Two AI cloud companies each borrowed at two very different prices in 2026, and the price changed with who stood behind the payments. CoreWeave’s $8.5 billion loan of March, rated A3 by Moody’s and secured by GPUs and a customer contract that Bloomberg reported was Meta’s, has a fixed-rate tranche at approximately 5.9%. In June CoreWeave sold senior unsecured notes at 9.625%.
IREN’s $3.65 billion facility of June, which it said was “Anchored by Microsoft’s offtake”, cost a blended 6.00%. Its August financing for GPUs at its Mackenzie site in British Columbia carries a fixed 9.0%, its annual report says. By our arithmetic, the gap is about 3.7 percentage points at CoreWeave and 3.0 points at IREN, for the same borrower.
Short explainer · GPU debt rates
A vertical news Short narrated from this piece’s headline and opening line.
Business Pill · Financial Conditions
A one-minute explainer of financial conditions: how the cost and availability of credit move across markets. It teaches the general idea only and says nothing about any company in this piece.
The key insight: By our arithmetic, CoreWeave’s unsecured 9.625% notes cost about 3.7 percentage points more than the approximately 5.9% fixed tranche of its Meta-backed loan, and IREN’s 9.0% Mackenzie financing costs 3.0 points more than its 6.00% Microsoft-anchored facility.
A post on X on 29 October by @nodiligence, written after reading Anthropic’s S-1, says that “the lenders have already answered it with their pricing”. The post lists 6% with Microsoft behind the contract, 7.75% for SpaceX, over 11% for Volta’s loan against an Anthropic offtake, 16% for CoreWeave’s unsecured notes and 6.15% for CoreWeave when a Meta contract is attached. He argues that “lenders are pricing the GPU itself at roughly zero”.
We checked each figure against company releases, filings and the reports behind our earlier pieces. Four match a published source. They are IREN’s 6.00% Microsoft-anchored facility, the roughly 7.75% Semafor reported SpaceX’s safer chip-backed debt was expected to price at, Bloomberg’s report that Volta’s loan yield went past 11%, and Reuters’ report that the yield on CoreWeave’s bonds due 2031 surged past 16%. For CoreWeave’s Meta-backed loan, CoreWeave’s release gives approximately 5.9% on the fixed-rate tranche and SOFR plus 2.25% on the floating tranche; we did not find 6.15%.
The releases we read describe the collateral as GPUs and do not all name the chip, so we have not checked the post’s point that every deal is against the same Nvidia silicon.
Rate on GPU-linked debt by who stands behind the payments, 2026: CoreWeave’s Meta-backed loan (fixed tranche ~5.9%, release; Meta link per Bloomberg), IREN’s Microsoft-anchored facility (6.00%, release), Lambda (6.78%), SpaceX (~7.75% expected, Semafor), IREN Mackenzie (9.0%, 10-K), CoreWeave’s unsecured 2032 notes (9.625% coupon, 8-K; 14.8% yield on 14 Oct, CNBC) and Volta (above 11%, Bloomberg). Issue or marketed rates; not like for like.
What Was Posted on X
after going through anthropic’s S-1 i’ve been working out who actually carries the GPU depreciation bill, and the lenders have already answered it with their pricing.
every one of these deals is against the same nvidia silicon.
— it costs 6% with microsoft behind the contract, — 7.75% for spacex with starlink revenue behind it, — over 11% for volta’s loan against an anthropic offtake with no bank backstop, and — 16% for coreweave’s unsecured notes. coreweave itself borrows at 6.15% when a meta contract is attached.
the chip is the same in all of them and only the name on the rent changes, so lenders are pricing the GPU itself at roughly zero while the owners carry it on their books for six years.
full piece on who ends up holding those last years is coming soon.
@nodiligence wrote after reading Anthropic’s S-1 that lenders price GPU deals by the name on the rent, listing 6% with Microsoft behind the contract, 7.75% for SpaceX, over 11% for Volta, 16% for CoreWeave’s unsecured notes and 6.15% for CoreWeave with a Meta contract; this piece checks those figures.
The three independent labs (OpenAI, Anthropic, xAI) use ~22% of the worldโs AI compute today, but theyโve signed most of the long-term contracts that kick in 2027โ28. A big chunk of the AI buildout rests on three companiesโ ability to keep paying.
The account @dennisihong posted that the three independent labs (OpenAI, Anthropic and xAI) use about 22% of the world’s AI compute today but have signed most of the long-term contracts that kick in during 2027 and 2028, and that a big chunk of the AI buildout rests on those three companies’ ability to keep paying. The figures are the poster’s.
CoreWeave: The Same Borrower at Two Prices
CoreWeave said on 31 March that its DDTL 4.0 facility was “the first investment-grade rated financing secured by HPC infrastructure and an associated customer contract”, rated A3 by Moody’s and A (low) by DBRS. The release names the customer only as a “leading AI enterprise”; Bloomberg reported that the loan is backed by a Meta contract. It says the facility “includes a floating rate tranche financed at SOFR + 2.25% and a fixed rate tranche financed at approximately 5.9%” and matures in March 2032.
CoreWeave’s 8-K of 18 June describes $1,250 million of 9.625% Senior Notes due 2032, guaranteed on a senior unsecured basis by certain subsidiaries. By our arithmetic, that coupon is about 3.7 percentage points above the March loan’s fixed tranche. On 14 October CNBC reported that the notes yielded 14.8% to maturity, about 8.9 points above the March tranche by our arithmetic. That yield is a secondary-market level, not what CoreWeave pays on notes already issued. Our CoreWeave bond piece has the detail, and our cash-to-debt piece has its balance sheet.
IREN: Microsoft Behind One Loan, Not the Other
IREN’s release of 1 June says its $3.65 billion facility was a $2.10 billion US private placement at a fixed rate equivalent to SOFR+2.13% and a $1.55 billion delayed draw term loan at SOFR+2.25%, with a blended cost of debt of 6.00%. It says the facility “is secured against the GPUs and associated contracted cash flows”, and that Fitch and DBRS rated it A and A(low).
IREN’s annual report describes financing of up to $2.4 billion signed in August with Blue Owl Capital Corporation among the parties, to finance GPU servers at its Mackenzie data center facilities. It says borrowings and notes “bear interest at a fixed rate of 9.0 % per annum” and mature 30 months after funding. Business News Today reported that the Mackenzie financing supports deployments for non-investment-grade counterparties.
The Other Rates on the Chart
Lambda said on 1 October that it closed a $1 billion investment-grade delayed draw term loan at a fixed 6.78%; its release, as we recorded it, did not name a customer. Semafor reported on 22 October that the safer chip-backed piece of SpaceX’s $40 billion GPU debt was expected to price around 7.75% and the lowest-rated portion about 8.3%. Elon Musk posted on 25 October: “SpaceX closed $40B of debt for GPUs, but we’re investment grade and have Starlink & launch revenue.” Our SpaceX piece and our count of October compute debt have both.
Bloomberg reported on 15 October that JPMorgan Chase and Goldman Sachs pushed the yield on Volta Infrastructure Holdings’ $5 billion loan past the 11% banks had earlier floated. It said the debt is unrated, that Anthropic has already signed a six-year agreement for compute capacity at the site, and that the rate will change if Volta gets certain ratings. Morgan Stanley’s Carolyn Campbell said in a podcast, according to Bloomberg, that the highest-rated data center bonds now yield closer to 7%, up from below 5% at the start of 2026, as our data center bond piece covers.
Who Stands Behind the Payments
The backstop can come from the customer’s supplier or investor as well. Reuters, reporting The Information’s story on 29 October, said Nvidia is in talks to offer Nscale billions of dollars in credit guarantees for a West Virginia data center Anthropic plans to use. A post on X by Rohan Paul, citing The Information, said the unsigned deal would back the building’s value rather than its lease or chips, letting Nscale borrow at investment-grade terms. Our Nscale piece has the update; Nvidia’s 10-Q already gives a maximum gross exposure of $108.5 billion under guarantees.
According to the Wall Street Journal, as reported by Reuters, Google agreed to guarantee billions of dollars of Anthropic’s lease and power-payment commitments if the startup defaults, for a Texas campus. Anthropic’s S-1 says it has relied on “arrangements with financing sources or partners who guarantee the lease obligations”. Our Texas loan piece has the detail.
The Structural Read
As we read the releases, the companies tie their lowest-cost debt to a customer contract: IREN says its facility was anchored by Microsoft’s offtake, and CoreWeave says DDTL 4.0 was secured by HPC infrastructure and an associated customer contract.
As we read the October reports, the highest rates sit with new or unrated borrowers: Bloomberg says Volta’s loan is unrated and that its rate will change if Volta gets certain ratings.
As we read the Nscale reports, a supplier’s guarantee is another way to put a stronger name behind the payments: the X post summarising The Information says Nvidia’s support would let Nscale borrow at investment-grade terms.
IREN, release, 1 June 2026
“Anchored by Microsoft’s offtake, the transaction received Fitch and DBRS ratings of A and A(low) respectively”
Three Implications
FOR LENDERS CoreWeave’s and IREN’s releases name a customer contract, or contracted cash flows, alongside the HPC infrastructure or GPUs as security for their investment-grade-rated, lowest-cost facilities.
FOR AI LABS AS TENANTS Bloomberg reported that Anthropic signed a six-year agreement for capacity at Volta’s site, whose unrated loan was priced past 11%, and Reuters reported Nvidia is in talks to guarantee Nscale debt for a site Anthropic plans to use.
FOR INVESTORS The rates come from different dates and instruments, so the gaps compare published terms; they do not measure how lenders value any single GPU.
The analysis puts it this way: “It is a sovereign-scale financing problem dressed in tech-company clothing — and the financial architecture is starting to invert.”
As we read it, the rates in this piece show one part of that architecture: debt with a large customer contract or guarantor behind it priced near 6%, debt without one at 9% or more. The analysis does not discuss these 2026 financings.
What Is Not Established
The rates are not like for like. They come from deals struck between March and October 2026, when base rates moved; IREN said it reached 6.00% notwithstanding higher base rates since its initial underwriting commitment. They mix secured loans, unsecured notes and a marketed loan yield, and the 14.8% is a market yield, not a coupon. The SpaceX rates are Semafor’s report of expected pricing, not final terms.
The Meta link to CoreWeave’s loan is Bloomberg’s; CoreWeave’s release does not name the customer. The Nscale terms come from one post summarising The Information, whose article we have not read. The 3.7, 3.0 and 8.9 point gaps are our arithmetic.
This Week In The AI Supercycle
Execution Fault Lines, Concentrated Demand, and Geopolitical Firewalls
This week’s issue: the AI build is not short of demand for its parts. It is short of cheap money, and its end revenue rests on a narrow base of large, often venture-funded buyers.
CoreWeave’s Meta-backed loan carries approximately 5.9% on its fixed tranche and its unsecured notes a 9.625% coupon; IREN’s Microsoft-anchored facility costs a blended 6.00% and its Mackenzie financing a fixed 9.0%. By our arithmetic, those gaps are about 3.7 and 3.0 percentage points for the same two borrowers.
A note on sourcing. On 30 October 2026 we read CoreWeave’s and IREN’s releases, IREN’s 10-K on EDGAR, Bloomberg’s reports on Volta and on Nvidia and Nscale, Reuters’ report on The Information’s story, and the X posts by @nodiligence and @rohanpaul_ai; other figures come from our earlier pieces, which name their sources (CNBC, Semafor, CoreWeave’s 8-K, Lambda’s release, Reuters). We did not contact the companies. The 3.7, 3.0 and 8.9 point gaps are our arithmetic.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
