Internal documents from the world's largest cloud computing provider have revealed an awkward reality behind the AI arms race: billions of dollars worth of NVIDIA (NVDA) latest-generation chips are sitting idle in Amazon warehouses because they cannot be plugged in.
According to internal Amazon documents and a recording of an employee meeting obtained by Business Insider, Amazon.com (AMZN) cloud division AWS is temporarily storing NVIDIA GB300 racks in "intermediate warehouses" because target data centers "have room for more racks but not enough available power to turn them on."
Internal emails show that Amazon alone expected to receive approximately $5 billion worth of GB300 racks in August, and continues to receive roughly $3 billion worth each month thereafter, maintaining this pace "for the foreseeable future."
AWS Chief Financial Officer John Felton joked to Prasad Kalyanaraman, the executive overseeing AI infrastructure buildout, at an internal employee meeting in September: "You turned our company's free cash flow negative. Congratulations." Kalyanaraman did not dispute this and said the company's free cash flow is expected to remain negative for "roughly the next few years."
In July, when Amazon reported its second-quarter earnings, its trailing free cash flow had clearly turned negative, recording negative $7.6 billion, the first time in years.
The impact of this situation on the market is already evident. Chips sitting idle in warehouses cannot generate any computing supply, leading to persistent scarcity of online computing power, with GPU rental prices remaining elevated, forming a classic inflationary bottleneck — massive amounts of capital, labor, and energy are consumed in advance, while output is delayed by months or even years.
Early Receipt: Chip Race Forces Payment Upfront
Amazon's previous practice was to pay only after equipment was actually installed in data centers, with suppliers bearing the financing cost of in-transit equipment. However, internal documents show that Amazon has introduced a new process called "Early Rack Receipt," acquiring ownership before racks even arrive at data centers.
Media interpretation suggests that competition for NVIDIA GPUs is so intense that Amazon must lock in hardware before installation locations are even available, lest competitors seize them first. This means Amazon now pays first and powers up later, and any installation delay will push back the point at which these assets begin generating revenue — while the cash has already flowed out.
The report also noted that at least one AWS supply chain team has no "clear forecast path" for "how to get racks into data centers." In response, an Amazon spokesperson said the company's usual practice is to temporarily store hardware near deployment sites while waiting for power to be ready, calling this "best commercial practice."
Research firm Bernstein assessed that power permits, cooling systems, electrical equipment, construction, and hardware each advance at different paces, making it "difficult to perfectly align the timing."
Power Bottleneck: The Core Constraint on Data Center Construction
The root cause of the chip backlog lies in severe delays in power infrastructure. Kalyanaraman revealed at an internal meeting that in West Texas, Amazon had to fund the construction of worker housing due to shortages of plumbers and electricians, allowing construction workers to live on-site for two to three years until project completion.
According to Cleanview Research data, Amazon currently has 44 gigawatts of data center capacity under development across 16 states, more than triple its existing operating scale of 13.9 gigawatts. And that 13.9 gigawatts is already the largest in the industry, 69% more than Microsoft and roughly twice that of Google and Meta.
AWS even has a dedicated internal dashboard called "Secured Power at Risk" that tracks transmission bottlenecks, curtailment requirements, and "use it or lose it" clauses — if a project misses milestone deadlines, Amazon could lose grid allocations it has already secured.
Kalyanaraman also revealed that Amazon had promised CEO Andy Jassy in 2023 that AWS would emerge from capacity constraints by 2027. Now, Amazon expects demand to still exceed available capacity in 2027.
Chips Pile Up in Warehouses, Yet GPU Rents Rise
By conventional logic, a large backlog of chips in warehouses should push down computing prices. But the opposite is happening.
The reason is simple: GPUs sitting in crates produce zero GPU hours. What is truly scarce is "powered computing," and its price continues to climb.
Rich Privorotsky, head of Goldman Sachs' Delta-One trading desk, noted in an internal report on October 16:
"Relative to the available supply of investable powered computing, there is too much capital looking for a home. Repricing is a process the market will continue to experience, sometimes gradually, sometimes violently, and it could evolve into systemic risk. The interesting contradiction is that spot computing pricing remains extremely strong. B200, H200, H100, and even older A100s all maintain substantial rental rates. Computing that theoretically should depreciate quickly is still holding value. The scarcity of available computing is undeniable, and this is structurally favorable for those who already hold computing power."
The Silicon Data B200 spot rental index tracked by Goldman Sachs (non-hyperscale cloud) stands at $6.01 per GPU hour, close to the historical record of $6.11 in March and up about 40% from the January low of $4.29.
$8 Billion Chip Sale: Balance Sheet Relief or Something Else
The above background provides a new interpretive lens for a recent noteworthy Amazon transaction. According to a Financial Times report on October 2, Amazon is seeking to sell approximately $8 billion worth of NVIDIA Grace Blackwell chips to external investors through a special purpose vehicle (SPV), then lease the chips back. The SPV plans to finance through debt and sell up to 10% equity, with Amazon holding no stake in the entity, with the official goal being a more "asset-light" balance sheet.
However, comparing this transaction with internal documents, the timeline is quite intriguing: in August, Amazon launched the "Early Rack Receipt" process, receiving approximately $5 billion in GB300 racks that month, followed by about $3 billion per month; in September, the AWS CFO confirmed at an internal meeting that the company's free cash flow had turned negative; on October 2, Amazon sought to sell $8 billion in chips externally. The sum of the expected delivery amounts for August and September exactly equals the SPV size.
Financial Times sources said the chips involved in the SPV have been deployed across more than a dozen data centers in five states, and the report did not clarify whether they are GB300s. Regardless, investors buying into the SPV are essentially purchasing a continuously depreciating asset — Grace Blackwell will soon be replaced by NVIDIA's next-generation Vera Rubin architecture, and every month a chip sits idle in a warehouse means one month of its five-year economic life is wasted without generating any revenue.
Who Is Paying Interest on These Dust-Collecting Chips
Microsoft CEO Satya Nadella came close to publicly acknowledging this problem last November, saying: "The biggest problem we face now is no longer excess computing power, but power," and admitting that Microsoft "might actually have a bunch of chips sitting in inventory that I can't plug in."
By April of this year, about half of US data centers originally planned to launch in 2026 faced cancellation or delay, with only about 5 gigawatts actually under construction out of 16 gigawatts planned.
According to media records, Kalyanaraman's own principle is: "If we can spend capital as late as possible, that's better for us." The AI arms race has completely overturned this logic: Amazon now rushes to spend early, buying chips it cannot use, just to prevent competitors from taking them first — and competitors, having obtained them, most likely also cannot use them.
Extending this logic to every hyperscale cloud provider and emerging cloud service provider competing for the same batch of NVIDIA allocations yields a warehouse idle problem measured in tens of billions of dollars. Most of this inventory is recorded under "construction in progress," and before officially being put into use, it does not even enter depreciation calculations, leaving no trace on financial statements...
From a broader perspective, this AI infrastructure construction boom is increasingly relying on debt financing. Every month a chip remains unpowered means paying one month of interest on a zero-return asset. Meanwhile, computing rents continue to climb, with costs ultimately passed on to every AI lab, enterprise, and startup renting GPUs.
Comments