Monday, September 21 — Recap · 7 Episodes

AI demand is colliding with energy, computing, and physical-deployment constraints.

All-In's Naveen Rao estimates Google AI services use 12 gigawatts versus roughly 40 gigawatts for all U.S. data centers, with energy near half of token-serving costs; he warns available energy could be exhausted in about three years. The Real Investment Show says AI capital spending could rise from $421 billion to $754 billion and roughly $1.1 trillion in 2027, while Odd Lots argues manufacturing and robotics deployment—not software alone—will determine the AI race.

Middle East disruptions could lift near-term oil prices even as longer-term supply points lower.

WTFinance's Simon Hunt says damaged Gulf infrastructure could keep flows below normal for almost five years, pushing gasoline toward $6.50 and diesel to $8-$9 per gallon in his scenario. The Wolf Of All Streets sees a near-term geopolitical premium but expects surpluses, rising production, and Chinese EV adoption to drive dramatically lower prices within five years; Gareth Soloway sees support near $94-$95 after oil fell below $96, while The Real Investment Show expects crude to stabilize around $90-$93.

Elevated earnings and AI spending assumptions leave equities vulnerable to a growth slowdown.

The Real Investment Show says consensus implies about 30% earnings growth next year, followed by 11%, 10%, and 8%, while hyperscaler spending drives much of the revenue cycle; a reversal could compress margins and multiples. It cited Goldman Sachs's 8,000 year-end S&P target and $375 of 2026 earnings alongside the risk of mean reversion, while Gareth Soloway expects the S&P to reach about 8,000 by year-end and sees a bullish flag.

Bitcoin's chart has improved, but speakers still frame a new high as conditional rather than inevitable.

The Wolf Of All Streets says Bitcoin made a higher high and closed above its 50-week moving average, potentially forcing out bears, but notes it has underperformed major U.S. equity indexes year to date and remains roughly twice as volatile as the Nasdaq. Gareth Soloway calls the bear market over after a new high and sees successive higher highs, while cautioning that he would buy pullbacks rather than chase resistance; the show says a 30% S&P decline could reopen the downside case.

Higher rates threaten demand, but several speakers see duration benefiting if growth and inflation cool.

The Real Investment Show says the Fed's unanimous 25-basis-point hike came with 16 of 18 officials projecting at least one more hike, and it favors gradually adding five- to ten-year Treasuries. The Wolf Of All Streets says markets may be pricing roughly 80 basis points of further hikes too aggressively, while Simon Hunt forecasts 10-year yields falling from 5.3% to 3% or slightly lower next year as recessionary pressure builds.

Robinhood sees tokenization turning markets into a 24/7, composable trading system.

Bankless reports $2.1 billion of tokenized assets on Robinhood Chain, about 700-750 million transactions, and 14 million wallets, with 70% of tokenized-stock volume occurring on weekends or outside exchange hours. The guest expects blockchain and traditional markets to converge within a few years, though exemptions, asset limits, unavailable U.S. stock tokens, and state licensing still constrain access.

U.S.-China competition is increasingly framed as a contest over manufacturing and deployment capacity.

Odd Lots argues that reserve-currency power and strategic influence depend on industrial output, contrasting China's production of 2,600 ships with U.S. missile inventories that could be depleted within seven days of a war game. The show says domestic manufacturing carries higher energy, labor, capital, and tariff costs, while vertical integration and robotics may improve resilience despite weaker near-term economics.

New hardware architectures aim to narrow AI's enormous efficiency gap with biological computing.

All-In's Naveen Rao says the human brain operates near 20 watts, while conventional computing remains roughly 10 billion times above the thermodynamic efficiency limit; his 4D prototype uses about 5,500 nanojoules per generated image and could eventually improve power efficiency 1,000-fold. He argues that migrating existing models would require substantial conversion compute, making software and hardware co-design a major hurdle.