2026-09-04 — ByteDance locks in a $29.6 billion unsecured loan, the year’s second-largest Asian financing after SoftBank’s, to fund AI buildout

• ByteDance secured a $29.6 billion syndicated loan from nearly 30 banks — Citigroup and JPMorgan coordinating, with Chinese, US, European and Singaporean lenders subscribing, Chinese banks taking over 60% of the total — after upsizing from an initial $20 billion target on more than $30 billion of investor demand. The three-year facility (with options to extend two more years) is unsecured — ByteDance pledged no collateral. Officially “for general corporate purposes,” sources say it mainly funds AI data-center investment in Southeast Asia, where ByteDance is the offtaker for capacity Bloomberg says maps to up to $70 billion of 2026 AI capex. This is the same debt-financed-buildout pattern this thread already tracks in SoftBank’s $40 billion March raise and Anthropic’s $15 billion credit facility (09-03), now with a Chinese hyperscaler on the same instrument type. (Investing.com, relaying Reuters, TechNode)

2026-09-02 — BOJ Governor Ueda hints at a September hike and the yen surges on intervention alert, a day after the 10-year JGB crossed 3%

• Bank of Japan Governor Kazuo Ueda said after the G20 finance meeting that the board would “decide on policy with upside price risks in mind” at its September 17-18 meeting, and markets read it as tacit acknowledgement of pricing that reached roughly 94% odds of a hike; board member Hajime Takata separately left a larger-than-25bp move open. The yen, which sat at 160.1 per dollar on 09-01 with intervention odds rising, reversed sharply — Bloomberg has it surging “with traders on high alert for further intervention,” Nikkei Asia has it briefly touching the low-158 level. This is the cross-border transmission leg this map’s context file twice deferred: the largest foreign holder of Treasuries hardening toward a hike while its own long end sits at a 30-year high. ⚠️ Dated 09-02 and available all day; it sat in this map’s own collector buffer 108 times under the four watchlist terms added on 09-01, and was not read until the 09-03 finalize. (Bloomberg, Bloomberg, Nikkei Asia)

2026-09-01 — Japan’s 10-year JGB crosses 3% for the first time since 1996, and the US Treasury Secretary publicly presses Tokyo to hike

• Japan’s 10-year government bond yield crossed 3% on 09-01, its first time above that line since 1996, while US Treasury Secretary Scott Bessent went on CNBC to press Tokyo and the Bank of Japan publicly for rate hikes. Reporting on the exchange flagged the specific transmission risk: Japan is the largest foreign holder of US Treasurys, and a yen-defence intervention funded by selling them would push US long-end yields the same direction this thread already tracks domestically. This is the escalation of the cross-border leg this thread first recorded on 08-19, when a JGB push toward its highest yield since 1996 took SoftBank, Kioxia, Arm and Tokyo Electron down together in one Tokyo session — the mechanism is the same, the yield has now actually crossed, and a US cabinet officer is now an actor in it rather than an observer. attention/capital-context.yaml has twice flagged its own reading as “almost entirely US-Fed-framed” with a proper BOJ/JGB leg deferred to a /week-scale addition; this is that deferral coming due. (CNBC, “Japanese borrowing costs hit 30-year high as Bessent says Tokyo may intervene to boost yen,” 2026-09-01)

2026-08-21 — Broadcom’s SPV goes to the debt market for $70-80bn to build Anthropic’s chips

• CNBC reported that lenders are structuring a roughly $45bn senior tranche and a roughly $35bn junior tranche — $70-80bn total — routed through a special-purpose vehicle to fund AI chip capacity for Anthropic, with Apollo and Blackstone in talks to participate. It extends the $35bn Broadcom/Apollo/Blackstone partnership announced in June (custom chips plus networking, targeting more than 20 GW of compute to top AI labs by 2028) at roughly double the size. (CNBC, Yahoo Finance)
• What is new today is the structure, not the story. The raise itself broke 08-20 on Bloomberg at a vaguer “$60-100bn”; today’s reporting adds the SPV and the senior/junior split, which is the part that decides who absorbs a loss first. Reported totals still span $60bn to $100bn across outlets — treat the range as the fact and any single number as one outlet’s draft, since a spread that wide usually means a deal still being shopped rather than sized. (SiliconANGLE)
• This is the thread’s thesis arriving at scale. A junior tranche near $35bn is too large to place quietly, so whatever spread it prices at becomes the first public market estimate of AI-buildout credit risk for lab compute — the number this thread was opened to wait for.

2026-08-20 — Nebius’s convertible-notes raise grew on syndication, and closed with a debt-for-equity swap on the side

• Nebius priced its convertible-notes offering upsized from the $4.5B announced 08-19 to $5.0B — $3.0B of 0.50% notes due 2030 and $2.0B of 4.50% notes due 2034 — at 40% and 45% conversion premiums over its $223.90 closing price, netting roughly $4.94B (up to $5.68B with the purchasers’ option). The first instance on this thread’s record of a neocloud’s AI-buildout debt raise growing between announcement and pricing rather than shrinking or holding — the opposite direction from Nvidia’s OpenAI-guarantee trajectory ($250B → under $120B → $105B) and CoreWeave’s own DDTL, which was flexed wider on spread but not on size. Concurrently, Nebius exchanged $800M of its existing 2029/2031 convertible notes for ~15.8M Class A shares with select holders — swapping old debt for equity dilution on the same day it adds new, larger debt. Settlement expected 2026-08-24. (SEC 6-K/exhibit 99.1)

2026-08-19 — The rate channel goes cross-border, then reverses at home

• Japan’s own version of Monday’s story: a 10-year JGB push toward its highest yield since 1996 broke Tokyo’s AI-proxy stocks together, with a sell-side analyst naming the mechanism explicitly. Nikkei 225 closed −2.54% to −2.56% (~67,460); SoftBank Group −8%+, Kioxia −9-13%, Arm −6% in US trading, Tokyo Electron −4.4%. A Daiwa Asset Management strategist: “higher yields will push up borrowing costs for hyperscalers, triggering market doubts over capital expenditure prospects.” This is the same transmission channel this thread opened to watch, appearing a second time in five days on a different continent. (cryptobriefing.com, TradingKey)
• Hours later, the US leg reversed via policy intervention rather than market stabilization. Treasury at least doubled its long-bond buyback size (10-20yr and 20-30yr sectors, $2B → “at least” $4B per operation), an unscheduled escalation. The 30-year fell from Tuesday’s 5.32-5.34% intraday high to 5.196%, the 10-year to 4.647%, and US equity futures/stocks rallied. This directly answers the “what to watch next” note logged here 08-18 — the market-test of the next investment-grade order book never had to happen, because Treasury acted first. (CNBC)

Watch: Whether AI-buildout debt is becoming its own underwritten, rated, publicly-watched risk category rather than a footnote inside each borrower’s story. Track: whether rating agencies publish an explicit methodology for compute-collateralized paper, whether any of it prices wide of comparable real-asset debt, the first covenant breach or restructuring, and whether Nvidia’s CDS moves with issuance volume. The open question is whether GPU compute behaves like an infrastructure asset with a 20-year life or like equipment on a 3-4 year depreciation clock — the two produce very different recovery rates in a default.

2026-08-18 — The rate channel reaches the AI trade, with no story attached — then the close adds a dollar figure to what “off-balance-sheet” means

Rebuilt at finalize with the day’s close and a same-day credit-market follow-up; the morning/afternoon reads are unchanged from the first pass.

• The 30-year Treasury extended to 5.32-5.34% intraday (a fresh 19-year+ high) and the semis broke together, before easing into the close. Live quotes at 10:09 ET: NVIDIA −2.29% at $219.79 (prior close $225.01) · AMD −5.2% at $479.90 (prior close $506.00) · Micron −5.4% at $957.00 (prior close $1,011.75), handing back the $1,000 it crossed on Monday inside one session. Closing levels: S&P 500 7,691.76 (−0.69%) · Nasdaq 26,289.71 (−1.33%) · Dow 53,343.40 (−0.22%) — Nasdaq deepened through the day, Dow recovered some ground. The 30-year eased off its intraday peak to close near 5.29%. (Yahoo Finance live, TheStreet)
• The co-movement is the evidence. An idiosyncratic story moves one name; a discount rate moves the whole duration bucket. No company-specific news attaches to any of the three. This is the transmission channel this thread was opened to watch, appearing for the first time without a narrative in front of it — Monday’s repricing was visible only in the long bond while equities shrugged.
• By the close, memory/storage extended further than the megacap names: Marvell ~−8%, Seagate ~−9%, SanDisk ~−9.1%, Western Digital ~−5 to −7%. A WSJ footnote analysis circulating the same day put nine major tech companies’ combined off-balance-sheet AI-related commitments at roughly $3 trillion — a materially larger aggregate than the ~$1.65T five-hyperscaler figure this map logged from the Meta 8-K read earlier in the week (🔄 Map changes, 08-18 digest). Worth carrying as the same accounting-boundary question at a different, larger scale rather than a contradiction — the two counts likely differ in which firms and which obligation types they include. (Yahoo Finance — WSJ report sends memory stocks down)
• ⚠️ Excluded as unverified: a Jefferies note recirculating this morning framing AMD’s Helios/ROCm as taking Nvidia share. The AMD event it discusses was 07-22/23, its price-target figures conflict across sources, and AMD fell harder than Nvidia today, which is the opposite of what that thesis predicts.
• What to watch next, in order: whether the next investment-grade tech issue’s order book holds against the ~36% initial-order pull rate logged in the week to 08-17. Equity duration reprices first because holders can sell; credit follows with a lag if the move is durable. 08-19 update: the answer arrived same-window, from the other direction — Treasury itself intervened (see today’s digest) rather than the market testing an issuer’s book.

2026-08-17 — Long end to a 19-year high while oil sits still

• The 30-year Treasury yield reached 5.31% at 2:14pm ET, its highest since June 2007, after opening the session at 5.28%; the 10-year sat at 4.71–4.72% and the 2-year eased to ~4.17% — a bear steepening rather than a parallel shift. (Yahoo Finance live blog)
• The cross-asset divergence is the signal: crude did not move. The US–Iran framework formally expired the same day and Trump threatened to bomb Oman, yet Brent gave up its early gains and held near $89–90. A geopolitical supply shock lifts crude and breakevens together; a fiscal/supply repricing lifts long nominal yields alone. Today gave the second pattern cleanly, pointing at term premium — issuance into a debt stock near $40T, and inflation persistence — rather than at the Gulf.
• The supply mechanism is concrete: global bond issuance crossed $5 trillion Monday, over a month faster than last year’s record. US investment-grade ran $136B in two weeks with August on track for a monthly record, and $1.4T year-to-date — 9% above the pace of 2020, which finished at a record $1.75T. Blue-chip borrowers are funding “everything from acquisitions to the artificial intelligence buildout,” and investors pulled ~36% of initial orders on average this week after final pricing was squeezed, twice the prior week’s rate. (Bloomberg)
• Why it bites on this thread specifically: AI datacenter financing is long-duration and discounts off the long end, not the policy rate. The PORTS-Pike leases Nvidia guaranteed today attach from 2028 and run 20 years. A persistent rise in term premium raises the hurdle rate for exactly that cash-flow profile, independent of what the Fed does in September.

2026-08-15 — Bond traders converge on the same concern Burry raised alone two days ago: $70B of “shadow credit backstops”

• Bloomberg reports bond traders are now pricing roughly $70 billion of off-balance-sheet “residual value” guarantees across the AI buildout — contingent obligations that let a chipmaker or hyperscaler backstop its own customers’ financing without the guarantee showing up as a liability on its own books. Named exposures: Nvidia can backstop up to 25% of its own $500B compute-financing platform (per Jensen Huang directly); Broadcom is backstopping Anthropic’s $35B “Big Sky” chip deal; Meta’s “Beignet” ($27B) and “Sopaipilla” ($13B) data-center debt deals are cited as the template these newer guarantees are copying. Moody’s warned “multiple such transactions over a short period” pose a primary risk to a guarantor’s own credit profile; DoubleLine’s Mariya Entina called the ratings treatment “gaming the system”; CreditSights called the structure “pro-cyclical” — “nearly costless in the boom phase, but becomes most relevant in a severe, abrupt downturn.” The first time this thread has multiple named credit-desk sources converging on the same specific concern Michael Burry raised alone on 08-13. (Bloomberg, Yahoo Finance, mirror)
• The same day, Nvidia as Lender logged Nvidia narrowing its own OpenAI Ohio guarantee from up to $250B in talk to under $120B, phase-one only — the first concrete instance of a backstop actually shrinking rather than growing, even as the aggregate shadow-guarantee figure across the ecosystem (Broadcom, Meta) keeps expanding. See this digest’s interpretation sidecar for the full mechanism read.

2026-08-13 — The debt shows up one layer up (a chipmaker, not a neocloud), then a named bear case, a cross-hyperscaler cash-flow number, and the first junk-rated print

• AMD is raising up to $5 billion in its largest-ever dollar bond sale, and Reuters/Yahoo Finance both frame it against AMD’s own AI-chip agreements and its $5B equity stake in Anthropic specifically — the first time in this thread’s record that debt has surfaced at the chipmaker layer rather than the neocloud/data-center layer (CoreWeave, Lambda) that has carried every prior entry. The four-tranche offering (3- to 10-year maturities, led by Barclays, Bank of America, Citigroup, JPMorgan, Morgan Stanley, and Wells Fargo) is officially for “general corporate expenses, which may include the repayment of debt” — AMD has $875M maturing next month — so the AI/Anthropic framing is press narrative around a nominally generic raise, not an earmarked use of proceeds; worth tracking whether that framing holds up or whether AMD itself ever ties debt explicitly to the stake. Reads directly against Circular Financing (ai lens), whose entity list (Nvidia/ OpenAI/Oracle) doesn’t yet include AMD/Anthropic despite this being the same shape of question. Priced the next morning (08-14) at $4.75B across the four tranches, with the longest tranche tightening ~0.25pt from initial talk to +0.9pt over Treasuries; AMD shares +5.6% on the print. (Yahoo Finance/Bloomberg, Bloomberg, pricing)
• A Jefferies note (strategist Christopher Wood) quantified the capex-vs-cash-flow tension across all four largest hyperscalers at once for the first time on this thread’s record: combined free cash flow fell to $7B in Q2 2026 from $60B in Q4 2025, against $165B of combined quarterly AI capex. Jefferies’ own read is not bearish (cloud revenue +38% y/y is framed as the return side), but the compression number itself generalizes what this thread had so far only seen company-by-company (Oracle’s FCF, CoreWeave’s debt service). (BusinessToday)
• Michael Burry named Nvidia’s 08-10 $500B Wall Street financing platform directly, calling its credit structures “abnormal” and its revenue “largely financed through circular arrangements” that are “off-balance-sheet and undisclosed” — this thread’s first named, specific bear case rather than general AI-boom skepticism. He also flagged private-equity-owned insurers filling balance sheets with illiquid, AI-tied asset-backed paper — the same structural shape as the Jane Street bond below. See this digest’s interpretation sidecar for the full mechanism read (held at plausible, both directions argued). (Tekedia)
• A data-center developer sold $2.25B of green bonds — rated below investment grade by Moody’s — backed by a 15-year Jane Street lease on a 149-megawatt Oklahoma facility; this thread’s first sub-investment- grade AI-datacenter asset-backed print. Priced at 99.5 cents on the dollar to yield almost 9%. Sustainable/green-labeled data-center bond issuance is already up sharply industry-wide (Bloomberg: $11.2B in 2025 vs $7.6B in 2024) — a junk-rated tranche inside that growth is exactly the shape of data point this thread exists to catch. (Bloomberg)

2026-08-12 — A second neocloud shows the same circular-financing shape, and closes the day with a first: an investment-grade rating

• Lambda’s Nvidia-GPU loan priced at a final $926 million — led by Morgan Stanley, nearly $2B oversubscribed — to buy the Nvidia GPUs that Nvidia itself is both financing and leasing back, and picked up a Moody’s Baa2 rating: the first investment-grade Term Loan B a private neocloud has carried. Final terms (SOFR+300bps, tightened 75bps during syndication, 99.5 cents, maturing 2030-12-31 with full amortization) came in bigger and tighter than the $917M/S+375bps preliminary terms reported earlier the same day. The loan funds Lambda’s side of a ~$1.5B, 18,000-server contract under which Nvidia leases back the compute it sold — meaning Nvidia sits simultaneously as Lambda’s investor, chip supplier, largest customer, and counterparty to the debt that pays for all three. The Bank for International Settlements flagged this exact structure — echoing Google’s own TPU model — in a June 2026 warning that vendor-financed circular deals are “typically poorly disclosed, with risks of the same asset being pledged multiple times.” Second confirmed transaction-level data point in this thread, after CoreWeave’s own loan — and the Baa2 rating is the first sign a rating agency, not just syndicate demand, is willing to treat this exact structure as investment-grade; the watch question (20-year infrastructure asset vs. 3-4-year depreciating equipment) is now testable across two borrowers, not one. (The Next Web, Bloomberg, Yahoo Finance/Lambda release)

2026-08-11 (later) — CoreWeave’s own Q2 print puts backlog growth and financing cost in the same quarter for the first time

• CoreWeave’s Q2 earnings are the first data point in this thread where backlog growth and financing-cost growth show up in the SAME company’s own numbers at once: net interest expense rose 140% YoY ($640M vs $267M) — faster than 112% YoY revenue growth — in the same quarter backlog grew to roughly $129B including new Q3 commitments. It is the first real-world test of whether the SOFR+550 spread that priced three days before this print reflects CoreWeave-specific risk or a broader AI-infrastructure debt repricing — this thread’s central open question. Also disclosed: over $10B in new unsecured debt/ converts and a separate $3.1B DDTL carrying a 1.35x DSCR covenant from 09-30-26 — the first hard financial covenant this thread has recorded anywhere in the AI-buildout debt stack. Shares still jumped 14% after hours and guidance was raised, not trimmed — see the full print on CoreWeave Bet. (SEC 8-K, CNBC)

2026-08-11 — Opened (ben-steer)

• Opened after three data points landed on the same day at three very different scales, none of which had a home that could hold a market-structure question. Until now all three were folded into Nvidia as Lender, which is an ACTOR thread — what Nvidia is doing — and could not carry what the debt is actually worth. (ben-steer, 2026-08-11)

• Scope is deliberately narrow: the borrowing, not the buildout. Capex volume belongs to Hyperscaler Capex, vendor equity stakes to Nvidia as Lender, and asset-manager ownership to Asset Managers Build. This thread is only about how the debt is priced, rated and recovered.

2026-08-10 — AI debt priced four ways in one day

• Bloomberg Opinion launched “AIndicators,” a free credit-market framework built specifically to flag whether lenders are mispricing AI infrastructure — the first dedicated public tracker treating AI-buildout debt as a category worth its own instrumentation rather than a line inside each issuer’s credit.

• JPMorgan led a $441M debt facility for Global AI, a two-year-old AI-infrastructure firm — small next to the day’s headline numbers, and that is the point: conventional syndicated debt is now reaching companies with almost no operating history, on the strength of the asset class rather than the borrower.

• Nvidia’s $500B+ memorandums of understanding with six asset managers propose special-purpose-entity bonds collateralized by GPU capacity itself, with Goldman Sachs — the only bank among the six — leading public debt issuance. These are MOUs, not final agreements, and no transaction has been named. (NVIDIA Newsroom)

• A fourth data point, found independently during the 08-11 finalize pass rather than flagged by any of the above: CoreWeave closed its own $2.6B delayed-draw term loan (JPMorgan/MUFG-led) at SOFR+550 — a 10.44% yield-to-maturity — after guidance was flexed wider from an initial S+425-450. This is the first CONFIRMED transaction-level coupon in this thread’s record, not commentary or an unconfirmed report — it resolves CoreWeave Bet’s own 07-30 flag (“exact new yield not independently confirmed”) and lands three days ahead of CoreWeave’s Q2 earnings call. (SEC EDGAR 8-K/credit agreement, CoreWeave press release via AOL)

2026-08-09 (late catch, added 2026-08-11) — Moody’s flags banks’ own AI-vendor dependence as a risk category

• Moody’s warned that banks’ rapid AI adoption is building a concentrated “vendor dependence risk” on a small set of foundation- model and cloud providers — naming OpenAI and Anthropic specifically — that could let a single outage cascade across the financial sector and eventually let dominant vendors dictate pricing to the banks that depend on them. A different axis from the rest of this thread (which tracks how AI-buildout DEBT is priced): this is banks’ own operational dependence on AI vendors, the demand-side mirror of the financing-side risk above. Surfaced by this digest’s coverage critic as a genuine miss against the original 08-09 digest; logged here at its true event date rather than reopening that already-closed digest. (The Guardian, NewsBytes)

2026-07-27 — The prior signal this thread was split out of

• Nvidia’s credit default swaps printed a record single-day widening as the market repriced it as a guarantor rather than a vendor — the earliest market-side evidence that the buildout’s financing structure carries its own risk, logged at the time inside Nvidia as Lender and carried here as this thread’s starting observation.