AI's $100 Billion Inventory Pile: Bubble Warning or Boom?

What is happening

Shares in the companies that make the chips, servers, networking and power equipment for AI data centres have fallen hard. Of the 191 electronics and components makers we rate, the typical one fell 10.5% in the week and 23.9% in the month to 4 November, and some fell much further: Intel lost 39% in the week and Marvell Technology 36%. At the centre of the worry is not the shares but what each company holds in its warehouses: its inventory, the finished goods and parts waiting to be sold. Together, the eight companies in this report held about $102bn of it at their latest quarter-ends, about double the $50bn of a year earlier.³⁰ Dell's has nearly tripled in a year, to $21.3bn.¹⁹ Nvidia holds $31.6bn,²² and AMD $11.5bn, the most in its accounts since 2018 measured against its sales.¹⁷ A useful way to judge such a pile is to ask how many days it would last if the company kept selling at its current pace; AMD's would last about 170.¹⁷ When that number climbs, either a company is stocking up for orders it can see coming, or it is holding goods nobody wants as quickly as it hoped. The fear is the second: that this inventory was ordered for demand that is now fading. The other side says it is the first: parts for sales already booked, held by companies that still cannot build fast enough. We set the two sides against each other, look at which share prices already carry the fear, and name the signals that would tell one story from the other before any company's accounts do.

The fear

The first part of the fear is simple arithmetic. When inventory grows much faster than sales, goods are arriving quicker than they leave, and that is what a glut looks like before anyone calls it one. Dell's inventory rose 195% in a year while its sales grew 58%, so a pile that would once have lasted about 27 days of sales now lasts about 52.¹⁹ Vertiv, which makes the power and cooling equipment that keeps data centres running, saw its inventory rise 129% against sales up 61%, taking it to about 112 days from 79.²¹ Cisco's rose 80% against sales up 18%, to about 84 days from 54, the most in its accounts since 2018.¹⁸ Dan Niles, who runs the Satori Fund, has argued that chip supply and demand are 'a lot closer than most companies think': Dell, he points out, could not have tripled its inventory in a year if parts were as scarce as the chipmakers say.¹

The second part concerns the customers, who are taking delivery of hardware faster than they can switch it on. Amazon, according to internal emails from August, set up an 'Early Rack Receipt' scheme to accept about $5bn of Nvidia's GB300 server racks, and roughly $3bn more each month, before its data centres were ready to run them; some of the hardware sits in warehouses because the sites lack the electricity to power it.³ Amazon called staging hardware near its sites 'best business practice' and disputed that it had no plan to bring them online.³ Then the biggest spenders paused their ever-rising forecasts. Microsoft cut its spending forecast for the year to $170bn from $175bn, and Alphabet left its own unchanged, both below what analysts had expected after quarters in which every update went up.⁴ For a hardware maker, a customer that stops raising its plans is the first sign that the orders behind today's inventory may not be followed by more.

The third part is about promises and money. Nvidia's commitments to buy from its own suppliers rose from $16.1bn to $95.2bn in twelve months. The investor Michael Burry called that 'troubling' and 'not business as usual', and compared it with Cisco before the dot-com crash, which later had to write down both inventory and supplier commitments when demand collapsed.⁶ Broadcom's chief executive Hock Tan has warned that the AI sector can face funding constraints, and likened some investors' strategies to gambling.⁵ In the same weeks JPMorgan and Goldman Sachs limited how much of a loan to the AI start-up Volta Infrastructure they would hold, and the stock-market listing of Firmus Grid, a data-centre operator backed by Nvidia, collapsed.⁵ At AMD, the investment writer Daniel Sparks notes, inventory grew by about $3bn in three months while the cash the company generated shrank to $554m; AMD, he argues, is 'paying up front for next year's sales', and if customers delay their projects that $11.5bn of inventory 'might become a problem'.² Money has also become dearer in the country where most of these data centres are being built. The US Federal Reserve raised its policy rate to 4.00% in September, and the yield on ten-year US government bonds, a benchmark for long-term borrowing, stood at 4.95% on 3 November, 0.82 points higher than a year earlier. Where data centres are paid for with loans, that raises the cost of every rack.

The case that it is overdone

The strongest answer starts with timing. In the last chip glut, inventory piled up only after sales had stopped growing; this time it is building while sales growth is still speeding up. Nvidia's sales grew 106% in the quarter to July, against 56% a year earlier, and AMD's 45% in the quarter to September, against 36%.²⁶ In 2022, by contrast, Nvidia's sales growth fell from 53% to 3% over the first half of the year, and only then did its inventory jump, from 94 days of sales to 212, as its sales fell 21%. AMD's went from 69 days to 143 as its growth turned from 70% into a fall of 18%, and Micron's from 120 to 239 as its sales fell 47%.²⁵ A pile that grows alongside accelerating sales is a different animal from one that grows because sales have stopped. Nvidia's inventory rose 111% in a year, about as fast as its sales; at about 120 days it is above its usual range of 80 to 95, but not wildly so.²² Its finance chief Colette Kress said the build was to prepare for the launch of Vera Rubin, its next generation of AI systems, and that most of its supplier commitments fall in the next three years and will be used 'to build the products that we need'.¹²

The second answer is that the customers are still spending more, not less. Microsoft, Alphabet, Amazon and Meta together spent about $202bn on buildings and equipment, mostly data centres, in the September quarter, against about $97bn a year earlier, and more than in the quarter before.²⁸ Amazon alone spent $77.3bn, against $35.1bn a year earlier, and its chief executive Andy Jassy said it expects to 'spend meaningfully more in 2027'.¹¹ Alphabet kept its forecast for this year at $195bn to $205bn, and its finance chief Anat Ashkenazi said that spending would 'increase significantly in 2027'.¹⁰ Meta narrowed its forecast to $135bn to $145bn from $130bn to $145bn, lifting the bottom of the range.²⁹ A pause in raising forecasts is not a cut in spending.

The third answer comes from the companies holding the inventory, which point to orders already in hand. Dell booked a record $60.9bn of AI server orders in the quarter to July, ended it with a record $95bn backlog of orders not yet delivered, and raised its forecast for AI server sales this year to $74bn from $60bn.¹³ AMD's finance chief Jean Hu said its inventory rose 'to fulfill our strong data center demand', and its chief executive Lisa Su said demand is higher than supply and that AMD has secured the wafers, packaging, memory and other parts 'needed to support much higher growth next year'.⁷ Measured against the $17bn of sales AMD expects next quarter, its inventory would last about 137 days, fewer than the 149 of a year earlier.²⁷ Cisco said its pile reflects demand for its Silicon One networking chips and higher memory prices.¹⁴ At several companies the inventory is already shrinking against sales. Super Micro's has fallen to about 79 days from 115 a year earlier,²⁰ and its finance chief Mark Long said customers paying in advance funded most of the build, made for revenue expected in coming quarters, against a record backlog.⁸ Arista's has fallen to about 220 days from 240,²³ and Micron's stands at about 132, against 239 at the end of 2022.²⁴ Further down the chain, the contract manufacturer Celestica set out a first forecast of $40bn of revenue for 2027 and said its customers' forecasts for 2028 are 'meaningfully higher'; it cut its cash-flow forecast for this year to $460m from $600m precisely because it is putting money into inventory to meet that growth.⁹ And memory, the part most prone to gluts, is still short: TrendForce forecast that contract prices for ordinary DRAM chips would rise another 10-15% in the fourth quarter and NAND flash 15-20%, because AI servers keep memory scarce, though at a slower pace than in earlier quarters.¹⁵

Where the two meet

Each side has a reply to the other's best evidence. Those who fear a glut answer the record orders above by saying a backlog is only as good as the customer's ability to switch the hardware on and pay for it, and they point to the Amazon racks. Their sharper reply is history. In mid-2022 the chip shortage became a glut in parts of the market within a few weeks, as demand for PCs and phones fell and manufacturers who had stockpiled during the pandemic began cancelling orders; Micron cut production and admitted the turn had taken it by surprise.¹⁶ The companies insisting they were short of parts then were not lying. They simply could not see the moment their customers stopped needing more, and the hardware makers now paying up front for inventory, as AMD and Celestica are, would carry the risk if that moment came again.

The other side's sharper reply turns Microsoft's trim against the fear. Its finance chief Amy Hood gave too little capacity as the reason, not too much: outside providers of graphics-chip capacity had delays that push some of it into the first half of 2027.⁴ On that reading, spending that slips is spending deferred rather than cancelled, and the warehoused racks belong to a customer in a hurry rather than one with a surplus. Super Micro's chief executive Charles Liang says shortages of key parts, even graphics chips, are still delaying deliveries across the industry.⁸ So both sides agree on the facts that matter. Hardware is being shipped faster than it is switched on, the customers are still spending more each quarter, and the hardware makers hold record orders and record inventory at once. The disagreement is over what the waiting hardware is waiting for: electricity and buildings that are coming, or demand that is not.

What the prices assume

A share price can be turned into a statement about the future. The usual measure is the price-to-earnings ratio, the number of dollars investors pay for each dollar of a company's yearly profit. The typical listed company trades at about 19 times. A company priced well above that is one whose investors expect profits to grow fast enough to catch up with the price; one priced below it is one they expect to stall or shrink. So we ask one question of each price: if it stood still, how fast would profits have to grow each year for five years to bring the ratio down to the typical 19? That is what the price already assumes. It is not a forecast, and it is not a measure of what a company did last year.

Read that way, three of the cast already carry the fear in their prices. Super Micro trades at 6 times earnings after falling 31.9% in a month, and its price assumes profits shrink by 20% a year for five years. Micron trades at 11 times after a fall of 23.9%, and its price assumes profits shrink 10% a year. Micron's case has a twist: its profits have soared from a low base, with sales up 256% in a year, and a low ratio on profits at a peak is how investors in memory have long priced the fear that the peak will not last. Dell trades at 22 times after a 35.1% fall in a month, and its price needs profits to grow only 3% a year, against a recent pace of 34%. In all three the shares have fallen hard and the price asks for little or shrinking growth: the glut is already largely in them.

Nvidia looks similar on the figure and is not. Its price needs profits to grow 6% a year, at 25 times earnings, but its shares have fallen less, 16.3% in the month. That low figure comes from earnings that have roughly doubled in a year, up 102%, not from fear; the question for Nvidia is whether those earnings hold. At the other end sits AMD. Even after a 24.7% fall in the month it trades at 106 times earnings, and its price still needs profits to grow 42% a year for five years. Its earnings have recently grown far faster than that, but that is its past, a boom partly from a low base, not what the price asks next. Arista's price needs 21% a year at 48 times earnings, Vertiv's 16% at 38 times, and Cisco's 11% at 31 times, though Cisco's earnings over the past year have been flat. These prices still assume years of growth, so the fear is not in them, and a glut would cost them most.

The people who forecast these companies' profits have so far been raising them. In the 30 days after their latest results, analysts raised their forecasts at least 31 times for Nvidia, 23 times for Cisco and 14 times for Dell, with none cut, and 33 times for Micron with one cut. In the single day after results on 3 November, they raised them at least 28 times for AMD, 19 for Arista and 10 for Super Micro, again with none cut. The numbers show where the fear already sits, and that the forecasters have not joined it; whether the forecasters are right is another matter.

The companies

Nvidia designs the graphics chips at the heart of almost every AI server and holds the largest pile of the cast in dollars; its results on Tuesday 17 November, after the US market closes, are the next check on whether the inventory built for Vera Rubin is turning into sales.

Advanced Micro Devices, Nvidia's main rival in AI chips and a large maker of the processors inside servers and PCs, is the company that has said most plainly it is buying ahead for next year, which makes its pile the sharpest test of whose story holds.

Cisco Systems sells the networking equipment that links computers together, and increasingly the Silicon One chips it designs for the big cloud companies. Part of the growth in its pile is dearer memory rather than more units, so what tells its story is whether the extra Silicon One it has committed to buy is matched by orders from those cloud companies. Its results on Wednesday 11 November are the first check among the cast.

Dell Technologies builds the servers that house the AI chips. Its pile has grown fastest of the cast, which made it Dan Niles's exhibit for the fear, and its record backlog makes it the other side's exhibit too. Its results on Tuesday 24 November will be the next check on whether that backlog is turning into deliveries or being pushed back while the inventory waits.

Super Micro Computer, a server maker that grew up assembling AI machines quickly for the cloud companies, is where inventory now runs the other way, as set out above. Because customers' advance payments funded most of its build, their money rather than its own is at risk first if demand turns, and whether those payments keep coming is the thing to watch.

Vertiv makes the power and cooling equipment that data centres need before any server can be switched on. That gives it an unusual place in the argument: if the bottleneck is power, Vertiv sells part of the cure, yet its own pile is among those growing fastest against sales. What tells its story is whether the sites it equips are connected and running.

Arista Networks makes the high-speed switches that move data inside AI data centres. Its inventory has been shrinking against sales for three years, which puts it on the other side's ledger. Its exposure is its price, which still assumes years of fast growth, and so depends on the cloud companies' plans for 2027 as much as any.

Micron Technology makes memory chips, the part of this industry most prone to boom and bust, and it was the company caught out by the 2022 turn. Memory is where any glut would show first, which makes memory's contract prices the signal closest to its own fortunes.

What would settle it

The answer will show up in places that move before any company publishes its accounts. The first is what the big cloud companies plan to spend in 2027, which Alphabet says it will detail with its fourth-quarter results.¹⁰ Those plans decide whether the hardware piling up has somewhere to go. A clear rise across Microsoft, Alphabet, Amazon and Meta would mean the inventory was built for real demand; plans that stand still would leave the hardware makers holding parts for orders that moved. What matters is less the size of any one forecast than whether the four keep raising them, since it was the pause in raising, not a cut, that unsettled the market.

The second is the price of memory, the contract prices noted above. Memory is bought ahead of building, it goes into every kind of server, and it is the part of the chain that turned first in 2022. If AI demand is real and still outrunning supply, those prices keep rising. If hardware is piling up unused, buyers stop ordering memory before anyone reports a fall in sales, and the rises slow to a halt or turn into falls.

The third is the queue: how long customers wait for servers and chips, whether orders are cancelled or pushed back, and whether customers keep paying in advance. Waiting times that shorten, advance payments that shrink or cancellations that appear would mean supply has caught up with demand; waiting times that stay long while inventory rises would mean the inventory is parts on their way to a customer, not surplus. Closely tied to it is power. If the Amazon racks are switched on as new sites come online, they were early; if more hardware joins them, the bottleneck is hardening.

The fourth is money. The hardware is ultimately paid for by data-centre builders, and some of them borrow. Whether banks keep lending to them, and at what cost, separates a customer that is waiting for electricity from one that cannot pay. More cases like the Volta loan and the Firmus Grid listing noted above would point to the fear; loans taken up readily would point the other way. The companies' next results will check whether inventory keeps outrunning sales, but they will report what has already happened in these four places.

Where they stand with us

As at 31 October, AMD, Arista Networks and Micron Technology hold our gold medal, our highest rating. Nvidia, Cisco, Dell, Super Micro and Vertiv hold silver. Across the wider group of 191 electronics and components makers we rate, 37 companies hold gold.

The judgement that is left

What is left is one question: is the inventory waiting for buyers, or are the buyers waiting for power? Of the four signals, the queue tells the two apart most cleanly. A glut and a bottleneck look alike in a warehouse and alike in the cloud companies' spending; they differ in whether customers still have to wait. Inventory that rises while waiting times stay long is inventory on its way to someone; inventory that rises while waiting times shorten is inventory nobody is queuing for. The difficulty is timing. In 2022 the turn came within weeks, and the companies' own words were among the last things to change, so the judgement rests less on what any chief executive says about 2027 than on how much weight to give the physical evidence that moves first. It is also a judgement about how long the gap between a rack arriving and a rack running can last before patience runs out: for the companies priced for years of growth it has to close, while for those already priced for a glut the question is whether the price was too harsh.

Notes

1. TIKR article by Gian Estrada, published on Yahoo Finance, quoting Dan Niles, 3 November 2026. finance.yahoo.com ↩
2. The Motley Fool (Daniel Sparks), 4 November 2026. fool.com ↩
3. Business Insider report, as carried by Investing.com, 20 October 2026. investing.com ↩
4. CNBC (Jordan Novet), 29 October 2026. cnbc.com ↩
5. Bloomberg (Abhishek Vishnoi and Winnie Hsu), republished on Yahoo Finance, 13 October 2026. finance.yahoo.com ↩
6. Sherwood News (Hyunsoo Rim), reporting Michael Burry's post (older; essential background for the bear case), 26 February 2026. sherwood.news ↩
7. AMD's third-quarter call on 3 November 2026, Motley Fool transcript, 4 November 2026. fool.com ↩
8. Super Micro's first-quarter (fiscal 2027) call on 3 November 2026, Motley Fool transcript, 4 November 2026. fool.com ↩
9. Celestica third-quarter 2026 results release (SEC filing), 26 October 2026. sec.gov ↩
10. CNBC, reporting Alphabet's third-quarter results, 28 October 2026. cnbc.com ↩
11. CNBC, reporting Amazon's third-quarter results, 29 October 2026. cnbc.com ↩
12. Nvidia's second-quarter (fiscal 2027) call on 26 August 2026, Motley Fool transcript (older; Nvidia's next results are due later in November), 31 August 2026. fool.com ↩
13. Dell second-quarter fiscal 2027 results release (SEC filing; older, Dell's next results are on 24 November), 1 September 2026. sec.gov ↩
14. Cisco annual report (Form 10-K) for the year to 25 July 2026, filed with the SEC, 2 September 2026. sec.gov ↩
15. TrendForce press release, 30 September 2026. trendforce.com ↩
16. Reuters, republished by Cyprus Mail (older; background on the 2022 glut), 13 July 2022. cyprus-mail.com ↩
17. AMD's quarterly accounts (results of 3 November 2026 and its SEC filings since 2018), days of stock worked out by Market Medals, 3 November 2026. sec.gov ↩
18. Cisco's annual report (SEC filing), days of stock worked out by Market Medals, 25 July 2026 (quarter end). sec.gov ↩
19. Dell's quarterly accounts (SEC filing), days of stock worked out by Market Medals, 31 July 2026 (quarter end). sec.gov ↩
20. Super Micro's results release (SEC filing of 3 November 2026), days of stock worked out by Market Medals, 3 November 2026. sec.gov ↩
21. Vertiv's quarterly accounts, days of stock worked out by Market Medals, 30 September 2026 (quarter end). sec.gov ↩
22. Nvidia's quarterly accounts (SEC filing), days of stock worked out by Market Medals, 26 July 2026 (quarter end). sec.gov ↩
23. Arista's quarterly accounts, days of stock worked out by Market Medals, 30 September 2026 (quarter end). sec.gov ↩
24. Micron's annual report (SEC filing), days of stock worked out by Market Medals, 3 September 2026 (quarter end). sec.gov ↩
25. Nvidia's, AMD's and Micron's quarterly accounts (SEC filings), days of stock worked out by Market Medals, 2022 to 2023. sec.gov ↩
26. Nvidia's and AMD's quarterly accounts (SEC filings), 3 November 2026. sec.gov ↩
27. AMD's results release, days of stock worked out by Market Medals, 3 November 2026. sec.gov ↩
28. The four companies' quarterly cash-flow statements (results releases and SEC filings), added up by Market Medals, 21 to 29 October 2026. sec.gov ↩
29. Amazon's and Meta's results releases, 29 October 2026 (Amazon); 21 October 2026 (Meta). sec.gov ↩
30. The eight companies' quarterly accounts (SEC filings and results releases), added up by Market Medals, 5 November 2026. sec.gov ↩