will
22:42
Josh: continue this trend of taking over broader and broader people.
22:44
Josh: Because when you think about how many monthly active users google
22:47
Josh: has it's like gigantic they're one of the few in the
22:50
Josh: world that actually has more than than chat gpt and open ai
22:53
Josh: and if they can convert all of these services to
22:56
Josh: pack ai into it into one coherent service
22:59
Josh: and package that's incredible we talk about a lot of times like with meta for
23:03
Josh: example the they had their awesome hardware product but no one really wanted
23:07
Josh: to use it because the ecosystem sucked well google has like the best ecosystem
23:10
Josh: ever it's funny on my even on my iphone i have an iphone hardware with software
23:14
Josh: from Google because their software is so superior.
23:17
Josh: So as they're able to integrate these top tier models into all the products,
23:20
Josh: this is a serious shift. And I'm very bullish on Google.
23:25
Ejaaz: Yeah, I don't think we're going to see this trend reverse. We already know that
23:28
Ejaaz: Apple is going to be using a Google Gemini-based model in their phone for Siri.
23:34
Ejaaz: That's right, yeah. And we know that we're going to start seeing a lot of Gemini-based
23:38
Ejaaz: apps just kind of like appear in our regular day-to-day.
23:41
Ejaaz: One other final point is like when you compare like Google versus OpenAI,
23:46
Ejaaz: remember OpenAI still isn't profitable.
23:48
Ejaaz: Google is massively, massively profitable, so they don't need to turn on ads.
23:52
Ejaaz: They don't need to kind of like like demean the
23:56
Ejaaz: user experience in any way they could just keep giving you the stuff for free
23:58
Ejaaz: and gaining millions and millions more monthly active
24:01
Ejaaz: users whereas open air at some point is going to turn on ads and when they turn
24:05
Ejaaz: on ads it's going to be an inferior performance it's going to be a inferior
24:08
Ejaaz: product to some extent and that might shift to more people using google gemini's
24:13
Ejaaz: products and google knows this so they're willing to just kind of sit back they
24:16
Ejaaz: own kind of every infrastructure layer and they're just going to see how things play out.
24:20
Ejaaz: But I think that is it for today's episode.
24:24
Ejaaz: Super exciting to kind of like see where Google and NVIDIA ultimately end up.
24:28
Ejaaz: In my opinion, I think that both companies, to your point earlier,
24:32
Ejaaz: Josh, are going to do extremely, extremely well.
24:34
Ejaaz: It's a positive sum game. And the fact of the matter is there is not enough compute.
24:37
Ejaaz: There's not enough energy to feed the compute that both of these companies are pushing out.
24:42
Ejaaz: So I think we're just going to see both these companies grow into two of the
24:45
Ejaaz: largest and most valuable companies in the world. Now, I need to take a quick
24:50
Ejaaz: victory lap for all listeners here who aren't subscribed and who haven't rated our show just yet.
24:56
Ejaaz: We released an episode about the bull case for Google.
24:59
Ejaaz: What was it? Two months ago now. And a lot of it has now played out right now.
25:03
Ejaaz: If you'd invested in Google back then, you would have participated in the hundreds
25:06
Ejaaz: of billions of dollars that their market cap is up relative to NVIDIA right now.
25:10
Ejaaz: Now, I'm not going to say that we triggered it. Maybe we did.
25:12
Ejaaz: Maybe we didn't. But if you want to hear more bull cases like this or more alpha
25:16
Ejaaz: in advance, subscribe to us, rate us, give us a thumbs up, give us feedback. We love it.
25:22
Ejaaz: And we will hear more from you or we will, you will hear more from us rather
25:25
Ejaaz: on the next episode. See you then.