Andy Jassy’s Shareholder Letter, Meek Mill Joins the AI Race| Diet TBPN

There is a ton of AI news. Andy Jassy released the 2025 shareholders letter. Amazon is known for fantastic shareholders letters dating back to 1997. I thought he did a good job of sort of resetting the AI narrative. There's this AI lab horse race going on. Of course, we've been covering it all week. Anthropic's Mythos preview and Project Glasswing launched on Tuesday, quickly followed by news today that OpenAI also plans to deliver a model with advanced cybersecurity capabilities to key internet infrastructure providers, and there's this debate going on over how and when these models will roll out. I think this is gonna be an ongoing trend. I don't think cybersecurity is the last model capability that will be slowly delivered to key companies first. Cybersecurity is a perfect fit for powerful coding agents, and people have been digging into exactly how some of these zero-day exploits, some of these bugs and vulnerabilities were discovered, and it makes a ton of sense that if you have a model that's fantastic at coding, it can basically try try every single coding exploit, try new coding exploits across a huge number of open source packages, submit pull requests, and generally harden the internet infrastructure that we all rely on. So in general, it seems like the rollout of Mythos, although people are disappointed because they wanna play with the latest and greatest model, even if it's very expensive, it seems like it's having a positive effect and should be a bit of a white pill for containing powerful models. I wouldn't be surprised if we see something similar happen in biosafety. Now, the biosafety AI research loop is a little bit longer because you might have to go to the lab. It's not entirely existing in a computer, in a virtual machine that you can spin up and just brute force reinforcement learning against. But you could imagine if a model develops capabilities, maybe in the next run, maybe in, middle of this year. If a model becomes powerful enough to design a harmful virus or something like that, you would want the lab that creates that model to deliver that to the scientific community and companies that can protect the population against the development of new harmful biological viruses, just like we're protecting the internet against cybersecurity viruses. And so I, I'm not exactly sure where this all goes, what the other what the 10 steps are, but in general, it seems like there's going to be a pattern of a powerful model becomes capable of doing something that it makes sense to share with the particular community that can defend against that new capability. And then the entire community needs to make sure that that capability is carefully under control before releasing a version of that model that can still, you know, in the biosafety example, like you still want a model that can help you learn about biology, learn about how viruses are made, how they mutate. This is important for education and augmentation of scientists from, you know, high schoolers all the way to professionals. But the most advanced technology, it makes sense to put it in the hands of people that can actually have a real impact on that immediately before rolling it out broadly. But Andy Jassy sort of zoomed out and was stepping back from the horse race because Amazon's a partner with basically everyone in the ecosystem and has their own models. So he shared this shareholder letter, zooming out on the state of AI, the state of Amazon's plans. And he shares a bunch of very interesting anecdotes about his personal life. He says, when I graduated from college, I wanted to be a sportscaster. After sending my resume reel to many small markets around the United States and only getting 2 nibbles, I settled on doing sports production at a major network. To make extra money, I coached my former high school soccer team and worked at a retail golf store. 6 months later, a college classmate convinced me to interview at the consumer products company where he worked. and I spent 3 years as a product manager there. I left that job to try some of my own businesses. After deciding these businesses weren't my calling, I tried short stints at sales and investment banking before going back to grad school and ending up at Amazon 3 days after my last final exam in May 1997. Not exactly a straight line, he says. AWS followed lots of squiggly lines too. And of course, Andy Jassy is the, by and large the creator of AWS. That was his major project during his tenure. Founder. Yeah, deserves so much credit in building that business. The original vision included storage, compute, payments, and human intelligence. They had a product called Mechanical Turk where you could go and dispatch a specific task. You would have to build sort of a web UI. It was the original sort of data capture tool, but it could also be used for little things like manual translation tasks. People weren't really using it for customer support tickets, but data labeling, data labeling before you really needed mass data labeling. But it was that type of business. And there was always a question when Scale AI started, like, is this just Mechanical Turk? Is this business process outsourcing? Obviously that went on to a fantastic outcome with Meta, but it became a different thing and quickly moved towards what we see in the expert networks and data collection that's much more nuanced than a single task on demand. But AWS, this was in the vision, and they wound up pulling away from that and sort of refocusing. So Andy Jassy says, "Some of those, e.g., storage and compute became linchpins in AWS. Others didn't, others didn't succeed. We didn't initially plan a database service, and when we built one, our first attempt failed to get traction. We went back to the drawing board and built new relational and non-relational databases, which have resonated well and become core to millions of AWS applications. When we launched EC2, our compute service, it was a single instance in one availability zone, Linux only, with no auto-scaling load balancers, balancing, block storage, or private networking. Over the time, we added those capabilities and hundreds more services. And you see that when you go into the AWS dashboard, it is chock-full of different products. AWS was initially attractive to startups. Companies like DoorDash, Dropbox, Pinterest, Slack, and Stripe were among many that built their businesses on AWS. Pundits said enterprises and governments would never use AWS for anything substantive. But in 2008, Netflix, which was delivering DVDs in the mail almost a decade at that point, decided to move all of their applications to AWS. Then came big commitments from GE, Intuit, and others, and eventually the CIA chose AWS as their partner to build their classified cloud. Growth came fast and furious, and as it accelerated, so too did our capital expenditures, CapEx, with a dilutive impact on free cash flow. And so he's, he's starting to make the case that there's this trade-off between free cash flow and CapEx. And this will come back when he talks about his plans for taking advantage of the AI boom. So he says, at our 2014, more than 12 years ago, AWS operating plan review, the discussion started with a senior leader at the company musing, tell me why we're doing this business. Like, why are we doing this? It was confusing and it was very different from being an online bookstore and then eventually the everything store. But it felt different. But then eventually it started working. He's reflecting on his career and the twists and turns in building AWS. He says most long-term endeavors do not follow a linear straight line. Up and to the right. Progress jumps around. It'll zig up, then sometimes stall or zag down or force you back to the starting line. Sometimes it feels like you're running in circles, but the path is rarely straight. That's because the world is complex and new technology, business model invention, competitors, global issues, or people and cultural shifts can come into play. When we're in the middle of some of the biggest inflections of our lifetime, e.g., AI, robotics, space, industrialization, geopolitical and military conflict. And just as proficient golfers need to be skilled across driving, approach shots, chipping, and putting, durable companies must be adept at managing different elements of inflections. He talks about how broad Amazon is these days— retail, logistics, AWS, ads, Kindle, Alexa, pharmacy. There are too many new efforts in flight to mention them all. The long push to, to bring robotic automation acquiring Kiva in 2012. He also talks about all of the benefits that Amazon has brought to rural customers, but who are often deprioritized by logistics. This is a classic thing when you launch an e-commerce site. Do you apply your flat rate shipping to Alaska? Because every time someone ships something to Alaska, it's going to charge you an arm and a leg. And he's also focusing on closing the digital divide in rural communities. So things like high-speed internet access, internet, LEO. He's been talking about Amazon LEO, the low Earth orbit satellite network. He makes this point that Amazon must be willing to pursue parallel paths when it's unclear what will best drive the trajectory, the desired trajectory. He says 2 is greater than 0. When I was a kid, I used to go to New York Ranger games, New York Rangers games with my dad. I loved hockey and it was a high-quality time together. I looked up to my dad, still do, and hung on his every word. One game My dad noticed that one of the Rangers defensemen, Dallas Smith, had gone missing from the bench and stood up and exclaimed, "Where's Dallas?" To which a nearby fan said, "In Texas, moron." So yes, it's very weird that his name is Dallas, I guess. Is that the confusing? So yes, it's fairly obvious that 2 is greater than 0, but too often companies focus on what looks most tidy instead of ensuring they have enough efforts in play to achieve an important outcome. Let's go back to fast delivery in our retail business. We know how much customers crave it, and we see higher order completion rates when delivery promises are faster. So he talks about, you know, all the things that we know about Whole Foods and rolling out more stores, rolling out more just investments in the core Amazon business. But then he goes on to talk about artificial intelligence. And he says, if there's an obvious path to changing your trajectory, take it and run. But most new jumps aren't forward like that. There's invention and experimentation required, and pursuing multiple paths gives you the best chance to find it. When you identify disproportionate inflections, bet big. Choosing which inflections are truly seminal versus just interesting requires judgment. Reasonable people can disagree. But if you believe you found one of these disproportionate shifts, you want to invest as aggressively as you possibly can. This will create investment spikes that will invite scrutiny, but the game changers don't typically accommodate smoother investment horizons. One of these seminal shifts, of course, he says, is AI. He says we've never seen a technology more quickly adopted than AI. When ChatGPT launched in November of 2022. It reached 100 million users in 2 months, 4 times faster than TikTok and 15 times faster than Instagram. ChatGPT already has over 900 million weekly active users. Both OpenAI and Anthropic have revenue run rates reportedly approaching $30 billion. These are breathtaking numbers for companies this soon after their commercial launches. And he has this great throwback to Thomas Edison and the dawn of electricity. A lot of people are looking for comparisons for what AI pattern matches to. And he goes back to the first commercial power station, which launched in 1882. So at the time, most people understood the first commercial power station as a way, a better way to light a room. That was going to be the main benefit of electricity. It was going to replace lamps and you were going to have electric lights. But what they couldn't see was that electricity would eventually reorganize every factory, home, and industry on Earth. And he says that AI may have a comparable impact. The difference is that electricity took 40 years to get where it was going, and AI appears to be moving 10 times faster. And he says that Amazon's smack dab in the, in the middle of this land rush, and companies are choosing AWS for AI. 3 years after AWS launched commercially, it had a $58 million run rate, 3 years in. And Amazon was already a big business at that time. So, you know, he's talking about the twists and turns of launching AWS. It's a, you know, a lot of products.

Seems so quaint.

It's very quaint to be 3 years in, have the backing of, you know, a massive company and still 3 years in only hit $58 million run rate. He compares it to 3 years into the AI wave, which of course started in 2023 basically, and now we're in 2026. AWS's AI revenue run rate is over $15 billion in Q1 of 2026, which is nearly 260 times bigger than AWS was at the same point, and it's ascending rapidly. And so he highlights a bunch of reasons why customers are choosing AWS for AI. Obviously, they have a product in sort of every single category now, model building, high-performance inference, lower-cost inference that runs on Trainium, their custom silicon, agent building, secure environments, et cetera. But he says that AWS could actually be growing faster, and they are in fact limited on capacity. So AWS added 3.9 gigawatts of new power capacity in 2025 and expects to double total power capacity by the end of 2027. And AWS is monetizing that capacity as fast as it's installed. In Q4 2025, AWS reported 24% year-over-year growth and a $142 billion revenue run rate. That's a lot of absolute growth. And yet we still have capacity strains that yield unserved demand. As an aside, two large AWS customers have already asked if they could buy all of the Graviton instance capacity in 2026, which is their widely adopted custom CPU chip. And he says, we can't agree to these requests given other customers' needs. You can't just say, okay, we're giving everything to one company, but it gives you an idea of the demand.

Touches on a bunch of different stuff. I think this is what people wanted out of Jassy.

Yeah.

Like about a year ago, right?

Yeah.

He was getting, you know, people were constantly kind of chirping at Amazon leadership and wanting some kind of like basically lay of the land like this to show that Amazon had a broad understanding and actually had, you know, real leadership and was committed, committed to being a real player.

Well, in other Amazon news, Amazon Pharmacy to offer Eli Lilly and Company's new GLP-1 pill Foundeo via same-day delivery. Shiel Monod says this thing is going to fly. Amazon is up 3%. I don't know if it's on this news.

It's hard to say. Amazon's actually almost up 5% today. I would expect that to be because of the letter, But people tend to buy a lot of GLP-1 drugs too.

There's so much skepticism about fly-by-night peptide operations. Where are they being compounded? Are they safe? Do they contain even what you think they contain? And giving consumers an option that's something as, you know, established as Amazon with all of the guardrails that they have in place feels like a very positive move for consumers.

Before we jump into the next story, Jassy did share a letter to shareholders from 1997.

Yeah.

And there was— I'll read the first two paragraphs. To our shareholders, Amazon.com passed many milestones in 1997. By year end, we have served more than 1.5 million customers, yielding 838% revenue growth to $147.8 million and extended our market leadership despite aggressive competitive entry But this is day one for the internet, and if we execute well, for Amazon.com. Today, online commerce saves customers money and time. Time and money, save both. Tomorrow, through personalization, online commerce will accelerate the very process of discovery. Amazon uses the internet to create real value for its customers, and by doing so, hopes to create an enduring franchise, even in established and large markets. Well, they certainly did that. And the letter goes on to talk about how it's all about the long term. So even back then, growing like absolute crazy, still, you know, wanting people to think about, think in decades. And so Jassy doing that now, and they have a good track record.

I mean, Amazon's had decades of drawing down on free cash flow to invest in the future, and they've had a long history of communicating that effectively to the community and to the shareholders.

The level to which people miss the ads. This is never— I don't think this has ever happened before.

Yeah.

In the history of content, of podcasting.

Yeah.

We miss the ads too.

Yeah. All we have now is— Well, there's a debate on the timeline about OpenAI's new cybersecurity product. There was a post that was deleted and then Axios issued an update. Basically, the question was, uh, OpenAI is launching a new model rumored to be called Spud, trained on Blackwell, sort of similar, you know, big model, lots of capabilities. Uh, and, uh, OpenAI has been working on cybersecurity products. Will they gate the rollout similarly? Are they running the same playbook? Will they, uh, take a different path? At what point will they make their models available? and people are going back and forth. And Andrew Curran is sort of clarifying here that the new model and the cybersecurity product are separate and only the special, the cybersecurity specialized model will have a limited release, not the new model itself. So it looks like a general public release for SpUD. Dan Shipper shared some extra commentary around this. The Axios story floating around about OpenAI limiting the release of their newest model SpUD isn't true. He just spoke to OpenAI and it appears the story conflated two things. They do have a cyber product. They are testing with a trusted tester group, but this is not the same thing as Spud. The Axios story has now been updated.

Yeah, there's a very strong argument to never release a model publicly that is specifically optimized for cyber.

Yeah.

Because like you're just inviting a bunch of like, think about like the teenagers out there. I now get to be a super powerful hacker, even though you're just kind of like, probably should just be vibe coding like a fun little app or something. Something like that, right?

Yeah.

And so—

You would think having some cybersecurity capabilities in the model would just be better if you're vibecoding a product, it becomes secure out of the box.

Sure. I'm just saying, like, having this super powerful, like, cyber-focused product shouldn't be the kind of thing that anyone— like, you should probably have to KYC to sell it, right?

Yeah, sure, sure, sure. Well, yeah, what do you think about this?

Yeah, I mean, also, I think when you zoom out on AI progress, it is very smooth. Where like people were talking about this with Mythos yesterday where like, well, actually, if you run even like an open source model and you actually run enough times, you actually can find similar exploits. And it's just like kind of the efficiency of the new model versus the smaller open source ones. I'm pretty sure right now I could take 5.4 and I could just run it a bunch of times and I could find a lot of exploits. Like people have the capabilities right now. Maybe it's like a little bit harder.

You might get flagged.

Sure.

Yeah.

But. And also, like, yeah, I think, I mean, at some point, like, a lot of the coding agents, I think you'll just see, like, in the, you know, in the system prompt or whatever of the coding agents, it'll be like, yeah, make sure you're checking to see if any, if there's, like, big security vulnerabilities. Even right now, you can just ask Codex, whatever you're building something, like, are there errors? Are there, you know, security problems? And it will solve them.

And this is the nature of every advanced model is that in order to understand how to fight something, you'd have to know how to build it. And this is true. I mean, I'm sure that the folks at CrowdStrike or Palo Alto Networks like could definitely, if they went black hat, it would be bad for everyone, even in the pre-AI era. And that remains true. Having at least as much economic incentive as possible to put the resources and the tokens and the inference budgets towards like white hat hacking is good. There's also somebody in the chat is making the comment that there's a lot of precedent for this with bug bounties and timed disclosures. So oftentimes white hat hackers will go to companies and say, we found a really bad vulnerability. We're giving you 90 days until we publish it, but it is in the public interest to know that this vulnerability exists broadly. So we have to release it, but also we want to give you the time to react to this. And so it feels like you're sort of holding that company hostage a little bit. It can be a little bit tense. If it's done carefully, it can be sort of a win-win.

But I think regardless, like, I think more KYC is probably the future, right? Even if you're just talking about, like, risk of distillation, stuff like this. Like, I think I've talked about this before where, like, yeah, at some point you need more KYC to ensure, like, people aren't just, like, training off the model or doing, like, nefarious things.

Yeah. Meek Mill pulled out a slide from Bill Ackman. The slide says a small percentage of songs are listened to. The overwhelming majority of music tracks receive zero or minimal engagement. AI is poised to exacerbate this reality. And so it shows that 0.2% of songs are culturally and commercially relevant. Commercially relevant is 10% of songs streamed 1,000 to 100,000 times. And then Meek is dropping that into what looks like Claude saying, you're not in that 88%, you already skipped the whole bottom of the pyramid, you're already in that top 0.2% of people know your name. They search for you specifically. They stream on purpose. That changes everything. You already have the hardest thing: attention. You don't need the algorithm to find you. Your fans already look for you. So while AI is flooding the market with millions of trash songs, this is AI throwing AI on the bus. This is crazy. Your catalog just sits there collecting streams because people want Meek, not just a rap song. And Dimes Square Holdings says, first in my bloodline to see a rapper with AI. Psychosis commenting on a deck from—

I think it's a, I think it's a reasonable, it's a reasonable analysis. It's a reasonable analysis.

I just think it's, I just think it's a little bit unfair for the AI to be dunking on the other AI.

Yeah.

This is kind of like a crabs, crabs in a bucket.

There's a whole story about this in the Journal the other day about a local group that's lobbying against data center development using AI tools, ChatGPT, like very heavily. To understand the legal code, how they can, how they can organize, who they should be calling. It sort of feels like a good use of AI in the sense that they're exercising their democratic rights properly through the correct channels. Like, it could be so much worse. I don't know. What do you think?

Yeah, well, hopefully, like, by using the model, like, models enough, they'll be like, oh wait, this is actually good.

Yeah.

Or like, or like there are some parts that I like narrowly. But do we need this data center in my neighborhood? Maybe not.

And they can also specifically on this, I think it's interesting, like among celebrities, it seems like rappers specifically are like really leaning in a lot more than other like types of celebrities. I see like Lil Baby. Yeah, Lil Baby, not little baby. Okay. There's a post yesterday about someone who got paid to like set up OpenClaw for him. Yeah, you've seen this a bunch. Yeah, it's just funny. Like you don't see actors talking about their—

Lil Baby. Was second after Baby Keem.

Yeah, that's right. That's right. I mean, Matthew McConaughey had some words about AI saying, oh yeah, when he was going to log in, he needs to log in. Well, no, that was separate. Like, wasn't it like months? It was like months, months, months and months later. He was like, this is coming. Like, you need to be prepared for this. You need to work alongside of it.

Yeah, but he's not talking about his specifically.

Yeah, exactly. Exactly.

Yeah. I guess Ben Affleck had that company, right?

Yeah, yeah, yeah.

But Yeah, Ben Affleck already got the 9-figure exit. He's good.

Yeah, but you would imagine that AI actually might be a useful tool for understanding the impact of building a data center in a particular community because everyone is debating the economic impact, how many jobs will be created, what's the environmental impact, and being able to crunch through all those trade-offs so that that the community gets to a net positive impact that the population can basically vote on and be happy about and say, yeah, we made this trade-off properly and we feel like we're getting the benefit. Because maybe it's generating a lot of tax revenue. Maybe the tax revenue is very durable. Maybe the tax revenue is going to the right place. Is it going to a tax refund? Or is it going towards a project that people don't actually support? And so there's a whole other question of, for the local community, is that tax revenue meaningful? Like what dollar value do they put on new tax revenues for the city? Depends on what the city's building or doing with that revenue. Here's another like self-referential thing, like, you know, AI, should AI be used to fight data center construction? Today, Meta began removing ads from attorneys who are seeking clients that claim to be harmed by social media while under the age of 18. And so there's been— these class action lawsuits have been a big— you see them on social media all the time. Usually it's some sort of data breach or some sort of random product that they're targeting you for. And I feel like many people just sort of scroll past them because the default class action is like you wait a long time and then maybe you get a check for $5 in the mail. But Meta has begun removing advertisements from attorneys. What is their justification for this? Lawyers across the country are now seeking new plaintiffs in the hopes of bringing a class action lawsuit that could result in lucrative verdicts. It's unclear if any of them have been backed by private equity, as the California lawsuit appears to have been. Axios has identified more than a dozen such ads that were deactivated today, some of which came from large national firms like Morgan Morgan. Almost all of them ran on Facebook and Instagram. Some appeared on Threads and Messenger, plus Meta's Audience Network. One such ad read, anxiety, depression, withdrawal, self-harm. These aren't just teenage phases. They're symptoms linked to social media addiction in children. Platforms knew this and kept targeting kids anyway.

Yeah. So Meta has something in their terms of service that says we also can remove or restrict access to content, features, services, or information if we determine that doing so is reasonably necessary to avoid or mitigate misuse of our services or adverse legal or regulatory impacts to Meta.

Huh.

It's basically like, we're not going to let you use our product to take legal action against us.

Yeah. I wonder where the class action recruiting will go next.

You know, it'll go everywhere else, maybe out of home, out of home.

Yeah.

Could go to podcasts.

I mean, it's such a broad base that, yeah. Well, YouTube is another—

It's relatively easy to reach Facebook and Instagram users.

Yeah, especially young people.

It's everyone.

It's everyone, yeah. So they just need to cast a wide net. But it'll be interesting to see how that develops.

Has anyone built an out-of-home advertising network for classrooms?

Classrooms? What do you mean?

Like teachers?

I don't think people want to advertise to kids.

Kidding, kidding, kidding.

We'll see you tomorrow. Leave us 5 stars on Apple Podcasts and Spotify. Sign up for our newsletter.