DEEP DIVE

COMPANION TO ISSUE #37 · 13 NOVEMBER 2026


Who Paid First

More and more of the AI build is funded by deposits, prepayments and guarantees. They don't trade and they don't pay coupons, and one of them just cost its owner $3.7bn.

The line was in SoftBank's consolidated financial report on Tuesday, against a subsidiary most shareholders had never heard of: Silver Bands 7 LLC. The description ran to eleven words, 'non-refundable advance payments made to secure manufacturing capacity and purchase equipment', and the number beside it was ¥457,030 million. About $2.85bn. No bond had missed a coupon. No lender had called anything in. Asked whether SoftBank had defaulted on its turbine orders, finance chief Yoshimitsu Goto said: 'There is no such default.'

He was right. Money was lost and nobody defaulted.

So the useful question isn't who has borrowed the most. It's who has paid first, who is holding that money, and what happens to it when plans change. This piece sorts the four ways money moves ahead of delivery, maps this week's examples, prices a second SoftBank trim in low, middle and high cases, and ends with a three-line check for any holding.

1. FOUR WAYS TO PAY AHEAD

Four kinds of money now move ahead of delivery in this build. They sit in different places in the accounts, and they fail in different ways.

Deposit or slot reservation

Who pays whom: A builder pays an equipment maker

Where it sits: Payer: advances, an asset. Receiver: customer deposits, a liability

What a loss looks like: Written off when the order is dropped

Customer prepayment

Who pays whom: A customer pays a cloud provider

Where it sits: Payer: prepaid asset. Receiver: deferred revenue

What a loss looks like: Late delivery; an unsecured claim if the provider fails

Guarantee

Who pays whom: A supplier backs its customer's lease

Where it sits: Off the balance sheet, in the commitments note

What a loss looks like: A payout if the customer can't pay

Supplier stake

Who pays whom: A supplier buys its customer's shares

Where it sits: Equity investments

What a loss looks like: A mark-down, slow to appear while the shares are unlisted

🌱 NEW TO INVESTING? HERE'S WHAT THIS MEANS Deferred revenue is a debt paid in service, not cash. When a customer pays a year ahead, the provider books a liability: it owes that customer a year of computing. Deliver late and the liability doesn't go away. It just sits there.

2. THIS WEEK'S MAP: WHO HOLDS WHOSE MONEY

Here is who paid whom, on this week's disclosures. Don't add the column up. These are different kinds of promise.

SoftBank

Who holds it: Power-equipment makers, reportedly GE Vernova above all

How much: $9.48bn paid by 30 Sep; $2.85bn written off; about $6.6bn still on the books

Status: Disclosed

CoreWeave's customers

Who holds it: CoreWeave

How much: About $9bn expected within three months

Status: Guidance

Nebius's customers

Who holds it: Nebius

How much: $10bn expected in 2026; $17bn–$20bn in 2027

Status: Guidance

Microsoft

Who holds it: IREN

How much: 20% of a contract worth about $9.7bn

Status: Reported contract terms

Seven chip and cloud companies

Who holds it: Anthropic

How much: $26bn of non-voting stock at the IPO price

Status: Filed; completes with the IPO

Nvidia

Who holds it: SB Energy, OpenAI's landlord in Ohio

How much: Lease guarantees capped at $105bn, plus $1.5bn of equity

Status: Filed

Two things stand out. The money runs in every direction: builders pay suppliers, customers pay builders, suppliers pay customers. And only one row has been tested so far. SoftBank's.

3. WORKED EXAMPLE: WHAT A SECOND TRIM WOULD COST SOFTBANK

SoftBank is the cleanest public case, because it has told us what walking away cost. Everything in the first table is disclosed. Everything in the second is assumption, so each line gets a low, middle and high value. The point isn't the answer. It's seeing what moves it.

Disclosed (SoftBank, to 30 Sep 2026)

Figure

Paid ahead by Silver Bands 7

¥1,500.9bn ($9.48bn)

Non-refundable advances written off

¥457.0bn ($2.85bn)

Contract changes and settlements

¥140.2bn ($0.88bn)

Total charge

¥597.2bn ($3.73bn)

Advances still on the books

about $6.6bn

Rebates on future purchases, not booked

up to $2.1bn over ten years, conditional

Share of the remaining $6.6bn given up

Low: 10%

Middle: 25%

High: 40%

Basis: Assumption. The first round wrote off about 30% of everything paid

Advances forfeited

Low: $0.66bn

Middle: $1.65bn

High: $2.64bn

Basis: Assumes the whole advance on a dropped order is lost, as in round one

Settlement costs on top

Low: $0.20bn

Middle: $0.51bn

High: $0.82bn

Basis: 31% of forfeited advances, the first-round ratio ($0.88bn ÷ $2.85bn)

Further charge

Low: $0.86bn

Middle: $2.16bn

High: $3.46bn

Basis: Forfeited advances + settlements

Further charge ÷ first charge

Low: 23%

Middle: 58%

High: 93%

Basis: Scenario comparison, not a forecast

Cost of carrying the $6.6bn, per year

Low: $0.34bn

Middle: $0.46bn

High: $0.59bn

Basis: At 5.1% (the 10-year Treasury yield), 7% and 9%. SoftBank's own rate isn't disclosed

What moves the answer. One assumption does nearly all the work: how much more SoftBank drops. The forfeiture rate matters too, and it's inferred. SoftBank calls the advances non-refundable, but neither it nor any supplier has said how many gigawatts were cancelled. One analyst's estimate is about 4GW out of more than 13GW reserved, close to the 30% of the money that was lost. If that's right, you lose what you paid on what you drop.

What could shrink it. The rebates, if SoftBank keeps buying. And need: it says its Ohio campus isn't being cut back, and the plan for that site includes at least 9.2GW of gas generation, so much of what remains may simply get used.

WHAT THE EXAMPLE SHOWS, AND WHAT IT DOESN'T It shows why Issue 37 set a write-offs test. A buyer that reserved in a shortage and trims in a slowdown takes the loss in one line, in one quarter, with no warning from the credit markets. The middle case would be a second charge of about $2.2bn. It doesn't show that a second trim is coming. That depends on SoftBank's 'thorough reassessment' of procurement, which nobody outside the company can see. February's results are the first check.

4. THE RECEIVER'S SIDE: CHEAP MONEY, WITH STRINGS

For whoever holds the cash, a prepayment is the cheapest funding going. CoreWeave expects about $9bn from customers within three months. Here's what borrowing it would cost instead.

On its next secured loan (guided 7% to 7.5%)

Rate: 7.25%

Interest a year: $0.65bn

Share of annualised Q4 interest guide: 15%

At its average cost of debt

Rate: 7.9%

Interest a year: $0.71bn

Share of annualised Q4 interest guide: 16%

At the yield on its 2032 unsecured notes, 11 Nov

Rate: 14.2%

Interest a year: $1.28bn

Share of annualised Q4 interest guide: 29%

So the saving is real. It comes with three strings.

It has to be delivered. More than $3bn of CoreWeave's $9bn relates to infrastructure that's already paid off; the rest is a promise of capacity. Nebius signed prepayments that more than cover its third-quarter builds, then cut its 2026 revenue outlook because data centres it doesn't own ran late. IREN can owe Microsoft delay credits if its Texas phases slip past early 2027.

Booked isn't collected. Nebius's deferred revenue rose $3.43bn in the quarter. Its receivables rose $3.76bn, and operating cash flow was negative $273m. Some of what customers have promised up front hasn't arrived yet.

A reservation isn't an order. GE Vernova ended September with 64GW of gas-turbine orders and 58GW of slot reservations. It turned 13GW of reservations into orders during the quarter, and says SoftBank's change has no impact on it either way. Fine. The 58GW is still the number I'd watch.

5. THE GUARANTOR: NVIDIA'S PROMISES AGAINST ITS CASH

Nvidia carries the build's largest book of promises, and until Thursday one man was its public face. These figures are from the second-quarter CFO commentary, the last before Tuesday's results.

Nvidia, quarter to 26 Jul 2026

Figure

Cash and marketable securities

$56.6bn

Guarantees, maximum gross exposure

$108.5bn, of which $105bn backs OpenAI's Ohio leases

Supply and capacity commitments

$279bn, up from $119bn a quarter earlier

Committed future equity investments

$25bn

Multi-year cloud service agreements

$29bn

Quarterly revenue, for scale

$96.2bn

Read the guarantee carefully. It isn't cash out of the door. Nvidia pays only if OpenAI fails to pay a lease, and then only the shortfall left after the landlord re-lets or sells. The first phase isn't expected to take effect until fiscal 2029, and the whole thing lapses if OpenAI earns a satisfactory credit rating. Against revenue of $96bn a quarter, that's bearable.

The open question is the next one. Nvidia holds an option to back about 3.8GW more at the same campus, and has since agreed to put $10bn into Anthropic's share sale. I don't know what a board with a new chair, an interim chief executive who is also the finance chief, and three House committees asking questions does with that sort of discretion. My guess is: less. It's only a guess, and Tuesday's guarantee table is the first evidence either way.

6. THE CIRCLE: WHEN SUPPLIERS FUND THE BUYER

Anthropic's sale closes the loop. Seven chip and cloud companies will buy $26bn of non-voting stock at the IPO price, led by Google and Nvidia at $10bn each. Anthropic's commitments to buy compute and infrastructure stood at about $895bn at the end of September, so a slice of that $26bn comes back to the same companies as orders. Last week's deep dive has the full table.

Two readings. For the suppliers, it's a stake in a customer, bought with money the customer will largely spend with them. For Anthropic, it's the reverse of SoftBank's position: about $84.6bn of net proceeds at the midpoint, and termination rights over roughly $213.8bn of those commitments. It promises a great deal and pays ahead rather less. That's the stronger seat, for as long as suppliers allow it.

7. WHAT TO WATCH

17 Nov · after close

Checkpoint: Nvidia Q3

What would change the picture: The guarantee table: above $108.5bn, unchanged, or new language about limits

18 Nov (reported)

Checkpoint: Anthropic prices

What would change the picture: Where in $70–$78 it lands, and the $26bn placement completing alongside

9 Dec

Checkpoint: Fed decision

What would change the picture: A half point raises the carry on every row of the map

Late Jan 2027 (expected)

Checkpoint: GE Vernova Q4

What would change the picture: Slot reservations against 58GW; the 130GW year-end target; any cancellations disclosed

Feb 2027 (expected)

Checkpoint: CoreWeave Q4; SoftBank Q3

What would change the picture: Prepayments received against about $9bn; a second charge on advances

31 Mar 2027

Checkpoint: Test date

What would change the picture: One more write-off of $1bn or more on money paid ahead, anywhere in the build

8. A THREE-LINE CHECK FOR ANY AI HOLDING

1 Paid ahead. Prepaid expenses, advances to suppliers and deposits, often tucked inside 'other assets'. Set the total against a year of capex. SoftBank's $9.48bn was in plain view before any of it was written off.

2 Been paid ahead. Deferred revenue, customer deposits, contract liabilities. Then look at receivables beside them. If both are climbing, some of that prepayment is still a promise.

3 Promised. Guarantees and purchase commitments. They live in the commitments note, not on the balance sheet. Set them against cash.

Then search the filing for 'non-refundable'. Three lines and one word, and you'll know which side of the deposit your holding sits on. I'd rather be wrong about this one. If I'm right, the next loss turns up the way SoftBank's did: in one line of a quarterly report, with nothing in the credit markets to warn you.

BEFORE YOU GO This piece backs up Issue 37's Investing Angle and its three tests: write-offs, deposits and slots. The weekly scoring lives in the Tripwire Tracker. Not subscribed? Subscribe

Disclaimer: This article is for informational and educational purposes only and does not constitute financial advice. This publication is not a registered investment adviser. The worked example uses labelled assumptions and is illustrative, not a forecast. Always conduct your own research and consult a qualified financial professional before making investment decisions.