Amazon is taking ownership of Nvidia GB300 server racks before the data centers or power capacity exist to run them. According to Business Insider, an internal email projected roughly $5 billion in GB300 rack deliveries in August alone, followed by approximately $3 billion per month after that, under a program AWS calls “Early Rack Receipt.”
For Amazon, the constraint is no longer just about securing chips. It is increasingly about having somewhere to actually run them.
Hardware First, Power Later
The “Early Rack Receipt” program shifts ownership of server racks to AWS earlier in the supply chain, before deployment sites are ready to receive them.
AWS assumes ownership of racks before facilities are prepared, meaning the company commits capital before it can generate revenue from the hardware.
Some AWS facilities reportedly had physical space for additional racks but lacked sufficient power infrastructure to activate them, leading Amazon to stage equipment in intermediary warehouses until sites were ready. Business Insider also reported that at least one supply-chain team lacked “a clear forecasted path” for moving specific incoming racks into data centers.
Amazon disputed the broader implication. A spokesperson described staging hardware near deployment sites as “best business practice,” and the company said AWS has long-established systems for tracking infrastructure assets and installation plans. The evidence Business Insider presented points to at least one team facing a planning gap at a specific point in time. That is not the same as the entire program lacking direction.
A $220 Billion Bet With No Guarantee of Timing
Amazon’s AI infrastructure spending has pushed the company into negative free cash flow territory, with no quick return in sight.
AWS CFO John Felton remarked during a September internal meeting, reviewed by Business Insider, that the AI expansion had made Amazon free-cash-flow negative. Prasad Kalyanaraman, Amazon’s vice president of infrastructure services, described the machine-learning and AI business specifically as “hugely free cash flow negative” and said that condition could persist for the next couple of years.
Amazon is projected to spend approximately $220 billion in capital expenditures this year, heavily weighted toward data centers, servers, power, and related infrastructure. That figure leads the other major cloud rivals cited in Business Insider’s reporting: Google at up to $205 billion, Microsoft at approximately $175 billion, and Meta at up to $145 billion.
Bernstein analysts summarized the situation in a note cited by Business Insider: “Timing is imperfect.” Dan Ives of Yorkville Ives put it more directly, writing that the market is “repricing anything that funds multi-year builds against revenue that arrives later.” Amazon’s position resembles a restaurant that purchased two years of ingredients before signing a lease.
The Strategic Logic Behind the Risk
Amazon is making an uncomfortable bet that securing scarce hardware now will outweigh the cost of deploying it later.
Competing cloud providers are pursuing the same Nvidia systems, and waiting risks losing access to components that may be unavailable later. CEO Andy Jassy has described AI as a once-in-a-lifetime opportunity and has argued against conservative investment in this cycle.
Amazon has absorbed front-loaded spending before, across AWS, fulfillment centers, and last-mile delivery. The coordination complexity here is considerably steeper. Electricity procurement, cooling systems, construction permits, and skilled labor all move on different timelines, and the shortages are tangible. Kalyanaraman told staff that AWS funded housing complexes near West Texas data-center sites to attract pipefitters and electricians for projects spanning multiple years.
Amazon tracks electricity commitments through an internal dashboard called “Secured Power at Risk,” which monitors grid constraints and provisions that could jeopardize contracted capacity if projects miss milestones, according to Business Insider.
Amazon reportedly expects AI demand may still outpace available capacity into 2027. Whether the warehoused hardware proves prescient or premature depends on how fast power infrastructure and construction can close the gap.
