Billions In Nvidia Chips Collecting Dust In Amazon Warehouses As AI's "Paperweight" Problem Goes Mainstream
Billions In Nvidia Chips Collecting Dust In Amazon Warehouses As AI's "Paperweight" Problem Goes Mainstream
For much of the past year, we have been pounding the table on what is probably the most underappreciated problem in the AI capex boom.
It goes like this: hyperscalers order hundreds of billions in Nvidia silicon, the chips get built, shipped and delivered on schedule... and then they sit in crates, because the data center they were meant for has no power, no transformers, or no roof. Until those chips are plugged in, they earn nothing, they depreciate anyway, and the scarcity they leave behind keeps the price of every GPU-hour that is online sky-high. In other words, a stockpile of unlit chips is the world's most expensive, and very inflationary, paperweight.
Now we have confirmation from inside the biggest cloud of them all.
According to a Business Insider report based on internal Amazon documents and a recording of an AWS staff meeting, Nvidia's newest GB300 racks, arriving at a clip of billions of dollars a month, have been staged at "intermediary warehouses" because the AWS sites they were meant for "had room for more racks but not enough available power to turn them on." Or, as we summarized it last night:
Billions in GB300s ... collecting dust in dark warehouses https://t.co/gfedEyPciR
— zerohedge (@zerohedge) October 20, 2026
The "Early Rack Receipt"
The details are worth walking through, because they show exactly how a chip shortage turns into a chip glut without anyone ever cancelling an order.
Per BI, internal emails from August show Amazon preparing a new "Early Rack Receipt" process for an expected wave of server racks carrying Nvidia's GB300 chips at US facilities. One email projected that Amazon would take delivery of about $5 billion of GB300 racks in August, followed by roughly $3 billion each month " for the foreseeable future ."
Normally, Amazon doesn't pay for racks until the gear is actually installed in a data center, which leaves the suppliers to carry the cost of equipment in transit (a very Amazonian way to conserve cash). Under the new process, however, Amazon takes ownership of the racks earlier in the supply chain, before they ever reach a data center . BI explains why: the scramble for Nvidia GPUs is so intense that Amazon has to compete harder for scarce components, so it grabs the hardware when it can, whether or not there is a socket waiting for it.
Translation: Amazon is now paying for chips sooner, and plugging them in later. Any delay in getting racks installed and connected to reliable power "potentially pushes back the moment when revenue starts flowing from these assets," as BI puts it, while the cash is already gone.
And it gets better: at least one AWS supply-chain team did not have " a clear forecasted path " for getting racks into data centers, according to one of the emails. Amazon's spokesperson pushed back, saying the company routinely stages finished hardware near deployment sites until power is available and calling the practice " best business practice ." Which is probably true... and also exactly what you would say about $3 billion a month of very expensive inventory sitting in a warehouse.
Bernstein, cited by BI, offered the sell-side's version of a shrug: power permits, cooling systems, electrical equipment, construction and hardware all move on separate schedules, and " timing is imperfect ."
"Congratulations"
The most quotable moment, however, came from AWS CFO John Felton, who at an internal September staff meeting had a joke for Prasad Kalyanaraman, the executive running much of Amazon's AI infrastructure buildout:
" You've taken us free cash flow negative as a company, so congratulations. "
Kalyanaraman did not argue. Amazon's machine-learning and AI business is " hugely free cash flow negative ," he said, and he expects the company's FCF to stay negative for " the next couple of years or so ." As a reminder, three months ago we reported that for the first time in years, Amazon's trailing free cash flow turned decidedly negative, -$7.6 billion, when it reported Q2 in July .
Two years ago, Wall Street was debating what Amazon should do with its cash pile; this year the company is spending $220 billion in capex, and the cash pile is something people reminisce about at staff meetings.
He also explained why the warehouses keep filling up. Amazon tries to commit capital at the "last responsible moment," when it has the clearest view of demand, but the forecast " is changing on a daily basis ," including when data centers will be ready and how much equipment AWS needs to keep on hand. An eight-week supply buffer might be routine for a smaller company, he said, but at Amazon's scale it can mean holding " billions and billions of dollars " of extra equipment inventory.
Some other gems from the BI piece:
Kalyanaraman told Andy Jassy in 2023 that AWS would "surely" be out of its capacity crunch by 2027. Amazon now expects demand to exceed available capacity in 2027 too.
An internal AWS dashboard is literally called " Secured Power at Risk ," tracking transmission bottlenecks, curtailment requirements and "use it or lose it" clauses that can cost Amazon its grid allocation if a project misses milestones.
In West Texas there weren't enough pipefitters and electricians, so AWS is paying to build housing complexes for workers who will live on site for the two or three years it takes to finish the projects.
Even the crates the servers ship in have become a material line item. " We have new constraints that keep popping up ," Kalyanaraman said.
“Per Cleanview Research, Amazon has 44 gigawatts of data center capacity under development across 16 states, more than three times the 13.9 GW it operates today. And that 13.9 GW is already the biggest fleet in the business: 69% more than Microsoft and roughly double Google and Meta . That’s a lot of sockets that will need a lot of power before the racks in the warehouses can fill them.
"A $1 Trillion Inventory Paperweight"
None of this will come as a surprise to regular readers. Back in June, when the debate was still about whether AI capex would ever earn a return, we argued that the market was looking at the wrong metric entirely:
ROI is not the right metric for AI. The right metric is what is the ROI assuming all the planned data centers are built and energized for the IRR period. SPVs have made funding a non-issue (for now). But if components remain unlit, they are a $1 TN inventory paperweight
— zerohedge (@zerohedge) June 5, 2026
Ten days later, after noting that it takes 2.5 years just to get the power transformers for a new data center, we pointed out that the US grid interconnection queue alone was already more than double the entire installed US power plant fleet :
It gets worse: the grid connection requests active at the end of 2023 were more than double the total installed capacity of the US power plant fleet (2,600 GW vs. 1,280 GW). It's well more than 3,000 GW. There will be $trillions in semiconductor paperweights over the next decade
— zerohedge (@zerohedge) June 15, 2026
And last month, as Oracle's Project Jupiter campus in New Mexico (built to deliver $300 billion of compute to OpenAI) ground to a halt under a force majeure declaration ( and looks increasingly like it may never be completed ), we asked the obvious follow-up question:
Would be a fun homework assignment to calculate how many thousands of B100 and petabytes of HBM chips are sitting in crates, unlikely to be used for years, yet pushing core PCE up by 0.5% this year
— zerohedge (@zerohedge) September 21, 2026
We weren't alone in suspecting it. Microsoft's Satya Nadella all but admitted it last November, when he said that " the biggest issue we are now having is not a compute glut, but it's power " and that Microsoft "may actually have a bunch of chips sitting in inventory that I can't plug in" ( we covered it here ).
By April, half of the US data centers scheduled to start in 2026 were set to be canceled or delayed , with only about 5GW of the planned 16GW actually under construction. Michael Burry went further last December, calling for "pics and proof" of Nvidia GPUs being warehoused in mass quantities . Back then it sounded like a conspiracy theory. Now we have Amazon's own emails describing an entire process built around receiving racks before there is anywhere to put them (and that's before even considering double (or triple) ordering).
And if Amazon, which by most accounts is the best (or among the best) data center builder on the planet, can't plug its chips in on time, maybe we should ask how everyone else is doing.
Idle Chips, Record Rents
Which brings us to the part of the story that matters for everyone who doesn't own Amazon stock: inflation .
Logic would suggest that a glut of chips in warehouses should crush the price of compute. The opposite is happening, for the simple reason that a GPU in a crate produces exactly zero GPU-hours. What is scarce is energized compute, and its price keeps going up.
Goldman's Rich Privorotsky, head of the bank's Delta-One trading desk, made precisely this point in his October 16 "GS Basics " note, ( available to pro subs ):
"There is too much capital seeking a home relative to the available supply of investable, energized compute . Repricing is a process the market will continue to go through, sometimes gradually, sometimes acutely, with the potential to become systemic. The interesting contradiction is that spot compute pricing remains incredibly strong. B200s, H200s, H100s, even ageing A100s continue to command meaningful rental rates. Compute that should theoretically be depreciating rapidly remains valuable. There is an undeniable scarcity of available compute , which is structurally bullish for those who already own it."
His chart shows the Silicon Data B200 on-demand rental index (non-hyperscale) at $6.01 per GPU-hour, within a hair of the March record of $6.11 and up 40% from January's low of $4.29 :
Goldman's equity research sees it the same way. In upgrading CoreWeave to Buy yesterday, Goldman's Gabriela Borges argued that compute pricing is " biased to the upside " over the next 24 months (a call HSBC promptly took apart, as we discussed in " HSBC Slams Goldman's CoreWeave Bull Case" earlier today).
Put the two together and you have the textbook definition of an inflationary bottleneck: enormous amounts of capital, labor, copper and electricity consumed upfront, with the output (compute) delayed for months or years, while the price of whatever capacity is available gets bid up . Every AI lab, enterprise and startup that rents GPUs pays that scarcity premium, and passes it on.
The chips sitting in Amazon's staging warehouses were supposed to be the supply response that brings those rental rates down. Instead they're collecting dust!
Revisiting Amazon's $8 Billion "Asset-Light" Chip Sale
Which also casts a rather different light on a story from three weeks ago.
On October 2, the FT reported that Amazon was seeking to offload about $8 billion of Nvidia Grace Blackwell chips to outside investors through a special-purpose vehicle. Amazon would then lease the chips back from the SPV, which would raise money through debt (expected to get an investment-grade rating off Amazon's AA), with up to 10% of the equity sold as well, so that Amazon "will not own any stakes in the entity." The stated goal: a more "asset-light" balance sheet.
At the time we focused on the obvious angle:
*AMAZON SEEKS TO OFFLOAD $8B NVIDIA CHIPS TO INVESTORS: FT SPV sale leaseback. More off-balance sheet gimmicks
— zerohedge (@zerohedge) October 2, 2026
Later that day we followed up with " Amazon, Broadcom Pile Into Off-Balance Sheet SPVs As AI Circularity "Undeniably Increasing" ", adding the Amazon deal to our running tally of hyperscaler, Nvidia and Broadcom off-balance sheet commitments, which by then had hit $3.1 trillion ( up $1.3 trillion in three months ). Sell the chips, rent them back, call it "asset-light."
But with the BI report in hand, the sale looks like it may have been about a lot more than accounting optics. Consider the timeline:
August: Amazon rolls out "Early Rack Receipt," slated to take title to ~$5 billion of GB300 racks that month and ~$3 billion a month after that, some of which have no powered data center to go into.
September: The AWS CFO "congratulates" the infrastructure chief for taking Amazon FCF negative.
October 2: Amazon is reported to be shopping $8 billion of Nvidia chips to outside investors.
Five plus three is, coincidentally, exactly eight : the projected August and September GB300 deliveries add up to the precise size of the SPV. Now, to be fair, the FT's sources said the chips in the deal have been "deployed" in more than a dozen US data centers across five states, and the report doesn't say whether they are GB200s or the newer GB300s. So maybe it's just a coincidence. But it is a very convenient one: a company that started taking ownership of billions in racks it couldn't power, and whose cash flow just turned negative, went looking for someone else to own (and finance) $8 billion of Nvidia silicon within weeks .
And whoever buys into that SPV is buying a depreciating asset. The FT notes Grace Blackwell "will soon be superseded" by Nvidia's Vera Rubin, while Amazon's filings assume each chip generation lasts at least five years. Every month a rack sits in a warehouse is a month of that five-year economic life burned without a single dollar of revenue (even if the accountants don't start the depreciation clock until it's plugged in). Then again, when your July $25 billion bond deal already ran into weaker demand for long-dated paper, as the FT also notes, selling chips to insurers and pension funds via an IG-rated SPV is probably the path of least resistance.
Past The "Last Responsible Moment"
Kalyanaraman's own philosophy, as recorded by BI, is that " if we can wait to spend the capital as late as possible, it's better for us ." The AI arms race has flipped that on its head: Amazon is now spending as early as possible, on chips it can't use, just to make sure a competitor doesn't get them first (where they are likely to also sit unused). Multiply that by every hyperscaler and neocloud fighting for the same Nvidia allocation, and you get a warehouse problem measured in the tens of billions, most likely parked in "construction in progress, " where, as we explained in " The Teaser Period ", it doesn't even start depreciating until it is placed in service. Out of sight, and out of the income statement... for now.
Our take: the glut gets bigger before it gets smaller. Much bigger. Nvidia keeps shipping, the "for the foreseeable future" $3 billion a month keeps arriving, and the bottleneck is still power, which no SPV can finance into existence on a 2.5-year transformer backlog. That is why we keep saying "behind the meter" generation has to become mandatory (think modular reactors, gas turbines, and batteries). And since all of this is increasingly financed with debt ( as we have been warning since last October ), every month a chip sits unplugged is a month of interest paid on an asset that earns nothing. In the meantime, the paperweights keep compute scarce, rental rates keep climbing, and the bill shows up in everyone's cost base. As for the capital markets, they have already started voting on the companies building the warehouses' worth of capacity with borrowed money:
Amazon is (for now) a very long way from that list. But if the best data center builder in the world needs warehouses, early rack receipts and an $8 billion SPV to keep the conveyor belt moving, we'd hate to see what's sitting in the crates of the neoclouds that rely on 15% bonds to pay for theirs.
Much more in the full Goldman Delta One's " GS Basics " note and CoreWeave upgrade note , both available to pro subs .
Tyler Durden
Tue, 10/20/2026 - 14:32