The Debt Moves Only If the Risk Does: Amazon's $8 Billion Chip Leaseback

Five hyperscalers have signed about $1.1 trillion of lease payments, undiscounted and mostly for data centres, on property they will rent and not own. That works for buildings. Amazon is now reported to be trying something similar with chips it already runs, where the accounting rules let the debt leave only if real risk leaves too. The five-year life Amazon set for a subset of its servers in 2025, if it covers these chips, decides how much.

TON618 Capital Research · As of 6 October 2026 · Download PDF

On 2 October the Financial Times reported that Amazon has been talking to investors about moving roughly $8bn of Nvidia Grace Blackwell chips, installed in more than a dozen US data centres, into a special-purpose vehicle funded with outside debt and up to 10% equity, and leasing them back so AWS keeps using them. The FT's own summary line is that the move "aims to improve the tech group's balance sheet health". Amazon has not commented, and no lease term, investor or price has been reported. Structures that keep AI financing off a sponsor's balance sheet have spread over the past year (§2). This note starts with a plainer question: why move chips off a balance sheet at all, when a bond is cheaper? It answers by measurement: what a vehicle changes that a bond does not, what the rules demand before that change counts, and how far $8bn moves each number it touches. Amazon's shares last traded at $251.53 on Friday 2 October, against a $248.23 close the day before.

Terms used throughout, defined once. Sale-leaseback: selling an asset and renting it back. Finance lease: a lease that is in substance a purchase, recorded like owning the asset with a loan; under ASC 842 a leaseback classified this way means no sale happened. Operating lease: a rental; the asset leaves the seller's books, and a smaller lease liability arrives. Committed term: the period the tenant cannot cancel, plus any renewal it is "reasonably certain" to take up. Residual value: what the chips are worth when the lease ends. Residual value guarantee: a tenant's promise to make up any shortfall in that value. SPV: special-purpose vehicle, the company that buys the chips. Free cash flow (FCF): Amazon's definition, operating cash flow less purchases of property and equipment net of proceeds from sales and incentives. Dose: capital spending per dollar of net income, the house measure of how much an income statement leans on its depreciation schedule; net income here excludes all of Amazon's non-operating income (mostly gains on its Anthropic investment), after a 21% tax. Rent curve: what a GPU-hour rents for, by chip generation. bp: basis points, hundredths of a percentage point.

A vehicle costs more than a bond, and it changes two things a bond does not. Amazon can borrow for two and a half years at about 5.3%. We estimate a vehicle that passes the accounting test would cost it about 8.3% (§6). Two things differ: what gets reported, and who owns the chips when the lease ends. The second is the proponents' reason in its own right (case 1 above), and §6 finds that most of the extra cost is its price.

The second row does not make $8bn of obligations disappear. The rent is itself a liability, about $6.6bn, so about $1.4bn net, 18% of the money raised, leaves the balance sheet, which is close to the $1.5bn outside equity pays for the chips' end value. §4 shows why the two match.

Three measured facts show what the reporting difference would touch. Free cash flow went from +$18.2bn in the year to June 2025 to −$7.6bn in the year to June 2026; in the first half of 2026 Amazon spent $98.4bn on property and equipment against $71.4bn of operating cash flow. The bond market filled the gap: Amazon raised $67.0bn of long-term debt in that half-year, its notes outstanding nearly doubled, from $68.0bn to $132.1bn, and its new-issue spreads widened between November 2025 and July 2026 (three-year money from 30bp to 40bp over Treasuries, ten-year from 55bp to 80bp), while two press tallies put its July order book below March's. And the rating is tied to both: Moody's moved its outlook on Amazon's A1 from positive to stable on 20 February 2026, citing the need for external funding, and said an upgrade would need consistent profit growth and positive free cash flow (as reported by Investing.com and Yahoo Finance).

Servers are the largest part of what Amazon buys, and the part not yet financed this way at scale. At Microsoft, Oracle and Meta, the leases signed and not yet started are primarily for data centres (§2), so a large volume of data-centre capacity is being rented from landlords, not built and owned. Amazon has $137bn of such leases across all its property. Servers and networking equipment have had no equivalent route at scale. They added more to Amazon's property at cost in 2025 than any other class ($59bn, against $32bn for land and buildings), and by our estimate spending on them equalled about 45% of that year's purchases of property and equipment, which at this year's $220bn guide is roughly $99bn. Amazon buys most of them with its own cash: property and equipment acquired under finance leases was $2.9bn in 2025.

The sections that follow measure how far $8bn moves each of those numbers, and §6 sets the answers side by side.

The build-out's building debt mostly sits on other people's books. The usual route is a lease. A developer borrows to build a data centre, and the tenant's signed rent commitment is what the lenders rely on. Nothing reaches the tenant's balance sheet until the lease starts. Five companies disclose about $1.1 trillion of payments on leases signed but not yet started, and some have added guarantees on top.

These are commitments, not borrowings, and the guarantees are ceilings. The lease figures are undiscounted payments over terms of up to 30 years, so they cannot be set against debt one for one, and each lease becomes a liability on the tenant's own books when it starts. The guarantees are maximums: about $28bn on an asset the sponsor itself uses (Meta's Hyperion), and about $178bn where Alphabet, NVIDIA and Broadcom stand behind someone else's payments. Meta calls payment "not probable", and Broadcom reports its backstop's fair value as "not material". What the table shows is narrower and, we think, hard to dispute. For buildings, it is routine for a landlord or a partner to own the asset while the sponsor stays committed to it, and Amazon has $137bn of such commitments of its own.

Meta shows how far the building route can be pushed. A long lease does reach the balance sheet once it starts: Microsoft carries $66.6bn of finance-lease liabilities, on leases that include its data centres. Meta's Hyperion structure avoids even that. Its leases run four years, renewable to twenty, and in place of a long commitment Meta guarantees the campus's value. The campus was financed with $27.3bn of bonds that do not appear in Meta's debt, and Meta reports a maximum exposure of $46.0bn.

Chips, at scale, are the exception. We searched SEC filings since 2024, one form type at a time, for sale-leasebacks of GPUs, servers or AI hardware, and read what the searches returned. Three companies report having sold AI servers and leased them back, and a fourth, Sangrix, reports the buyer's side of such a deal. The only large case sits on the seller's books. The last three rows of the table are financings that never sought a sale.

Four cases are too few to prove a rule, and three of them are tiny. The two leasebacks disclosed as five-year, xAI's third and the one Sangrix bought into, were booked as financings. The one three-year leaseback whose accounting is disclosed, Boost Run's, was booked as a sale, which fits the next section's arithmetic only if the servers had more than four years of economic life left at signing. None of these filings names the lease term as the reason: xAI's cites finance-lease classification without saying which test, and Sangrix's cites the operator's rights over the servers. So the pattern is consistent with the arithmetic, not evidence for it. Among the deals we could confirm, Amazon would be the first hyperscaler to sell installed chips it runs and lease them back. One detail in the reporting complicates even that: Reuters says the chips were "bought or leased" by Amazon, and chips Amazon already leases are not its to sell.

On the rules, the asset's life is the likely reason for the split. A four-year lease is a small slice of a building that lasts decades. The rules treat it as a rental, the tenant books four years of rent against an asset worth several times that, and Meta's filings show a guarantee of the building's value sitting beside its four-year leases. The same four years is most of a chip's life. The rent then repays most of the price, the rules call that a purchase, and a guarantee, which the test counts alongside the rent, only makes it more so. The next two sections put numbers on that.

The rule is short. ASC 842-40 says the buyer has not obtained control, so no sale has occurred, "if the leaseback would be classified as a finance lease or a sales-type lease". A lease is a finance lease if any one of five tests is met. Two matter here: whether the lease runs for "the major part of the remaining economic life" of the asset, and whether the present value of the rent equals "substantially all" of the asset's fair value. The standard's guidance offers 75% and 90% as "one reasonable approach" to those two phrases, and practice follows it.

The first test runs on a number Amazon chose. Deloitte's guide to the standard says it "would expect the economic life of an asset to correspond to its depreciable life used for financial reporting." Amazon's depreciable life for servers is five to six years. In January 2025 it cut "a subset of our servers and networking equipment" from six years to five, citing "the increased pace of technology development, particularly in the area of artificial intelligence and machine learning." The 10-K does not say which servers are in the subset. We assume AI accelerators, the stated reason for the cut, are in it; if they sit on six years instead, the limit moves out by three-quarters of a year. AWS made Grace Blackwell systems generally available on 9 July 2025, and Reuters describes the chips as ones Amazon "is installing", which points to a young fleet. We assume about a year old on average and show the sensitivity.

To pass with a committed term of four years or longer, Amazon would need its auditor to accept an economic life well beyond the depreciable life it chose, for AI reasons, in January 2025. That is possible, because the test asks about "one or more users", and AWS itself will still rent an A100 for $1.70 an hour from 7 October, nearly six years after launching it. It would also run against the direction Amazon itself moved in 2025: it is the only one of the six large AI-compute buyers in our depreciation work that has shortened a server clock.

Renewal options are the obvious alternative, and they only go so far. Meta's building leases run four years with options to renew for up to twenty. Amazon could write a short committed term with options to extend. Under the standard, a renewal counts toward the term only if the tenant is "reasonably certain" to exercise it, a threshold PwC puts at about 75% probability. A renewal at or above the initial rent "is not considered to be reasonably certain of exercise", PwC says, unless "the economic penalties the lessee would suffer by not exercising the renewal option are significant", so renewals at market or flat rent would probably keep the committed term short; a rent that steps down on renewal creates a presumption of exercise, which PwC says can be overcome; and fixed-price renewals covering the rest of the asset's life are contested, with PwC recording one view that they amount to a fixed-price buyback, which would end sale accounting, and another that they do not, and Deloitte saying they do not automatically preclude a sale. The open point for Amazon is the economic penalty: the standard counts the cost of removing the asset and its importance to the tenant's operations, and these chips sit in Amazon's own halls with AWS, on the FT's account, still using them. Either way, a market-rent renewal does not hand the risk back. The investors still own chips whose next rent is whatever the market pays.

Two more traps sit beside the term. Any option for Amazon to buy the chips back at a fixed price ends sale accounting outright, and even a fair-value option survives only if "alternative assets, substantially the same", are "readily available in the marketplace", a hard claim for the scarcest GPU on the market. And if the vehicle leans on Amazon, through a residual value guarantee or Amazon's control over re-leasing hardware that sits in its own buildings, Amazon may have to consolidate it under ASC 810, which brings the debt straight back. Who holds the equity matters for that test, and the reports disagree: Bloomberg says the vehicle could sell a stake of up to 10% to investors and Reuters that Amazon plans to offer it, while one outlet, Quartz, says Amazon would hold it. Amazon's 10-Ks and 10-Qs from 2019 to 2026 have never used the phrases "failed sale" or "residual value guarantee", so there is no Amazon precedent to read.

The second test turns on what the chips are worth when the lease ends. For the vehicle to buy $8bn of chips, its investors must get $8bn back, plus a return, from two sources: Amazon's rent and the chips' end value. The less the chips are worth at the end, the more the rent must cover, and above 90% the lease is a finance lease. The best evidence of a 2025 chip's rent in 2028 is what older chips rent for today. Ornn's compute price index, the series Polymarket's GPU-rent contracts settle on, publishes daily hourly rents by chip.

Two facts hold at once. Rent for the B200 rose 52% and for the H200 50% between July and October. Rent for the A100 fell 12%, and Polymarket settled it one bracket lower in September ($0.75–1.00) than in August ($1.00–1.25). The H100, three years on sale, rents for 34% of a B200. The A100, about six years on sale, rents for 11.5%. Fitted across the four generations on each day of the public three-month window, rent falls between 29.5% and 38.6% per year of age, 33.5% on average. AWS's own price list for reserved GPU blocks draws a similar curve: from 7 October it charges $14.21 an hour for a B200, $5.97 for an H100 and $1.70 for an A100, about 35% less per year of age. A list price is not a market price, but the two land close together.

The slope is a cross-section, and that cuts both ways. It captures what a chip earns once newer ones exist, which is what the vehicle will face. It does not capture whether today's frontier chip keeps a scarcity premium longer than its predecessors did. The B200's 52% quarter is the best evidence that it might, and Polymarket prices 77.5% on it still renting above $7.00 at year-end. That is why a 20%-a-year path runs through every table as a full scenario.

On our base inputs, both tests give out at about three years. We value a GPU as the present value of its future rent less the cost of running it, and price the vehicle tranche by tranche (§9). On the market's curve, a two-and-a-half-year lease puts the rent at 82% of value and a three-year lease at 89%; it reaches 90% at about 3.2 years on Amazon's borrowing rate (3.5 at the rate implicit in the lease). The term test on Amazon's five-year clock fails at 3.0. On the scarcity case's inputs, 20% annual decline and an eight-year life, a four-year lease passes both (86%; four years against a 5.25-year limit). In the 50% stress, even a two-and-a-half-year lease fails the value test (96%).

The end value is an estimate, and it has two soft inputs. The first is the running cost. A buyer of used chips must house and power them. We build that cost from published figures: a Grace Blackwell rack draws about 124kW for 72 GPUs, large Northern Virginia colocation asks $160–185 per kW a month, and Virginia industrial power costs about 10 cents a kWh. That gives about $0.66 per GPU-hour. A hyperscaler running its own halls may pay nearer $0.40, and a small buyer nearer $0.87.

The last two rows are a cross-check, and it widens the range in both directions. At the index rent, our method values a year-old Blackwell GPU at about $102,000. Published hardware prices are far lower: about $47,000 a GPU for a complete eight-GPU B200 system (CloudRift), and about $32,500 a GPU on Modal's estimate, relayed by SiliconANGLE, of $60,000–70,000 for a Grace Blackwell unit, which carries two GPUs. So the rent implies a value two to three times the price. There are two ways to read that. If a fleet earns much less than the spot index, which is plausible for capacity sold under long contracts, the end value falls to 17% of the price ($1.4bn), or 13% ($1.1bn) on the lower hardware price, and the value test binds at about two and a half years. If instead the chips really earn the index and the vehicle simply buys them near cost, their end value is 54% of the price ($4.3bn), or 78% ($6.2bn) on the lower hardware price, and the value test would not bind until about four years. Neither reading moves the term test, which stays at three years on Amazon's clock. The dollar figure is the uncertain part: our base estimate is $2.0bn, and the readings above run from about $1bn to about $6bn.

So a deal that passes looks like this. A committed term under three years; no fixed-price buyback or renewal; no material guarantee; and outside equity that owns the chips at the end. On our base estimate that end value is about $2.0bn of the $8bn on a two-and-a-half-year term and about $1.4bn at the three-year limit; a shorter term sells more, and the cross-check above shows how wide the band is. That slice is the chip risk that genuinely leaves Amazon, and under the rules the debt leaves only with it.

What comes off the balance sheet is about what investors pay for that slice. A passing leaseback brings a lease liability equal to the present value of the rent, about $6.6bn on a two-and-a-half-year term, so $8.0bn less $6.6bn, about $1.4bn net, leaves the balance sheet. Outside equity pays about $1.5bn for an end value of about $2.0bn, because it discounts that value at 13%; the only part of the price Amazon does not repay as rent is the part investors fund for the chips' end value, and the two differ slightly because the liability is measured at Amazon's borrowing rate, below the lenders' rate. At the three-year limit the liability is about $7.1bn and $0.9bn net leaves; on the scarcity case's inputs a four-year lease passes and $1.1bn leaves. On our base pricing, across all three of those cases, 80–90% of the money raised comes back as an obligation to pay rent. The figure moves with the same readings as the end value: on the two-and-a-half-year term it runs from about $0.7bn on the lowest reading, which just fails the 90% test (a passing sale moves at least $0.8bn), to about $4.6bn if the vehicle buys near hardware cost and the chips earn the full index.

If it passes, the whole $8bn lands in headline free cash flow. Amazon defines free cash flow as operating cash flow less "Purchases of property and equipment, net of proceeds from sales and incentives." The standard's cash-flow guidance puts a successful sale-leaseback's proceeds in investing activity, and the proceeds line is where Amazon reports cash from selling equipment, so a passing sale would be netted against capex there. A failing sale's cash goes to financing, which free cash flow ignores. In the last four quarters the proceeds line ran between $0.87bn and $1.13bn a quarter, so $8bn would be about eight times a normal quarter and impossible to miss.

It is not the free-cash-flow turn. The deal was still in talks two days after the third quarter ended, so the earliest it can appear is the fourth-quarter cash-flow statement. By then the trailing window drops the strong second half of 2025 (+$17.3bn of free cash flow) and adds a second half of 2026 in which, on the $220bn guide, Amazon would have to spend about $122bn on property and equipment to meet it, against roughly $75bn a year earlier. Our estimate, assuming second-half operating cash flow grows 20–40% on a year earlier (the first half grew 44%, against a first half of 2025 that grew 12%; the second half of 2025 grew 26%, so the same two-year pace implies about 28%): trailing free cash flow at December of about −$27bn (range −$18.5bn to −$36.5bn), and about −$19bn with a passing sale. If 44% holds instead, the December base is about −$15bn, and about −$7bn with the sale. Fitch, as reported by Investing.com in July, said capex growth could result in negative free cash flow of around $40bn a year in 2026 and 2027, on a $200bn capex assumption. On every one of these cases, the sale narrows the gap; on none does it close it.

Amazon used to publish two stricter measures. It stopped in 2025. Until its Q4 2024 release (filed 6 February 2025), Amazon reported free cash flow "less principal repayments of finance leases and financing obligations" and a second variant that also deducted equipment bought on finance leases. It explained them as giving "additional perspective on the impact of acquiring property and equipment with cash and through finance leases and financing obligations." Every release since Q1 2025 carries the headline measure alone. We rebuild the first from the cash-flow statement; the second cannot be rebuilt, because Amazon no longer splits equipment leases from property leases.

Our dose metric moves too, and unless the sale passes, on the §3 terms or on an auditor-accepted longer life, it should not. The dose nets the same proceeds line, so a passing sale would cut it from 2.34× to 2.23× with no change in what Amazon spent on machines. When we next score the hyperscalers' doses, we will show Amazon's with and without any sale-leaseback proceeds.

Over the lease, the lift is a pull-forward. In a passing structure Amazon receives $8.0bn at closing, then pays about $2.8bn a year in rent through operating cash flow for two and a half years. Net over the lease, free cash flow is up about $1.0bn, and Amazon no longer owns chips with an expected end value of about $2.0bn. In a failing structure nothing arrives in free cash flow, only the interest portion of each payment touches operating cash flow, and the chips stay Amazon's. For scale, $8bn is 4.7% of trailing net capex and 3.6% of the 2026 guide.

Amazon's own curve sets the benchmark. On the 1 October Treasury curve and its July new-issue spreads, widened by the 9bp that investment-grade spreads have moved since 21 September (high-yield spreads moved 58bp, from 2.66% to 3.24%), Amazon borrows for two and a half years at about 5.3%. Its 2029 notes last traded at 5.28%.

The vehicle pays more, and four comparable structures show how much. They are indicative, not like-for-like: Broadcom's tranches carry a Broadcom residual-value agreement and their tenors were not reported, and the A2 figure compares a coupon with a Treasury yield.

We take that range, 60–138bp with 78bp central, for the vehicle's debt, and a 13% required return for the equity that owns the chips' end value.

Most of that premium is the price of the risk, not of the debt. About $6.5bn of the $8bn is debt repaid from Amazon's rent, which lenders can price close to Amazon's own credit. The other $1.5bn is equity that owns about $2.0bn of uncertain end value and wants 13% for it. The table shows who gains or loses, against what the deal priced in, when the chips come back in 2029. It adds two variants: a failing version in which Amazon keeps the end value through a buyback or guarantee, and a 90%-debt version in which lenders are owed a final payment of about $0.8bn from the chips' value.

Each cell is the gain or loss against the end value the deal priced in, so an equity loss larger than the stake includes the return it expected. In a structure that passes, Amazon's exposure to the chips' value ends at signing and it remains a tenant owing fixed rent. If the scarcity case is right, Amazon pays about $0.4bn of extra rent to give away a gain of about $1.6bn. A thin equity slice strains the structure: with lenders repaid only from rent, debt reaches 90% of the price at a term of about 3.25 years, past the point where the sale fails, so a vehicle with 10% equity either keeps the lease short and asks lenders to take some chip risk, or it is a loan.

The price also gives a break-even. Against a bond at its own rate, Amazon's gain from a passing vehicle is the part of the price whose present value it never repays as rent, the $1.4bn net; its loss is the chips' end value, which it would otherwise own. On a two-and-a-half-year term the two balance at an end value of about $1.6bn, which the rent curve produces at a decline of about 37% a year, above the 33.5% average we use. The market's three-month window ran from 29.5% to 38.6%, so the break-even sits inside the range the market's own curve spanned this quarter: on the average slope the vehicle costs Amazon about $0.3bn in present value against a bond, and on the window's last day, 38.4%, the end value is about $1.5bn and the vehicle is already the cheaper route on residual grounds alone. The arithmetic is a condition, not a motive. A seller indifferent to the reporting and funding effects would need to expect rent to fall faster than 37% a year; a seller that values those effects at more than about $0.3bn needs no view beyond the market curve, though on the scarcity case's 20% path the loss is about $1.8bn, and the effects would have to be worth more than that. The buyer's side could have reasons that are not a price view either: an investor that can use the hardware itself in 2029 may pay for access to it, not only for its resale value.

Whether moving the debt works depends on who is counting. GAAP would let it go. Rating agencies need not. S&P's published criteria, dated 2019, add lease obligations to adjusted debt and say S&P may increase a lease liability "if we view remaining lease terms to be artificially short relative to the expected use of the lease asset", a clause that would apply if S&P expected Amazon to keep renewing. The same criteria add guarantees of third parties' debt, unless the guaranteed party is creditworthy enough. Moody's, in a February 2026 report quoted by Fortune, said of hyperscalers' off-balance-sheet lease commitments that it "may make a nonstandard adjustment to Moody's adjusted debt based on our expectation of likely cash outflows." Fitch's summary of its criteria says lease liabilities are "not classified as debt for most sectors", though its commentary on Amazon cites a rent-adjusted leverage measure. None of the three has commented on this vehicle, so their treatment is untested. The benefit is surest in reported free cash flow and reported debt, uncertain in rating-agency leverage, and real in one other respect: it opens a pool of lenders, such as the insurers and pension funds Bloomberg mentions, that buys rated asset-backed paper and not Amazon bonds.

Set against the measures in §1, $8bn moves some of them modestly, and sells one thing outright. Free cash flow: +$8bn in the quarter the sale closes, about $7.0bn paid back as rent over the next thirty months, and a December trailing figure near −$19bn in place of −$27bn. The bond market: $8bn not issued on the unsecured curve, about 12% of the $67.0bn Amazon raised in the first half. Obligations: about $1.4bn net fewer on our base estimate (band in §4), against $132.1bn of notes outstanding, or about 1%; taking $10bn off the balance sheet this way would mean running about $57bn of chips through such vehicles and finding outside owners for the end value of all of them. The rating: untested, and S&P's 2019 criteria say it may increase the liability on a lease whose term it views as "artificially short". What the deal does deliver in full is the first item on the proponents' list. About $2.0bn of exposure to what a 2025 chip is worth in 2029 passes to investors who chose it, for about $0.4bn of extra rent over the lease, most of it the price of that risk. So the measured answer to why a chip vehicle and not a bond is narrower than the phrase "off balance sheet": it is a way to sell chip risk and to bring cash flow forward, and only a small way to move debt. For servers to follow buildings off the balance sheet at scale, outside investors would have to own the chips' end value at scale, and the filings we found in §2 show no one doing that at scale.

The deal is a swap of risks, and each side's case can be laid out without guessing what anyone thinks. The table shows what each party holds today, what it would hold after a passing deal on our base terms, and the rent path on which it comes out ahead. The dollar figures are those of §5 and §6; the depreciation figures are new, and their assumptions are stated in §9.

Two readings fit these facts, and the reader can weigh them. In the first, the seller has the better information. Amazon's fleet utilisation and the rent its customers actually pay after discounts are not disclosed, and what the equity would see in diligence is not reported. Amazon has already shortened one server clock, for AI reasons, and runs the highest dose among Microsoft, Alphabet, Meta and itself in our July valuation (2.34×; Oracle's is higher). A passing deal converts an uncertain depreciation charge on $8bn of chips into a fixed rent, at an estimated extra cost of about $0.4bn. On this reading the equity is buying the risk of a further life cut, and the 13% return we assume for it is the price of not knowing whether or when one comes.

In the second, the buyer has the better information. An operator sees its own forward demand and the rents it charges, which no index records. The B200's rent rose 52% in a quarter, Polymarket prices 77.5% on it holding above $7 at year-end, and a buyer that can run the chips itself values them above what a financier would. On this reading Amazon pays about $0.4bn to fix a cost that the 20% path would have turned into a $1.6bn gain, and the operator secures scarce hardware at a price that returns 13% a year on the market's own curve before any scarcity premium.

Both readings are consistent with the same deal. They differ on the rent path, and the market's own range this quarter straddles the break-even that separates them. The terms will say which fits. A passive fund as equity, pricing the chips off the public index, fits the first; an operator that will run the chips from 2029 fits the second. A committed term well under three years fits either, since a seller that wants the risk gone and a buyer that wants the hardware would both ask for it; a term at the limit with market-rent renewals fits a deal shaped by the accounting rather than by either reading. The dated test follows.

Not in the Q3 numbers. Amazon's third quarter ended on 30 September, and the deal was reported as talks on 2 October. The Q3 release, expected after the close on Thursday 29 October (a date Amazon has not yet confirmed), and the 10-Q that follows cannot contain the cash. Watch three things there: any mention of the transaction in the 10-Q's liquidity or commitments discussion, whether management frames 2026 free cash flow as including equipment-sale proceeds, and the capex outlook. The settling document is the FY2026 10-K, due in early February 2027, with the fourth-quarter cash-flow statement.

Tripwires we will track: the 29 October call and the 10-Q (any reference to the vehicle); the identity of the equity holder, operator or fund, when named; Ornn A100 on 31 October (Polymarket: 53.5% on $0.75–1.00) and B200 (92.5% on $6.50 or more); any rating-agency comment on the vehicle; the FY2026 10-K lines above; and any statement by Amazon on the deal, which as of 6 October has not been made.