Media Report: Even Amazon Cloud Faces Power Shortages, Billions of NVIDIA Chips Gather Dust in Warehouses

Internal documents from the world's largest cloud computing company reveal an awkward scene in the AI arms race: billions of dollars' worth of Nvidia's latest generation chips are lying idle in Amazon warehouses because they can't be powered up.

According to internal Amazon documents and an employee meeting recording obtained by Business Insider, Amazon's cloud division AWS is temporarily storing Nvidia GB300 racks in "staging warehouses," because the target data centers "have space for more racks, but not enough available power to bring them online."

Internal emails show that Amazon expected to receive about $5 billion worth of GB300 racks in August alone, and to continue receiving about $3 billion worth per month thereafter, maintaining this pace for the "foreseeable future."

At an internal staff meeting in September, AWS CFO John Felton joked with Prasad Kalyanaraman, the executive in charge of AI infrastructure: "You turned our company’s free cash flow negative, congratulations." Kalyanaraman did not refute this, and said the company’s free cash flow is expected to remain negative for "the next few years or so."

In July this year, when Amazon released its Q2 earnings, its rolling free cash flow already turned distinctly negative, recording -$7.6 billion, the first time in many years.

This situation has already had an impact on the market. Chips sitting idle in warehouses provide no computing power supply at all, resulting in ongoing online compute scarcity, keeping GPU rental prices high, and forming a classic inflationary bottleneck—massive amounts of capital, labor, and energy are consumed in advance, while the output is delayed for months or even years.

"Early Rack Receipt": The Chip Battle Forces Advance Payment

Amazon’s usual practice was to pay for equipment only after it was physically installed in a data center, with suppliers covering the cost of devices in transit. However, internal files show that Amazon has rolled out a new process called "Early Rack Receipt," in which it takes ownership of the racks before they even arrive at the data center.

Media interpretations note that competition for Nvidia GPUs is so fierce that Amazon has to lock in hardware before even securing an installation spot to keep it from being snatched by rivals. This means Amazon now pays upfront, powers up later—any installation delays will push back when these assets begin generating revenue, but the cash has already flowed out.

The report also pointed out that at least one AWS supply chain team does not have a "clear predictive path" for "how to get racks into data centers." Amazon’s spokesperson responded that the company’s usual practice is to temporarily store hardware near deployment sites waiting for power to be ready, calling this "best business practice."

Research agency Bernstein determined: Power permits, cooling systems, electrical equipment, construction, and hardware all progress on different timelines, making "timing hard to align perfectly."

Power Bottlenecks: The Core Constraint on Data Center Construction

The root cause of the chip backlog is the severe lag in power infrastructure. Kalyanaraman revealed in internal meetings that in West Texas, Amazon had to pay to build worker housing for pipefitters and electricians due to shortages, allowing construction personnel to live onsite for two to three years before project completion.

According to Cleanview Research, Amazon now has 44GW worth of data center capacity in development across 16 US states—over three times its existing 13.9GW operating capacity. This 13.9GW is already the largest in the industry, 69% higher than Microsoft, and about double that of Google and Meta.

AWS even has an internal dashboard called "Secured Power at Risk," tracking transmission bottlenecks, power rationing requirements, and "use it or lose it" clauses—if a project misses milestone points, Amazon could lose its grid quota.

Kalyanaraman also revealed that in 2023, Amazon promised CEO Andy Jassy that AWS would emerge from capacity constraints before 2027. Now, Amazon expects demand in 2027 will still exceed available capacity.

Chips Pile Up in Warehouses, Yet GPU Rents Are Climbing

By conventional logic, massive chip stockpiles should drive down compute prices. Yet the opposite is true.

The reason is simple: GPUs sitting in crates generate zero GPU hours. What’s truly scarce is "powered compute," and its price continues to climb.

Goldman Sachs Delta-One trading desk head Rich Privorotsky wrote in an internal note on October 16:

"Relative to available supply of investable powered compute, there’s too much capital looking for a home. Repricing will be an ongoing process, sometimes gradual, sometimes drastic, and could morph into systemic risk. The irony is, spot compute pricing remains shockingly strong. B200, H200, H100, and even older A100 chips are holding considerable rental value. Compute which theoretically should be depreciating rapidly is still retaining value. The scarcity of available compute is undeniable, structurally benefitting those who already hold compute."

Goldman Sachs’ Silicon Data B200 spot rental index (non-hyperscale cloud) is at $6.01 per GPU hour, nearing the March record of $6.11, up about 40% from the January low of $4.29.

$8 Billion Chip Sale: Balance Sheet Relief or Something Else?

This background adds a new perspective to a recent Amazon transaction that drew attention. According to the UK’s Financial Times on October 2, Amazon is seeking to sell about $8 billion worth of Nvidia Grace Blackwell chips to outside investors via a special purpose vehicle (SPV), then lease the chips back for its own use. The SPV plans to use debt financing and sell up to 10% of equity; Amazon would hold no shares in the entity, officially aiming for a more "asset-light" balance sheet.

Yet, comparing the transaction to internal documents, the timeline is revealing: In August, Amazon launched the "Early Rack Receipt" process, taking in about $5 billion in GB300 racks, then around $3 billion per month after; in September, AWS’s CFO confirmed turning free cash flow negative; on October 2, Amazon sought to sell $8 billion in chips. The August and September delivery amounts add up precisely to the SPV’s scale.

Sources told the Financial Times that chips involved in the SPV have already been deployed at more than a dozen data centers across five US states, though the report does not specify if these are GB300s. Regardless, investors buying into the SPV are essentially acquiring a rapidly depreciating asset—Grace Blackwell is soon to be replaced by Nvidia’s next-gen Vera Rubin architecture, and every month these chips sit idle in warehouses, a month of their five-year economic life is wasted without generating any revenue.

Who’s Paying Interest on These Dusty Chips?

Microsoft CEO Satya Nadella nearly acknowledged this issue last November, saying, "The biggest problem we now face is not excess compute but a lack of power," and admitting Microsoft "may actually have a bunch of chips sitting in inventory that can’t be plugged in."

By April this year, about half of previously planned US data center projects starting in 2026 had already been cancelled or delayed; of the planned 16GW, only about 5GW was actually under construction.

According to media records, Kalyanaraman’s general principle was, "If we can spend capital as late as possible, it benefits us." The AI arms race has completely flipped this logic: now Amazon races to spend money early, buying chips it can’t yet use—just to keep competitors from getting them first. And if rivals do, odds are they’ll also be unable to use them right away.

If you extrapolate this logic to every hyperscale and emerging cloud provider competing for the same batch of Nvidia allocations, you get a warehouse idle inventory problem measured in the tens of billions of dollars. These stockpiles usually sit under "construction in progress" accounts and, until officially put into service, don’t even count toward depreciation—essentially invisible on the financials…

From a bigger-picture view, this AI infrastructure boom increasingly relies on debt financing. Every month chips remain unpowered means paying interest on a zero-yield asset. Meanwhile, compute rental prices keep rising, and costs are eventually passed on to every AI lab, enterprise, and startup renting GPUs.