Executive Summary

On Friday, November 6, 2026, word came that ING had taken roughly $500 million of its data center loans to market. Bloomberg had it first; CNBC, Reuters, and the data center trade press picked it up the same day. The portfolio holds twelve loans, and as a single block made up only of data center debt, the coverage called it one of the first of its kind. That much is the first fact the reporting confirms.

How it is being sold matters. The format is a bids-wanted-in-competition list: the holder posts what it wants to sell, and each interested buyer submits a sealed price by the deadline. The minimum bid increment is $10 million and the deadline is November 19, 2026. So at the time of writing there is no price. No discount, no clearing level, no buyer has been reported. An ING spokesperson declined to comment and the bank has not publicly confirmed the sale at all. Nobody on either side has said these loans have gone bad.

Only three of the twelve borrowers have been named. There is the $25 million tied to a New Mexico site Oracle is due to lease, plus Vantage Data Centers and CyrusOne. All three are development loans that paid for buildings and campuses. No outlet has written down what the other nine are. The figure below draws that boundary. Read the three known names as the character of all twelve and you are saying something the coverage never said.

The first of the three named loans is the one that catches the eye. Project Jupiter, which carries the $25 million, is a 1,400-acre site in Doña Ana County, New Mexico, with a design capacity of 2.45 GW. Stack Infrastructure and Borderplex Digital Assets are developing it, Oracle is due to lease it for Stargate use, and the original target for operation was 2028. It is also the site Oracle sent a force majeure notice to in September 2026. Pebblous looked at whose books that notice travels onto at the time, and two months later the loan against that site turned up on an auction list.

It helps to put the size on a scale. ING's total assets at the end of the third quarter of 2026 were €1,181.4 billion. The $500 million is about 0.04% of that, and even measured against the wholesale banking loan book it is roughly 0.15%. ING is also a heavyweight in data center finance, with more than 200 transactions behind it. A single bank walking away from AI is therefore a hard story to tell from this auction. At the same time, this $500 million does not stand for everything ING holds in data center loans. That total has not been disclosed.

What the coverage put in the background was not chips but elections and regulation. The sale came days after a large Democratic win in the midterms, with forecasts circulating that some states would slow data center construction or tighten the rules. That same week several data center operators lost more than 14% of their share price, and one tally put the evaporation across the S&P 500 at about $3 trillion in market value. The longer-running shift of credit risk off bank balance sheets and onto private credit, asset managers, and insurers came up alongside it. This section records that much and no more. The judgments come later.

Three days later, on the morning of November 9, Ellen DiMauro at Semafor published a short item titled "AI chip financing hits a wall." It runs four paragraphs and is the only primary account that could be read in full. Inside it are two different stories. Plenty of secondhand pickups welded them into one, so it is worth prying them apart first.

The first is secondhand. The piece records that whole loan portfolios are circulating cheaply, and that the ones secured by GPUs are breaking worst. The reason it gives is where this article starts: the worry that a loan will outlast the useful life of the technology behind it. Two tags hang off that sentence, though. Semafor itself attributes it to one person speaking to the outlet, and what it points at is not ING but several portfolios that go unnamed.

The second is the ING sale, which Semafor flags as picked up from Bloomberg. In the piece the two paragraphs are not joined by cause. They are simply adjacent. So the sentence "ING is dumping loans because of GPU depreciation" appears in no account anywhere, and it is not written here either. What is being followed is not causation but contrast: one side says it cannot price the thing while the other side posts a list to get a price, and both sides use one word for different lengths of time.

A third number travels with these two, $38 billion, and the event it belongs to sits well before November 9. It is the credit raised to build the Texas and Wisconsin sites Oracle leases, and it is not a bond but a loan syndicated among banks. The figure combines $23.25 billion for Texas and $14.75 billion for Wisconsin, with JPMorgan and MUFG as lead arrangers. Vantage Data Centers is the developer. Pricing sits about 2.5 percentage points over the benchmark, and the term is four years with two one-year extension options. The picture changes once the dates on the coverage are lined up.

The $38 billion of September 2025 and the $38 billion of October 2026 point at different states of the same deal. The amount still stuck at the later date is less than one thirty-eighth of the whole.

Put together, it comes out like this. Saying that $38 billion has gone undigested for more than six months runs against the coverage. Both accounts record that nearly all of it was placed, and the portion left to be marked down and moved is under $1 billion. The nature of the deal matters more. The collateral is buildings and the term is four years. So the Semafor sentence about a loan outlasting the chips does not attach to this construction loan. That sentence points at the lenders who took chips as security. Mixing the two sends every calculation downstream crooked.

What Semafor named as Wall Street's next test is something else: the $40 billion chip financing for SpaceX that Goldman Sachs and Apollo have begun marketing. The reported terms describe a structure in which SpaceX does not guarantee the residual value of the equipment and investors carry the depreciation risk, but those figures reached us through aggregators, so they are not pushed any further. Pebblous took up where the chips in that deal end up owned two weeks ago, so one line is enough here.

That leaves one question. To check the worry about a loan outliving its collateral against a number, you need to know how many years the collateral is good for. Finding where that number is written down is the next section.

The period over which a book writes an asset down year by year is its useful life. Working out what the collateral is worth four years from now comes back to that number. So we opened the filings submitted to the U.S. Securities and Exchange Commission for seven companies. The table below pairs two things: which line of the depreciation note the asset falls into and how many years are attached to that line, and what the same company calls the same asset elsewhere in the same filing.

Values checked directly in the seven companies' filings. The useful life column is the figure attached to the whole line, not to a GPU. The Nebius change was confirmed in the second-quarter 2026 report.

Start with the top five rows. The depreciation notes of Alphabet, Microsoft, Amazon, Meta, and Oracle carry no line called GPU. Each uses one line close to "server and network equipment," and racks, switches, cables, and ordinary servers sit inside it alongside the GPUs. Amazon's five years and Meta's 5.5 years are therefore not figures written about GPUs. From these five you cannot say how many years a hyperscaler depreciates a GPU over. All you can say is how many years it depreciates the asset class that contains GPUs over.

The bottom two rows are the more interesting ones. Absent in those five does not mean absent everywhere. Nebius calls the asset a GPU outright in the body of its management discussion, writing that its main assets, GPUs, are currently depreciated over four years. CoreWeave writes only "technology equipment" in the table inside its depreciation note, but in the debt note that explains its borrowings it uses the phrase "GPU servers" and ties its borrowing capacity to the depreciated cost of those GPU servers. The loan agreement uses a name the accounting does not.

The two companies share something else. CoreWeave and Nebius do not rent GPUs to run their own work; they buy GPUs and rent them out. And when a bank takes chips as security, the book those chips sit on belongs to this side, not to a hyperscaler. The side that rents does not call the asset by its own name. The side that pledges it does.

Lay the three places over one another and this article's central observation falls out. One asset is written down in three places inside a single company under three different names: "server and network equipment" in the table of the depreciation note, "GPUs" in the management discussion, "GPU servers" in the loan agreement. Yet the number called useful life is officially posted in only the first of those. Where the name is most precise there is no number, and where the number sits the name is at its bluntest.

The figure below sets those three places side by side. An empty period in the middle and right boxes does not mean no number ever appears there. It means that the place that tabulates how many years each asset class gets is the left box alone. The four years Nebius states in a single sentence of prose is the exception, and because it lives in prose rather than in a table, it cannot be lined up against the same column in another company's filing.

3.1Four Different Periods Fold Into the Word "Life"

The naming problem does not stop at the asset. It runs through the period too. Writing on this subject uses one word for four different things, which is why comparisons that cannot hold keep showing up. The four get separated below, and from here to the end they are kept apart.

The lease term row carries two values because buildings and construction loans split there. A pre-lease on a building runs 12 to 15 years, while the construction loan in the previous section matures in four.

With this table in hand, one common misreading comes apart. A good deal of writing sets Nvidia's eight years next to the bank's three to four and calls the gap 2.7 times, but the two figures come off different rulers. The eight years is a seller's forecast of how long a chip stays useful, which puts it near the fourth row; the three to four years is how long a lender waits to be repaid, which is the second row. Divide one by the other and the comparison is already wrong. None of which should be taken to mean the two numbers sit comfortably together. That is what the next section is for.

With no official line carrying the name, everyone writes their own figure. That is what happened. For the same generation of GPU, the seller, the buyer, and the lessor each use a different number, and there is no place where those numbers meet. The table below records who wrote which number where, together with which of the four periods from the previous section it belongs to.

Rows marked as filings were checked directly in the originals submitted to the Securities and Exchange Commission. Rows marked as aggregator reached us through coverage because the underlying material could not be opened. Microsoft's 2 to 6 years spans too wide a band to be plotted as a point on this ladder.

On one line, the numbers take on a shape. The left end belongs to the lenders and the right end to the chip seller. The companies that keep the books cluster in between. And the four-year maturity of the construction loan from the previous section happens to run right between the left cluster and the middle one.

Even so, among circles alone the value of the collateral still swings hard. Take an asset of 100, depreciate it on a straight line, assume no residual, and the result looks like the table below. The end of year four is the point to look at because that is where the construction loan from the previous section matures, and because that is usually where someone asks what the collateral is carried at.

An illustrative calculation, not any company's figures. It assumes an asset of 100, straight-line depreciation, and no residual value.

It is the same object. Yet one period written into a book leaves the collateral at nothing at the end of year four, or leaves half of it standing. If the side taking security works with three to four years while the side keeping the books writes six, two people are negotiating over one asset holding different numbers. When the chief executive of Ares said, as Semafor rendered it, "I, for one, personally do not know how to underwrite the decline and depreciation curves on that type of technology," the remark did not mean there is no data at all. It comes closer to meaning that several numbers are available and the grounds for choosing among them are not public.

4.1The Hand That Writes the Estimate Moved One Way

Since 2020 the five companies have revised the useful life of servers and network equipment several times. Nearly all of those revisions extended the period. There is one shortening: Amazon pulled some assets back from six years to five with effect from January 1, 2025. Amazon is also, as it happens, the company that had stretched five years into six in January 2024. A year later part of that went back where it started.

On the very same day Meta went the other way, extending most of its servers and network assets to 5.5 years with effect from January 1, 2025. Same date, same kind of asset, opposite direction. Set the two disclosures next to each other, though, and what catches the eye is less the direction than whether a reason was given at all.

Checked directly in the body of both companies' annual reports. Commentary circulates that attributes Meta's change to operating experience, but that phrase does not appear in the disclosure itself.

The side that shortened wrote down a reason; the side that extended did not. The shortening cost $1.0 billion of net income that year, and the extension added $2.59 billion. Writing an estimate, it turns out, cannot be a neutral act. That is not an accusation that either company cooked its books. Useful life is an estimate management is supposed to make, and a change applies going forward rather than restating the past. What is observable is that the estimate carries discretion and that the discretion leaned one way.

Outside the five, the hand moved the same way. Nebius, the one company that had written the name GPU and a period together against the asset, extended the useful life of those very GPUs from four years to five in January 2026. The change cut depreciation expense for the first half of 2026 by $86.1 million and lifted net income by $75.7 million. Naming the asset properly does not make the number written against it any more conservative.

One thing has to be added to keep this fair. Extending the period did not make depreciation expense itself fall. Meta's depreciation on servers and network assets rose year after year, from $7.32 billion in 2023 to $11.34 billion in 2024 and $13.36 billion in 2025, because the buying grew faster than the period stretched. At CoreWeave, depreciation and amortization as a share of revenue climbed from around 45% in the first half of 2025 to 54.4% in the second quarter of 2026. The hand writing longer periods and the pace at which the asset base grows are working at the same time.

4.2Counting the Same Thing Three Times Gives Three Numbers

While preparing this article we counted the useful-life revisions since 2020 three times. Looking at the same filings, the three counts came out at 13 changes, 14 changes, and 21 asset lines. The data was identical all three times. What differed was the decision about what counts as one change. When a company revises three asset categories on the same day, nothing in the public material answers whether that is one change or three.

So the count is left unsettled. It is more accurate to write that the figure comes out as 13 changes or as 21 lines depending on how the unit is drawn. One fact held across all three passes and can be stated flatly: the only shortening is Amazon's January 2025 change. This is an observation rather than a confession of modesty. The subject of this article, that the grain of a label changes the tally, happened once more while the article was being written.

The situation is much the same for what the pledged hardware fetches on the secondary market. What is public is asking prices rather than cleared prices, and for the same model in the same period the sources diverge by a factor of two or three. Nor does the price only fall. During the mid-2024 shortage used cards changed hands above new ones, and one tally has the one-year contract rental rate for an H100 rising nearly 40% from its October 2025 low through March 2026. Saying there is no price is therefore wrong. The accurate statement is that there is no standardized reference price fit for valuing collateral.

4.3One Number Erases an Order-of-Magnitude Spread

Why it is risky to summarize an asset with a single period shows up in measured records. Meta researchers published a study of two large machine learning clusters watched over eleven months, covering more than 150 million GPU-hours on an A100-normalized basis and four million jobs. The abstract states that they fitted a model estimating mean time to failure as a function of GPU scale from that data. As reported, hardware-caused failures account for under 1% of jobs but consume roughly 19% of GPU run time. And the failure rate is not pinned at one value: quiet stretches and spikes show up inside the same cluster.

None of that says GPUs break easily. It says the failure rate varies by an order of magnitude from stretch to stretch, and summarizing it as one annual figure erases that variation. The shape is exactly the shape of writing a depreciation curve as a single number. A Princeton cluster study that separately counted the hours when a seat was taken but no work was being done is the sibling record from the same house. That one measures occupancy; this one asks whether a single number can stand in for the whole.

There is a case where the same trap actually sprang. The Llama 3 training report records 466 interruptions over 54 days of running 16,384 H100s. Of those, 47 were planned and 419 unexpected, and of the 419, GPU-related causes account for 58.7% while GPU failures themselves account for 148, or 30.1%. Much of the coverage carried this across as "419 GPU failures." One category was absorbed into a coarser label and the number went wrong. It is a miniature of the whole problem.

The remark quoted above was made on October 29, 2026, at the third-quarter earnings call of Ares Management. The speaker was chief executive Michael Arougheti, and the sentence has a clause in front of it. Arougheti first says this is not meant to cast aspersions on GPU financing, and only then says the firm cannot do it. Carrying the remark over as AI pessimism therefore moves away from the source. What matters is that the thought did not stop there.

On the same call Arougheti put capital expenditure on AI and digital infrastructure through 2030 at more than $5 trillion and then cut that block in two: roughly $3 trillion for chips and GPU financing, and roughly $2 trillion for data center buildings together with the power generation and transmission attached to them. Ares, the call said, would skip the first and take the second. The reasons given for avoiding the chip side were complexity and concentration: deals on that side tend to bunch in one place.

The shape of the asset Ares said it would buy is specific: data centers in places like Tokyo, London, Atlanta, and Amsterdam, pre-leased in full to investment-grade hyperscalers on 12- to 15-year contracts. The call described this as frankly a simpler underwriting with attractive relative value, and said it generates attractive returns regardless of the ultimate technology outcomes. That last condition is the point. For value to hold whatever the technology does, what carries the value has to be the contract rather than the technology.

This is not a withdrawal but a splitting of the asset and a renaming of the pieces. Ares separated what it could measure from what it could not, then bought only the measurable side. For context, the same quarter put assets under management at $691 billion, up 16% year on year, with fee-related margins improving as well. This was not the tone of a firm backing out of the market.

An odd picture forms at this point. The branch Ares said it would buy is the branch ING has put up for sale. Both are data center loans; one side is buying and the other selling. It looks like a contradiction and is not. The two are measuring different things. Ares is measuring the credit of an investment-grade tenant already inside an operating building, on a 12- to 15-year contract. What ING has listed is development lending against sites still under construction, and permitting, grid connection, and schedule slippage all remain live there. One of the three named loans is the site that received a force majeure notice two months ago.

So two different things sit under the single name "data center loan," in the same shape as the racks and the GPUs sitting inside one line called server and network equipment in the previous section. One side holds only what it could split out and measure; the other posts twelve loans together and takes bids. The side that split can say what it bought. The side that bundled receives a single price.

What separates a competitive bid list from an ordinary sale is how the price gets made. Banks usually sell loans quietly, one at a time. A counterparty is found, a price is agreed, terms are written, and that price stays between the two parties. Posting twelve loans as one block and collecting sealed prices from several buyers by a deadline is, for that reason, rare. The format does not only close a trade. It manufactures a number.

The number it manufactures does not stay on ING's books. Other lenders holding comparable data center loans mark their own positions against the most recent comparable price. Give an asset class with no public comparable a single comparable, and that number travels onto the books of people who never bid. This is the real reason November 19 matters. What appears that day is not only ING's sale proceeds but a clearing level for the asset class.

Which means this section has to record the facts on the other side as well. What is public right now is a list and a deadline. No price and no terms have come out, nobody on either side has said these loans went bad, and ING has not publicly confirmed the sale. There is no report that bids will land below par, and none that they will land above. Until that number exists, calling this auction a loss is saying something the record does not contain.

There is also little reason to read a bank selling data center loans as a distress signal in itself. Investment banks originate loans like these expecting to distribute them later, as people in the market explain it, and credit risk has in fact been moving off bank balance sheets toward private credit, asset managers, and insurers for several years. The $38 billion construction loan above was, in the end, almost entirely placed with more than two dozen banks and investors. As a market grows, so do the channels for moving its risk around.

The open question is what happens after the number exists. Once a price is printed, people start valuing other assets with it. The thing to ask then is not whether the price is high or low but where it is written down what the price is the price of. The number received for twelve bundled loans carries no such line.

That is the U.S. record. What about Korea? In the tax code, in the corporate disclosures, and in the guarantee documents, the same slot is empty. The shape of the emptiness is a little different, though.

Appendix 5 of the Enforcement Rules of the Corporate Tax Act sets standard useful lives by asset type in a table. Servers and GPUs fall under "tools, instruments, and fixtures," with a standard useful life of five years and a band of four to six years a company may choose within. That table contains no line called "server" and no line called "GPU." Move to the Rules on National Accounting Standards and there is no table of figures at all, only the principle that the period must be estimated objectively and reasonably. Listed-company disclosures follow suit. Naver, Kakao, KT, and Samsung SDS all record a range of three to six years against umbrella categories such as "machinery" or "fixtures," and none breaks GPUs out onto a line of their own.

Statutes were checked against the originals at the Korean Law Information Center and disclosures against each company's own report. A claim circulates that Samsung SDS treats GPU servers as five-year assets, but it comes from a spoken answer at an earnings briefing rather than the body of the annual report, so it is not in this table.

The last row is the one that catches. Naver has guaranteed about $12.9 billion (KRW 18.74 trillion) of payment obligations for Naver Cloud usage fees, an amount equal to 64.73% of its equity. One of the things that guarantee document says is that the residual value of GPUs is excluded from the guarantee. The name appears there once. What that value is, and who measures it by what method, is written nowhere. Excluding something from a guarantee would seem to require knowing what it is worth, and here the party excluding it does so without knowing. Pebblous took up the structure of that guarantee and the equity stake beside it in October.

The size of the sums is worth pinning down too. Korea Ratings has projected that once Naver finishes executing its GPU investment, consolidated depreciation from 2028 onward will rise by roughly $1.4 billion (KRW 2 trillion) a year. The company's consolidated depreciation last year was $469 million (KRW 680 billion), so that is more than a tripling. How many years that $1.4 billion gets spread over is exactly the number this article has been following, and the line it attaches to is called "machinery."

One thing in Korea runs the other way, though. Somebody measured, and left a record. An operational report published in 2026 by SK Telecom, Upstage, Lablup, NVIDIA Korea, and Vast Data is public. It covers 504 B200s across 63 nodes monitored for 55 days, with the operational logs themselves spanning 73 days. Across 224 multi-node training sessions, ten GPU failures were identified by XID code, and when a failure hit, automatic retry succeeded 33.3% of the time against 12.5% for manual human recovery, a factor of 2.7. The books carry no name for that asset, yet how often it stopped and how it came back is on the public record.

Public sector standards we could not confirm. We looked for a standard setting the useful life of public AI infrastructure, could not open a published document, and have no grounds for saying none exists. Rental rates, on the other hand, are posted. The monthly rate to rent a single H200 is published at roughly $280 (KRW 400,000-odd). Korea's current state is that the price of borrowing the asset sits in a table while the price of owning it does not.

At the center of this episode is not a financial product but the grain of an asset-class label. Racks, switches, cables, and GPUs all go into one line called server and network equipment, so the depreciation computed from that line explains none of the objects inside it. This is the shape DataClinic meets over and over in customer data. When a label is too coarse, the figures aggregated under it answer no question at all, while the report shows a tidy single row. It is why the first condition of AI-Ready Data is neither the model nor the volume but what that row points at.

8.1The Name Comes First, the Number Second

For data quality, this episode leaves four things behind. First, pour several things into one row and what was poured in stops being visible. No GPU figure emerges from the depreciation notes of the five hyperscalers, however long anyone looks, because the asset is not in there under its own name. Second, let one word stand for four periods and every comparison made with that word turns hollow. Divide eight years by three to four and call it 2.7 times, and two rulers have already been mixed. Separate the names before comparing.

Third, count the same thing three times and three numbers come out, which is what happens when the unit of a count is undefined. The counts of useful-life revisions made while preparing this article, splitting into 13, 14, and 21 lines, are that in the flesh. Fourth, the hand that writes an estimate moves with where it sits. One shortening among a dozen-odd changes, and in the two January 2025 revisions the side that shortened cut its net income for the year while the side that extended added to it. That is why a quality review asks who wrote the number and whom the number favors.

8.2Five Questions to Put to Your Own Organization

What this article wants to leave behind is a set of questions rather than a conclusion. Five is enough, and all five hold outside financial institutions as well.

• How many kinds of object sit inside one asset-class row in our books, and whose life is the average life written against that row?
• Which of the four do we mean by "life": the period the books write down over, the period we wait for money, the period the contract pays, or the gap between generations? Are those four mixed inside a single report?
• Who wrote the number, and who benefits if it is written long?
• Does what we took as security, or carved out of a guarantee, have a name on it? If the document says only that residual value is excluded, nobody has a method for measuring that value yet.
• Is our observation window shorter than the life we assumed? Three years of observation behind a six-year figure is extrapolation, not measurement.

8.3Fix the Label and the Asset Gets Priced Again

Earlier Pebblous pieces asked whether a price was right or who ends up carrying the risk. This one asked where the period that feeds those calculations is written down. The answer was that under that name it is written almost nowhere. Pessimism is the wrong place to stop, though. Section 5 showed the case running the other way. When there is no method of measurement, deals do not halt; the asset gets split and the pieces get renamed. Ares cut $5 trillion into $2 trillion and $3 trillion and bought only the measurable side. That was not surrender but a corrected label.

The Korean consortium in section 7 stands in the same place. The asset has no name in the books there either, yet somebody watched 504 cards for 55 days, counted the failures, and left a public document behind. Where someone measures, a record appears. Where a record appears, it can be given a name, and once it has a name it can be priced again. So the next question is not whether AI infrastructure is expensive. It is whether what we took as security has a line by that name.

Sources are grouped by kind. Items 1 and 2 are the coverage this article stands on, and items 5 through 11 are the filings behind the useful-life table and the ladder. Items 12 through 14 are the measured operational records behind sections 4.3 and 7, and items 15 onward are statutes, Korean material, and earlier Pebblous pieces.

Coverage

• 1.Ellen DiMauro, "AI chip financing hits a wall," Semafor (November 9, 2026). semafor.com — the only primary account obtained in full. The two quotations in section 2, the mention of the $40 billion SpaceX financing, and the English wording of the Ares chief executive's remark all come from this piece.
• 2.Coverage of ING's data center loan bid list (November 6, 2026). Bloomberg had it first, with CNBC, Reuters, and the data center trade press picking it up the same day. The size, the loan count, the minimum bid increment, the deadline, the three named borrowers, and the declined comment in section 1 all come from this group. The original sits behind a paywall and was confirmed through the pickups.
• 3.Ares Management third-quarter 2026 earnings call transcript (October 29, 2026). Published transcript — the $5 trillion, $3 trillion, and $2 trillion figures in section 5, the 12- to 15-year pre-leases, and the quarterly metrics come from here. This is a published transcript rather than one issued by Ares itself.
• 4.The reporting trail on the $38 billion construction loan — Financial Times (May 5, 2026) and Bloomberg (April 15 and October 12, 2026). The five rows of the table in section 2 come from that trail. Both outlets are paywalled and were confirmed through pickups and summaries.

Filings (U.S. Securities and Exchange Commission)

• 5.Alphabet annual reports FY2025 and FY2023 — FY2025 10-K. The six-year life for server and network equipment and the 2023 revision history are both in this filing.
• 6.Microsoft annual reports FY2026 and FY2023 — FY2026 10-K. Computer equipment at two to six years, and across the full document the word GPU never appears in a depreciation context.
• 7.Amazon annual reports FY2025, FY2024, and FY2020 — FY2025 10-K. The January 2025 shortening in section 4.1, the $1.4 billion of added depreciation, the $1.0 billion of lost net income, and the verbatim reason are all here.
• 8.Meta annual reports FY2025 and FY2024 — FY2025 10-K. The extension to 5.5 years, the $2.92 billion of avoided depreciation, the $2.59 billion of added net income, and the absence of a stated reason were all confirmed in this document.
• 9.Oracle annual reports FY2026 and FY2025 — FY2026 10-K. The six-year life for computer, network, machinery and equipment and the June 2024 revision were taken from the FY2026 document.
• 10.CoreWeave annual report FY2025 and second-quarter 2026 report — FY2025 10-K · Q2 2026 10-Q. That the line in the depreciation note is called technology equipment while the phrase GPU servers appears in the debt note, that borrowing capacity is tied to the depreciated cost of those servers, and the 54.4% depreciation share of revenue all come from here.
• 11.Nebius annual report FY2025 and second-quarter 2026 report — FY2025 20-F · Q2 2026 6-K. The sentence "Our main assets, GPUs, are currently depreciated over four years" in the body of the management discussion and the change from four years to five are in these two reports. The annual report described the change as under review; completion was confirmed in the quarterly report.

Measured Operations

• 12.A. Kokolis et al., "Revisiting Reliability in Large-Scale Machine Learning Research Clusters" (arXiv:2410.21680, submitted October 2024, revised February 2025). arxiv.org — the eleven months, two clusters, four million jobs, and 150 million A100-normalized GPU-hours were confirmed directly in the abstract. The figures for under 1% of jobs against 19% of run time, and the variation in failure rate across stretches, are given as reported because the body could not be opened.
• 13.D. Kang et al. (SK Telecom, Upstage, Lablup, NVIDIA Korea, Vast Data), "From Detection to Recovery: Operational Analysis on LLM Pre-training with 504 GPUs" (arXiv:2605.09370, submitted May 2026, revised June 2026). arxiv.org — the 504 B200s, 55 days of monitoring, 224 multi-node sessions, ten identified failures, and 33.3% automatic recovery rate in section 7 come from here.
• 14.Llama 3 team, "The Llama 3 Herd of Models" (arXiv:2407.21783). arxiv.org — the breakdown of 466 interruptions in section 4.3 comes from this paper. The numbers themselves were confirmed through aggregators.

Statutes and Korean Material

• 15.Corporate Tax Act Enforcement Rules, Appendix 5, "Table of Standard Useful Lives and Useful Life Ranges for Buildings and Other Assets" — Korean Law Information Center. The five-year standard life and four-to-six-year band for tools, instruments, and fixtures are here, and there is no line named server or GPU.
• 16.Rules on National Accounting Standards — Korean Law Information Center. The basis for the statement that the rules set an estimation principle without any table of figures.
• 17.Annual reports of Naver, Kakao, KT, and Samsung SDS — checked for the depreciation ranges on tangible assets. The statement in section 7 that none breaks GPUs onto a separate line rests on these reports.

Related Pebblous Posts

• 18.When an AI Data Center Runs Late, Who Pays for the Wait? — cited in section 1. It covered the New Mexico site now on this auction list two months ago.
• 19.SpaceX Pays for the AI Chips, and Apollo Owns Them — referenced in one line in section 2. It covers the deal Semafor named as the next test.
• 20.Naver Builds the AI Factory, and Nvidia Owns a Piece of It — the Korean anchor for section 7. The clause excluding residual value from the guarantee was confirmed in that piece.
• 21.GPU Cluster Fully Booked, Compute Cores Mostly Idle — the sibling record for section 4.3. That one measures occupancy; this one takes up the grain of the summary.