Latest / In the Loop AI / Why the Markets Ignore Amazon: AWS, Robotics, and AI Chips (Trainium 3)

Transcript

Why the Markets Ignore Amazon: AWS, Robotics, and AI Chips (Trainium 3) · In the Loop AI

0:00
Ejaaz: The most contrarian AI bet of 2026 is the same company that you buy toilet paper from.
0:06
Ejaaz: Many people will mistake Amazon as just an e-commerce company,
0:09
Ejaaz: but they're secretly a frontier AI lab.
0:11
Ejaaz: For example, did you know that they manufacture their own AI chips that are
0:15
Ejaaz: as good as NVIDIA's GPUs, but 50% cheaper, which means that companies like Anthropic
0:19
Ejaaz: and OpenAI, which have signed deals with Amazon, save tens of billions of dollars
0:24
Ejaaz: training their frontier models. But that's not all.
0:26
Ejaaz: Amazon's compute platform, which accounts for 50% of their operating profits, AWS.
0:31
Ejaaz: They're running the same playbook for AI now, serving AI cloud to any company
0:35
Ejaaz: that wants to inference or train their own models.
0:38
Ejaaz: And finally, Amazon has a secret up their sleeve which no one's talking about.
0:42
Ejaaz: Robots. For over a decade, Amazon robots have helped them scale manufacturing
0:45
Ejaaz: and factory automation to the tune of tens of billions of dollars,
0:49
Ejaaz: which make them the perfect company to design and build the robots of tomorrow.
0:53
Ejaaz: Amazon is easily a $5 trillion company hidden as a shopping platform.
0:59
Josh: It's funny you mentioned the shopping platform part because when everyone,
1:02
Josh: myself included, thinks of Amazon, they very clearly think of shopping platform.
1:06
Josh: And for the right reason, I have some pretty unbelievable stats.
1:09
Josh: So last year, Amazon shipped 6.3 billion packages, which is 17 and a quarter
1:14
Josh: million per day, which is 200 per second that gets shipped.
1:18
Josh: So that equates to about 30% of all the US parcels were one company.
1:23
Josh: And that means one out of every three and a half packages was shipped by amazon
1:25
Josh: so in terms of the amount of
1:27
Josh: mass and atoms that they're moving throughout the universe throughout the world
1:32
Josh: like as it relates to the other mag 7 companies they're moving the most amount
1:35
Josh: of stuff just raw stuff and you have to imagine that once you start to apply
1:38
Josh: ai to this in terms of efficiency gains and improvements like you're mentioning with robotics,
1:43
Josh: The amount of stuff can really be optimized quite a bit and have a meaningful impact on the business.
1:48
Josh: But what we're seeing with the stock price that's on the screen here tells a
1:50
Josh: very different story, which is basically flat in a year where every company
1:55
Josh: in the world that was building an AI at Amazon's size went up an outrageous amount.
2:00
Ejaaz: I think the story behind this is Amazon's just very misunderstood.
2:04
Ejaaz: Like, to your point, in the last year, it's gone up 1.4%, which is just insane.
2:10
Ejaaz: It doesn't even beat inflation. doesn't beat inflation, which is just insane.
2:14
Ejaaz: And listen, there are theories as to why this might be, but we're here to tell
2:17
Ejaaz: you the story why Amazon is basically the biggest AI beach ball underwater that
2:22
Ejaaz: is about to pop up in 2026.
2:24
Ejaaz: And some of my foundational thesis behind this, Josh, is that the stuff that
2:29
Ejaaz: they focus on is really unsexy.
2:32
Ejaaz: It's operational. And they're about to do the same for AI.
2:35
Ejaaz: Like, think about it, like sorting, fulfilling packages, delivering it to people
2:38
Ejaaz: isn't really a sexy thing.
2:40
Ejaaz: And then if you talk about compute and serving compute to different companies,
2:44
Ejaaz: again, I don't really care about it. But
2:46
Ejaaz: What most people don't realize is that the top software companies in the world run on AWS.
2:52
Ejaaz: That's why when AWS servers went out a few weeks ago, the world couldn't function.
2:56
Ejaaz: I couldn't use X. I couldn't scroll my favorite social media platforms,
2:59
Ejaaz: Josh, because it ran on AWS.
3:01
Ejaaz: So most people don't realize this. And they're about to do the same for AI.
3:04
Josh: Yeah. And one last thing I want to mention on this chart, since we have it on
3:06
Josh: the screen, is that little PE ratio number, the price to earnings ratio,
3:10
Josh: it's down to 35 now. And for reference, back in 2018, Amazon was trading at
3:15
Josh: over a 200 times price to earnings ratio.
3:18
Josh: So the multiples have really compressed a lot. But this is a much more mature
3:22
Josh: company, which actually has a vector of growth, which is adding AI to the mix
3:27
Josh: to increase this efficiency.
3:28
Josh: So EJS, you wrote all about this. It's in the newsletter that we published.
3:31
Josh: But you actually created a proper article going through the bull case for Amazon.
3:35
Josh: And I think we're going to spend a good amount of time kind of going through
3:38
Josh: the outline that you framed here for why you believe it to be,
3:42
Josh: I mean, as the title says, the most asymmetric bet of 2026.
3:46
Ejaaz: And for those of you who are wondering, hey, what is this article?
3:49
Ejaaz: What is this newsletter? It's the Limitless Newsletter.
3:51
Ejaaz: And if you were subscribed to it, you would have seen this article about a week and a half ago.
3:55
Ejaaz: So again, if you want the best alpha in AI, you have to subscribe to our newsletter.
4:00
Ejaaz: But yes, Josh, one of the first things or one of my first arguments as to why
4:04
Ejaaz: Amazon and there's an asymmetric bet here, is this thing called AI chips.
4:09
Ejaaz: You heard of them, Josh? You know, these little GPUs thing?
4:12
Josh: You come across them? I might have heard of GPU, a little TPU, a whole bunch of U's.
4:16
Ejaaz: All right, all right, okay. So you've heard of NVIDIA, it sounds like.
4:19
Ejaaz: It sounds like you've heard of Google as well, as you know we're very bullish on Google here.
4:23
Ejaaz: But what most people don't realize is Amazon created their own chips.
4:27
Ejaaz: And it wasn't like they did this yesterday.
4:29
Ejaaz: They did this 10 years ago, Josh, when they acquired a company called Annapurna Labs in 2015,
4:35
Ejaaz: which marks the start of their training and build of AI chips to help them with
4:41
Ejaaz: machine learning inference, stuff like they was figuring out way back when,
4:45
Ejaaz: when it came to recommend systems on their shopping platform.
4:48
Ejaaz: Now, if you fast forward today, in the last few weeks, Josh,
4:51
Ejaaz: they released this chip called Tranium 3. It's part of their Tranium series
4:55
Ejaaz: of chips, which are used to build and scale large LLMs or rather large AI models.
5:02
Ejaaz: And Tranium 3, obviously, you know, the first question that pops into your head
5:05
Ejaaz: is like, well, how does this compare to NVIDIA?
5:07
Ejaaz: Well, let me give you a few stats to kind of whet your appetite,
5:10
Ejaaz: Josh, which is it is four and a half times more powerful than Tranium 2,
5:15
Ejaaz: but it's also four times cheaper than Tranium 2.
5:18
Ejaaz: So if you net both of those together, you get like a 10X kind of chip here, right?
5:23
Ejaaz: But then it can also store five times more data than the previous chip that they had.
5:28
Ejaaz: So all of these combined together gives you about 80 to 90% of the same performance
5:33
Ejaaz: as NVIDIA's latest GPU, Blackwell.
5:37
Ejaaz: So I'm not just talking about NVIDIA's second, third, or fourth generation.
5:40
Ejaaz: I'm talking about the latest generation that they have right now. So automatically,
5:45
Ejaaz: Josh, if I was a Frontier AI lab that is spending hundreds of billions of dollars
5:48
Ejaaz: each year on NVIDIA's GPUs and you suddenly give me an option to spend 50% of that bill. So I save...
5:56
Ejaaz: 50 billion dollars why wouldn't i use this chip
5:59
Josh: Yeah it sounds like a pretty good deal and i think we should
6:02
Josh: probably start by outlining what exactly this chip is because there's a difference between
6:05
Josh: nvidia's gpu google's tpu and then training which is something totally different
6:11
Josh: and they fit if i'm not mistaken into two kind of basic categories so gpus which
6:15
Josh: is nvidia's blackwell system that is a graphical processing unit it's kind of
6:19
Josh: a general purpose supercomputer you can use it for a lot of different things,
6:23
Josh: graphic being one of them,
6:24
Josh: but also the matrix math required to do AI training as another.
6:28
Josh: With TPUs like Google's making and these Amazon chips through Tranium,
6:32
Josh: they're more focused on specific types of math.
6:35
Josh: They are not general purpose. They're narrowly focused on AI training and AI
6:39
Josh: inference. And that's where you get a lot of the efficiency improvements.
6:42
Josh: So like we're seeing on the screen, 4.4 times higher improvement,
6:45
Josh: four times more memory bandwidth, like the specs are insane,
6:48
Josh: but they cannot be used for everything.
6:50
Josh: So it's a very specific type of customer that wants these types of chips.
6:53
Ejaaz: Exactly. So if you were to kind of like compare the two NVIDIA GPUs and Amazon's training chips,
6:59
Ejaaz: NVIDIA's GPUs can be used for a lot of broad use cases when it comes to training models.
7:04
Ejaaz: Like there are several different types of ways to train an AI model.
7:07
Ejaaz: If you weren't sure which one to use, you would probably use an NVIDIA GPU because
7:11
Ejaaz: it would just be consistent across all of those things.
7:13
Ejaaz: But for Amazon's training chips, you need to specifically know exactly how you're
7:17
Ejaaz: going to train a specific model and then it ends up being cheaper.
7:21
Ejaaz: So it's for like a highly specialized type of AI lab that wants to train their own model.
7:26
Ejaaz: And what's interesting about this, Josh, is it's not just kind of the specific
7:31
Ejaaz: architecture of how this chip is designed, it's also very much the cost.
7:35
Ejaaz: I have a table pulled up here, and if you can take a direct look at it,
7:40
Ejaaz: which compares Tranium 3 to Google's TPUs and NVIDIA's latest chip,
7:45
Ejaaz: the Blackwell, and it is half the cost of NVIDIA's Blackwell.
7:49
Ejaaz: That is excluding NVIDIA's margin that they add on top of this, Josh.
7:53
Ejaaz: So if you look at this, the average cost of an NVIDIA chip to manufacture is
7:58
Ejaaz: around $6,500 to $7,000, but they end up selling it for $40,000.
8:03
Ejaaz: That's the price that Amazon has to pay to buy these chips. That's the cost
8:07
Ejaaz: that OpenAI has to pay to buy these chips.
8:09
Ejaaz: So I can think of two things here. If Amazon has an almost as good chip,
8:14
Ejaaz: then it's going to largely kind of be a more attractive chip to buy for Frontier
8:18
Ejaaz: AI Labs, but it's also going to cut into NVIDIA's margins drastically.
8:22
Ejaaz: So you start seeing Amazon and Google TPUs being able to eat into the market
8:26
Ejaaz: monopoly that NVIDIA has.
8:28
Josh: Okay, so interesting. I want you to kind of help me understand this,
8:30
Josh: because a lot of this has been new information to me. I got intro to this through the article, and,
8:35
Josh: What I understand is that Amazon's chips, not TPUs, are better on a per watt
8:42
Josh: basis. They're more efficient. They're more cost effective.
8:44
Josh: But I guess my question is, if Amazon were to start making these available to
8:48
Josh: the public tomorrow, would they outsell Blackwell?
8:51
Josh: Would people be more interested in these or would it just be a very specific audience?
8:55
Ejaaz: No. OK, so there's two things I want to bring up with you that kind of like
8:59
Ejaaz: sway people's decisions when they're choosing between the chips.
9:02
Ejaaz: Number one, an NVIDIA Blackwell chip on its own and also an Amazon training
9:07
Ejaaz: chip on its own isn't useful, Josh.
9:09
Ejaaz: You need to stack them into this thing called clusters.
9:12
Ejaaz: Now, for an NVIDIA chip, you only have to put 72 of these chips together to
9:17
Ejaaz: get the same performance per watt.
9:19
Ejaaz: But with an Amazon chip, you need to stack 144 of them together.
9:23
Ejaaz: So it's a higher volume thing.
9:25
Josh: Okay, so it's about double the amount of chips per little cluster that we have
9:29
Josh: in a rack. and do those ships are they more expensive too or are they like cost
9:34
Josh: better because it seems like that's a lot more complexity for like 80% of the efficiency.
9:39
Ejaaz: They're cheaper. So on this table right now, they're half the cost.
9:42
Ejaaz: That's where the $3,000, the $3,500 come from.
9:46
Ejaaz: So you may have more chips, which may take up more volume in your data center,
9:50
Ejaaz: but they are cheaper to run at a cost kind of inference, right?
9:54
Ejaaz: The other thing, Josh, is you might have noticed I said they're not as good
9:58
Ejaaz: as NVIDIA GPUs. They're around 80% to 90% as good.
10:02
Ejaaz: And the reason why there's that difference is because of NVIDIA's software moat.
10:07
Ejaaz: So this is something called CUDA or Compute Unified Device Architecture.
10:12
Ejaaz: So basically, if you have the chips, that doesn't solve your entire problem.
10:16
Ejaaz: You need software to be able to make these chips run really coherently together in their clusters.
10:21
Ejaaz: NVIDIA has the stronghold of this, Josh. Any AI lab that is using NVIDIA GPUs,
10:26
Ejaaz: which is the majority of AI labs, run the CUDA software system.
10:30
Ejaaz: And typically, this has locked customers into using NVIDIA.
10:34
Ejaaz: So let's say they're interested in using Google's TPUs, they may not necessarily
10:39
Ejaaz: still want to jump to use Google's TPUs because the software isn't the same.
10:43
Ejaaz: They would have to rewrite their entire code base.
10:46
Ejaaz: Amazon saw that and thought, hmm, I bet I could make this easier.
10:50
Ejaaz: And so they released this thing called a Neuron SDK, which now allows you to
10:54
Ejaaz: copy and paste your code base from your NVIDIA instance into your Amazon GPU
10:59
Ejaaz: or your Amazon Tranium chip in a few clicks.
11:02
Josh: Maybe this is a good time to point out the fact that
11:05
Josh: this is a really big deal for amazon even if they don't sell
11:08
Josh: these chips anywhere um because aws is such a huge story we were looking up
11:13
Josh: before this episode was recorded how much of the internet is run by aws and
11:17
Josh: it's about a third and another fascinating thing that you kind of pointed out
11:21
Josh: in the intro but um as of last quarter i believe um aws
11:26
Josh: revenue was something like less than 20% of the total company,
11:31
Josh: but it accounts for close to 70% of the actual profit share.
11:36
Josh: I think it's like 66% as of the third quarter.
11:38
Josh: So this small part of Amazon's business accounts for over half of the total
11:43
Josh: profit every single quarter that comes in.
11:45
Josh: And with these plans to scale by using these new shifts, by using this gigantic
11:49
Josh: data center that we're seeing on the screen that we're going to get into,
11:52
Josh: there is a very clear trajectory for amplifying the one part of the business
11:55
Josh: that actually matters the most for their top line.
11:58
Josh: And we were talking earlier, it's kind of similar to what Costco does with their
12:01
Josh: membership, where Costco's business operates on very thin profit margins.
12:05
Josh: They really don't make a lot, if any money, on the actual goods sold.
12:09
Josh: A lot of that revenue comes from that membership, from the lock-in.
12:12
Josh: And Amazon, what they have with AWS, is this unbelievably profitable engine
12:17
Josh: that is becoming much more efficient using these new Tranium chips.
12:20
Josh: But here on screen, and we have this really crazy looking data center.
12:23
Josh: So maybe you could tell us more about what's going on here.
12:26
Ejaaz: Yeah, I mean, listen, we're not strangers to crazy data center setups.
12:30
Ejaaz: We've spoken about Elon Musk's Colossus 2. We've spoken about Meta's super data
12:36
Ejaaz: center that they're building.
12:36
Ejaaz: Basically, all the top companies are spending tens of billions of dollars.
12:39
Ejaaz: And Amazon is no stranger to this either. They've invested $11 billion and they're
12:43
Ejaaz: going to invest another $20 billion next year to build out their Indiana data
12:47
Ejaaz: center campus to create around 2.2 gigawatts of compute.
12:51
Ejaaz: That's equitable to about 1 million homes worth of energy by the end of next
12:56
Ejaaz: year, which is just an insane goal to kind of like figure out.
12:59
Ejaaz: And to your point, Josh, like if they're able to pull off what they did with
13:03
Ejaaz: AWS for AI compute specifically,
13:07
Ejaaz: There is kind of like no reason for me to think at least why they wouldn't eat
13:11
Ejaaz: into all the other NeoCloud's valuations. We've spoken about, what's his name?
13:15
Ejaaz: Leopold investing in this company called CoreWeave, which was one of the top NeoCloud providers.
13:20
Ejaaz: And he invested to the tune of like, I think it was like 350 billion,
13:24
Ejaaz: almost $400 billion into this company.
13:26
Ejaaz: If Amazon just eats, like switches this on at the end of next year,
13:30
Ejaaz: they like companies who are already running on AWS will just switch to the AI
13:35
Ejaaz: version of AWS. it kind of doesn't make sense for them to kind of flip here.
13:39
Ejaaz: And it's why I think they have such a kind of sticky mode. They help with all the unsexy stuff, Josh.
13:45
Ejaaz: I don't know if you saw this rollout of something called AI Factory.
13:49
Ejaaz: Did you catch this by any chance?
13:51
Josh: No. AI Factory sounds interesting. We like factories. Okay.
13:54
Ejaaz: So let me lay this out for you. A lot of enterprises and even governments want
13:59
Ejaaz: to create their own versions of AI models, but they run into two main issues.
14:04
Ejaaz: Number one, they don't know who to buy the chips from. Should they go to NVIDIA?
14:08
Ejaaz: Should they go to Google? Should they go to Amazon? I have no idea.
14:11
Ejaaz: And then two, they don't want to set it up themselves, right?
14:14
Ejaaz: But they also don't want to rent compute directly through AWS. Why?
14:19
Ejaaz: Because, you know, they have some private information. They don't want to leak information.
14:23
Ejaaz: They want Amazon to own the information. They want to hold it on their own private
14:26
Ejaaz: service. So Amazon looked at this and said,
14:28
Ejaaz: Okay, we're rolling out a service called AWS Factory, and here's what we offer.
14:32
Ejaaz: If you want NVIDIA GPUs, we got you. If you want Amazon Tranium GPUs or chips,
14:38
Ejaaz: which are, by the way, 50% cheaper, we also got you.
14:40
Ejaaz: If you want a hybrid or mix of both of these things, we've got you.
14:43
Ejaaz: And what they do is they build the server racks, Josh, for them.
14:47
Ejaaz: They basically build out a data center for them, and this benefits them in so
14:51
Ejaaz: many ways because they have the software and they have the hardware,
14:54
Ejaaz: and they have 50% less the cost if they run AWS or Amazon Tranium chips.
14:59
Ejaaz: So it's like an all-in-one package where they have lower latency,
15:02
Ejaaz: they save on costs, and they have premium frontier intelligence. Pretty cool.
15:06
Josh: But there's an important distinction
15:07
Josh: there where they're not building the factories for the company.
15:11
Josh: They're building their own infrastructure that they're lending to the company, right?
15:14
Josh: So if we're considering a company like Oracle, whose job is to build data centers
15:19
Josh: for companies, you can think about the Project Stargate project with OpenAI,
15:23
Josh: they are building actual infrastructure that OpenAI will own.
15:26
Josh: What Amazon is doing is they are building the infrastructure for you,
15:30
Josh: but they're just leasing it to you. They still own the underlying core infra.
15:35
Josh: So these other companies are essentially bankrolling the build-out of these
15:38
Josh: factories, but doing so in a way that doesn't incur a lot of debt.
15:42
Josh: Like they already have customers here.
15:43
Josh: And the switching costs, I was looking because I was curious as it relates to
15:47
Josh: these big cloud providers. Like AWS is 30%.
15:51
Josh: Azure and Google Cloud are another 33%. And then the rest is just kind of this
15:55
Josh: mix of things. So a third of the companies are already, they trust the security
16:00
Josh: of Amazon. They already use Amazon.
16:02
Josh: They have their whole custom software stack built on Amazon.
16:04
Josh: And in the world that they can just extend this to integrate AI into their offering,
16:09
Josh: well, now 33% of the internet, they just get to direct AI offering built in.
16:13
Josh: And that's like a pretty powerful thing.
16:15
Ejaaz: Yeah, Josh, you know which other little small company this reminds me of?
16:19
Josh: Google. Who's that?
16:21
Ejaaz: Google already had their roots sewn into so many different products.
16:25
Ejaaz: Basically, anyone who's ever graced the internet has come across a Google product,
16:29
Ejaaz: whether it's Gmail, Android, Play Store, whether it's Google Search itself.
16:34
Ejaaz: Amazon is the same bedrock for this, except it's just, hey, we've got the compute
16:39
Ejaaz: that you're going to run your number one website or internet product on.
16:43
Ejaaz: And they're converting the same users, as you said, into AI users.
16:47
Ejaaz: I just think it's a complete no-brainer.
16:49
Ejaaz: And yeah, to kind of build on your analogy, they're not building the entire data center for you.
16:54
Ejaaz: You're going to still have to buy the warehouse and supply it with energy and
16:57
Ejaaz: electrical grid, but they've got everything else for you.
16:59
Ejaaz: You don't have to have the upfront AI capex costs that, for example,
17:03
Ejaaz: OpenAI has when they're committing $1.4 trillion over the next five years to buy these GPUs.
17:08
Josh: Okay, so we've covered the chips, we've covered the cloud, Now we have the physical
17:11
Josh: infrastructure, the world of atoms. This gets to our 6.3 billion package statistic
17:16
Josh: where Amazon, of all the Mag7 companies, they just move the most amount of stuff through the universe.
17:22
Josh: And through that, they benefit, they stand to benefit a lot from automation,
17:26
Josh: particularly as it relates to robots.
17:28
Josh: So here we have news that they have three quarters of a million robots already deployed.
17:32
Josh: EJS, what are their plans going forward that are part of your bull case as it
17:35
Josh: relates to kind of robotics and warehouses?
17:37
Ejaaz: Okay, what I'm about to say is going to sound controversial.
17:40
Ejaaz: Because I have said that Tesla is the number one robot company so many times, but I was wrong.
17:47
Ejaaz: It's going to be- Oh, wow, that is a hot take. It's going to be Amazon.
17:50
Ejaaz: They already have almost a million of these things out there. And listen, listen.
17:54
Ejaaz: Okay. They might not be the humanoid robots that grace your home,
17:58
Ejaaz: that help you with the laundry.
18:00
Ejaaz: I have laundry in my thing that needs to get sorted right now.
18:03
Ejaaz: Wish I had one of those, right?
18:04
Ejaaz: It's not going to be a robot car that takes me wherever I need to from A to
18:09
Ejaaz: B with minimal accidental type stuff.
18:13
Ejaaz: But they're going to be the robots that scale very important things,
18:16
Ejaaz: such as manual labor, Josh, which is like, you know, a multi-trillion dollar
18:20
Ejaaz: industry or global sector as itself.
18:23
Ejaaz: If you are able to cut down costs by 50 to maybe even additional percent,
18:28
Ejaaz: why wouldn't you do that?
18:30
Ejaaz: Amazon is the perfectly positioned company for that. They already have a million of these robots.
18:34
Ejaaz: They are making very, very aggressive cuts on their labor force.
18:38
Ejaaz: I don't know if you saw, but like they cut 30,000 jobs about two months ago,
18:42
Ejaaz: and they're aiming to scale that up to 600,000 warehouse workers by 2023,
18:47
Ejaaz: which is honestly quite scary to hear for a lot of people, I'm sure, who are in these jobs.
18:51
Ejaaz: But I think those job roles will essentially evolve. But it basically makes
18:55
Ejaaz: Amazon the perfect company to design and build these automation robots of the future.
19:00
Ejaaz: Again, they're doing the unsexy stuff, the behind-the-scenes work.
19:04
Ejaaz: They're not consumer-facing robots, but it's still a multi-billion-dollar industry.
19:08
Josh: We're going to have to agree to disagree on that Tesla robotic statement, I think.
19:12
Josh: But I think there is an important difference to outline in the two different
19:16
Josh: approaches the companies are taking.
19:17
Josh: So Tesla very much treats the factory as the robot, and the robots that they're
19:23
Josh: going for are more humanoid general purpose or as it relates to transportation.
19:26
Josh: What Amazon is doing and what we're seeing here is...
19:29
Josh: Sure, there's some robots that look like humanoids, but a lot of the robots
19:31
Josh: that are going to be in these factories, they're narrow-purpose robots.
19:35
Josh: They're good for one task. It's a single arm that moves a specific way very quickly.
19:39
Josh: And I think that's, when we talk about robots, it's important to understand
19:42
Josh: that there are narrow-band robots and general-purpose robots like humanoids.
19:46
Josh: And what Amazon most likely stands to benefit from the most,
19:49
Josh: at least in the shorter term, is these narrow-band robots like these arms that
19:52
Josh: you're seeing on the screen that are really, really efficient at doing one specific
19:56
Josh: task. or whether it be those things that are rolling on the floor that can move all these boxes around.
20:00
Josh: So as it relates to that type of robotics, Amazon is, they probably have the
20:05
Josh: strongest case to be made on how much they can profit from putting those into
20:09
Josh: their product because the factory very much is their product.
20:12
Josh: How many packages they can ship per minute, if they can get that up from over, what was the number?
20:16
Josh: Some crazy number, but whatever 200 per second packages, if they can get that
20:20
Josh: up to 250, I mean, that's a huge increase in improvement.
20:23
Josh: So I think as it relates to factories, this automation is gonna be huge for them.
20:27
Ejaaz: I have a question for you Josh. I noticed you said that it'll help them profit for their own products.
20:32
Ejaaz: Do you ever see Amazon selling this type of robot to other factory manufacturers
20:37
Ejaaz: in completely different sectors?
20:39
Josh: I would hope not. But what I imagine Amazon does is uses this to create more
20:45
Josh: of a platform for to entice sellers.
20:47
Josh: So a similar business, which I'm kind of thinking, like people use Shopify a
20:51
Josh: lot to create web stores and to sell things.
20:54
Josh: Shopify is good at creating the tooling. It's good at creating the actual website.
20:57
Josh: It's good at pairing customers to consumers, taking care of all the infrastructure.
21:01
Josh: But Amazon is the layer that sits beneath that and can actually handle the logistics for you.
21:07
Josh: So if you're shipping physical goods, I suspect that Amazon won't sell these
21:10
Josh: robots to other companies.
21:12
Josh: They will just, again, like they're doing with the data centers,
21:14
Josh: they will build the infra and then offer the services to anybody who wants,
21:18
Josh: because that's where they get the most amount of profit.
21:19
Josh: So if you're a seller on Amazon who wants to know, well, what are the sales
21:24
Josh: going to be like in November and December during the holiday season?
21:26
Josh: Can you project that for me?
21:27
Josh: How much do I need to make? Okay, how much should I send to your warehouse so
21:31
Josh: that you're able to get out all the orders on time?
21:33
Josh: And as they get more of this intelligence, as they start to build more of an understanding.
21:37
Josh: And this gets to the consumer side too, where they understand their customer.
21:40
Josh: They know what the customer is shopping for. They know how to sell them ads.
21:43
Josh: You get this really fully vertically integrated experience as a seller on Amazon
21:47
Josh: that you just can't get anywhere else.
21:49
Josh: So if they're building these robots for their own warehouses,
21:52
Josh: they should keep them and make everyone else use them because that total vertical
21:56
Josh: integration where they understand the customer, they have all the automation,
21:59
Josh: it creates the best experience for everybody.
22:00
Ejaaz: I see it. I just don't,
22:03
Ejaaz: know if I can fully believe it just yet, because like in the same way that they
22:07
Ejaaz: rent compute or like kind of CapEx, you know, data hardware to different people,
22:11
Ejaaz: I feel like they would probably do the same for their robots as well.
22:15
Ejaaz: But I saw Newsleek a few months or like last month, Josh, I don't know if you
22:20
Ejaaz: saw this, that they're planning to take on the U.S. Postal Service.
22:25
Ejaaz: For those of you who don't know, Amazon pays the U.S.
22:28
Ejaaz: Postal Service or rather facilitates $6.6 billion of revenue by using the U.S. Postal Service.
22:34
Ejaaz: Amazon just didn't want to scale the delivery kind of service to the extent
22:38
Ejaaz: of the U.S. Postal Service.
22:40
Ejaaz: Now they're in the market of actually doing that. So we might end up with a
22:44
Ejaaz: company that is just all consuming and using all the goods for themselves.
22:47
Ejaaz: I don't think they'll do it for chips, but maybe they'll end up doing it for robots.
22:51
Ejaaz: And Josh, I guess the final point that I want to make and that we had in this
22:56
Ejaaz: essay is Amazon is just really good at understanding what they're good at and
23:01
Ejaaz: not treading outside of that line. So what do I mean by that?
23:05
Ejaaz: Well, many people think that, okay, if you're a frontier AI company,
23:09
Ejaaz: you should have a bleeding edge model.
23:11
Ejaaz: The funny part about this is Amazon has arguably the weakest AI model that's
23:16
Ejaaz: out there, but it's good for their in-house use.
23:19
Ejaaz: So they have a series of models, Josh, and it's called Nova.
23:23
Ejaaz: But have you ever used Nova, Josh? Have you ever kind of heard of it?
23:26
Josh: No, I have not used. I actually haven't engaged with any Amazon AI yet. This is a new frontier.
23:31
Ejaaz: Okay. Well, the reason for that might be because it is a speech-to-speech model,
23:37
Ejaaz: meaning that you speak to it and it speaks back to you.
23:40
Ejaaz: So you might be like, well, when the hell would I ever do that?
23:43
Ejaaz: It's primarily in customer support. That's where they've used their AI model.
23:47
Ejaaz: So they trained a very small but hyper-efficient AI model to kind of facilitate
23:51
Ejaaz: and backlog all of that kind of stuff. So they've been able to reduce their
23:55
Ejaaz: kind of reliance on human customer support for that and save costs in that end.
23:59
Ejaaz: The other side of things is Amazon is really good at focusing on enterprise customers.
24:04
Ejaaz: And so they produced an AI model software service, which basically allows any
24:09
Ejaaz: enterprise to train their own AI model using Amazon's model architecture so
24:15
Ejaaz: you can train it on its own proprietary data. Why would you want to do that as an enterprise?
24:18
Ejaaz: Well, you have private data, you don't want to give it to OpenAI or you don't
24:21
Ejaaz: want to give it to Google, but now you have a private instance where you can
24:23
Ejaaz: just run it on Amazon's model product. And that's really cool.
24:26
Ejaaz: And that brings me to the final point, Josh, which is like kind of Amazon's
24:30
Ejaaz: secret ability here, they've invested in some really big companies.
24:34
Ejaaz: In fact, you might've heard of Anthropic.
24:36
Ejaaz: Actually, I think you told me this on a previous episode.
24:39
Josh: Yeah, a little known fact that most people don't know is that Amazon owns about
24:44
Josh: 20%, give or take a few percentage points of Anthropic.
24:47
Josh: And it's funny because when I was considering the bear case,
24:50
Josh: the reason to be skeptical of Amazon, one of the big things was comparing the
24:54
Josh: other major cloud providers. We have Google Cloud Services, which is,
24:57
Josh: I mean, a huge entity that runs on Gemini.
25:00
Josh: It benefits from Gemini. Then we have Microsoft Azure, which is Microsoft's
25:04
Josh: cloud service provider that partners with OpenAI and they receive 100% of the IP from OpenAI.
25:09
Josh: But then I was like, well, you know, Amazon actually does have their own big
25:12
Josh: dog in their corner, which is Anthropic.
25:15
Josh: They have this huge partnership with Anthropic where I'm sure a lot of resources
25:18
Josh: are being shared, but Anthropic is also training using these new chips,
25:22
Josh: which is fascinating because right now, Now, Anthropic has the most unbelievably
25:25
Josh: great coding model in the world.
25:27
Josh: So something is working behind the scenes.
25:30
Josh: And I guess time will tell us to see how it bleeds out into the rest of the
25:33
Josh: industry. But they do have some big guns in their corner.
25:36
Ejaaz: Kind of like going on the theme of the bear case, Josh, I have to like put the
25:41
Ejaaz: realistic framing on this.
25:43
Ejaaz: Amazon's chips are awesome. They're almost as good as NVIDIA. They're 50% cheaper.
25:47
Ejaaz: But there's one major constraint, which actually Google faces as well.
25:53
Ejaaz: There's a little company in Taiwan called TSMC.
25:56
Ejaaz: Some of you might have heard about it. But they require Taiwan Semiconductor
26:01
Ejaaz: Manufacturing Company to be able to build the chips for them.
26:06
Ejaaz: Google relies on them to build their TPUs as well.
26:09
Ejaaz: Can you take a guess at which company owns around 90% of the capacity for TSMC next year.
26:14
Josh: Oh, I'm going to guess there's only one correct answer to this question,
26:18
Josh: and that is NVIDIA, the owner of all chips and GPUs.
26:24
Ejaaz: Yeah, so even if Amazon wanted to, let's say they created a hit product with
26:29
Ejaaz: these chips and everyone wanted to use it, they couldn't even fulfill demand
26:32
Ejaaz: because NVIDIA holds the stronghold.
26:34
Ejaaz: And so they have to wait until the end of 2027 where TSMC would have scaled
26:39
Ejaaz: enough capacity at that point to be able to service that.
26:42
Ejaaz: So they're going to start to look for alternative providers.
26:46
Josh: Yeah, is that a real constraint? Because, okay, so I'm of two minds here where
26:49
Josh: Amazon seems to be undervalued just on a relative basis. Their price to earnings ratio is very low.
26:55
Josh: Wall Street has basically priced them as break even throughout the course of
26:58
Josh: the year, while a lot of the comparable companies have gone up like 80, 90, 100 percent.
27:03
Josh: But then I hear things like this and I'm like, well, is this bull case that
27:06
Josh: we're laying out actually even possible? Because we do have the TSMC kind of
27:10
Josh: monopoly situation with NVIDIA.
27:12
Josh: There are alternatives. I know Samsung is working on building another chip architecture.
27:17
Josh: Is that really the nail in the coffin? Like, is it possible for them to see
27:21
Josh: this upside growth without TSMC? or is it really just relying on TSMC?
27:26
Ejaaz: So a consistent trend for NVIDIA that owns the monopoly of capacity on TSMC
27:32
Ejaaz: is they want to be able to train frontier AI intelligence.
27:37
Ejaaz: Amazon's approach has been very clear. They want to be the cheaper alternative.
27:41
Ejaaz: Once you've scaled and built your frontier intelligence, we're the cheaper chip
27:46
Ejaaz: for you to use to scale your product in general.
27:48
Ejaaz: Inference is where they plan to make most of their money is my suspicion.
27:53
Ejaaz: The other way I would address this is, I do think, it's not going to be immediate,
27:57
Ejaaz: but eventually, over the next couple of years, companies like Samsung and other
28:01
Ejaaz: competitors are eventually going to provide an alternative to TSMC.
28:06
Ejaaz: Now, critics will be jumping down my throat on that one because they're going
28:09
Ejaaz: to say, well, for the last decade, nothing has changed.
28:11
Ejaaz: And I will just respond to you that AI wasn't really a big thing over the last
28:15
Ejaaz: decade. It's only over the last couple of years.
28:18
Ejaaz: And if big companies like Google, like Amazon are completely constrained because
28:23
Ejaaz: of this one company and NVIDIA maintains the market monopoly there,
28:26
Ejaaz: it's going to force companies like Samsung to create an alternative.
28:30
Ejaaz: We're already seeing this. We had an episode this week or rather last week on
28:34
Ejaaz: China copying the key components of a company called ASML to try and fill this gap.
28:40
Ejaaz: I just, if I was a betting man, which I am, and I own Amazon stock,
28:45
Ejaaz: you know, just putting it out there, I think a competitor is going to come into the ring.
28:50
Josh: Yeah, it seems like, well, the interesting thing is Samsung actually beat TSMC
28:54
Josh: to the two nanometer chip architecture, which was a really big deal.
28:58
Josh: And now they're going to partner with Apple and the largest companies are going
29:02
Josh: to get their chips from somewhere else.
29:04
Josh: Is it going to be too long? EJS, 2027, we're going to have like AGI on Mars
29:09
Josh: by then. So can they survive?
29:11
Josh: Like, is that not too long? I don't know. So I guess maybe we could wrap this
29:15
Josh: up by kind of where you currently stand.
29:17
Josh: How long do you think these things will take to play out?
29:20
Josh: How much upside do you think is really possible in the near term?
29:24
Josh: How poorly has the market mispriced?
29:26
Josh: Is this just like a bubble underwater that is waiting to pop out?
29:29
Josh: How does that all look kind of going forward into 2026 based on your understanding of it all?
29:34
Ejaaz: Okay, so in short, I think it's bullish because even though they are reliant
29:39
Ejaaz: on TSMC's capacity, they are still going to produce in the order of five to
29:44
Ejaaz: six million Amazon chips next year. In fact, Google's going to do the same as well.
29:49
Ejaaz: And one simple problem remains, Josh.
29:52
Ejaaz: There's not enough GPUs, TPUs, or Tranium chips to satisfy demand right now.
29:58
Ejaaz: So even with NVIDIA's capacity, they're sold out.
30:02
Ejaaz: So companies are looking for other types of chips to fulfill demand.
30:06
Ejaaz: That's why OpenAI last week signed a $10 billion deal with Amazon.
30:09
Ejaaz: That's why Anthropic signed a 1 million chip deal with Amazon. It's the same thing.
30:15
Ejaaz: If Amazon creates all the chips, they're going to end up selling it.
30:18
Ejaaz: Now, over the long term, are they going to win? Yes or no? it remains to be seen.
30:23
Ejaaz: You know, everyone thinks TSMC has a stronghold. I think there's eventually
30:26
Ejaaz: going to be a competitor, which makes Amazon and Google way more competitive to NVIDIA's mode.
30:31
Josh: If you build it, they will come. And Amazon is building it. And they're building
30:34
Josh: a lot of it. And they have won the least out of everybody.
30:38
Josh: So just based on that math alone, there's got to be some hidden upside there. Like they're trading.
30:43
Josh: There's such a cash generating business. Everybody who's watching this video
30:46
Josh: has interacted with Amazon once in their life.
30:48
Josh: It's just it's a great company that will probably stand to benefit from ai when
30:52
Josh: and at what scale we don't know but that is part of the fun we will follow it
30:56
Josh: along as we go is just are there any parting thoughts before we we wrap up here.
31:00
Ejaaz: Actually yes i have one final bit of law we started off the episode critiquing
31:04
Ejaaz: amazon stock pricing it's flat one of the main reasons that critics give is
31:09
Ejaaz: because they don't really believe in the ceo andy jassy i don't think i agree
31:13
Ejaaz: with that he's led the company to kind of like create quarterly earnings upon upon upon a time.
31:19
Ejaaz: They're doing really well. They've grown a lot. But there's a small percentage chance, Josh.
31:24
Ejaaz: That Jeff Bezos reclaims the throne, similar to how Sergey Brin did so with Google. Why not?
31:30
Ejaaz: He's outspokenly said that AI is the most important technological shift over the next decade.
31:36
Ejaaz: He started a company that's focused on creating AI lab stuff for robotics and stuff.
31:40
Ejaaz: It feels kind of weird for him to just leave his baby behind.
31:45
Ejaaz: He knows he already has the vessel there. Why wouldn't he come back?
31:48
Josh: Okay, well, it's certainly the same thing as Google. Google had their founders
31:51
Josh: disappear. Sergey Brin came back and the stock is up 80% in the 12 months that followed that.
31:56
Josh: So there is something, there is a precedent for this happening.
31:59
Josh: I mean, we have, Bezos is off, he's building rockets with Blue Origin, he's doing robotics.
32:03
Josh: He seems very busy. If he comes back to Amazon, I'm sure that would be bullish to some extent.
32:07
Josh: What the extent is, we don't know, but we will be here to cover it as always
32:12
Josh: as we go along this journey.
32:13
Josh: That will wrap up our episode today on the Amazon Bullcase. If you enjoyed this
32:17
Josh: episode, please do not forget to share it with a friend.
32:19
Josh: Or if you haven't go rate it five stars on your
32:22
Josh: favorite podcast player there is a special ask which
32:25
Josh: is actually i'm not even sure this isn't ask this is a value add
32:28
Josh: to your life this is the alpha drop that comes two times a week in our newsletter
32:32
Josh: every wednesday is a thought piece like the one that we covered today and every
32:35
Josh: friday is a short roundup it's five bullets on the most interesting things that
32:39
Josh: we found really cool this week in the world of ai and frontier technology so
32:43
Josh: if that seems interesting we have the link in our description to go sign up it's a
32:46
Josh: stack it's really easy it integrates with all the places you watch you get
32:49
Josh: your emails from and i don't know i think it's pretty interesting and
32:52
Josh: you do hear the podcast episodes before they come out because that's normally
32:55
Josh: where we curate our ideas before presenting them in long form so would highly
32:59
Josh: recommend that concludes our first episode this week we still have two more
33:03
Josh: to go so stay tuned for that there's a lot of interesting stuff happening and
33:06
Josh: thank you so much for watching as always and we will see you guys on the next one see ya.