Andy Jassy's shareholder letter, published Thursday, runs thousands of words and opens with a story about wanting to be a sportscaster. But buried past the hockey anecdotes and New Zealand band references is the real pitch: Amazon plans to spend roughly $200 billion in capital expenditures this year, mostly on AI infrastructure, and its free cash flow just cratered 71% to fund the bet.
The Graviton playbook, applied to GPUs
The most revealing section of the letter isn't the $200 billion figure, which Amazon already disclosed during its February earnings call. It's Jassy drawing an explicit line from Graviton to Trainium, framing Amazon's custom AI chips as the sequel to the company's successful effort to displace Intel in cloud computing. Graviton, he notes, now runs across 98% of AWS's top 1,000 EC2 customers. He wants the same trajectory for Trainium against NVIDIA.
"The same story arc is unfolding in AI," Jassy writes, which is a polite way of telling NVIDIA that Amazon sees their partnership as transitional. He's careful to say AWS will "always have customers who choose to run NVIDIA" and will keep making AWS the best place for those chips. Diplomatic, sure, but the spending tells a different story.
Amazon's custom chips business, which bundles Graviton, Trainium, and the Nitro networking card, has hit a $20 billion annual revenue run rate with triple-digit year-over-year growth. Then comes the number Jassy clearly wants investors focused on: if the chips unit were a standalone company selling to third parties, he estimates its annual run rate would be around $50 billion. That's a figure designed to make you think about NVIDIA's market cap, not Amazon's operating margins.
How real is $50 billion?
It's worth being skeptical here. The $50 billion figure is a hypothetical, not a revenue line. Amazon currently monetizes Trainium only through EC2 instances, not as standalone chip sales. Jassy is essentially saying: if we sold chips the way NVIDIA does, here's what it would look like. That's a thought exercise, not accounting.
The $20 billion run rate is more concrete but still bundles three different chip families. How much of that is Trainium versus the more mature Graviton? The letter doesn't say. And internal documents from early 2024, previously reported by Business Insider, showed Trainium adoption among AWS's largest customers at just 0.5% of NVIDIA GPU usage. That was over a year ago, and Trainium2 has since shipped at scale for projects like Anthropic's Project Rainier, which uses over 500,000 of the chips. But the gap between "Anthropic uses it" and "enterprise customers are switching" remains large.
Trainium2 offered roughly 30% better price-performance than comparable GPUs, according to Jassy. Trainium3, which started shipping in early 2026, pushes that to 30-40%. He says Trainium3 is already "nearly fully-subscribed" and that customers have reserved significant capacity on Trainium4, still about 18 months out.
The cash flow problem nobody's ignoring
Free cash flow dropped from $38.2 billion to $11.2 billion in 2025. Operating cash flow actually grew 20% to $139.5 billion, but property and equipment purchases surged by $50.7 billion. The entire letter is structured as an argument that this is fine.
Jassy's defense follows a familiar template: AWS has to spend money on land, power, buildings, chips, and servers 6-24 months before it can bill customers. The assets last decades (30+ years for data centers, 5-6 years for compute gear). So the FCF compression is temporary, a function of growth outpacing monetization. "We've been through this cycle with the first big AWS growth wave, and liked the results."
Maybe. But $200 billion is not the first AWS growth wave. Amazon spent $131.8 billion on capex in 2025, up from $83 billion in 2024. The 2026 plan represents another 52% jump. At some point, "we've done this before" stops being reassuring when the numbers keep doubling.
Jassy points to customer commitments as proof this isn't speculative. The OpenAI deal, worth over $100 billion, anchors the claim, alongside "several other customer agreements completed (and unannounced), or deep in process." AWS's AI revenue hit a $15 billion annual run rate in Q1 2026, which Jassy describes as 260 times larger than AWS itself was at the same point in its development.
What the letter skips
For all its length, the letter avoids some obvious questions. There's no mention of how Trainium compares to Google's TPU Ironwood, which launched recently and has its own aggressive pricing story. No acknowledgment that NVIDIA's CUDA software ecosystem remains a massive moat that price-performance alone doesn't overcome. Jassy references the Graviton-Intel parallel repeatedly, but Intel's software lock-in was nothing compared to CUDA's.
AWS also reported 24% year-over-year growth on a $142 billion revenue run rate in Q4 2025. Strong, but Jassy admits the company still has "capacity constraints that yield unserved demand." Two large customers apparently asked to buy all of AWS's Graviton capacity for 2026. The humble-brag underscores a real tension: Amazon is spending $200 billion partly because it can't serve demand fast enough, which is either the best possible reason to spend or a convenient justification for capex that's taken on a life of its own.
Amazon's stock dropped around 11-15% after the February capex announcement. The shareholder letter reads like Jassy's attempt to stop the bleeding, complete with a closing declaration that reads less like corporate communication and more like a dare: "We're not going to be conservative in how we play this."
Trainium4 is expected roughly 18 months from now. By then, we'll know whether the Graviton playbook actually translates to AI silicon, or whether Jassy just wrote a very long letter about a very expensive hypothesis.
