AI Bubble Thesis

Summary

PSE’s “AI Bubble Thesis” frames the 2025–2026 AI-led tech rally as a late-cycle speculative mania structured like a Jenga tower: the visible top layer is the language-model / US-listed AI stock story (NVIDIA, OpenAI, Anthropic, the hyperscaler CapEx wave) that Western media obsessively narrates, but the structural support beneath sits on three additional fronts — physical AI infrastructure (data centers, transformers, gas turbines), rare-earth supply chains, and electricity generation — each of which has a single brittle dependency, most of them on China. Pull one brick (a rare-earth export halt, an infrastructure delay, an energy bottleneck, or investors collectively asking “when do we get a return on this CapEx?”) and the whole tower comes down in classic Winners Curse Phase fashion. Darren Wilson (BBB Postcard #36, 5 June 2026) explicitly analogises it to the railway mania of the 1800s — “everyone piled in, then one day someone did the sums and found they didn’t add up.” The thesis sits inside PSE’s broader 18.6-year cycle framework: AI is the named asset class of this cycle’s peak, equivalent to the dot-coms in 2000 and securitised real-estate finance in 2007.

Core Claims

• 2026-06-05-bbb-postcard-36-ai-media (2026-06-05): “Somewhere in that AI Jenga tower is a weak link, and if you pull that brick out, the whole lot will basically come down.” — Darren Wilson — confidence: high
• 2026-06-05-bbb-postcard-36-ai-media (2026-06-05): “This is just like the railways back in the 1800s… everyone just felt that they were missing out. They all dived in. And then one day someone decided to see if the sums added up. And when they don’t, you get the almighty crash.” — Darren Wilson — confidence: high
• 2026-06-05-bbb-postcard-36-ai-media (2026-06-05): “What precisely are investors paying for with these incredible multiples for AI-related stocks?… When’s our return?… One day we’re all going to wake up and realize there’s no return anytime soon for the money put in. And that could very well be the tipping point that precedes an almighty crash.” — Darren Wilson — confidence: high
• 2026-06-05-bbb-postcard-36-ai-media (2026-06-05): “Data centers are driving some of the hottest land markets globally” — AI CapEx is, structurally, a land play and fits cleanly inside the 18.6-year cycle framework. — Darren Wilson — confidence: high
• 2026-06-05-bbb-postcard-36-ai-media (2026-06-05): On energy as the decisive constraint — Hank Paulson (cited): “The biggest potential drag we have is not having enough electricity to power our data centers.” Wilson: “China has more renewables than Europe, the United Kingdom, and the US combined.” — confidence: high
• 2026-06-05-bbb-postcard-36-ai-media (2026-06-05): Pentagon now making “direct equity investments in rare earth miners after 15 years of failed attempts to build supply chains outside China” — confirms the structural-dependency layer of the Jenga thesis. — confidence: high
• 2026-06-10-bbb-postcard-37-spacex-ipo-distortion (2026-06-10): Darren Wilson’s analysis of the SpaceX IPO reinforces the AI bubble thesis by highlighting the speculative excess and market manipulation inherent in such large, late-cycle listings. He frames the IPO as fundamentally altering stock markets to benefit a select few, connecting this market distortion back to the broader real estate cycle implications and speculative peaks. — Darren Wilson — confidence: high
• 2026-07-22-bbi-july-2026-qa-transcript (2026-07-22): Patel frames data centres as fundamentally a real-estate story: “they are very locationational investments” — must be near population centres and internet cable routes, cannot be in the Nevada desert. Land was ~5% of data-centre cost three years ago; now “significantly more.” Overbuilding already evident: Facebook renting out server capacity = “classic sign they’ve overdone it.” The 1840s railway-boom parallel is reinforced by concrete numbers: 20% of the S&P 500 are semiconductor stocks (highest weighting ever, exceeding dot-com peak); 7 companies = 34% of the index and 70% of all profits. Patel adds an accounting-vulnerability layer: hyperscalers depreciating AI chips over ~7 years despite potential 6-month obsolescence = overstated earnings; chip companies recognising revenue immediately = “double whammy” on both sides. Wilson: OpenAI offered 5% stake to US government — “AI is not making any profits… you’ve just given everyone a share of the losses.” — confidence: high
• 2026-07-22-bbi-july-2026-qa-email (2026-07-22): Stacey provides concrete data-centre land-price evidence: $360,000/acrepaidvs$25,000 normal (14.4× premium); $16Mcashvs$3M normal (5.3× premium); $500K/year option payments to keep land off market. Over 1,000 hyperscale data centres worldwide, capacity doubling every 4 years, 54% US-based. Equinix (REIT) flagged as a data-centre REIT to watch. — confidence: high
• 2026-07-23-gann-23-cpi-integrity-second-half-construction-boom (2026-07-23): Anderson identifies the second-half construction boom as channelling through AI/data-centre infrastructure spending. Tim Cook (Apple): semiconductor cost increases “unlike anything he had seen in any area in over 40 years.” Elon Musk: “the biggest price jump in anything I’ve ever seen.” Import prices ex-fuels at 4.2%, computer/electronic components +7.5% — “a long period of deflation in the cost of computers and electronic components is reversing.” The AI infrastructure boom is now an inflation driver, not just a speculative excess signal. — confidence: high

Mechanism / How It Works

The Four Layers of the Jenga Tower

Wilson’s framing (BBB Postcard #36) decomposes the AI rally into four interdependent layers, each a potential failure point:

1. Super-intelligence (the model layer). Language models (ChatGPT, Claude) drive the visible IPO and CapEx mania. Trillion-dollar valuations for OpenAI and Anthropic listings (see IPO Mania) ride on this layer. Ironically, many Western models were trained on Chinese open-access content (e.g. People’s Daily English), so even the headline US lead is partially Chinese-data-dependent. US still leads, but the gap is narrowing toward parity.

2. AI infrastructure (the land play). Data centres are now driving the hottest land markets globally — a direct expression of the 18.6-Year Real Estate Cycle manifesting through a new asset class. NVIDIA’s market cap is comparable to the GDP of a mid-size country. But US infrastructure has structural bottlenecks: transformer lead times of 3–5 years, gas turbines taking ~10 years, and 8,000 transformers already imported from China. Without continued Chinese hardware flow, US data centre buildout slows materially.

3. Rare earths (the materials layer). China controls gallium nitrate and most heavy rare earths through a single mine — Bayan Obo in Inner Mongolia. Export controls have been tightening since “Liberation Day” tariffs (April 2025). The Pentagon’s recent direct equity investments in domestic rare-earth miners come after 15 years of failed attempts to build non-Chinese supply chains — confirming that this is a time-constrained problem money alone cannot solve.

4. Energy (the decisive front). Hank Paulson: insufficient electricity is the largest single drag on US AI ambition. China has more renewables installed than the EU, UK, and US combined; the US grid is reaching breaking-point household-price stress. China is also positioning to become the world’s dominant clean-energy exporter — the next major export wedge after computers and cars.

The Two Failure Modes

Wilson identifies two distinct ways the tower can fall:

(a) Structural brick-pull. One of layers 2–4 fails first — a rare-earth export halt, a multi-month transformer delay, a grid-capacity failure during a heat-wave summer. The supply-side collapse forces a CapEx writedown across the sector. This is the geopolitical / physical failure path.

(b) Demand-side reckoning. Investors collectively ask “when do we get a return on this CapEx?” — the classic late-bubble question that Wilson explicitly analogises to the 1800s railway mania. When the answer is “not for years, and possibly never at scale,” the multiples compress catastrophically. This is the financial / behavioural failure path and is the more likely peak signal in PSE’s framework, because it is the one that maps cleanly to historical cycle peaks.

Why This Fits the PSE Cycle Framework

• Naming the cycle’s asset class. Every 18.6-year cycle peak has a defining speculative asset: 1929 utilities/Florida land, 1973 Nifty Fifty, 1989 Japanese real estate, 2000 dot-coms, 2007 securitised mortgages. 2026’s asset class is AI — both the equities and the underlying data-centre land.
• Concentration in winners. A handful of names (NVIDIA, the hyperscalers) carry the index. This is the Market Breadth Divergence pattern in equity-market form.
• Mega-IPO timing. The OpenAI/Anthropic/SpaceX wave (~$4T combined, June 2026 onwards) is the IPO-Mania expression layered directly on the AI bubble (see IPO Mania).
• Geopolitical entanglement. The cycle’s geopolitical layer (Geopolitical Cycle) is now plumbed directly through the AI supply chain — US-China interdependence on rare earths, chips, and infrastructure is no longer optional, and any escalation propagates straight into AI capex.

Key Evidence

Headline scale (mid-2026):

• NVIDIA market cap ≈ GDP of a mid-size country.
• ~$4TimpendingIPOwave(SpaceX$2T target 12 June 2026; OpenAI ~$1T;Anthropic$1T) — see IPO Mania.
• US hyperscaler CapEx setting new records each quarter.

Structural bottlenecks (mid-2026):

• Transformer lead times in US: 3–5 years.
• Gas turbine lead times: ~10 years.
• 8,000 transformers imported from China to date.
• Pentagon direct equity stakes in rare-earth miners (post 15-year failed indigenisation effort).

Energy constraint:

• US grid stress reaching breaking point; household electricity prices rising.
• China renewable capacity > (EU + UK + US) combined.
• Hank Paulson on record citing electricity as the largest US AI drag.

Historical analog:

• 1840s UK railway mania: heavy CapEx, speculative subscription, eventual realisation that route economics did not pencil out → catastrophic price collapse. Wilson explicitly invokes this as the closest pre-electronic analog.
• Wilson (BBI July 22 2026) extends the railway parallel with hard numbers: 20% of the S&P 500 are semiconductor stocks (highest weighting ever, exceeding dot-com peak); 7 companies = 34% of the index, 70% of all profits. “If it becomes a race to the bottom, how is it possible for 20% of the S&P 500 to create the profits to justify the earnings?” The dot-com parallel is also explicit: Cisco traded at 50× revenue in March 2000; “the majority of people that built the internet got wiped out but it did not wipe out the internet. Amazon and Google started their empire buying cents on the dollar.” Wilson expects the same pattern: the AI thesis will be correct long-term but current equity holders will be wiped out.
• Patel (BBI July 22 2026) adds the accounting-vulnerability angle: hyperscalers depreciating AI chips over ~7 years despite potential 6-month obsolescence = overstated short-term earnings. Chip companies recognising revenue immediately. “You’re getting this sort of double whammy of overstated earnings and current earnings on both sides of that equation.” This makes S&P earnings themselves a late-cycle vulnerability, not just a strength.
• Stacey (BBI July 22 2026 email) provides concrete land-price evidence: data-centre companies paying 14.4× normal acreage prices ($360Kvs$25K/acre), with $500K/year option payments to keep land off the market. This is the Land Speculation mechanism operating through the AI infrastructure channel — the same dynamic that drove the 1920s Manhattan skyscraper boom and the 2006 housing land rush, now manifesting through data-centre land assembly.
• Anderson (Gann #23, July 23 2026) identifies a new transmission channel: the AI infrastructure boom is now driving consumer inflation through electronics prices. Tim Cook’s Apple price increases and Musk’s “biggest price jump” quote confirm that semiconductor costs are propagating into the broader CPI. The second-half construction boom is “intensifying” and “boosting prices in areas we don’t immediately associate with inflation.” This connects the AI bubble thesis directly to the Financial Conditions framework — AI-driven inflation keeps the Fed hiking, which is what breaks the cycle.

Applications

• Cycle-position confirmation. AI bubble intensity is a primary peak-detection signal, alongside IPO Mania, Market Breadth Divergence, and the Bubble Index & Rule of 20.
• Sector rotation overlay. PSE’s Sector Rotation thesis already anticipates rotation out of US tech into industrial / resource / banking — the AI bubble thesis sharpens this by identifying the trigger (CapEx-return reckoning) rather than just the direction.
• Watch the Jenga bricks. Operationally, four signals are tracked:
  1. Rare-earth export-control headlines (China policy moves).
  2. Hyperscaler earnings calls — when CapEx growth peaks and when questions about return on AI investment intensify.
  3. US grid stress events (regional outages, transformer-shortage news).
  4. Post-IPO performance of the SpaceX / OpenAI / Anthropic listings — first-day pop and 30-day action.
• Trade timing. Wilson does not call an immediate top; the tower can stand for some time. Anderson’s parallel “8–9 months” framing (Gann #13) is consistent — the thesis is intensification of late phase, not immediate peak.

Evolution Over Time

• 2023–2024: Initial wave of generative-AI optimism; NVIDIA breaks out; first hyperscaler CapEx surge. Treated by PSE as a sector story, not yet a cycle marker.
• 2025: AI CapEx becomes the dominant equity story. “Liberation Day” tariffs (April 2025) trigger Chinese rare-earth export controls that begin to expose the supply-chain layer.
• May–June 2026: PSE explicitly names the AI rally as the cycle’s defining bubble. Anderson (Gann #13) ties the $4T IPO wave to it; Wilson (BBB Postcard #36) provides the structural / Jenga-tower decomposition and the railway-mania historical analog. The thesis is now an actively tracked PSE indicator.
• July 22, 2026 (BBI Q&A): The thesis is significantly deepened across three dimensions. (1) Semiconductor concentration risk — Wilson’s 20% / 34% / 70% S&P statistics exceed even the dot-com peak, making the index itself a structural AI bet. (2) Accounting vulnerability — Patel’s chip-depreciation analysis reveals that S&P earnings are overstated on both the buyer and seller sides of AI CapEx, meaning the earnings strength underpinning valuations is itself a bubble artifact. (3) Land-price evidence — Stacey’s $360K/acre data-centre land premiums confirm the AI boom as a direct expression of the 18.6-year land cycle, with the same speculative land-assembly dynamics as 1920s Manhattan or 2006 housing. Anderson (Gann #23, July 23) adds the inflation channel: AI infrastructure spending is now driving electronics price increases that will show up in CPI, connecting the AI bubble to the rate-hike cycle that breaks it. The “free AI from China” October 2027 shock scenario is raised as a potential triggering event. [Sources: BBI July 22 2026, BBI July 2026 email, [[2026-07-23-gann-23-cpi-integrity-second-half-construction-boom|Gann #23]]]

Contradictions & Open Questions

• Is AI actually different? The “this time is different” claim — that AI productivity gains will eventually justify the CapEx — cannot be ruled out a priori. Wilson’s thesis treats this as the bull-case tail risk. Resolution will be empirical, via post-IPO earnings vs. consensus.
• Timing of the demand-side reckoning. PSE does not specify when investors will collectively ask “where’s the return?” The signal is qualitative; quantitative triggers (operating-margin compression at the hyperscalers, AI-product revenue disappointment) need to be defined. Patel’s July 22 BBI accounting analysis adds a specific quantitative vulnerability: chip depreciation schedules (~7 years) vs. actual chip lifespans (~6 months) = the earnings overstatement is a ticking clock. When the market reprices these depreciation assumptions, earnings compress.
• “Free AI from China” as October 2027 shock. Anderson raises a speculative but high-impact scenario: what if China releases the world’s fastest AI and gives it away for free at October 2027? “That would put the earnings of American AI companies at zero.” This is the geopolitical failure path layered onto the demand-side reckoning — the trigger “will come out of left field” and will be “simple” because the system will be maximally leveraged when it hits. Not a base case, but a tail risk worth tracking.
• Cooperation vs. fragmentation outcome. Wilson explicitly considers two paths: (a) US-China cooperation forced by interdependence (tie outcome, slower AI progress), or (b) China goes alone and in ~10 years no longer needs the West. Path (a) is bearish for current US-listed multiples; path (b) is catastrophic for them. Either path bends against the current bull thesis on a multi-year horizon.
• Asset-class substitution risk for the cycle peak. If AI is not the defining asset class of this cycle peak (e.g. if a different mania emerges between now and the absolute top), the IPO-and-CapEx-concentration signal weakens. Currently no competing candidate is visible.

Related Concepts

• IPO Mania — the public-markets pricing layer of the AI bubble; OpenAI/Anthropic/SpaceX wave is the AI mania’s IPO expression.
• Winners Curse Phase — the AI bubble is the most visible Winners-Curse asset of the current cycle.
• 18.6-Year Real Estate Cycle — AI data-centre land demand makes the AI rally fit cleanly inside the cycle’s real-estate substrate.
• Geopolitical Cycle — US-China rare-earth / chip / energy interdependence is the geopolitical-cycle expression of the AI Jenga tower.
• Real Estate Cycle Peak — AI bubble is part of the peak-detection signal stack.
• Sector Rotation — AI-bubble collapse is the catalyst PSE expects for rotation out of US tech.
• Market Breadth Divergence — narrow AI-stock concentration drives the divergence at the index level.
• Bubble Index & Rule of 20 — quantitative valuation indicators complementing this qualitative bubble thesis.