Credit Just Wrote the Script for the AI Buildout

What Needs Your Attention Today

60-Second Regime Matrix

Regime: rate headwind, spending supercycle, credit on watch. The 10-year closed Friday at 5.24% — down a touch on the week, still up ~1.2 points in 2026 and just below 24-year highs. Moody's now says the five giant cloud companies will push US hyperscaler capex above $700B in 2026 — six times the 2022 level — and $870B in 2027, with the lack of electricity as the hard constraint through 2027. Its warning is explicit: higher capital intensity and debt levels could trigger a "reassessment of creditworthiness" if profit growth fails to materialize. The bond market is already flinching — spreads widening, median equity prices declining.

Phase: the financing negotiation. Friday's Firmus Grid collapse is the concrete data point: a $30 billion valuation ask, refused. Only ~42 megawatts of roughly 1 gigawatt of planned capacity actually running. The underwriters cut the price by a quarter and the book still wouldn't clear. Meanwhile TSMC's September numbers — NT$511.86B, up 54.6% year over year — confirm demand is very much alive. The tension is no longer "will AI demand arrive?" It's "who funds the wait?"

What hasn't changed: the software/hardware split from Friday's session (covered in Saturday's issue): Snowflake +7.4% and Datadog +7.1% rallied while Nvidia fell 0.5% and semi ETFs sagged. OpenAI's revenue-math debate (gross vs. net — the $50B vs $70B gap) still awaits Anthropic's pre-Thanksgiving S-1. Posture: watching the debt market, not the tape.

The $70B figure now has a full subscriber-text trail. Bloomberg (Oct 9) reports OpenAI expects to reach or exceed $70B in annualized revenue by end of 2026 — up from ~$50B at the end of September, driven by enterprise growth — with the numbers shared alongside a fundraising round reportedly targeting $30B+ at a $1.4T pre-money valuation. The gross-vs-net quirk is now documented in detail: Anthropic counts as revenue the gross amount of certain sales through cloud partners like Amazon, while OpenAI recognizes only its net share through partners including Microsoft — so OpenAI's revenue can look smaller even if the underlying business is the same (Anthropic's own annualized revenue hit $65B by end of July). "The confusion over the revenue figure largely reflects differences in how OpenAI and Anthropic calculate comparable revenue," wrote Bloomberg Intelligence analysts Mandeep Singh and William Tong. The market read it as another financing event: tech shares led the S&P 500's Thursday losses, the Nasdaq 100 fell 1.4%, and the chip gauge sank 3.4%. Timing notes for the file: OpenAI has pushed its IPO to at least next year (AI safety focus); Anthropic is expected to go as soon as November, with talks underway with UAE funds including Abu Dhabi-based MGX. Last raise: $122B in March at an $852B valuation — the step-up to $1.4T is the number doing the repricing work this week. The action item: if the $30B raise lands at $1.4T, it resets the private-mark valuation bar every AI-infrastructure deal is priced against — Anthropic's expected November IPO becomes the public-market cross-check, and the tell to watch is how the $42B senior notes syndicate.

The Macro Vector: The Buildout Is Borrowing Its Future

Read the weekend's deal coverage side by side and the pattern is unmistakable: the AI buildout is moving from balance-sheet spending to bespoke debt. Per Bloomberg and the Wall Street Journal, Broadcom is arranging financing for the custom AI chips it's developing with OpenAI — ~$30B per Bloomberg, $50B+ per the Journal, two estimates of the same early-stage effort, the final figure still movable. Reuters and the FT report SpaceX is seeking $40B ($10B bank loans + $30B investment-grade debt, Apollo leading) to buy Nvidia chips, expected to close in 2027. The Journal has Oracle discussing a structure with Apollo and Goldman Sachs in which a separate company, set up to hold the debt so it doesn't show on Oracle's own books, would buy the chips for a 1-gigawatt data center and lease them back. Bloomberg separately reports a $60B AI-chip financing package, described as a separate deal — $42B of debt that gets paid back first if things go wrong (senior-secured), backed by a promise from Broadcom that the chips won't be worthless when the lease ends (residual-value support), plus $18B of riskier debt that gets paid back later (junior), led by Blackstone.

One correction worth making explicit, since our own morning framing left it fuzzy: that $60B package funds ANTHROPIC's buildout, not OpenAI's. Bloomberg's Oct 2 full-text reporting names the beneficiary, and the syndicate lines up accordingly — Bank of America, Citigroup, and Morgan Stanley are poised to send syndication letters for the $42B Class A senior-secured tranche, and Broadcom is providing residual-value support that, in the words of people familiar with the talks, "effectively backstops that portion of the financing" (Bloomberg, Oct 2). Blackstone is leading the $18B Class B junior tranche and has committed $9B of its own funds, syndicating the rest; other banks were being invited into the senior tranche as of Friday, and part of the senior notes is expected to be refinanced into the bond market later. The credit read changes with the beneficiary: the senior lenders' risk is underwritten against chips with a guaranteed floor value, not against a lab's promises. What would undermine that read: the residual support covering only a sliver of the $42B, or the senior notes failing to syndicate at target terms. (The parallel data point: Nvidia's August partnership with six finance houses including Blackstone to mobilize $500B+ for AI — the institutional conveyor belt is already running.)

Two more details from Bloomberg's full text (Oct 7) make the financing question concrete. First, the Broadcom/OpenAI number is nowhere near final: "No formal process has begun and plans could change, though early talks indicate the two companies could seek to raise around $30 billion of debt for the OpenAI chips," one person familiar told Bloomberg — the WSJ's $50B+ is a different early-stage estimate of the same effort, not a second deal. The talks sit atop a real hardware program: the two companies unveiled the custom "Jalapeno" chip alongside Broadcom CEO Hock Tan in June, first accelerators go into use late this year, wider deployment follows in 2027, and Broadcom expects to supply 5 GW of chips to OpenAI in 2028 — making OpenAI its second-biggest custom-chip client (Anthropic is the largest in 2027: 5 GW next year, 10 GW the year after). Second, Broadcom set up a platform with Apollo and Blackstone as anchor investors back in June to finance 20+ GW of computing capacity through 2028 — hundreds of billions needed. The plumbing for bespoke AI debt already exists; the weekend's question is who fills it. SpaceX's $40B chip bid belongs in the same queue: Bloomberg confirms talks with banks have begun, and the deal isn't expected to close until 2027.

None of these deals is closed. Every one is debt — not equity, not cash flow. Morgan Stanley's estimate, per the Reuters reporting, puts total external financing needed for AI infrastructure at $1.5 trillion through 2028. This is where the analogy to earlier buildouts breaks before it starts: the railroads comparison (a Brookings Papers draft by Columbia's Stijn Van Nieuwerburgh, per AllFacts, puts the 2025–2032 US AI buildout at 3.63% of GDP per year, vs. 2.24% for US railroads in 1870–1890) flatters the story with growth, but railroads were financed with bonds that were sold against land grants and monopoly freight franchises — visible, enforceable collateral. A data center's collateral is a GPU that loses value the day it ships. Broadcom's residual-value support on that $60B package is, in plain language, someone promising the chips won't be worthless when the lease ends. That's the actual frontier of AI credit: pricing the depreciation of intelligence.

And here is Moody's cold arithmetic underneath it: a data center takes 12–24 months between initial spending and revenue generation. The hyperscalers are borrowing at ~5% to spend during a window when the payback hasn't started. That's not a fraud thesis — Moody's explicitly says AI demand is "substantially exceeding supply." It's a timing thesis. The claim is precise: the buildout needs roughly $300B a year in AI revenue (Goldman's figure, per Saturday's issue) to justify ~$800B of annual capex, and the bridge between now and then is borrowed. What would prove it wrong: TSMC's Thursday call showing margin expansion and raised guidance — revenue turning into profit ahead of schedule. What's priced in (a number): Firmus's IPO book, cut 25% and still abandoned. The market just told you what it charges for AI-infrastructure risk when the collateral is megawatts on paper.

The Anomaly: Firmus Grid

Firmus Grid was supposed to be the clean AI-infrastructure IPO: Nvidia-backed, contracts with hyperscalers including OpenAI, raising up to $5.5B (including the greenshoe) at ~$30B valuation (A$43.7B per Nikkei Asia) — Australia's biggest listing in nearly 30 years. Then investors did the math: roughly 42–46 megawatts of about 1 gigawatt of planned capacity actually running, a $10 billion debt facility already on the table (led by Blackstone), and a valuation that had jumped from $10.5B to $30B in two months. Investors closed the order book on the marketed price Thursday, the company cut it by a quarter (A$11 to A$8.25), and demand still wouldn't clear it. Friday the IPO was pulled; Mas Group, a 3.2% investor, fell as much as 30% in Sydney. Bloomberg now reports the company is seeking a $2–3B private round instead.

Bloomberg's full reporting (Oct 8) sharpens the anatomy of the miss. The joint leads were Bank of America, JPMorgan, Morgan Stanley, and Morgans Financial — the tier-one syndicate couldn't clear the book even after the 25% cut. One of Australia's biggest pension funds, UniSuper, declined to participate outright. The $2B of existing commitments included Nvidia and Blackstone; Jane Street was already a shareholder; Coatue Management had led a $505M round in April. The company's statement was terse: "Having considered recent market volatility and prevailing market conditions, the board determined that the terms on which the offer could proceed would not appropriately reflect the strength of the company's business and long-term growth outlook." The line the weekend will remember came from UniSuper's chief investment officer John Pearce: "We think that Firmus indeed has a compelling story. It just doesn't have a compelling valuation." That's the whole credit question in one sentence — and the macro footnote is that roughly $500B of new debt has been priced this year to fund AI infrastructure, so the private-round fallback isn't a cheap seat either.

The anomaly is the point. Nothing about AI demand changed between the IPO launch and its withdrawal — TSMC was reporting a record quarter the same week. What changed was the price of risk: institutional investors decided, in the middle of a bookbuild, that they wouldn't pay $30B for megawatts on paper at 5%+ rates. That is the weekend's honest signal, and it rhymes exactly with Moody's note: demand substantially exceeds supply, and the financing terms are tightening anyway. Both can be true. The market can love the theme and reject the structure.

On the demand side, Bloomberg's write-up (Oct 8) gives the record quarter more texture than the monthly disclosure: Q3 revenue of NT$1.49T, up 51% year over year — above TSMC's own guidance — with shares hitting a record, and high-performance computing (the AI bucket) accounting for more than 75% of sales in the first half of 2026. The corroborating voices: AMD's Lisa Su says chip demand will stay sharply higher for years; Samsung just posted an almost nine-fold jump in operating profit on a memory windfall; Hon Hai (Foxconn) beat quarterly sales estimates too. TSMC is even considering further expansion in Texas — on top of its $265B campus pledge — and has added $100B to its Arizona investment. The honest framing, via Bloomberg Intelligence's Charles Shum: "The bigger questions at the 3Q briefing concern 2027 and beyond." Thursday's full results and Q4 guidance remain the test Moody's clock cares about.

What to Watch

News (headline, link and one-line tag only):

• US hyperscaler capex to top $700bn in 2026, investors fear overbuild — Moody's Data Center Dynamics (Moody's Ratings) · 2026-10-09 · Credit
  6x the 2022 level; capex eroding FCF, borrowing rising; "reassessment of creditworthiness" risk if profits lag; $870B more in 2027.
• AI chip spending fuels debt financing wave; Broadcom eyes tens of billions for OpenAI TrendForce (Bloomberg/WSJ) · 2026-10-08 · Debt
  Broadcom arranging ~$30B (Bloomberg) to $50B+ (WSJ) for OpenAI's custom chips — early-stage, movable; the $60B Broadcom/Blackstone package is also moving.
• SpaceX seeks $40 billion financing led by Apollo to buy Nvidia chips Reuters/FT · 2026-10-06 · Debt
  $10B bank loans + $30B investment-grade debt, Pimco in the mix; expected close in 2027. Morgan Stanley: AI infra needs $1.5T in external financing through 2028.
• Nvidia-backed Firmus confirms IPO withdrawal after weak demand Financial News (company statement to CNBC; Nikkei Asia) · 2026-10-08 · IPO
  $5B float targeting ~$30.6B valuation pulled Friday; only ~42 MW of ~1 GW running; company says terms didn't reflect long-term outlook.
• Nvidia-backed Firmus seeks major reset after IPO collapse TradingView/GuruFocus (Bloomberg) · 2026-10-10 · Debt
  Now discussing a $2–3B private round with existing investors after the float failed; a $10B Blackstone/Coatue debt facility sits behind it.
• TSMC revenue hits record NT$1.49 trillion TechNode (Reuters) · 2026-10-09 · Demand
  Q3 ~$46.7B, ~50% YoY, beating analyst consensus (NT$1.46T LSEG SmartEstimate) and TSMC's own $44.6–45.8B guidance. Full results Oct 15.
• Broadcom lines up $50B+ private-credit deal for OpenAI chips AI Weekly (WSJ) · 2026-10-07 · Debt
  Apollo and Blackstone among lenders in talks; targeted close by year-end; Oracle separately negotiating an Apollo/Goldman 1GW off-balance-sheet structure.

Earnings & data: US CPI Wednesday 8:30 ET (last print before the Oct 27–28 Fed meeting — a hot print with the 10Y above 5.3% is the de-risk signal); ASML Wednesday; TSMC's full Q3 results and Q4 guidance Thursday — margins and guidance matter more than revenue now. Big banks kick off earnings season.