Amazon's CFO had one line for his infrastructure chief in September: "You've taken us free cash flow negative as a company, so congratulations." John Felton said it to Prasad Kalyanaraman, and the joke describes the actual state of AI infrastructure in 2026.
The racks arrive faster than the power does.
$5 billion of NVIDIA GB300 deliveries landed in August alone, then about $3 billion a month after that.
Business Insider reports that some AWS data centers have the physical space for those racks but not enough power to run them, so the equipment is staged in warehouses.
Amazon takes ownership and pays early, long before a single rack earns anything.
The bill shows up in free cash flow.
Roughly $50 billion in Q1 2024, about $1 billion in Q1 2026, negative $7.6 billion in Q2 2026.
Capital spending is targeted at $200 billion for 2026, and the CFO says the situation will not improve for a couple more years.
Now picture a two-person team with a model to test this quarter.
They read that number and conclude compute is closed to them.
It is not closed.
You do not need a building, a power contract, or a five-year capital plan to run a training job.
You need a GPU for the hours you actually use it.
RTX 3090s at $0.15/hr, billed for the time you run, and Buy Now Pay Later means zero upfront with payment after one month, so a test can start before the budget is approved.
What is the longest you have waited on power, a rack, or a purchase order before a GPU could actually run?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w43&utm_content=morning
Originally posted on linkedin_personal