▶ 8:50 · AI Infrastructure · Post-training inference demand will be 5-10x larger than pre-training compute

episode briefing

What The Best AI Investors Are Buying Right Now

2026-05-28 · 14 company · 10 thematic

sentiment

11 bull1 bear2 neu

speakers

Managing partner and CIO of Atreides Management, a technology-focused hedge fund. He previously ran the Fidelity OTC fund for years and is one of the most-followed voices on semiconductors, AI infrastructure and growth tech.

leopold aschenbrenner

Founder of the investment firm Situational Awareness and author of a widely read essay series on artificial general intelligence and national security. Previously worked on the superalignment team at OpenAI.

ejaaz

Co-host of Limitless Podcast (episode 210), a 4x/week show covering the frontier of AI and technology with a newsletter reaching hundreds of thousands. The podcast recently spun out as an independent entity.

now playing · AI Infrastructure

AI Infrastructuretailwindscore 9/10gavin baker

Post-training inference demand will be 5-10x larger than pre-training compute

The AI compute stack is shifting from pre-training to post-training (fine-tuning, reasoning, agentic workflows), which requires dramatically more inference compute; this structural shift dr…

AI Geopolitics & Export Controlstailwindscore 7/10gavin baker

China's energy abundance and manufacturing scale favor inference-focused chip architectures

China has surplus energy and domestic fab capacity, enabling a different AI infrastructure buildout optimized for inference rather than pre-training; this creates a parallel supply chain an…

AI Infrastructuretailwindscore 7/10gavin baker

Verticalized SLMs on device solve enterprise privacy and compliance barriers

Enterprises cannot send sensitive data (medical, financial) to cloud models; small language models running locally on devices (enabled by Apple silicon) unlock AI adoption in regulated vert…

Sovereign AItailwindscore 8/10gavin baker

Speed of physical deployment is the new moat in AI infrastructure

Bits move exponentially faster than atoms; the ability to permit, power, and build data centers in weeks rather than years is a defensible advantage — companies that compress deployment tim…

AI Hardware & Chip Architecturetailwindscore 8/10gavin baker

Tokens per watt becomes the decisive metric for AI lab procurement

As hyperscalers spend billions on compute, cost efficiency (performance per watt) determines which chips win; labs will switch to cheaper inference even at slightly lower intelligence, maki…

Space-Based Computetailwindscore 8/10gavin baker

Space-based data centers bypass terrestrial grid and permitting constraints

Terrestrial grids are at capacity and face 40% petition rates against new data centers; orbital compute via Starship offers unlimited solar energy and cooling, making space the next infrast…

AI Bubble / Capex Debatetailwindscore 9/10gavin baker

Supply constraints (TSMC, ASML, energy) prevent oversupply and keep AI CapEx cash-funded

Unlike the debt-fueled dot-com bubble, current AI CapEx is funded from hyperscaler free cash flow; hard physical bottlenecks in chip fabrication, packaging, and power prevent oversupply, ma…

Semiconductorstailwindscore 8/10ejaaz

ASML's EUV monopoly and 5-year backlog enforce a hard ceiling on chip supply growth

ASML is the single choke point for leading-edge semiconductor manufacturing; its inability to scale EUV output quickly means no amount of capital can rapidly increase chip supply, protectin…

Memory & Storagetailwindscore 8/10ejaaz

HBM scarcity gives SK Hynix unprecedented 70% margins and pricing power

High-bandwidth memory is the tightest sub-sector in AI hardware; SK Hynix allocates supply by price not volume, extracting monopoly rents and signaling a multi-year structural shortage that…

AI Infrastructuretailwindscore 9/10gavin baker

Watts and wafers are the two brick walls limiting AI acceleration

The fundamental constraints on AI growth are electricity (watts) and leading-edge chip fabrication (wafers); so long as demand exceeds supply for both, infrastructure providers will capture…

←

▶ 8:50 · AI Infrastructure · Post-training inference demand will be 5-10x larger than pre-training compute

episode briefing

Limitless Podcast

What The Best AI Investors Are Buying Right Now

2026-05-28 · 14 company · 10 thematic

sentiment

11 bull1 bear2 neu

speakers

gavin baker

leopold aschenbrenner

ejaaz

now playing · AI Infrastructure

AI Infrastructuretailwindscore 9/10gavin baker

Post-training inference demand will be 5-10x larger than pre-training compute

The AI compute stack is shifting from pre-training to post-training (fine-tuning, reasoning, agentic workflows), which requires dramatically more inference compute; this structural shift dr…

AI Geopolitics & Export Controlstailwindscore 7/10gavin baker

China's energy abundance and manufacturing scale favor inference-focused chip architectures

China has surplus energy and domestic fab capacity, enabling a different AI infrastructure buildout optimized for inference rather than pre-training; this creates a parallel supply chain an…

AI Infrastructuretailwindscore 7/10gavin baker

Verticalized SLMs on device solve enterprise privacy and compliance barriers

Enterprises cannot send sensitive data (medical, financial) to cloud models; small language models running locally on devices (enabled by Apple silicon) unlock AI adoption in regulated vert…

Sovereign AItailwindscore 8/10gavin baker

Speed of physical deployment is the new moat in AI infrastructure

Bits move exponentially faster than atoms; the ability to permit, power, and build data centers in weeks rather than years is a defensible advantage — companies that compress deployment tim…

AI Hardware & Chip Architecturetailwindscore 8/10gavin baker

Tokens per watt becomes the decisive metric for AI lab procurement

As hyperscalers spend billions on compute, cost efficiency (performance per watt) determines which chips win; labs will switch to cheaper inference even at slightly lower intelligence, maki…

Space-Based Computetailwindscore 8/10gavin baker

Space-based data centers bypass terrestrial grid and permitting constraints

Terrestrial grids are at capacity and face 40% petition rates against new data centers; orbital compute via Starship offers unlimited solar energy and cooling, making space the next infrast…

AI Bubble / Capex Debatetailwindscore 9/10gavin baker

Supply constraints (TSMC, ASML, energy) prevent oversupply and keep AI CapEx cash-funded

Unlike the debt-fueled dot-com bubble, current AI CapEx is funded from hyperscaler free cash flow; hard physical bottlenecks in chip fabrication, packaging, and power prevent oversupply, ma…

Semiconductorstailwindscore 8/10ejaaz

ASML's EUV monopoly and 5-year backlog enforce a hard ceiling on chip supply growth

ASML is the single choke point for leading-edge semiconductor manufacturing; its inability to scale EUV output quickly means no amount of capital can rapidly increase chip supply, protectin…

Memory & Storagetailwindscore 8/10ejaaz

HBM scarcity gives SK Hynix unprecedented 70% margins and pricing power

High-bandwidth memory is the tightest sub-sector in AI hardware; SK Hynix allocates supply by price not volume, extracting monopoly rents and signaling a multi-year structural shortage that…

AI Infrastructuretailwindscore 9/10gavin baker

Watts and wafers are the two brick walls limiting AI acceleration

The fundamental constraints on AI growth are electricity (watts) and leading-edge chip fabrication (wafers); so long as demand exceeds supply for both, infrastructure providers will capture…