1HeadlinesWhat does this mean?
Stage 3 (forced deleveraging), unchanged from the last report; earliest downgrade 11 October. AI economics: Watch, unchanged.
The bad
- The investment-grade bond market has split into AI and non-AI credit.3 AI-related issuers were trading around 115bp spreads versus 78bp for the broader investment-grade market.1
- Investors were deliberately demanding more yield from hyperscalers, data-centre operators and chip-related borrowers despite their generally strong credit ratings.
- SoftBank was launching more than $11bn of junk-rated bonds, partly to finance its OpenAI investment, with indicated yields around 9–10%.2
The good
Not what it looks like
- The AI premium does not mean lenders expect AI borrowers to default. The reason appeared to be primarily supply and concentration: hyperscaler gross issuance was forecast to rise sharply, while some institutional investors were approaching exposure limits once direct corporate bonds and associated data-centre vehicles were considered together. That is discrimination rather than contagion.
2ScorecardWhat does this mean?
| Pillar | Status | Trend | Why |
|---|---|---|---|
| AI equities | Watch | Unchanged | The semiconductor index closed 15.0% below its 52-week high on 21 September.4 The wider market remained risk-on and semiconductor demand was still strong. |
| Rates and macro | Elevated | Unchanged | US 10-year at 4.96%, inside the Elevated band.7 No new rates evidence in this report. |
| Broad credit (the firewall) | Calm | Unchanged | Investment-grade spread 77bp and high-yield 266bp, both well inside Calm.312 In the comparison with AI issuers the broader investment-grade market was at 78bp, so lenders are discriminating rather than pulling back from everyone.1 |
| AI-specific credit | Elevated | Unchanged | AI-related issuers were trading around 115bp versus 78bp for the broader investment-grade market, with investors deliberately demanding more yield from hyperscalers, data-centre operators and chip-related borrowers despite their generally strong ratings.1 The conventional bond market itself is now attaching an identifiable AI premium, alongside Project Jupiter’s troubled syndication. |
| AI financing flow | Elevated | Unchanged | SoftBank was launching more than $11bn of junk-rated bonds, partly to finance its OpenAI investment, at indicated yields around 9–10%.2 Hyperscaler gross issuance was forecast to rise sharply, and some institutional investors were approaching exposure limits once direct corporate bonds and associated data-centre vehicles were considered together. |
| Leverage and liquidity | Elevated | Unchanged | No new forced-selling evidence in this report; status held from the 20 September report. |
| Private credit | Elevated | Unchanged | No new private-credit evidence in this report; status held from the 20 September report. |
| AI economics | Watch | Unchanged | No new hyperscaler capex or cash-flow evidence in this report; status held from the 20 September report. Semiconductor demand was still strong, and the AI premium reflected supply and concentration rather than expected defaults. |
3Change since last reportWhat does this mean?
Compared against the 20 September 2026 report, back-scored the same way. Scores for this report were back-scored on 4 November 2026 from its own text and from FRED and Yahoo Finance closing data, because it predates the scorecard format.
| Last report (20 Sep) | This report | Change | |
|---|---|---|---|
| Stage | 3 (evidence 11 Sep) | 3 (evidence 11 Sep) | Unchanged. No new forced-selling evidence; the 11 September entry evidence still holds, so the earliest downgrade is 11 October. Stage 2 is now confirmed beneath the headline. |
| AI economics | Watch | Watch | Unchanged. No new capex or cash-flow evidence; semiconductor demand was still strong. |
| AI equities | Watch | Watch | Unchanged. The semiconductor index closed 15.0% below its 52-week high, against 18.5% at the last report’s close.4 |
| Rates and macro | Elevated | Elevated | Unchanged. The 10-year closed at 4.96%, against 5.01% at the last report’s close.7 |
| Broad credit | Calm | Calm | Unchanged. Investment-grade flat at 77bp; high-yield 266bp from 268bp.312 |
| AI-specific credit | Elevated | Elevated | Unchanged. Worse inside it: the AI premium in mainstream investment-grade bonds (115bp against 78bp) is new, and it moves the repricing out of troubled project finance and into conventional credit.1 |
| AI financing flow | Elevated | Elevated | Unchanged. Worse inside it: SoftBank is launching more than $11bn of junk bonds at indicated yields of 9–10%, and investors are nearing exposure limits on hyperscaler paper.2 |
| Leverage and liquidity | Elevated | Elevated | Unchanged. No new evidence. |
| Private credit | Elevated | Elevated | Unchanged. No new evidence. |
Net: no pillar status moved, but the chain did. The AI premium in investment-grade bonds is a second independent sign of AI-specific credit stress within a month of Project Jupiter, so Stage 2 is now confirmed beneath the Stage 3 headline. Pressure built inside AI-specific credit and AI financing flow; equities improved inside Watch, and the firewall held.
4Transmission chainWhat does this mean?
- 1 AI equity correction Confirmed
- 2 AI-specific credit stress Confirmed
- 3 Forced deleveraging Confirmed 11 Sept Earliest downgrade 11 Oct
- 4 Firewall breached Not present
- 5 Systemic liquidity event Not present
| Stage | Status |
|---|---|
| 1 — AI equity correction | Confirmed. Tripwire met on 3 September and not yet cleared. The index is back within 20% of its high (15.0% below on 21 September) but has not yet closed inside the line for five consecutive days.4 |
| 2 — AI-specific credit stress | Confirmed. Entry evidence from this report: two independent signs within a month: Project Jupiter’s debt failing to syndicate on 18–19 September, and AI-related investment-grade issuers trading around 115bp against 78bp for the broad market on 22 September.1 |
| 3 — Forced deleveraging | Confirmed. Entry evidence confirmed from the 13 September report: JPMorgan curbed lending to Situational Awareness after its margin call and forced liquidation, reported 11 September. No new forced-selling evidence; earliest downgrade 11 October. |
| 4 — Firewall breached | Not present. Investment-grade 77bp and high-yield 266bp, far below the 130bp and 450bp tripwires.312 The AI premium is discrimination, not contagion: the broad market has not followed. |
| 5 — Systemic liquidity event | Not present. No funding-market malfunction reported. |
5JudgementWhat does this mean?
This was meaningful deterioration: the investment-grade bond market has split into AI and non-AI credit.3 It matters because the repricing was no longer confined to troubled project finance or weaker private-credit structures. The conventional bond market itself was attaching an identifiable AI premium, around 115bp for AI-related issuers against 78bp for the broader market, and investors were demanding it from hyperscalers, data-centre operators and chip-related borrowers despite their generally strong credit ratings.1
The reason appeared to be primarily supply and concentration rather than expected defaults. Hyperscaler gross issuance was forecast to rise sharply, and some institutional investors were approaching exposure limits once direct corporate bonds and associated data-centre vehicles were considered together. At the riskier end, SoftBank’s more than $11bn of junk-rated bonds, partly to fund its OpenAI investment, at indicated yields around 9–10%, shows what the weaker AI borrowers are paying.2 However, the wider market remained risk-on and semiconductor demand was still strong.
The significant distinction was therefore:
Previously: individual AI financings were showing stress while mainstream investment-grade credit remained largely indifferent.
Now: mainstream bond investors were systematically pricing AI-related investment-grade borrowers differently from the rest of the corporate market.
That was discrimination rather than contagion. Contagion would be the 78bp moving too.
6Key indicatorsWhat does this mean?
Market data as at the close on 21 September 2026, compared with the close used by the 20 September 2026 report (18 September 2026). Reported figures are carried forward from the report that last cited them until a newer one appears.
| Indicator | Latest | Previous | Change | Tripwire | Source |
|---|---|---|---|---|---|
| AI equities | |||||
| PHLX Semiconductor Index (SOX) | 12,433; 15.0% below its 52-week high (22 Jun) | 11,922 | +4.3% | Stage 1 when more than 20% below its 52-week high for five closes | Confirmed, Yahoo Finance ^SOX4 |
| Nasdaq 100 | 30,482 | 29,644 | +2.8% | Context | Confirmed, Yahoo Finance ^NDX5 |
| KOSPI | 7,008 | 6,894 | +1.6% | Context | Confirmed, Yahoo Finance ^KS116 |
| Rates and macro | |||||
| US 10-year Treasury | 4.96% | 5.01% | −0.05pp | Elevated 4.75–5.25%; Stressed above 5.25%, each for five closes | Confirmed, FRED DGS107 |
| US 30-year Treasury | 5.29% | 5.34% | −0.05pp | Context | Confirmed, FRED DGS308 |
| Brent crude | $116.15 | $119.66 | −3.51 | Context | Confirmed, FRED DCOILBRENTEU9 |
| US dollar index (broad) | 119.40 | 119.51 | −0.1% | Context | Confirmed, FRED DTWEXBGS10 |
| Yen per US dollar | ¥157.46 | ¥156.87 | +0.59 | Context | Confirmed, FRED DEXJPUS11 |
| Broad credit | |||||
| ICE BofA US investment-grade spread | 77bp | 77bp | ±0bp | Watch at 100bp; Stage 4 above 130bp for five days | Confirmed, FRED BAMLC0A0CM3 |
| ICE BofA US high-yield spread | 266bp | 268bp | −2bp | Watch at 350bp; Stage 4 above 450bp for five days | Confirmed, FRED BAMLH0A0HYM212 |
| ICE BofA CCC & lower spread | 1,077bp | 1,083bp | −6bp | Early warning only | Confirmed, FRED BAMLH0A3HYC13 |
| ICE BofA CCC & lower effective yield | 15.50% | 15.59% | −0.09pp | Early warning only | Confirmed, FRED BAMLH0A3HYCEY14 |
| AI-specific credit | |||||
| AI investment-grade premium over the broad index | ~115bp vs 78bp broad IG | — | first reading | n/a | Reported, Reuters1 |
| Oracle 5-year CDS | Not yet reported in the series | — | — | n/a | — |
| Meta 5-year CDS | Not yet reported in the series | — | — | n/a | — |
| CoreWeave 5-year CDS | Not yet reported in the series | — | — | n/a | — |
| SoftBank 5-year CDS | Not yet reported in the series | — | — | n/a | — |
| Named data-centre project debt | Project Jupiter (~$18bn) quoted 89–91c (19 Sep report) | Project Jupiter (~$18bn) quoted 89–91c | no new reading | n/a | Carried from the 19 September report |
Yahoo Finance ^SOX, daily closes to 21 Sep.
FRED DGS10, daily closes to 21 Sep.
FRED BAMLC0A0CM, daily closes to 21 Sep.
FRED BAMLH0A0HYM2, daily closes to 21 Sep.
Faint vertical lines mark each of the last 10 reports. Dashed red lines are the published thresholds.
What would change our view
Would make it worse (up toward Stage 4): The broad investment-grade spread starting to move with the AI premium rather than staying put; high-yield above 450bp or investment-grade above 130bp for five consecutive trading days would breach the firewall. A failed or pulled SoftBank or hyperscaler deal would show the exposure limits biting, and a second forced-selling event would restart the Stage 3 clock.
Would make it better (down to Stage 2): A month with no new forced-selling evidence takes the stage off 3; the clock runs from 11 September, so the earliest that can happen is 11 October. The AI premium narrowing as issuance is absorbed would point back toward pricing for supply rather than risk.
Would move the AI-economics axis: A sign that the premium reflects expected defaults rather than supply, or any weakening in semiconductor demand or hyperscaler capex, would move it worse; continued strong chip demand would keep it at Watch.
7SourcesWhat does this mean?
- Reuters — Corporate bond buyers get picky with flood of AI debt (2026-09-22) —
reconstructed-match - 9fin — SoftBank launches $11bn-eqvt bond offering to power AI ambitions (2026-09-21) —
reconstructed-match - Federal Reserve Bank of St. Louis (FRED) — ICE BofA US Corporate Index Option-Adjusted Spread (continuous) —
official-data-series - Yahoo Finance — PHLX Semiconductor Index (^SOX) price history (continuous) —
market-data-series - Yahoo Finance — Nasdaq 100 (^NDX) price history (continuous) —
market-data-series - Yahoo Finance — KOSPI (^KS11) price history (continuous) —
market-data-series - Federal Reserve Bank of St. Louis (FRED) — 10-Year Treasury Constant Maturity Rate (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — 30-Year Treasury Constant Maturity Rate (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — Crude Oil Prices: Brent - Europe (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — Nominal Broad U.S. Dollar Index (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — Japanese Yen to U.S. Dollar Spot Exchange Rate (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — ICE BofA US High Yield Index Option-Adjusted Spread (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — ICE BofA CCC & Lower US High Yield Index Option-Adjusted Spread (continuous) —
official-data-series - Federal Reserve Bank of St. Louis (FRED) — ICE BofA CCC & Lower US High Yield Index Effective Yield (continuous) —
official-data-series
Not investment advice. Scoring, tripwires and the persistence rule are set out in How to read the AI Market Watch; terms are explained in the explainers.