My thoughts on Amazon's letter to their shareholders
Andy Jassy’s Letter Is the Closest Thing to a Bubble Deflation We’ve Had
For a while, I held the same quiet unease a lot of people were holding in the back of their minds. AI was attracting mind-boggling investment, enormous hype, and an almost religious fervour that it would reshape everything. And yet, the actual revenue numbers were vague. You heard about compute spend, about headcount, about infrastructure buildouts. What you heard less about was customers actually paying for AI in ways that showed up clearly on a balance sheet. That is what you call a bubble, and it was getting harder to argue against. Then Andy Jassy published his annual shareholder letter1 for Amazon’s shareholders.
The number that stopped me was AWS’s AI business now running at over $15 billion in annual revenue. That’s real revenue growing at triple digit percentages year over year. That is not what bubbles look like at this stage. Bubbles are characterised by investment that outpaces demand, by valuations that float on future promises rather than present transactions. What Jassy described is enterprises actually paying for AI infrastructure, at scale, right now. The hype preceded the revenue, yes. But the revenue is arriving.
I am pretty sure most of us have been skeptical about seeing numbers like this so soon. I was wrong to be. And I think being honest about that matters because calling something a bubble can feel like the smart, cautious take. It lets you sound careful without actually having to change your mind.
The second thing Jassy confirmed is something I have been watching closely for a while and have been particularly interested with. It’s the not so commonly acknowledged fact that hyperscalers are quietly moving to custom, in-house silicon.
This is not a small shift. For years, the assumption was that Nvidia owned inference and training, and that the hyperscalers would remain customers indefinitely. That story is changing. AWS is now running more of their inference workloads on their own custom chips—Trainium, the GPU, and Graviton, their CPU—and the performance case is becoming undeniable. Trainium2 offers roughly 30% better price-performance than comparable GPUs. Trainium3, which just started shipping, pushes that another 30–40% beyond Trainium2, and is nearly fully subscribed already. Trainium4 hasn’t even hit broad availability yet and still a significant portion is already reserved.
What made me pause was Jassy’s own framing of what this means at scale. Amazon expects Trainium to save them tens of billions of dollars a year in hardware costs, and make their profit margins significantly better than if they were buying chips from someone else.
He also noted that if Amazon sold its chips on the open market the way rival chip companies do, that business would be worth roughly $50 billion in annual revenue. Right now, nearly all of it is internal. But the implication is sitting there, readable between the lines.
I have been saying for a while that the hyperscaler custom silicon story was under-appreciated. Jassy’s letter is the clearest public confirmation I’ve seen that this thesis is real, it’s scaling, and it’s only going to deepen.
None of this means AI will be universally transformative in every domain, on the timelines people claim. There is still plenty of noise. There are still real questions about where and how the value compounds across the stack.
But the specific concern that AI is all investment and no revenue is harder to sustain after a letter like this. The revenue is real. The silicon advantage is real. And at the layer where it matters most, the infrastructure layer that everything else runs on, the numbers are already large and growing fast.
Thanks for reading!
https://www.aboutamazon.com/news/company-news/amazon-ceo-andy-jassy-2025-letter-to-shareholders