AWS's embarrassment: billions in Nvidia GB300 chips gathering dust in warehouses — space aplenty, power none

The most awkward scene of the AI arms race has been exposed by internal documents: Amazon, the world's largest cloud provider, is stacking billions of dollars of Nvidia's newest chips in "intermediate warehouses" — not because it can't buy them, but because it can't plug them in. Per internal files and a staff-meeting recording obtained by Business Insider, AWS is holding Nvidia GB300 racks because target datacenters "have space for more racks but not enough available power to turn them on" — with about $5 billion of racks expected in August alone and roughly $3 billion a month thereafter^[1].

"Congratulations, you've taken our free cash flow negative"

The internal meeting has a black-comedy flavor. AWS CFO John Felton quipped to AI-infrastructure executive Prasad Kalyanaraman at a September all-hands: "You've taken our company's free cash flow negative — congratulations." Kalyanaraman didn't dispute it, acknowledging free cash flow is expected to stay negative "for the next few years or so" — in the July-reported Q2, Amazon's rolling free cash flow came in at negative $7.6 billion, its first in years^[1]. The payment cadence has changed too: abandoning the pay-on-installation convention, AWS introduced an "Early Rack Receipt" process taking ownership before racks even reach the datacenter — read as GPU-arms-race logic: lock the hardware first, wait for power later; any installation delay pushes back revenue-generating time while the cash is already out. At least one supply-chain team admitted having "no clear forecast path" for getting racks into datacenters^[1].

The power bottleneck: 44GW under construction, still short in 2027

The root cause is severely lagging power infrastructure. Kalyanaraman disclosed that in West Texas, shortages of plumbers and electricians forced Amazon to fund workers' housing compounds so crews could live on-site for the two-to-three years before projects finish; per Cleanview Research, Amazon has 44GW of datacenter capacity in development across 16 US states — more than triple its 13.9GW operating base, which is already the industry's largest, 69% bigger than Microsoft's and roughly double Google's and Meta's. AWS even runs a "Secured Power at Risk" dashboard tracking transmission bottlenecks and use-it-or-lose-it grid-allocation clauses; Kalyanaraman also revealed he promised CEO Andy Jassy in 2023 that AWS would emerge from capacity tightness by 2027 — today, the expectation is that demand will still exceed available capacity in 2027^[1]. Bernstein's summary: permits, cooling, electrical work, construction and hardware each advance on its own clock, and "timing is hard to align perfectly"; an Amazon spokesperson said staging hardware near deployment sites pending power is "best business practice"^[1].

The paradox: more chips, higher GPU rents

By ordinary logic, a chip glut should depress compute prices; the reality is the opposite — a GPU in a crate produces zero GPU-hours, and what's truly scarce is powered compute, whose price keeps climbing: a classically inflationary bottleneck, with capital, labor and energy consumed up front while output slips months or years out^[1]. Goldman's Delta-One desk head is cited in support; Microsoft's plight shows Amazon isn't alone — per the market monitor, a Microsoft datacenter with 23,040 chips is stuck on the same power problem^[2]. For the supply chain, two implications: first, the gap between Nvidia's shipment numbers and actually-available compute is widening — the demand signal is distorted; second, the boom in electrical equipment, transformers and grid upgrades may outlast the chip cycle itself — when the chips in warehouses are waiting for watts, the market's attention will eventually shift from GPUs to kilowatts.

What to watch

First, Amazon's capex and free-cash-flow language at its upcoming earnings — if early-rack receipts keep growing, the negative-cash-flow trough deepens. Second, whether "Secured Power at Risk"-style power clauses start appearing in more cloud providers' risk disclosures. Third, order books at grid and electrical-equipment companies — the bottleneck's most direct beneficiaries. Fourth, GB300 rack energization progress — the day the warehouses clear is the day compute supply actually inflects^[1][2].