AI Ecosystem Stress Test Results ST-2026Q3-01 (July–September 2026, Preliminary)

at.AI / Published November 2, 2026

Run ID: ST-2026Q3-01
Type: Preliminary
Period covered: July–September 2026
Model version: v1.0
Data cut-off: November 2, 2026
Test horizon: 2026Q4–2028Q3 (8 quarters)

This is a preliminary release. July–September results for NVIDIA, CoreWeave and SoftBank are not yet reflected, and data coverage is about 70%. The full version (ST-2026Q3-02) will be published in late November 2026 and will replace the figures on this page. All figures are model estimates based on our own estimates and assumptions from public information, and are not investment advice. $bn = billions of US dollars.

Key findings

The weak side is funding, not demand: Demand (S1) and price (S2) shocks stay at M3 even at L2. Credit market shutdown (S3) and a shorter GPU refresh cycle (S4) reach M4–M5 at L2.
Current trends alone are M3: In a scenario that mildly extends current price declines and credit deterioration (S8), total impact is about $89bn. CoreWeave, the Stargate site SPVs, Oracle's AI unit and the neoclouds run out of cash, concentrated between October–December 2026 and January–March 2027.
CoreWeave runs out of cash even without a shock: In Base its cash turns negative in April–June 2027, with a shortfall of about $7.5bn through July–September 2028. This matches the bond market, where its 2032 notes trade below 80.
OpenAI and Anthropic stay green in every scenario: July–September revenue beat expectations, and cash on hand plus Anthropic's IPO proceeds (provisional) are ample. But dependence on payments from these two is concentrated in the supply-side players above.
The combined scenario (S7) remains M5: Total impact is about $402bn, slightly better than about $419bn with the v1.0 default inputs. Oracle's equity raises and SoftBank's refinancing helped, while Oracle's higher capex made S4 worse.
The gap versus markets: bonds agree, stocks reflect something else: Model expected losses are below 1% of each company's market cap. NVIDIA's 11% fall from its peak (about $550bn) cannot be explained by losses in the circular structure; we read it as a reassessment of growth and valuation multiples.
The monthly checkpoint met the trigger for an ad hoc test: CP-2026-10 shows two reds and four yellows, and this run doubles as that ad hoc test.

Monthly checkpoint CP-2026-10

Two reds and four yellows. This meets the rule that one or more reds, or three or more yellows, trigger an ad hoc stress test. Demand-side indicators are green; the reds and yellows are concentrated in prices and credit.

#	Leading indicator	This month	Yellow / red threshold	Signal	Basis
1	Tokens processed, year on year	4.0×	1.7 / 1.5× or below	Green	Google's first-party API volume up 4× year on year; OpenRouter flat
2	Annual decline in realized unit price	50%	35 / 50% or more	Red	Anthropic's token prices down about 50% year on year
3	GPU rental rate, 3-month change	+11% (est.)	−20 / −35% or below	Green	H100 index at $2.88, up 44% since December 2025
4	AI-related credit spread change	About +350bp (est.)	150 / 300bp or more	Red	Oracle CDS at 151bp; CoreWeave 2032 notes yielding about 15%
5	AI lab valuation in latest round vs. previous	1.41×	1.1 / 1.0× or below	Yellow (manual)	Value is green, but overridden because OpenAI shelved its 2026 IPO ("delay to plan")
6	Hyperscaler capex guidance revisions	0%	−10 / −20% or below	Green	Microsoft, Alphabet, Amazon and Meta held or raised guidance
7	Share of cloud revenue from AI labs	45%	40 / 50% or more	Yellow	Estimated from lab dependence in Oracle's and CoreWeave's backlogs
8	DRAM contract prices, month on month	+4%	−10 / −20% or below	Green	Q4 contract price forecast of +10–15%, spread by month
9	AI stock basket, decline from peak	−20% (est.)	−20 / −30% or below	Yellow	SOX −13% in eight trading days, NVIDIA −11%, CoreWeave −24% in a week
10	Negative rating/lending events	1	1 / 2 or more	Yellow	Oracle issued a force majeure notice on its Jupiter site (New Mexico) on September 24

"Est." values are converted from public data. Indicators 4 and 9 will be replaced with measured values in the full version. Indicator 5 is a manual override.

Magnitude by scenario

Horizontal bar chart of total impact by scenario ($bn). S1 demand L2 81.0 (M3), S2 token price L2 76.6 (M3), S3 credit shutdown L2 188.3 (M4), S4 shorter GPU life L2 334.1 (M5), S5 OpenAI deferral L2 179.1 (M4), S6 guarantee calls L2 72.8 (M2), S8 current trend 89.0 (M3), S7 combined 401.9 (M5) — Total impact by scenario (ST-2026Q3-01, stress test model v1.0). Total impact T = additional funding need vs. Base + exposure-layer losses.

Demand (S1) and price (S2) shocks stay at M3 even at L2, because strong token volumes and capex guidance in July–September leave headroom on the demand side.

The heavy side is funding. If credit markets close for four quarters (S3-L2), four players (CoreWeave, the Stargate site SPVs, Oracle's AI unit and the neoclouds) run short of cash between October–December 2026 and January–March 2027. If the GPU refresh cycle shortens (S4-L2), extra capex adds SpaceXAI, turning five players red.

In no scenario do OpenAI or Anthropic newly turn red. The magnitude comes not from a lab failing but from cash shortfalls among those supplying compute to the labs, and from losses at their lenders and investors.

Risk points

Five of the nine funding-layer players show a signal. Each cell shows "shortfall ($bn) · first quarter cash turns negative." Yellow means cash remains but is below half of one quarter's outflow.

Player	Base	S8 current trend	S3 credit L2	S4 GPU refresh L2	S7 combined
CoreWeave	Red 7.5 · 2027Q2	Red 25.7 · 2026Q4	Red 36.4 · 2026Q4	Red 40.7 · 2026Q4	Red 50.2 · 2026Q4
Oracle (AI unit)	Yellow	Red 6.6 · 2027Q1	Red 27.5 · 2027Q1	Red 57.6 · 2027Q1	Red 95.3 · 2026Q4
Stargate site SPVs	Yellow	Red 5.4 · 2027Q1	Red 21.3 · 2027Q1	Red 30.5 · 2027Q1	Red 34.0 · 2026Q4
Neoclouds (aggregate)	Green	Red 3.2 · 2027Q1	Red 17.6 · 2027Q1	Red 20.2 · 2027Q2	Red 25.6 · 2027Q1
SpaceXAI	Green	Yellow	Green	Red 26.7 · 2028Q1	Red 8.5 · 2028Q3
OpenAI, Anthropic, SoftBank, Mistral	Green	Green	Green	Green	Green

Quarters are calendar quarters; 2026Q4 means October–December 2026.

CoreWeave is the first point of failure: Interest is about $0.9bn a quarter, and if new borrowing stops, cash runs out that quarter. Even in Base its funding assumes refinancing.
Oracle's AI unit has the largest shortfall: Capex is $28.5bn a quarter and needs both bonds and equity (ATM). In S7, where both stop, the shortfall is $95.3bn.
The Stargate site SPVs are exposed to timing gaps between site completion and OpenAI's payments: The force majeure notice at the Jupiter site widens that gap.
SpaceXAI is vulnerable to higher capex: It expects $35–40bn of capex in October–December, and a faster GPU refresh would exhaust its cash in 2028.

Gap versus stock and bond markets

Model expected loss is the probability-weighted average of losses across scenarios, with Base 55%, light (S1-L1) 25%, medium (S7) 15% and severe (S3-L3) 5%. Market caps are as of July 14, 2026 and will be updated in the full version.

Holder	Expected loss ($bn)	Market cap ($bn)	% of market cap	Market moves
Oracle	5.3	614.6	0.86%	CDS at 151bp, highest since 2009; 2035 bond spread 187bp
Microsoft	3.8	3,100	0.12%	Capex guidance held at about $175bn
Amazon	2.1	2,600	0.08%	Trailing 12-month free cash flow −$42.4bn
NVIDIA	1.7	5,000	0.03%	Down 11% from peak (about $550bn)
Alphabet	0.3	4,500	0.01%	Signaled a sharp increase in 2027 capex
SoftBank	2.2	n/a	—	Fell 11% on September 14 (OpenAI shelved its 2026 IPO)
CoreWeave	0.3 (holdings)	n/a	—	Down 24% in a week; 2032 notes below 80 (yield about 15%). Own shortfall shown above

Stocks are pricing in far larger declines than losses in the circular structure: NVIDIA's fall is about 300 times the model's expected loss. It is better read as a reassessment of future growth and valuation multiples, and of the capital constraints flagged by Broadcom CEO Hock Tan.
Bonds are watching the same places as the model: Bond yields and CDS are rising at CoreWeave and Oracle's AI unit, the players the model flags red. On where company-level risk sits, the model and the market agree.
The gap does not mean the market is wrong: The model measures only losses arising from equity, purchase commitments, loans and guarantees, not lower future profits from lost orders. For the full version we will consider adding a column that converts backlog (RPO) declines into profit.

In S7, the holders with the largest losses are Oracle ($33.1bn), Microsoft ($25.0bn), private credit (Blackstone, Blue Owl and others, $18.2bn), bond investors ($15.1bn) and Apollo and others ($14.0bn).

Items to watch

When	Event	What to look for	Related shocks
Week of Nov 9	Anthropic IPO pricing	Amount raised (model assumes $30bn) and valuation; a test of labs' access to capital markets	S3, S5
Nov 10	CoreWeave Jul–Sep results	Refinancing and new borrowing, interest vs. EBITDA, 2027 funding plan	Red in Base, S3
Mid-November	SoftBank Jul–Sep results	Remaining bridge loan, margin loans, collateral value of OpenAI stake	S3
Nov 19	NVIDIA Aug–Oct results	Data center revenue, customer concentration, disclosure of lease guarantees and investments	S1, S6
Late November	Full version ST-2026Q3-02	Recalculation reflecting the four items above	All
Early December	Monthly checkpoint CP-2026-11	Whether the reds in prices and credit persist for a second month	S2, S3
Mid-December (expected)	Oracle FY27 Q2 results	Capex and free cash flow, continued ATM equity issuance, recovery at the Jupiter site	S3, S4
Ongoing	OpenAI's next round and IPO timing	Whether valuation holds; payment schedules to Oracle, Azure and AWS	S5
Ongoing	Oracle CDS, CoreWeave 2032 notes	Whether Oracle CDS exceeds 200bp or CoreWeave notes fall below 70	S3

Oracle CDS at 200bp and CoreWeave notes at 70 are markers set for this run. If either is crossed, we will run an ad hoc test with S3 raised to L2.

Input updates and provisional items

The model structure is unchanged from v1.0; only inputs were replaced with July–September 2026 disclosures. Amounts are annual.

Player	Main updates	Basis
OpenAI	Revenue raised to $44bn (2026), $85bn (2027) and $125bn (2028)	Revenue run rate about $70bn (September); April–June revenue $6.7bn
Anthropic	Revenue $53bn, $100bn and $140bn; IPO proceeds of $30bn assumed provisionally	July–September revenue $16.8bn, adjusted operating margin 11%
SpaceXAI	Revenue $25bn, capex $60bn; Apollo-led $40bn GPU financing booked as leases	AI unit July–September revenue $7.6bn; capex about $20bn a quarter
Oracle (AI unit)	Capex $95bn, higher interest; $5bn a quarter of ATM equity added	FY27 Q1: capex $28.5bn, free cash flow −$5.4bn, backlog $664bn
CoreWeave	Debt of $51.6bn; interest raised to $3.6–5.0bn a year	Debt at end of June; July–September interest guidance $0.86–0.94bn
SoftBank	Partial refinancing of the bridge loan and new facility reflected	$25.9bn of the $40bn repaid; new $11.87bn facility

Oracle's AI unit figures are estimates that allocate the AI share of company-wide disclosures. The aggregate neoclouds and Stargate site SPVs remain low-precision estimates based on press reports. The full version (ST-2026Q3-02) will replace the following provisional items.

Results not yet reported: July–September figures for CoreWeave (Nov 10), SoftBank (mid-November) and NVIDIA (Nov 19).
Provisional funding: Anthropic's IPO proceeds, and OpenAI's cash on hand and next round.
Estimated monthly indicators: Credit spreads and the AI stock basket, to be replaced with measured values.
Market caps: July 14 values updated to the latest, with SoftBank and CoreWeave added.

Sources

Figures are from press reports and analysis articles retrieved on November 2, 2026. Cross-checking against primary filings (earnings releases, SEC filings) will be done in the full version. For model definitions, see the stress test model v1.0.

Back to reports ← Stress Test Reports and Monthly Checkpoints