AWS AI revenue and massive capex

Amazon is saying the quiet part out loud: the artificial intelligence boom is no longer just a cost center for Amazon Web Services. Andy Jassy said Amazon Web Services is now running at more than $15 billion a year in artificial intelligence revenue, the first time Amazon has put a hard number on that business. (reuters.com) He paired that with an even bigger number: about $200 billion in capital spending in 2026, with most of it going into artificial intelligence infrastructure. That means data centers, servers, networking gear, and the power systems needed to keep large models running around the clock. (cnbc.com) Amazon Web Services is Amazon’s cloud rental business, where companies pay to use computing instead of buying their own machines. In full-year 2025, that unit produced $128.7 billion in sales and $45.6 billion in operating income, which is why Amazon can fund a buildout this large. (aboutamazon.com) The new artificial intelligence number is big, but it is still only a slice of the cloud business. Amazon said Amazon Web Services reached a $142 billion annualized revenue run rate in the fourth quarter of 2025, so $15 billion suggests artificial intelligence is already a meaningful workload, not the whole machine. (aboutamazon.com) (reuters.com) Jassy also pointed to a second business hidden inside the first one: Amazon’s homegrown chips. He said Graviton, Trainium, and Nitro together are now running at more than $20 billion a year of annualized activity, up from more than $10 billion that Amazon cited for its custom chip business in February 2026. (finance.yahoo.com) (fool.com) Graviton is Amazon’s main central processor, the general-purpose brain that runs ordinary cloud workloads. Amazon says Graviton-based servers can deliver up to 40% better price performance than comparable x86 systems, which is how Amazon tries to make cloud computing cheaper while keeping margins healthy. (aws.amazon.com) Trainium is the chip built for training large artificial intelligence models, which is the expensive step where a model learns from giant piles of data. Amazon has been pitching Trainium as a lower-cost alternative to scarce Nvidia graphics processors, and in 2023 it said customers including Databricks were already using Trainium systems for large-scale model training. (press.aboutamazon.com) Nitro is the less glamorous piece, but it is the hardware that offloads networking, storage, and security jobs from the main server. That matters because every task Nitro handles is one less task stealing cycles from the chips customers are actually paying to use. (aws.amazon.com) The backdrop is that Amazon’s cash flow already took a hit from this buildout. GeekWire reported Jassy said free cash flow fell from $38 billion to $11 billion last year after a $50.7 billion increase in capital spending, mostly tied to artificial intelligence infrastructure. (geekwire.com) That is why Amazon is suddenly disclosing numbers it used to keep vague. Investors have spent two years asking whether cloud companies were buying chips first and figuring out the business later, and Amazon’s answer is that demand is already large enough to justify building at industrial scale. (cnbc.com) (reuters.com) The clearest read on this is not that Amazon found a side business worth $15 billion. It is that Amazon is trying to own every expensive layer of cloud artificial intelligence at once: the data centers, the rented computing, and the chips inside the machines. (aws.amazon.com) (aboutamazon.com)