Who gets hurt
Bar = share of cushion used; tick = failure line. Click a row for the why.
Where the losses land
Lenders, bondholders and funds the Bloomberg chart leaves out. Modeled as absorbing losses, not failing.
This tool is for informational purposes only. See the disclaimer at the bottom of the page.
AI research tool
AI Network Stress Test
Pick a company in the AI deal network and assume it fails. The model traces every money tie we could source (stakes, contracts, loans, leases, guarantees) to show who gets hurt, how badly, and whether the damage stops there.
- Data
- Published by
- SEM Wealth Management
TradersBlog.SEMWealth.com
Demand for AI capacity
Assumptions (change any of them)
knock-on failures
real losses to others
money that won't come back
write-downs
stakes marked to zero
used-GPU price drop
The network
Click any company to fail it. Lines show where the damage travels; thicker means more money.
Annuity issuers
Athene's cushion
Three of the largest fixed and indexed annuity sellers, from their own statutory filings. Holdings in companies on the map are modeled as ties; the rest (data center securitizations and ordinary big tech bonds) are listed but not tied to a company the model can fail. An insurer's "cushion" here is its capital and surplus: the money it holds beyond what it has set aside for policyholders. A loss that uses part of the cushion is a hit to the company, not to policyholders.
Which failure hurts the network most?
Each company failed on its own, with the assumptions above. Solid bar = real losses to everyone else; pale extension = write-downs. Color = verdict. The rating column is the market's read on how likely it is. Click a row to open that scenario.
SystemicSpreads, then stopsContained, big company woundedCompany problem
Most exposed
How often each company is wounded or knocked over when someone else fails.
| Company | Wounded | Fails | Worst case |
|---|
Which assumption matters most?
| Assumption | Now | Nudge toward worse | Across its full range | Tipping point toward worse | Eases at |
|---|
How to read this: each assumption is moved on its own across its full range while the others stay where you set them. "Nudge toward worse" moves it one tenth of its range in the bad direction (for example, re-sale from 40% to 34%, or contract years from 5 to 5.9) and shows the extra real losses. A tipping point is the setting at which another company falls; "eases at" is where one stops falling. The demand switch is treated here as a continuous re-sale rate. The line in the middle column shows real losses from the low end of the range (left) to the high end (right); the dot is your current setting.
Every tie in the model
| Money flows from → to | Tie | Amount | Per year | Status | Source |
|---|
How the model works
The model fails one company (or two), then walks through every tie that company has. Each tie turns into a dollar hit for the company on the other end. Anyone whose hits use up their cushion fails too, and the model walks their ties in the next round. It repeats until nobody else falls.
Two kinds of hit
- Real losses are money that won't come back: unpaid loans, guarantees that get called, promised investments that never arrive, lost contract revenue, and the cost of replacing a lost supplier. These are what knock companies over.
- Write-downs are stakes marked to zero. They hit reported earnings and the stock price, but not the cash a company needs to pay its bills. They only count toward failure for SoftBank, which is a holding company that borrows against its stakes.
The cushion
Cash and short-term investments plus one year of operating cash flow. For private companies that don't publish this (OpenAI, Anthropic, the private neoclouds) it is an SEM estimate and flagged as one. Project companies built for a single tenant (Oracle's Stargate sites, Meta's Hyperion) have no cushion: if the tenant fails, they fail.
How each tie becomes a hit
- A customer fails: the supplier loses the contract. We count up to the years set in Assumptions, then take off what gets re-sold to someone else (70% when capacity is sold out, 40% normal, 10% in a glut). Only part of lost revenue is a real loss: 70% for borrow-to-build companies (neoclouds, landlords, Oracle), 50% for others. Purchases with no contract behind them count one year.
- A supplier fails: the customer pays extra to replace it (25% of a year's spend when sold out, 10% normal, nothing in a glut) and loses any prepayment.
- A borrower fails: lenders lose the loan minus recovery. Defaults: 25 cents for unsecured bonds, 20 for converts, 50 for secured loans, 60 for loans secured by a project. Loans backed by GPUs recover less when used-GPU prices fall. Loans backed by an investment-grade customer contract recover at least 85 cents.
- A guarantee is triggered: the guarantor pays a share of the cap (25% by default). Nvidia's $105B Ohio guarantee is triggered by an OpenAI default, not SB Energy's.
- The used-GPU fire sale: a failed company's GPUs get sold cheap. The price drop is the fleet dumped as a share of roughly $700B of GPUs installed since 2023, times two (four when severe), capped at 60%. Companies that borrowed against their GPUs take the markdown as a real loss; the rest take it as a write-down. Half the dumped chips replace a new purchase, which costs Nvidia and AMD sales.
The assumptions, one by one
Every default below is a starting point, not a measurement. Where we could anchor a number to history or to today's market, we say so and link the source. The "What matters most" tab shows which of these moves the results most.
Annuity issuers
Athene (owned by Apollo), Corebridge's American General Life and American Equity (owned by Brookfield) are the three largest indexed annuity sellers that publish statutory filings detailed enough to list every bond they own. SEM read each company's full year-end 2025 bond list (8,648, 10,763 and 5,723 bonds, each matched to the filing's total to the dollar) and screened it for AI-linked names, then added each company's 2026 quarterly purchases where they are published (Athene and American Equity; Corebridge posts only annual statements).
- What is modeled: holdings in companies on the map (xAI/SpaceX, SoftBank, Oracle, CoreWeave, Meta's Hyperion project). These were moved out of the "Private credit" and "Bondholders" totals so nothing is counted twice.
- What is listed but not modeled: data center securitizations (pools of leases to many tenants) and ordinary bonds of big tech and chip companies the model does not track as borrowers.
- The cushion is capital and surplus. Athene's U.S. company holds only $4.6B against $373B of assets because much of its business is reinsured to Bermuda affiliates while the assets stay in Iowa. Athene's rating-agency exhibit allocates $22.4B of Bermuda capital to that business; the switch shows both views.
- What the filings do not show: mortgage loans are listed by city and state only (Athene says about $1.4B of its loans are on data centers), and billions of privately rated, code-named deals (for example Apollo's "AP" deals) do not say what is inside them.
What the verdicts mean
- A company problem: nobody else fails and no major company is badly hurt.
- Contained, but a major company is wounded: nobody else fails, but a load-bearing company loses more than a third of its cushion.
- Spreads to landlords and lenders, then stops: smaller companies fail, but none of the load-bearing ones.
- Systemic: it knocks over a load-bearing company (Oracle, Nvidia, the big four, Broadcom, AMD, OpenAI, Anthropic, SoftBank or SpaceX).
What this does not capture
- Stock-market contagion. Look-alike stocks will sell off in sympathy (Nebius, IREN and Applied Digital fell 3-5% on Oct 15). The model tracks money, not mood.
- Links with no public dollar figure. Google's TPU purchases from Broadcom, the OpenAI and Meta chip deals with AMD, OpenAI's 10 GW with Broadcom, and most Nvidia sales by customer are real but unsized. Each scenario counts how many it skipped. Results understate the damage to Broadcom and AMD in particular.
- Timing. Everything is treated as if it happens at once. Many leases and guarantees (Oracle's $288B of unstarted leases, Nvidia's Ohio guarantee) only become live as buildings finish in 2027-2029.
- Bankruptcy law. Landlords' lease-rejection claims are capped in bankruptcy, and a healthy buyer often takes over the data centers (Exodus 2001). That would soften the hit to customers and landlords.
- Who is likely to fail. The model answers "how bad," not "how likely." Use the rating and the bond-market signals for that.
Sources and caveats
Ties come from 10-Qs, 10-Ks, 8-Ks and S-1s first, then Reuters, WSJ, Bloomberg, FT and The Information. Each tie carries a status: funded or signed (in a filing or company release), contingent (an "up to" ceiling that depends on events), reported (press, not confirmed), or estimate (SEM's own arithmetic, explained in the note). In rev2 a separate checker re-sourced the 111 ties that drive the results (every tie of $5B or more, plus every soft one): 66 confirmed, 32 corrected, 13 could not be verified. The ties tab shows which is which. To avoid double counting, three ties stay in the list but out of the math: Broadcom's $42B standby lending to Anthropic (it protects the same lease payments as its $29B backstop), the $21B of Broadcom rack orders (lender-funded vehicles buy those racks), and the "$50B" Anthropic/Fluidstack program (an announced plan, not a contract). Market values are Oct 14-15, 2026 closes; balance sheets are as of June 30, 2026 unless noted. Ratings marked * were not re-checked. On Bloomberg's chart but with no sized money tie: Harvey and Ambience Healthcare (OpenAI Startup Fund, which uses outside investors' money), Cursor (now owned by SpaceX), and the Intel to Mistral line (no deal found).