Andy Jassy's 2025 Amazon shareholder letter is the clearest, most grounded take on where AI is headed that any major tech CEO has published this year. In it, Jassy compares AI to the dawn of electricity, flags AWS as the dominant cloud platform for the AI era, and makes the case for aggressive capital spending even as free cash flow takes a hit. If you've been wondering what Amazon's actual AI strategy looks like — this letter is it.
What Did Andy Jassy Say About AI in 2025?
Jassy opened his shareholder letter by zooming out from the week-to-week AI model horse race and instead offering a decade-scale view. His central analogy: when the first commercial power station launched in 1882, most people thought electricity was just a better way to light a room. What they missed was that electricity would eventually reorganize every factory, home, and industry on Earth. Jassy believes AI has a comparable — possibly larger — impact ahead, and that it's moving roughly ten times faster than electricity did.
He backed that up with staggering adoption numbers. ChatGPT reached 100 million users in just two months after its November 2022 launch — four times faster than TikTok and fifteen times faster than Instagram. As of the letter, ChatGPT already has over 900 million weekly active users. Both OpenAI and Anthropic are reportedly approaching $30 billion in annualized revenue. Jassy's point: this is not a normal technology cycle.
He also revisited Amazon's own history of squiggly-line progress — from his early career ambitions of becoming a sports broadcaster, to the chaotic early years of AWS, to the moment in 2014 when a senior Amazon leader sat down at an operating plan review and asked out loud, "Tell me why we're doing this business." The lesson: transformational bets rarely look clean from the inside. That applies directly to Amazon's current AI push.
How Fast Is AWS AI Revenue Actually Growing?
The numbers Jassy shared are hard to wrap your head around. Three years after AWS launched commercially, it had a $58 million annualized revenue run rate. Three years into the AI wave — which Jassy dates to roughly 2023 — AWS's AI revenue run rate has already crossed $15 billion in Q1 2026. That's approximately 260 times larger at the same point in the cycle.
Overall, AWS reported 24% year-over-year growth in Q4 2025, with a $142 billion annualized revenue run rate. And here's the kicker: Jassy says AWS is still capacity-constrained. They added 3.9 gigawatts of new power capacity in 2025 and expect to double total power capacity by end of 2027. Two large AWS customers reportedly asked to buy all of Amazon's Graviton custom CPU capacity for 2026 — a request Amazon had to decline because of obligations to other customers. That kind of demand signal is extraordinary.
Jassy's message to investors skeptical about the capex spending: this is the same trade-off Amazon made with AWS in its early years, and it paid off. The game-changers don't accommodate smooth investment horizons. When you find a disproportionate inflection, you bet as aggressively as you possibly can.
What Is OpenAI's Cybersecurity Model and Is It Public?
There was a brief wave of confusion this week when an Axios story suggested OpenAI was planning to limit the release of its newest model, rumored to be called Spud, trained on Blackwell hardware. That story was later updated after OpenAI clarified the situation to journalist Dan Shipper: the new general model Spud is on track for a public release. The limited rollout applies only to a separate, cybersecurity-specialized model that is currently being tested with a trusted tester group.
This mirrors what Anthropic did with its Mythos preview and Project Glass Wing earlier in the week — releasing advanced capabilities in a controlled way to organizations that can immediately put them to defensive use. The pattern is becoming clear: when an AI model develops genuinely dangerous dual-use capabilities, the responsible move is to deliver it first to the community that can defend against those capabilities, not to release it publicly all at once.
For cybersecurity specifically, this makes intuitive sense. A sufficiently powerful coding model can probe open-source packages for zero-day exploits, try novel attack vectors at machine speed, and submit pull requests that harden infrastructure — all things that white-hat security teams desperately need. Giving that capability to the defenders first, before releasing it broadly, is a meaningful head start.
Why Are AI Labs Limiting Their Most Powerful Models?
The short answer: because some capabilities are dangerous enough that putting them in the hands of defenders first is genuinely better than a wide public release. The longer answer is that this is becoming a deliberate pattern, not a one-off PR move.
The debate around staged rollouts is real. On one hand, researchers and developers are frustrated — they want access to the latest and most capable models, even at high cost. On the other hand, a cybersecurity-optimized model released publicly is essentially a powerful hacking tool available to anyone, including teenagers who should probably be building fun apps with Cursor instead. The case for Know-Your-Customer (KYC) verification for high-capability models is getting stronger as model power increases.
There's also a precedent in the security world: bug bounties and timed disclosures. White-hat hackers routinely find critical vulnerabilities and give companies a 90-day window to patch before public disclosure. It's tense, but it works. AI labs are essentially adapting that model for the deployment of powerful AI systems.
Could AI Biosafety Models Be Next After Cybersecurity?
Jassy didn't go here explicitly, but the logic extends naturally. If AI models become capable of designing harmful biological agents — something that may arrive within the next model generation or two — the same staged rollout logic applies. You'd want that capability delivered first to the scientific and public health community that can develop countermeasures, not released broadly.
The biosafety loop is slower than cybersecurity because it involves physical lab work, not just compute cycles in a virtual machine. You can't purely brute-force reinforce-learn your way through biology the way you can through code. But the principle is the same: pair the dangerous capability with the people best positioned to contain it, before the capability becomes widely accessible.
This is an emerging framework for how the most powerful AI gets deployed — and it's worth paying attention to, because cybersecurity is almost certainly not the last category where it applies.
Is AI Killing the Music Industry or Helping Artists?
Rapper Meek Mill surfaced an interesting data slide this week from investor Bill Ackman showing that only 0.2% of songs are culturally and commercially relevant, 10% get meaningful streams, and 88% receive essentially zero engagement. Ackman's point was that AI will flood the market with more low-quality content and make that bottom 88% even harder to escape.
Meek's response? He pasted the slide into Claude and got back a pointed analysis: he's already in the top 2%. People search for Meek Mill specifically. They stream on purpose. AI flooding the market with generic rap tracks doesn't threaten artists with genuine cultural gravity — it just makes the gap between real artists and noise wider. It's a reasonable take, even if there's something a little funny about an AI model throwing other AI-generated content under the bus.
The broader pattern here is real: in a world where content volume explodes, brand and authentic identity become more valuable, not less. The artists who have already built real audiences are insulated in a way that new entrants are not.
Why Is Meta Removing Class Action Lawyer Ads?
Meta this week began deactivating ads from law firms seeking plaintiffs in class action lawsuits alleging that social media harmed minors. The ads — running on Facebook, Instagram, Threads, and Meta's Audience Network — came from firms including Morgan & Morgan and targeted parents of children who experienced anxiety, depression, or self-harm allegedly linked to social media addiction.
Meta's justification comes from its own terms of service, which allow the platform to remove content when it determines doing so is necessary to avoid adverse legal or regulatory impacts to Meta. In plain English: they're using their platform rules to limit advertising that helps people sue them. Whether that holds up to scrutiny — legal or public — remains to be seen. The law firms will simply advertise elsewhere, including YouTube, podcasts, and potentially out-of-home. The cases themselves are not going away.
