AI Value Chain Risk: What Actually Breaks, and Where the Data Already Is

As of 2026-09-06. Educational and analytical only. Not investment advice. Numbers below come from an internal research memo built on a reproducible 28-name fact pack (yfinance, 2026-09-06) plus sourced reporting and SEC filings; the judgment layered on top is point-in-time. Nothing here is a buy or sell instruction. All volatility figures are realised, not implied — the pack ran on a Sunday, when option quotes are dead and any implied-vol reading is solver junk. The implied-versus-realised comparison is the entire edge in premium selling, so nothing here is actionable without a market-hours run.

This article was generated by AI — levelbox's deep-dive tooling — and reviewed before publishing. The data is point-in-time and can be wrong; if you spot a mistake, please tell us.

The question we keep getting asked is what breaks upstream if OpenAI or Anthropic fails. On the evidence, neither is close to failing, and ten minutes with their balance sheets says so. The fragility is real, but it sits one layer out, in the financing structure built on top of the labs, and unlike the labs themselves that structure reports daily.

This is the third piece in an arc. The AI Buildout, Three Ways framed it as a capex story before a demand story. From Capex to ROI Capture covered the market beginning to answer that question layer by layer. This one asks who actually absorbs the damage if the answer comes back badly.

The labs are not the weak link

Anthropic was running at a $47B revenue run rate as of May 2026 and filed confidentially for an IPO on 1 June. OpenAI raised $110B in February 2026, closed a record $122B round on 31 March, and filed on 8 June. Both are loss-making at scale. Neither is short of capital, and the "OpenAI faces bankruptcy" genre is generally a loss figure quoted without the balance sheet beside it.

Every realistic distress path here runs through a down round, a renegotiation, or absorption by a strategic buyer. Liquidation is not on the table for either company, which means the interesting question is what happens to everyone who wrote contracts against them.

The number that moved by more than half

OpenAI cut its own infrastructure commitment from about $1.4T to about $600B in February 2026, then raised it back to about $750B by July. Five months, two revisions, more than half the number, while the company was raising record sums and preparing to list.

Nothing was wrong at the time, which is the uncomfortable part. Anyone who financed physical capacity against that commitment holds a contract the other side revises at will, and the revision mechanism has now fired twice at full strength. A mechanism that fires at full health is a demonstrated behaviour of the system rather than a tail event. What makes it dangerous is not its probability but who absorbs it — the lenders, lessors and levered operators one layer down, whose obligations do not get revised in sympathy.

Eisman is nervous, still long, and not shorting

Steve Eisman has made the closest public version of the counterparty thesis, and it deserves an accurate hearing.

Credit where it is due: Eisman put this dependency on the map. While most commentary was still arguing about whether AI capital spending was too large in aggregate, he named the specific thing that carries the risk — two loss-making private companies sitting underneath everyone else's revenue — and he put numbers on it. He laid it out in Is OpenAI the Achilles' Heel of the US Economy? on his Weekly Wrap. The framing of this piece follows his. Where we differ is narrow, and it is about where to look for confirmation, not about whether the risk is real.

His claims, on CNBC's Fast Money on 13 August: the Achilles heel of AI is dependence on OpenAI and Anthropic, which he puts at roughly 70% of AI revenue across Microsoft, Amazon, Google and Oracle, and 25-35% of those companies' cloud revenue. If either lab stumbles, cheaper Chinese open-weight models could start a price war across the industry. About half of Oracle's roughly $600B backlog is OpenAI. It is one trade, so a capex cut anywhere takes the market down everywhere. His own position is worth stating precisely, because it is easy to caricature in either direction. He is still long Nvidia and other tech, in his words "less long 'cause I've gotten nervous, but I am not short." He is not calling a top, and he is not recommending anyone buy chipmakers either. Back in late June he did argue for owning suppliers over hyperscalers; by late July he had sold his Alphabet position and moved to cash, so treat that earlier view as dated rather than current. What he says he is waiting for is financial data once the labs go public, because before subprime he had monthly delinquency figures deteriorating in front of him, and "there is no such data set with respect to AI." The labs are private. Economics have to be inferred from funding rounds and disclosed commitments.

His 70% and 25-35% figures have not been confirmed against a primary disclosure by us, so treat them as an experienced practitioner's estimate rather than a filed number. And his caution has been right so far, which matters more than whether his stated reason for it holds up.

The warning signs show up in bonds before stocks

His point is about borrower-level data, and on that he is right. But the subprime signal was never only the borrower's accounts. It was also the credit performance of the paper funding the borrower, and that series does exist for AI, is published daily, and has started to move.

Indicator	Reading
Incremental debt as a share of hyperscaler capex	9% (FY24) → 32% (LTM to Jun 2026)
Oracle credit rating	S&P cut to BBB- (July 2026) on surging capex and negative cash generation
Oracle committed leases	~3× its FY27 capex guidance
Hyperscaler IG spreads, 2-4yr	30bp → 40bp over Treasuries
Hyperscaler IG spreads, 5-7yr	50bp → 60bp
30-year Treasury	5.323% on 18 August, a 19-year high

In a capital spending bust, credit reprices before equity. The equity market argues about whether the revenue arrives; the credit market prices whether the interest gets paid, continuously, and it does not need a narrative to do it. Watching a semiconductor multiple for evidence of an AI unwind is watching the lagging indicator.

Spreads at 40bp and 60bp are historically tight in absolute terms. Investment grade widens to 150-200bp in genuine stress. So the dashboard has moved roughly a third wider off a very tight base, with one BBB- downgrade attached. Saying credit is already breaking would overstate it badly. The correct claim is narrower and still worth having: the instrument exists, it is observable daily, it has started to move, and it will move before equity does.

Every long-dated bond funding a GPU cluster competes for the same pension, insurance and sovereign money that buys 30-year Treasuries. The buildout is issuing into the long end, which raises the discount rate on the long-duration assets it is priced on and the cost of its own next tranche. Self-limiting, and it does not need a recession to bind.

Two deals that work exactly like 2007 did

CoreWeave's DDTL 4.0, closed 30 March 2026, is the first investment-grade-rated GPU-backed financing. Senior tranche A3, junior Ba2. The collateral is rapidly depreciating single-purpose hardware whose resale market is the same industry that would be in distress in any scenario where the loan is actually tested. Tranching a correlated, depreciating asset pool to manufacture a senior investment-grade rating is the structured-credit machinery of 2005-2007 applied to a shorter-lived asset. It has not featured much in the public discussion of AI credit risk.

Nvidia's 8-K of 17 August 2026 discloses a $105B residual-value guaranty backstopping the lessor in the Ohio arrangement if OpenAI defaults on the lease. OpenAI has contractually agreed to indemnify Nvidia for anything it pays out. That indemnity is worthless in exactly the state of the world that triggers it: if OpenAI cannot pay the lease, it cannot honour the indemnity either. The protection is perfectly correlated with the event it protects against. Textbook wrong-way risk, and the same structural error as buying default protection from a counterparty who fails when the reference entity does.

Solvency and drawdown are different axes

Collapsing them is the central error for anyone selling puts on this theme.

This is the value trap in a different costume: a low multiple on peak-cycle earnings is a warning rather than a cushion. Solvency asks whether the enterprise survives, and is set by capital structure and whether revenue is contracted or discretionary. Drawdown asks how far the equity falls, and is set by how much of the current multiple is priced off AI-contingent growth. A solvent company can fall 70%. An insolvent one can look cheap the whole way to zero.

That gap is why we keep arguing that your stock positions are tail risk too. Underneath both sits asset specificity — how re-deployable the asset is if AI demand halves. A gas turbine serves any load. A transmission interconnect serves whoever shows up. A GPU serves AI or very little else. A purpose-built 200MW shell in a remote county serves AI or becomes a stranded asset.

Our own screener has a gap here. CoreWeave returns eligible on this repo's gate with debt-to-equity of 10.3 and a free-cash margin of −1.20. The eligibility test checks revenue scale and quality; it does not test balance-sheet fragility. That is now a backlog item.

What a capex bust does to solvent suppliers

Cisco fell from $80.06 on 27 March 2000 to $8.60 on 8 October 2002, about −89%. It did not close at a new record until roughly 10 December 2025. Twenty-five years and eight months, in a company that was never insolvent, never impaired, and grew throughout. For an allocator that is arguably worse than a bust, because a bust gets recognised and recycled while this looks survivable the entire way down.

JDSU is the cleaner cautionary case. Its successors, Viavi and Lumentum, sit together about 82% below the 2000 peak twenty-six years on, and Lumentum is riding today's AI-optics boom while that remains true. The technology winning eventually does not mean the shareholder does. Nortel simply went to zero.

Memory is where this cycle is loudest right now, and we have looked at what that fat premium is actually pricing and at what it means when implied vol sits below realised in those same names. The analogy gets pushed further than it carries. For a name like Nvidia the central estimate is its own history: −57% across 2018-19 and −66% across 2021-22, both times as a solvent, unimpaired, high-quality company, recovering within roughly sixteen to eighteen months each time. It currently trades on a 14.9x forward multiple, which is a very different starting point from Cisco at well over 100x in 2000. Call it −60% and back inside eighteen months for the base case, −89% and a generation for the tail. The distribution contains both, and nothing in the fundamentals separated Cisco from Nvidia at their respective peaks. Both looked like quality survivors, and both were.

A drawdown of that size with an eighteen-month recovery is survivable if you sized for it and genuinely wanted the shares. The tail version is not survivable at any size.

Power: the forecast is already being cut

The popular repositioning is to pivot into power. On this snapshot that trade is late rather than early.

On 10 December 2025 the EIA cut its 2026 US generation growth forecast from 3.0% to 1.7%, and ERCOT's 2026 demand growth from 15.7% to 9.6%. The stated reason was how little large load had actually come online. Separately, only about 50-60% of datacenter capacity scheduled for the next one to two years is expected to arrive on time. So the power complex has already started de-rating, on a load forecast being marked down in real time, well before any AI unwind arrives to do it.

The physical demand story underneath is genuinely resilient, and that is a different claim from the equity multiple built on it. Roughly 60-75% of forecast load growth survives an AI halving, because electrification, reshored manufacturing and baseline growth do not care whether a lab raises another round.

State-level protection is also better than assumed and unevenly applied. Ohio and Virginia have both adopted an 85% minimum-take structure for large-load datacenter customers, so the customer pays for 85% of contracted capacity whether or not it is used. Illinois has deposit rules only, with no minimum take.

Where this leaves a premium seller

The hypothesis this work started with was to find names with contracted revenue wearing implied volatility inflated by AI association. Safe underlying, rich premium, both halves of a wheel thesis holding at once.

It did not survive contact with the data.

Regulated utilities in the pack carry 30-day realised volatility of 13-19%. They are genuine defensives that pay an options seller close to nothing. Everything across the AI layer carrying 45%+ is carrying it for a reason the fact pack can see: debt-to-equity of 3.9 at Oracle, 3.7 to 5.8 across the merchant power producers, 10.3 at CoreWeave. In this chain the safe names do not pay, and the paying names are not safe.

One candidate came closest, and it is offered as an illustration of how the lens scores rather than as a pick. Constellation reads at debt-to-equity of 0.8 against merchant peers at 3.7 to 5.8, with 0.35 realised volatility and nuclear contracts written with hyperscaler counterparties — a theme we walked through separately in getting paid to wait on nuclear. Its liquidity sub-score is 0.25, among the worst in the group, so the premium may not be harvestable at size. That needs a market-hours run to confirm and might not clear it.

The thing being looked for is low leverage, non-AI revenue, pricing power, and enough volatility to pay for the obligation taken on. Nothing in this value chain has all four. The useful answer is probably to stop looking inside the chain rather than to settle for the least-bad name in it.

On the names above

Every ticker above is a worked example of how a lens reads a name in one point-in-time fact pack. None of it is a shortlist, a recommendation, or a comment on the quality of any business. A different day, a different delta or a different account size surfaces different names entirely, and a name that reads marginal here is outside the parameters of one screen rather than a bad company. For the current, full list scored on live numbers at your own delta, run the levelbox.ai screener.

None of this changes in isolation from the weather. Market regime sets what you are selling into, toppy and bottomy markets change what a sensible strike looks like, and how much delta to sell is the lever that connects the two. Worth remembering too that delta is not your chance of assignment, which matters more than usual when the base rates above are this wide.

Where a cash-secured put is involved, the obligation is to buy the stock at the strike, so the maximum loss is the strike less the credit received, per contract, and it is realised in exactly the scenario where the name falls hardest. The historical base rates above exist to put a number on how hard that can be, in companies that were never in any trouble at all.

Verify every figure against primary filings before you act on anything. And remember the caveat carried throughout: these are realised volatility readings from a Sunday pack, and the implied-versus-realised comparison that would make any of it tradeable has not been run.

---

Analytical and educational, not investment advice. Figures were current as of 6 September 2026 and move daily. Sourcing runs from SEC filings (Nvidia's 17 August 8-K, CoreWeave's 10-K and 10-Q), official statistics (EIA, December 2025), rating agency actions (S&P on Oracle, Moody's on DDTL 4.0), state utility dockets, and reputable reporting for the funding-round and spread figures. Eisman's revenue-share estimates are his own and have not been confirmed against a primary disclosure.