it’s boring, but it’s what bursts the bubble

Finance-Focused Summary (Markets / Investing / Macro / Portfolio)

Core thesis

• The video argues the so-called “AI bubble” is less likely to pop because AI fails, and more likely to pop because AI delivery is late—meaning the payoff arrives after investors/financiers stop waiting.
• The key bottleneck is energy delivery and infrastructure, not just AI demand or AI adoption.
• It frames the risk as two competing clocks:

1. Clock 1 (Power clock): How quickly the U.S. can supply enough electricity and grid capacity for AI data centers.
2. Clock 2 (Chip clock): How long AI hardware (chips/servers) remains economically cutting-edge before it becomes obsolete.

---

Scale of the wager (AI capex concentration)

• Big Tech AI infrastructure spend:
  • 2024: a little over $200B
  • 2025: almost $400B
  • “This year” (unnamed): estimates > $700B
• Claim: AI-related spending is concentrated heavily in large tech firms and is carrying parts of the broader economy.
  • AI accounted for “more than half” of U.S. real GDP growth last quarter (attributed to Deutsche Bank messaging).

---

Macro / investing implication

• Investors are currently funding losses and buildouts because they expect future returns.
• The bubble “pops” if capital rotates away due to delayed delivery—even if AI ultimately works.

---

Clock 1 Details: Electricity Supply Chain Constraints

Interconnection queue constraints

• PJM Interconnection backlog (wait times):
  • Grew from < 2 years (2008) to > 8 years today

Electricity consumption

• Data centers already consume > 4% of U.S. electricity
• Expected to reach ~20% by 2035 (about 1/5)

Power generation equipment constraints

• Gas turbines: “roughly 40% of U.S. electricity”
• Only a few manufacturers; GE Vernova is named:
  • Order backlog: ~80 GW → 116 GW
  • CEO claim: “largely sold out through the end of 2028”
• The claim (as stated): developers ordered nearly 3x more turbine capacity than global manufacturing can produce in a year.

Transformers and transmission lead times

• Transformers: > 2.5 years lead time, with an estimated ~30% supply shortage
• New transmission lines: ~6.5 years average completion time (permits/backlogs delays cited)

---

PJM Capacity Auctions (Mechanism Linking Costs to Consumers)

What PJM does

• PJM Interconnection serves 67 million people across 13 states (about 1 in 5 Americans).
• PJM runs annual capacity auctions to secure “resource adequacy,” an insurance-like mechanism so the grid can handle extreme peak days.

Auction price escalation and impact

• 2024 clearing price: about $29 per megawatt-day
• Next auction: just under $270 (>830% spike)
• Regional bill change: $2.2B → $14.7B
• The auction then faces a federal price ceiling (a legal maximum):
  • The auction hits the ceiling repeatedly
  • In Dec 2025, it clears short (not enough supply even after paying the max)
• Household cost example:
  • Maryland average bill up ~$21/month (nearly ~$250/year)

Data-center cost attribution (Monitoring Analytics)

• Of $16.4B auction cost:
  • ~$6.5B traced directly to data centers
• “40 cents of every capacity dollar” traced to data centers across the last three auctions
• Cumulative: > $23B (as stated)

Investment implication embedded in the argument: Even if data center demand is “funded,” the grid cannot scale instantly, so costs can flow to consumers—creating political/regulatory pressure and margin risk that could worsen capital market sentiment.

---

Clock 2 Details: Chip/Server Obsolescence + Earnings Accounting Risk

Economic obsolescence timing mismatch

• Chips don’t “age well” economically:
  • Cutting-edge AI chips stay cutting-edge about 2 to 6 years (range cited)
  • Grid equipment lasts decades
• Risk: stranded/allocated costs shift to customers if data centers scale back or chips become obsolete.

Depreciation / earnings mechanics (Michael Burry cited)

• Burry’s claim: Big Tech’s record earnings may be partly driven by aggressive depreciation assumptions (i.e., extending the “useful life” of servers).
• Mechanism described:
  • Longer useful life → lower annual depreciation → higher current profits
  • Shorter useful life → higher depreciation → lower profits
• Burry estimate:
  • Understated depreciation by ~$176B between 2026 and 2028
  • Implies:
    • Oracle earnings overstated by ~27% by 2028
    • Meta earnings overstated by ~21% by 2028

Example revisions/assumption changes

• Amazon:
  • Jan 2024: servers from 5 → 6 years
    • Reduced depreciation by $3.2B
    • Added $2.5B profit
  • ~12 months later: walked estimates back 6 → 5 for subsets (cited increased pace of AI tech)
• Meta:
  • Extended life of servers/network assets to 5.5 years
  • Reduced 2025 depreciation costs by $2.9B
  • Added about ~$1 to EPS (as stated)

Investment implication embedded in the argument

• If chips become obsolete faster than expected but depreciation assumptions are too optimistic, reported earnings could later be revised downward—potentially amplifying “bubble” dynamics.

---

Timeline / Probability of Schedule Risk

Goldman Sachs (as cited)

• Historically: ~72% of data center projects arrive within the next year
• Adjusted for current conditions:
  • Next year promised capacity on time: ~60%
  • Two years out: ~50%
• Framing: “Flip a coin” for whether AI progress arrives on schedule.

Bloomberg energy research (as cited)

• By 2033, forecast power gap = 63 GW
• Context: operational data centers add up to a little over 40 GW
• Framed as: “gap half as large as every operational data center” while the grid has an 8-year wait list.

---

Overall Causal Chain (As Presented)

1. AI success boosts demand
   • “Data centers drove roughly half of U.S. electricity demand growth last year.”
2. The grid and supply chain lag
   • Interconnection queues, turbines, transformers, and transmission constrain throughput.
3. Chip cycles tighten the timing mismatch
   • Faster AI improvement → chips become obsolete sooner → the payoff clock accelerates.
4. If the clocks fall out of sync
   • Infrastructure arrives late or mismatched to the intended AI generation → ROI weakens → potential capital flight.

---

Disclosures / Disclaimers

• No explicit “not financial advice” statement appears in the provided subtitles.
• The only “disclosure-like” element is a sponsor mention (Chat LLM by Abacus AI), not a financial disclaimer.

---

Instruments / Entities / Assets Mentioned (Explicit)

• GE Vernova
• Oracle, Meta, Amazon, Nvidia
• PJM Interconnection
• Gas turbines
• Transformers, transmission lines
• Capacity auction (capacity market mechanism)
• Data centers
• AI chips / semiconductors
• No ETFs/bonds/stock tickers are provided as symbols in the subtitles.

---

Methodology / Framework Explicitly Shared

• Two-clock framework (timeline risk model):
  • Clock 1: power infrastructure delivery speed (interconnection queues + generation equipment + grid equipment + transmission)
  • Clock 2: chip/server economic obsolescence speed (2–6 years)
• Depreciation / earnings mechanics:
  • “Useful life” affects depreciation expense → affects reported profits/earnings.

---

Presenters / Sources Mentioned (End)

• George Saravelos, Head of FX research at Deutsche Bank (quote/message attributed)
• Michael Burry (investor associated with The Big Short)
• Goldman Sachs (schedule/on-time projections cited)
• Bloomberg (energy research forecast cited)
• Monitoring Analytics (PJM cost attribution cited)
• Abacus AI (video sponsor: Chat LLM by Abacus AI)