Five ways to fund AI. Only two create interest expense

In twenty-two days this summer, four companies raised $52.35 billion to build or equip AI capacity. Two others moved capacity without paying anything at all.

They used five different instruments. The money is interchangeable; where it lands on the financial statements is not.

How these were found

Not from news. We read the filing histories of twenty-one companies across the AI build-out and filtered by item number — the code that tells you what a Form 8-K is about before you open it.

Twenty-one filings carried one of those markers between 15 July and 19 August. Four of them were false positives, and saying which matters as much as the list itself:

• Amazon’s 424B3 is the merger prospectus for its acquisition of Globalstar. Not a financing.
• Digital Realty’s and IREN’s 424B7s register resales by existing holders. Both documents say the company is not selling and will receive no proceeds.
• Equinix’s $5.5bn revolving facility is borrowing capacity. Nothing is owed and no interest accrues until it is drawn. Counting it would have inflated the total by more than a tenth.

The five instruments

Alphabet    public bonds, 9 tranches       $25.00bn
Intel       common stock                   $20.00bn
AMD         senior notes                    $4.75bn
CoreWeave   secured term loan               $2.60bn
─────────
$52.35bn
Core Scientific  warrants to its customer    no cash
NVIDIA           guarantee of OpenAI leases  up to $105bn contingent

Only Alphabet, AMD and CoreWeave will report interest on any of this. That is $32.35bn of the $52.35bn — 61.8%. The rest arrives without an income-statement line.

The same dollar, priced three ways

Alphabet, AMD and CoreWeave all borrowed within eleven days of each other, for broadly the same purpose, and paid very different prices.

The gap is 3.70 points. On $2.6bn drawn in full, that difference is worth about $96m a year against what Alphabet would pay on the same balance.

CoreWeave’s facility says more than its rate does. It is a delayed draw loan, available until December 2026 and maturing September 2031, secured on substantially all assets of the borrowing subsidiary, guaranteed by the parent, and subject to a debt service coverage ratio of at least 1.35x. Alphabet’s bonds carry none of that. The 8-K states the purpose plainly: to finance capital expenditure required to perform customer contracts, including GPU servers.

We have written up what Alphabet’s issue said through its maturity distribution and what AMD’s notes cost against its existing interest line separately.

The two that create no interest at all

Intel sold stock. 210,526,315 shares at $95.00, for gross proceeds of $20.0bn and $19.67bn before expenses. The last reported sale before pricing was $97.52, so the issue came at a 2.58% discount, with an option for a further 31.6m shares.

No interest, no maturity, no covenant. The cost is borne entirely by existing shareholders, and it does not appear on the income statement at all — it appears in the share count, which is the same mechanism that produced Intel’s $11.0bn quarterly loss earlier this year.

Core Scientific paid its customer in equity. On 27 July it leased 377 MW to AMD across three sites, plus 152 MW to a third-party operator, on fifteen-year terms with three five-year options. In connection with those leases it issued AMD a warrant for up to 30 million shares at $23.47, vesting at 12,222 shares per megawatt.

AMD leases                     377 MW
Third-party operator leases    152 MW
───────
529 MW  ×  12,222 shares/MW
= 6,465,438 shares vested immediately

The landlord did not borrow and the tenant did not pay a premium. The capacity was bought with dilution.

The one that lands nowhere

On 17 August, NVIDIA guaranteed the residual value of leases for OpenAI, up to a cumulative $105 billion — a structure we took apart when it was filed. NVIDIA borrowed nothing. Until the guarantee is called, it exists in a footnote.

AMD did a smaller version of the same thing ten days earlier, and it is in the Core Scientific filing rather than AMD’s own. Under Credit Support Agreements covering the third-party operator’s 152 MW, AMD holds the right to cure that tenant’s defaults and takes on defined obligations if the tenant fails materially.

Two chipmakers, three weeks apart, both standing behind a customer’s real-estate obligations so the customer could take space. Neither commitment is debt. Neither generates interest. Neither would appear in any screen built on borrowings.

What our own numbers would have missed

Five days before this, we published a panel of interest expense across 1,519 companies. It is useful and it is led by Alphabet, whose interest line rose 514% — but run against this month, that panel would eventually capture Alphabet, AMD and CoreWeave and record nothing for Intel’s $20bn, nothing for Core Scientific’s warrant, and nothing for $105bn of guarantees.

That is not a flaw in the panel. Interest expense measures interest expense. It is a reminder that “how leveraged is the AI build-out” cannot be answered from a debt total, because a third of the cash this month and all of the contingent commitments were arranged so as not to become one.

What would make this wrong

• This is twenty-one companies, not the market. We chose them. A large financing by a company outside that list does not appear here, and there is no reason to think the list is complete.
• Twenty-two days is not a trend. Four raises in three weeks is a snapshot. It says these instruments were all in use at once; it does not establish a shift from one to another.
• CoreWeave’s rate floats and its loan is a commitment, not a balance. The 9.15% is the rate at one fixing on undrawn money. What it eventually pays depends on SOFR and on how much it draws.
• The purposes are not all stated as AI. CoreWeave’s 8-K names GPU servers. Alphabet’s prospectus says general corporate purposes and we said so when we wrote about it. Intel’s use of proceeds is broader than data centres. We are grouping by timing and context, and that is a judgement.
• The $105bn is a ceiling, not an exposure. It is the cumulative maximum under the guarantee, and the guarantee extinguishes if OpenAI reaches a satisfactory credit rating.
• We did not read every exhibit. The credit agreements themselves run to hundreds of pages and are filed as exhibits. What is quoted here is from the 8-K narratives and the prospectus covers.

Check it yourself

Every document behind this article is linked below and none of them is paywalled. The filing scan is a script in this site’s repository, scan-ai-financings.mjs; it takes a list of CIKs and prints everything they filed with a financing item number, which is how the false positives surfaced.

If a figure here does not match a filing, tell us and it will be corrected on the article and on the corrections log, with the date.