In Episode 99, Benny & The Squirrel are joined by Brandon Carl. Is AI headed for its shale boom moment? Brandon is a rare operator - with a decade on Wall Street in fixed income derivatives followed by a decade building in AI. He is now the founder of Soundcast, and he argues that the loudest voices on BOTH sides of the AI trade are wrong.
We really enjoyed this conversation - it’s a banger!
The Slides
The Finance/AI Divide
Playing both sides of finance and AI reveals two smart camps that completely miss each other:
Finance people understand compounding but not compounding innovation.
AI people understand Jevons-style efficiency but not macro constraints like savings rates and financing equations.
Mispricings emerge where the two frameworks fail to meet.
Tech guys assume funding is “always there” after a 17-year bull market, while finance guys ride AI narratives without understanding the underlying mechanics.
Token Deflation and the Supply Glut Case (Minute 3)
Brandon frames the bull/bear split:
Bears: Michael Burry applying a 2008 playbook to “circular financing,” with the thesis resting heavily on how fast GPUs depreciate; Brandon argues GPUs last longer than bears assume, especially in a supply-constrained market.
Bulls: Jevons paradox means demand grows forever (Bill Ackman’s “infinity ROI” line); Brandon rejects this too.
Brandon’s base case: a traditional commodity-like cycle of shortage then glut, which is actually an optimistic view, not pessimistic. Analogy: British railways were overbuilt, capital holders got smoked, but society benefited for a century.
The key puzzle addressed: how token prices can fall while GPU rental rates rise.
Game-theory: if recursive self-improvement (RSI) is winner-take-all, you buy compute to deny rivals and slash prices to block new entrants; but this alone cannot explain Anthropic’s observed gross margins.
The fundamental answer is token yield: Algorithmic improvements (i.e., DeepSeek’s inference-serving updates) deliver 2 to 3 times more tokens from the same GPU and data center. Prices fell slower than yield rose, creating an arbitrage that justifies buying more hardware, and the end result of that is a supply glut.
The parallel the shale boom, horizontal drilling unlocking inaccessible reserves; same capital-heavy dynamics, same glut outcome.
The headline numbers:
A fixed unit of machine intelligence (”one unit of 100 IQ”) is deflating roughly 13x per year; nothing in history has deflated at that rate.
Extrapolating 13x annual efficiency gains, 25% annual data-center capacity growth, and a 25% shift of training to inference implies roughly 8,000x today’s token supply in three years.
Demand must still be financed out of roughly $115T of global GDP; AI takes share rather than conjuring new income.
Labor risk: challenges the lump-of-labor fallacy. Past technologies created new jobs because people learned faster than machines; this is the first time machines learn faster. Cannibalizing part of a roughly $40T global white-collar wage bill in 3 to 5 years guts the households with the highest propensity to consume, breaking the circular flow that funds Meta, Google, and Amazon free cash flow.
RSI vs AGI, Model Tiers (Minute 17)
RSI is not a walk-back from AGI; they are different axes: AGI is the level of intelligence, RSI is the rate of change.
Karpathy’s “jagged intelligence” idea: models are superhuman in some domains, subhuman in others, so AGI has become partly a marketing term.
Chollet and others argue LLMs are only quasi-symbolic reasoners; true symbolic reasoning could unlock up to six more orders of magnitude of efficiency, which dwarfs today’s price declines.
Enterprise behavior: the “token maxing” era is over; big banks and firms like Uber now issue token budgets. ChatGPT’s default “auto” mode is the answer to wasteful model selection, though routing is genuinely hard.
The “Jev” model launch: a deliberately non-general model, trained on reinforcement learning optimized for task completion rather than human-pleasing behavior, can be about 100x cheaper and faster. Extends the toolkit beyond the “nuclear sledgehammer” to screwdrivers and saws.
The profitability paradox: as efficiency compounds while prices lag, inference economics become wildly profitable; DeepSeek’s paper implied roughly $38 per hour per NVIDIA rack. The market is inflecting from supply-constrained to demand-testing, hence the scramble for new use cases (3D modeling, cyber, robotics) to fill an 8,000x supply curve.
Personal Agents & Data Security (Minute 33)
Personal-agent reality check: Sora/Muse-style products show weak retention (about 22% after two weeks) because agents require habit change, not entertainment. Brandon’s end state: personal agents end up largely free, bundled to sell devices or intercept shopping.
Data security and AAPL 0.00%↑
Enterprise buyers prioritize data security over price (JP Morgan containerization example); open-source models on personal machines have likely leaked 10x more data than any prior year.
Apple at 39x forward earnings is partly a privacy and on-device AI bet; notable as the rare mega-cap tech not spending down free cash flow on CapEx, even as Intel reports awful consumer PC results.
The “third leg” framework: alongside the real economy and the financial economy sits a new AI economy. Whether or not LLMs are a good business, a massive AI labor force gets built anyway, with major implications for labor’s wage-bargaining power. Mike Green’s framing: the Roman Empire ran on slaves; AI labor has no leisure, rights, or illness, a fundamentally different economic model.
Macro arithmetic: even the bullish 15 to 20% AI-driven GDP growth forecasts require a matching surge in money supply or velocity; velocity is the wrong candidate since US money velocity peaked in 1999 and has fallen since, like Japan.
Soundcast and the Reckoning Ahead (Minute 43)
Podcasts are the most information-dense medium (where Huang, Musk, Nadella actually share their thinking), yet there is no real search engine for them. Brandon applies his old “Palantir for sensitive data” playbook (he previously processed bank surveillance data for market-misconduct detection) to structure, attribute, and search podcast content.
Working with roughly 100 sub-agents daily Brandon’s usage ladder runs from level 0 (do it yourself) to level 4 (multiple concurrent projects); his stack is automated research, development and hardening, plus Cloudflare’s serverless platform to skip heavy engineering.
Vibe coding versus commercial software: roughly a 40:1 time ratio; two hours to prototype, two weeks to properly engineer, and the ratio is hard to compress further.
Coding craft shift: the industry is moving from artisan to industrialized phase, but AI is weak precisely at the foundational layers (architecture and database decisions). “Radioactive weeds”: about 10% of AI-generated code is quietly wrong, and models replicate those errors from your codebase over time. Printer-press analogy: what becomes scarce is editorial judgment, not output.
The labor cycle: first-order panic layoffs, second-order hiring of engineers to build automation, then 6 to 18 months of governance and RIF structure before labor data actually weakens, which is what we are now seeing.
XPO 0.00%↑ XPO Logistics example: an 18-month internal build, now generating large returns; ROI is real but slow, like any capital project.
The market timing question: when does the hated, puked-out 2002-style entry point arrive?
Goldman’s May forecast of 120 quadrillion tokens per month by 2030 was already outmatched: with current DeepSeek-class serving efficiency, roughly 10,000 quadrillion tokens per month is producible today.
Lab CEOs (Dario) must commit to compute purchases 24-plus months out while model efficiency compounds and demand is unknowable; odds of nailing supply are under 10%.
Because this buildout is debt-financed, overbuilding is not a soft equity loss but “balance sheet Russian roulette.” Even if CapEx stopped today, algorithmic gains would keep compounding on surplus supply.
Noam Brown’s observation: top researchers’ visibility has collapsed from about a year to about three months, making 2040 forecasts farcical.
Follow Brandon on X (@BrandonJCarl), essays at essays.brandoncarl.com, and Soundcast at soundcast.cc
If you act on anything provided in this newsletter, you agree to the terms in this disclaimer. Everything in this newsletter is for educational and entertainment purposes only and NOT investment advice. Nothing in this newsletter is an offer to sell or to buy any security. The author is not responsible for any financial loss you may incur by acting on any information provided in this newsletter. Before making any investment decisions, talk to a financial advisor.
