Wholestory

Hyperscaler bonds now trade at spreads similar to BBB-rated credit while rated above it — the market charging for risk the agencies have not marked. Barclays Private Bank's Michel Vernier, writing on 4 September, reads the fundamentals as sound: hyperscalers expect net debt below 1.0x earnings, and a conservative stress holds it near 1.4x — enough for negative outlooks or selective single-A downgrades, well short of BBB — against over $500 billion of liquidity. He puts 2026 issuance on track for $400 billion, after $160 billion in 2025 and $44 billion in 2024; others project nearer $250 billion, counting dollar bonds only. Barclays owns and recommends this paper.

Anthropic is set to finalise expanding its revolving credit facility to $15 billion — six times the one it signed last year — Bloomberg reported on 3 September, citing people familiar; nothing is filed and the terms could still change. Morgan Stanley leads, with Goldman Sachs, JPMorgan and Citigroup prominent; the same four lead the IPO. Barclays and Wells Fargo are expected in key roles, ahead of Bank of America, Deutsche Bank, RBC, UBS and seven more. The ladder is the point: about $1.25 billion sought from the most active lenders, $1 billion from the next tier, $750 million or less below that — rank on the loan tracks rank on the listing.

Lambda closed a $926 million senior secured term loan B on 27 August, priced a fortnight earlier, secured on the GPU servers the money buys and the cash flows those servers earn. Moody's rated it Baa2 — investment grade, on collateral the loan's own schedule assumes will wear out: it matures on 31 December 2030 and amortises fully in step with the contracted revenue and the hardware's expected useful life, at SOFR plus 3.00%, issued at 99.5. Morgan Stanley was lead left arranger and administrative agent and MUFG joint bookrunner, with Citizens Bank, Crédit Agricole and Wells Fargo as documentation agents. Lambda calls it the first broadly syndicated, investment-grade-rated term loan B by a private cloud provider, and its second large debt raise this year after a $1 billion facility in May. The customer whose contracted payments underpin the loan is described only as investment grade, and is not named.

US hyperscalers have sold more than $180 billion of dollar bonds this year, piling unusually far into the ultra-long end: 41 percent of their 2025-26 issuance runs beyond 15 years against 16 percent for the broad index, with Amazon paper reaching 2076 and Alphabet's £1 billion century bond — the first from a technology company since Motorola in 1997 — drawing £9.5 billion of orders. TwentyFour Asset Management, a fixed-income manager, argues this long-duration supply now competes with 30-year Treasuries for the same pension and insurance money, a marginal push on a long-end sell-off that US-deficit dynamics still dominate. Demand held firm, with recent 30-year deals 3.6 to 4.6 times covered. The reading complements the ECB's finding a week earlier of no crowding-out of European borrowers.

The largest data-centre lenders — JPMorgan, Goldman Sachs, Citi, Morgan Stanley and Wells Fargo, with Bank of America pledging $250 billion — are passing the credit risk to less-regulated buyers rather than holding it. Drawing on Morningstar, American Banker describes banks using syndication, significant-risk-transfer trades, loan sales and securitisation to push exposure onto insurers and private-credit funds while still originating: Morgan Stanley and KKR arranged a $9.7 billion warehouse facility for CyrusOne, and BofA provided $14 billion for the Oracle/OpenAI campus in Saline Township, Michigan. The Bank of England's July financial-stability report found the financing already spread across public debt, private credit, leveraged finance, securitised assets and SPVs. Morningstar's Arnaud Journois warns the migration to nonbanks cuts transparency and echoes 2008.

The European Central Bank has looked at what US hyperscalers are doing to Europe's corporate bond market, and found two things that do not point the same way. Crowding out has not happened: cover ratios for euro area issuers held firm through 2026 while hyperscalers' edged down, and some European borrowers simply timed their deals to avoid the days big tech came to market. But hyperscaler credit spreads are widening across every maturity on news of higher AI capex, while other issuers' spreads stay flat — investors repricing one sector, not the market. US big tech now accounts for just under 10 percent of gross new euro-denominated corporate issuance, with about €40 billion outstanding; Amazon's deal this year set an all-time size record. The ECB's own caution: AA- ratings for a new sector rest on revenue-growth and leverage assumptions that may not hold.

The question this page exists to ask — where the risk actually sits — now has a number. On data to 17 August, J.P. Morgan Asset Management puts disclosed hyperscaler data-centre lease obligations at about $1.4 trillion, of which roughly $1.1 trillion is off balance sheet, because a lease is recognised only once the data centre becomes operational. Separately, unconditional purchase commitments for chips, compute and power come to about $1.5 trillion. Underneath sits a structural shift: hyperscalers increasingly raise debt through ring-fenced project-finance vehicles secured on a single project's cash flows rather than the parent, which preserves balance-sheet flexibility and makes leverage harder to read. Reported leverage, the firm notes, is no longer the measure that matters.

The cash bond market has repriced the buildout. Hyperscaler spreads widened about 30 basis points this year against 2 for the broader dollar investment-grade index, and duration-adjusted they now trade near 105 basis points against roughly 80 for similarly rated paper. The widening accelerated in July after larger — and, J.P. Morgan Asset Management notes, less well-flagged — issuance, and is sharpest at long maturities where the buyer base is thinnest. Supply explains the scale: hyperscalers have gone from 2 percent of dollar investment-grade issuance across 2022–24 to an expected 9 percent this year, with $219 billion issued so far. The firm suggests the long-end issuance may itself be lifting long-dated Treasury yields.

Eagle Point Credit Management, a $14 billion investment firm, is providing about $1.3 billion of a roughly $16 billion project-finance package for Nexus Data Centers' first campus — a gas-fired site at Hubbard, Texas, with Anthropic as primary tenant. Eagle Point's piece is the single largest in the loan, but it is mezzanine debt, below the safest tranches; the mezzanine portion recently closed, with banks led by Morgan Stanley arranging the larger facility. The deal is a template for how this borrowing is structured. It firmed up only after Anthropic won the tenancy and Google agreed to guarantee the senior debt — the same backstop Alphabet has extended to data centres nationwide, which reassures lenders while concentrating the risk on one balance sheet. And construction debt like this is repaid by issuing more: Nexus may sell high-yield bonds later this year to refinance the bank facility, which would require its first credit rating. The figures come from a person familiar with the private deal, not a filing.

Two of the largest institutions in the bond market published measurements four days apart that found the same thing: the buildout's borrowing has grown big enough to change what an ordinary investment-grade bond fund holds. BNY, whose custody business gives it a direct view of investor flows, reported total US investment-grade issuance of nearly $1.5 trillion in the year to 12 August — the largest on record — with hyperscalers more than 12% of it and behind eight of the ten biggest deals of 2026, and their share of the Bloomberg aggregate investment-grade index up 60% in a year. The same companies are more than a fifth of the S&P 500, so the same credits now sit on both sides of a conventional portfolio. PIMCO put the level at nearly 5% of the dollar investment-grade index against 1.2% of the euro one, on record net index-eligible supply of $120 billion in dollars and €22 billion in euros, and found that six of the eight non-financial American investment-grade companies raising more than $20 billion this year were AI-related. The mechanism is mechanical: index-tracking money buys what is issued, so a borrower large enough to move the benchmark is bought by funds that never formed a view on it. Both firms attach limits to their own findings. BNY calls the absolute index share still modest, notes that indexing is far less prevalent in credit than in equities, says late July's simultaneous widening across the hyperscalers likely overshot and has since retraced, and reports that its own flow data did not show the investment-grade fund outflows described elsewhere. PIMCO reads the widening as supply pressure and investor fatigue concentrated in the dollar market rather than deterioration at any company — the same issuers' euro bonds held up better, and the share of the dollar index trading wide of the benchmark is at levels last seen in April 2025 while the euro market barely moved. Both are interested parties: PIMCO owns this paper and BNY custodies it.

The price of protection against a hyperscaler default reached its highest levels on record. Oracle's credit-default-swap spread widened past 200 basis points, against about 40 a year earlier, on a debt load of roughly $130 billion. Societe Generale calculates that five-year spreads across the US hyperscalers now imply a cumulative default probability of about 7%, against roughly 4.5% for investment grade as a whole — though it says Oracle, whose own implied probability exceeds 16%, materially inflates the group figure. None of these companies has ever defaulted on its debt. The counterweight is the size of the market doing the pricing. Clearing data show credit-default swaps on sixteen technology firms traded a combined average of $637.5 million of notional a day in the second quarter of 2026 — about 4% of the $16 billion traded daily across all corporate and sovereign credit — on an average of 54 trades a day, a third of them in Oracle alone; fourteen of the remaining names averaged single-digit daily trades, and Apple's contracts did not trade at all. Reuters columnist Jamie McGeever reads the widening as demand for a cheap hedge rather than an expectation of failure: "These spreads aren't necessarily widening because investors believe these companies will default, but because CDS is a relatively straightforward and cheap vehicle to hedge against their increasing exposure to the AI sector." On a market this thin, a handful of hedging trades sets a price that everyone else then reads as a signal. Reuters is unreachable to direct retrieval; its column is recorded from the Globe and Mail's licensed carry of it.

The SEC's staff determined in a letter that a major subset of data-centre securitisations are not "asset-backed securities" in the traditional sense — because a data centre is a physical facility rather than an amortising instrument like an auto loan — and therefore need not meet the disclosure and risk-retention requirements imposed after the 2008 financial crisis. Those rules had required a deal's sponsor to keep a slice of the debt on its own books to align its interests with investors; the exemption removes that constraint for facility-backed data-centre bonds (data-centre CMBS, backed by mortgages, is unaffected). The request that prompted it, filed by Latham & Watkins on 23 July 2026, sets out both the argument and the market it concerns. Its argument is that data centres "endure well beyond the tenor of the related securities" and "may appreciate in value", so they do not "by their terms convert into cash within a finite time period" and are not the self-liquidating financial assets the statute describes; and that payments to investors depend as much on an operator's ability to hold down running costs as on any lease. Its disclosures: the market has issued more than $50 billion of debt cumulatively since the first such deal in 2018, and these transactions are typically written at a loan-to-value ratio of no more than 70% of appraised value, with repayment anticipated at about five years and a final maturity of twenty-five to thirty. Investors have recourse to the assets, but generally none to the sponsor or operator except for fraud, wilful misconduct or gross negligence. Data-centre asset-backed issuance has climbed to $15.5 billion in 2025 from $2.4 billion in 2020 and is on pace for a record in 2026; a Morgan Stanley note (Katy Huberty) reports data-centre ABS spreads still near year-to-date tights even as broader hyperscaler and data-centre corporate credit has widened, which it attributes to supply technicals rather than deteriorating fundamentals. The staff's own response letter is not publicly retrievable; the request it answered is.

Meta announced a venture with BlackRock to develop a data-centre campus in El Paso, Texas, in which the asset manager — not Meta — holds an 80% stake, and said part of BlackRock's investment will be funded with proceeds from a $12.5 billion debt financing. Meta describes the project as representing an investment of over $10 billion from Meta; CNBC sizes the whole campus at $14 billion, a figure the announcement does not reconcile. Mark Zuckerberg described it the following day as a one-gigawatt site. The structure matters more than the sum: debt raised inside a majority-outside joint venture sits outside the hyperscaler borrowing totals that rating agencies and bank research count, which is the same place Meta's roughly $27 billion Blue Owl venture for its Hyperion campus went in 2025. Meta's own long-term debt, on the balance sheet, rose to $83.664 billion from $58.744 billion in the six months to 30 June. Subsequent reporting fills in the structure's terms, though its arithmetic does not fully reconcile with the announcement: BlackRock-managed funds are described as contributing $4.9 billion in cash for the 80 percent stake against roughly $2.3 billion of Meta-contributed assets for its 20 percent, with Meta retaining sole occupancy under lease and providing a $13 billion residual-value guarantee over the first 16 years — Meta keeps the long-term asset risk while the upfront capital moves off its balance sheet. The dollar figures are carried from a single secondary account whose components exceed its own stated total; the announcement's own numbers remain the anchor. The debt itself priced on Monday 27 July, and it priced expensively: lenders demanded a higher interest rate than they had taken on a similar project a year earlier, in a year of technology borrowing heavy enough to leave debt investors fatigued and to push up the cost of capital across the industry. Alphabet's spending forecast the previous week had already helped drag down the price of Microsoft's and Amazon's bonds. The reporting is paywalled beyond its first paragraph, so no coupon, spread, tenor or rating is available and the 2025 comparable is not named — what is established is the direction.

Broadcom, together with Apollo and Blackstone's credit and insurance business as anchor investors, established the 'AI XPV Platform' to finance more than 20 gigawatts of Broadcom-designed XPU compute for frontier labs — named as including Anthropic and OpenAI — through 2028. The platform launched with an initial $35 billion tranche led by Apollo, in partnership with Blackstone, to fund Anthropic's previously announced expansion of more than one gigawatt of compute in Fluidstack-operated sites from mid-2026. It is a distinct structure from the Google-guaranteed Nexus/Hubbard $15 billion bank loan recorded separately: this capacity runs on Broadcom's custom accelerators rather than Nvidia GPUs, and it is funded through an Apollo-led private-credit vehicle rather than a bank syndicate — private credit and insurance balance sheets, not deposit-taking banks, carrying the exposure. The arrangement confirms and dates the roughly $35 billion Anthropic compute-financing structure previously flagged from secondary analysis (Chicago Booth).

Days after lining up the $15 billion loan for the Google-backed, Anthropic-leased Nexus campus in Hubbard, Texas, the Morgan Stanley-led bank syndicate is planning to push the debt off its own books — refinancing it in the bond market as soon as the loan draws, split into several bond offerings with some portion routed through the leveraged-loan market. The structure shifts risk onto investors: because Google's lease guarantee only takes effect once the data centre is complete, bondholders bear the construction-phase risk of delays and cost overruns, and the paper is expected to carry a speculative-grade rating. Bundling the centre with its on-site natural-gas plant into one financing further complicates pricing, as lenders must underwrite two different risks at once. The move is not isolated: banks had already spent months seeking buyers for more than $50 billion of construction debt on data centres leased to Oracle, and Meta's self-powered 'Project Walleye' had to offer higher yields to place its debt — early evidence that Wall Street's appetite for holding AI-infrastructure loans is cooling even as it keeps arranging them. This is deal reporting sourced to people familiar with the plans, not a filing; the Financial Times broke it and the substance is carried in full by Futu/wallstreetcn.

The debt side of the buildout is now the dominant theme in credit markets. Goldman Sachs Research estimates nearly $500 billion of AI-related debt has been issued so far in 2026 — with the five hyperscalers only about 40% of it, having sold $194 billion this year against $108 billion in all of 2025 — and reports that appetite is starting to strain: the spread on its AI-leadership credit basket has roughly doubled over twelve months, from 74 basis points toward twice that, and orders from insurers, among the largest buyers of long-dated investment-grade paper, roughly halved in the 30-year part of the AI bond complex between the first quarter and late in the second. The exposure also reaches well beyond funded debt: S&P Global has identified about $675 billion of signed-but-not-yet-commenced hyperscaler lease obligations that sit outside reported corporate debt, and a Chicago Booth stress test of the wider financing system (chips, project finance, private credit and GPU-secured lending) put potential losses at up to $140 billion — while noting that, unlike 2008, most first-loss positions sit outside the regulated banking system.

A consortium of banks led by Morgan Stanley is in advanced talks to lend about $15 billion to developer Nexus Data Centers to build a 1.6-gigawatt campus and an on-site natural-gas power plant in Hubbard, Texas, dedicated to Anthropic. Google, which is supplying the chips, has agreed to provide financial guarantees standing behind the power and lease obligations that underpin the financing; Evercore is advising Nexus. The arrangement extends the buildout's newest instrument — the guarantee — that already runs through Nvidia's reported backstop of OpenAI's Ohio leases and Google's earlier agreement to stand behind Anthropic's lease payments: a supplier's credit substituting for a customer's balance sheet. This is deal reporting sourced to people familiar with the talks, not a filing, and the $15 billion is a financing structure rather than a figure disclosed by Anthropic.

Moody's Ratings said AI capital intensity, leverage and off-balance-sheet commitments now threaten the credit quality of six hyperscalers — Alphabet, Amazon, CoreWeave, Meta, Microsoft and Oracle — reporting roughly $460 billion of direct debt across the group and $1.2 trillion of lease commitments, of which more than $820 billion covers leases not yet begun because the data centres are still being built. It counts those unrecognized commitments as debt-equivalent liabilities, and calls the pattern of hyperscaler backlogs resting on deals with pre-IPO labs that spend the money back on the same clouds an 'AI circularity' that 'injects a new dimension of concentration and credit risk.' Its capex projection for the group is now about $785 billion in 2026 and close to $1 trillion in 2027 — above the ~$700 billion for 2026 and $820 billion for 2027 it was being cited for days earlier, and a measure of how fast these estimates are moving. FactSet, counting five of the same companies and excluding CoreWeave, found incremental annual debt had risen to 32% of capex in the twelve months to June 2026 from 9% in fiscal 2024, with aggregate total debt near $700 billion; the two debt totals are not comparable, resting on different company sets and definitions, but they point the same way as Alphabet's cash-flow statement the same week. The counterweight, also FactSet's: total debt to EBITDA remains around 1x or below at four of the five, well inside downgrade thresholds, with Oracle the outlier. Neither firm took a rating action, and Moody's notes demand is real and accelerating.

S&P Global Ratings cut Oracle's long-term issuer credit rating to BBB- from BBB, leaving it a single notch above speculative grade, citing rising business risk and weaker cash flow from its AI-infrastructure pivot. S&P lifted its projection of Oracle's fiscal-2027 capital spending to $90–95 billion from $60 billion, and widened its projected fiscal-2027 free operating cash flow deficit to about $42 billion from $24 billion, with adjusted leverage reaching the mid-4x area. Oracle carries $167 billion of total debt against a $638 billion cloud backlog, roughly half of it tied to OpenAI alone — the customer concentration that turns one lab's fortunes into a rated company's credit risk. Oracle's shares rose 2.7% on the day. Moody's rates Oracle Baa2 with a negative outlook and took no action. S&P's own release could not be retrieved (HTTP 403 with no archived snapshot), so the action is recorded from Investing.com's and FactSet's readings of it, which agree on the notches, the reasoning and the date.

By its Q2 2025 filing CoreWeave carried $11 billion of debt at a blended ~11% rate (two loans at 11% and 15%), plus $34 billion of off-balance-sheet lease payments and ~$30 billion in remaining performance obligations, having lost over $600 million on $2.2 billion of first-half revenue with 71% of it from Microsoft. Bulls point to its take-or-pay backlog; short-seller Kerrisdale called it 'the poster child of the AI infrastructure bubble.' Its stock was still up 160% from its March IPO.

Meta priced a six-part $30 billion investment-grade bond deal — one of the largest US corporate bond sales ever — to term out AI and data-center capex, drawing a $125 billion order book. Alongside a ~$27 billion off-balance-sheet joint venture with Blue Owl for its Hyperion campus, it signaled that even the cash-richest hyperscalers had begun funding the buildout with debt and off-balance-sheet structures rather than cash flow alone.

CoreWeave's transformation into an AI cloud began with two loans from Magnetar Capital — $50 million in convertible notes in October 2021 and $125 million a year later — nearly all spent on Nvidia GPUs. The debt-financed, GPU-collateralized model it pioneered would later scale to billions and become a central exhibit in the bubble debate.