AI bubble — exposure ranking
If the AI capex cycle unwinds the way 1999 telecom did, who has the most to lose? Ranked by structural exposure across five dimensions: AI-revenue concentration, capex commitment overhang, vendor financing exposure, equity-stake impairment risk, and customer concentration. This is positioning analysis, not a prediction.
Tier 1 — Epicenter
EXTREME EXPOSURE
100% AI revenue, debt-financed buildouts, customer concentration in the very entities they're funded by.
CoreWeave (CRWV)
Exposure
- $20B+ Nvidia GPU commitments on the balance sheet
- $15B+ outstanding debt against assets that depreciate 30%/year
- Two-customer concentration: MSFT + OpenAI ≈ 77% of revenue
- If OpenAI's commitments to CRWV are renegotiated, no plausible substitute customer at scale
Failure mode
- Customer demand softens → utilization drops → debt covenants trip
- GPU residual values collapse (Blackwell/Rubin obsolescence)
- Goodwill impairment + revolver default within 6 quarters of contraction
Bear-case outcome: Equity zero or near-zero, debt restructuring, GPU fleet sold at discount to hyperscalers.
Other GPU clouds (Lambda, Crusoe, Applied Digital, Iris Energy)
Exposure
- Same Nvidia-funded buildout pattern
- Less customer diversity than CoreWeave
- Crypto-mining pivots to AI add execution risk
Failure mode
- Cannot compete with hyperscaler-owned GPU capacity post-bubble
- Consolidation or wind-down
Tier 2 — Chip-vendor reflexivity
VERY HIGH EXPOSURE
Revenue concentrated in customers they themselves funded; equity stakes carried at unrealized gains.
Nvidia (NVDA)
Exposure
- Top 4 customers ≈ 40% of revenue (MSFT, META, GOOGL, AMZN)
- ~10–15% of incremental data-center revenue traceable to Nvidia-funded counterparties
- Equity portfolio (CoreWeave + others) marked to fair value — would require significant write-downs in unwind
- $50B+ annual buyback program continues — reflexive support for own multiple
Failure mode
- Hyperscaler capex committee cuts (already started at Meta in Q1 cycles historically)
- In-house silicon (Google TPU, MS Maia, MS Cobalt, Amazon Trainium, Meta MTIA) takes share at the margin
- Inventory write-downs from forward purchase commitments
- Revenue could halve in 4 quarters as it did in crypto-bust 2022
Bear-case multiple: 8-10× P/E (vs current ~30×) on normalized data-center revenue. 60–70% drawdown plausible without bankruptcy risk.
AMD (AMD)
Exposure
- OpenAI warrant: 10% of company contingent on AMD stock price hitting milestones — circular reflexivity
- $10B+ MI300 buy commitment from OpenAI is the single largest AI-revenue lever
- Smaller customer base than Nvidia → higher concentration risk
Failure mode
- If OpenAI defers MI400, AMD AI revenue projection halves
- Warrants out of the money triggers re-rating
Broadcom (AVGO)
Exposure
- ~30% of revenue now AI-related
- Concentrated in 2-3 hyperscaler ASIC programs
- VMware acquisition added $130B goodwill — large impairment surface
Failure mode
- Hyperscaler ASIC delays cascade into multi-quarter revenue gaps
- Goodwill / TA already > 0.45 — flagged by this screener
Tier 3 — Stretched hyperscalers and AI-anchored balance sheets
HIGH EXPOSUREOracle (ORCL)
Exposure
- $300B counterparty risk on a single private customer
- Has to issue ~$50B+ debt to build the data centers
- Capex going from $7B/yr to $25B+/yr in two years
- OCI (Oracle Cloud) revenue projection requires the deal to perform
Failure mode
- OpenAI renegotiates → stranded data-center capacity
- Debt covenants on a $50B+ raise become problematic if EBITDA softens
- Goodwill from Cerner ($28B) impairment vulnerability simultaneously
Microsoft (MSFT)
Exposure
- OpenAI stake carried at fair value with substantial unrealized gain — would write down meaningfully in a bear scenario
- Azure AI revenue concentrated in OpenAI inference (Copilot, ChatGPT API resold via Azure OpenAI Service)
- $10B+ CoreWeave contract is at-risk inventory if customers slow
Failure mode
- Survives — diversified product, $250B+ free cash flow base
- But growth re-rate (from ~25× to ~18×) is a 30% drawdown
Meta (META)
Exposure
- Largest AI capex of any company without monetizing it directly
- Llama is open-source — defensive moat, not a revenue stream
- Reality Labs & AI together = $50B+ annual cash burn vs. core ad business
Failure mode
- Capital allocation pressure from shareholders à la 2022
- Ad business is the cash machine — survives but multiple compresses
Tesla (TSLA)
Exposure
- ~50–70% of equity value attributed to FSD/Robotaxi/Optimus narrative — none materially revenue-generating today
- Auto business margins compressed; AI optionality the bull case
- xAI funding rounds rely on overlapping investor base
Failure mode
- Multiple compresses to auto OEM levels (~10× P/E) → 70%+ drawdown
Tier 4 — Power, real estate, and infrastructure piggybacks
MODERATE EXPOSURE
These trades up on AI demand projections; they're cushioned by underlying utilities/REIT economics but are still re-rated above fundamentals.
Vistra (VST), Constellation Energy (CEG), Talen Energy (TLN)
- Re-rated 4–10× over 18 months on hyperscaler PPA announcements
- If AI capex disappoints, power offtake projections reset; equity gives back the multiple expansion (50–70% drawdown plausible) though the underlying utility business survives
Digital Realty (DLR), Equinix (EQIX), Iron Mountain (IRM)
- Pricing power from hyperscaler land-grab phase — that's the temporary part
- REIT cash flows survive; multiple compresses
Nuclear fuel: Cameco (CCJ), Centrus (LEU), NuScale (SMR), Oklo (OKLO)
- Pre-revenue (Oklo, NuScale) names down 80%+ in a base-rate AI slowdown
- Cameco / Centrus survive on uranium cycle; multiples normalize
Tier 5 — Public-market "AI wrapper" small/mid caps
EXTREME REVALUATION RISK BUT SMALL ABSOLUTE
These names re-rated on AI association without proportionate revenue. They survive any bubble unwind only as zombie equities.
Palantir (PLTR)
- Trading at ~70× sales — historical software outliers (Snowflake, Datadog at peak) topped at ~50×
- AIP product is the multiple driver; if commercial revenue growth decelerates below 30% it gets re-rated to ~15× sales (60–70% drawdown)
- Beneish/Piotroski on the screener look fine — this is a multiple risk, not an earnings-quality risk
C3.ai (AI), BigBear.ai (BBAI), SoundHound (SOUN), Recursion (RXRX)
- Persistent operating losses, ticker-driven retail bid
- 10× equity moves on every AI hype cycle, give them back on every reset
- Survival depends on continuing equity issuance — at-risk in a multi-quarter bear
SaaS rebrands (Salesforce, ServiceNow, Adobe, Workday, Snowflake)
- Each priced for AI-revenue expansion that hasn't materialized at promised pace
- Real businesses with moat — survive comfortably; valuation premium evaporates
If you compress this into one risk hierarchy
| Rank | Company | Worst-case drawdown | Survival risk? |
|---|---|---|---|
| 1 | CoreWeave + GPU cloud peers | −85% to −100% | Yes |
| 2 | Pre-revenue SMR / nuclear narratives | −80% to −95% | Yes |
| 3 | AMD | −55% | Stress |
| 4 | Oracle | −50% | Stress |
| 5 | Nvidia | −60% to −70% | No (size) |
| 6 | Broadcom | −45% | No |
| 7 | Vistra / CEG / Talen | −50% to −70% | No |
| 8 | Palantir | −60% | No |
| 9 | Tesla | −60% to −70% | No |
| 10 | Meta / Microsoft / Google / Amazon | −25% to −35% | No |
These ranges assume an AI-capex-led recession comparable to 2001's telecom unwind, not a 2008-style systemic event. Numbers reflect order-of-magnitude positioning analysis from publicly disclosed exposures, not a forecast. Actual outcomes will depend on whether AI revenue catches up to the capex curve and on hyperscaler discipline.