Latent Space: The AI Engineer Podcast

Episode: Claude Code for Finance + The Global Memory Shortage: Doug O'Laughlin, SemiAnalysis
Date: February 24, 2026
Guests: Doug O'Laughlin (SemiAnalysis), Host: John (Latent Space)

Episode Overview

In this episode, Doug O'Laughlin, analyst and co-founder at SemiAnalysis, joins John for a deep-dive into the explosive impact of AI coding agents (specifically Claude Code), the emerging "global memory shortage" and its impact on the semiconductor industry, and insights into adopting AI tools within high-stakes finance and tech research. The conversation moves fluidly from technical analysis to career advice, the transformation of information work by AI, and even Doug's personal journey. This episode stands out for its candid, high-signal look behind the scenes of both tech and finance at the dawn of widespread AI-powered productivity.

Key Topics & Insights

1. From Analyst to Tech Thought Leader

(00:00–12:02)

• Doug's Background: Doug shares his entry into finance as a value investor and how he "nerd sniped" himself into semiconductors after discovering ASML, seeing chips as science fiction come true.
  • "I found ASML, and I fell in love with it… It's the most important thing we've ever made." (04:06)
• Shifting Industry Perceptions: Once viewed as mature and unsexy, the semiconductor industry’s centrality to modern society (esp. as Moore’s Law slowed) was Doug’s “goated insight”.
  • "People just thought [semis] were an old mature business... But every single day, just making a new chip was like science fiction.” (06:51)
• Career-Defining Predictions: Doug explains how his early prediction of Moore’s Law ending and the rise of companies like Nvidia turned out “not only right, but at a magnitude nobody predicted.”
  • "We called it, we were right... but the thing that still shocks me is the magnitude of how right we are." (10:05)

2. Explosive Rise of Claude Code and AI Agents

(12:02–26:20, 34:07–55:46, 61:19–62:48)

• Adoption Inflection Point: December 2025, Doug experiences "Claude code psychosis," where new AI code agents start to “one-shot” complex multi-step financial tasks that would've taken a human analyst a day.
  • "It just started one-shotting everything... all these MVPs that used to take forever, just done. It’s a huge difference." (15:48)
• Benchmarking with Real Work: Doug ran actual case studies used for hiring at SemiAnalysis through Claude Code, seeing it beat poor human submissions and dramatically accelerate workflows.
• On AI-Generated Slop and Human-in-the-Loop: Claude (and other LLMs) still make mistakes, behave like “junior analysts,” and lack meta-level learning.
  • "This crap makes mistakes all the time... It's like a junior analyst... I don't think that meta-level learning is there yet." (00:00, 45:54)
  • "It's a game of hygiene now." (45:54)
• Multiplicative Amplification of Experts: LLMs massively increase research velocity for experts, but artisanal judgment is still indispensable, especially to prevent “slop.”
  • "It amplifies everyone who is an expert... but you have to still do something. You can't just slop it up." (47:22)
• Implications for Entry Level Work: Entry-level data analysis jobs are "at risk"—AI agents can outperform an average 22-year-old analyst and handle many rote information tasks.
  • “I just can’t imagine... that a 22-year-old would outperform a well-thought-out agentic system.” (48:19)
• Automation and Human Review: AI agents enable rapid, parallel exploration of ideas, with human review/hygiene as the critical bottleneck.
  • "As a human, you get one turn. With Claude, maybe you get three turns, and the review process is the thinking." (52:32)

3. The Changing Landscape of Information Work

(19:54–31:53, 61:19–66:03, 69:39–74:08)

• Claude Code as Bloomberg/Excel Killer: Doug is unequivocal that Excel, PowerPoint, Bloomberg, and other “human IDEs for information work” are approaching obsolescence for analysts—AI agents will assemble analysis, charts, and reports with superior speed and flexibility.
  • “I will never make a chart in Excel again.” (71:33)
  • "Claude Code or CoWorker or whatever is... going to destroy all [legacy information work].” (70:10)
• Microsoft is Most at Risk: As the world's dominant horizontal software provider, Microsoft's moat is at risk (“cannot paint a bigger target”).
• The Next Productivity Layer: Drawing analogies to the industrial revolution and railroad boom, Doug suggests that AI-driven information work is the next "layer of the economic cake"—suggesting white collar transformation of similar magnitude.
  • “It just feels like amazing, massive moments happen... All of knowledge work is going to change.” (60:13)
  • “It’s like the steam engine was invented... This is happening in our lifetimes.” (61:21)
• Memory Shortage as the New Bottleneck: The global AI GPU buildout is now supply-constrained on DRAM/HBM, akin to oil refiners in industrialization.
  • "The most obvious logical leg of the AI trade is effectively investing in memory capacity." (102:50)

4. Semiconductor, GPU, and Memory Market Dynamics

(62:56–111:13)

• The “Memory Mania”: DRAM and HBM are deeply supply-constrained due to years of underinvestment and disruptive demand from AI/LLM training/inference.
  • “We could see DRAM prices go up 100% again... The memory shortage is just crazy.” (101:33)
• TPUs vs. Nvidia: Google’s TPUs (esp. V7) are briefly more cost-efficient than Nvidia, but supply, software ecosystem, and talent dispersion matter—a window quickly closing as Nvidia’s supply chain and next-gen (Reuben) comes online.
  • "TPU is kind of like this replicatable pod... If you want market share, now’s the time." (96:37)
• Squeeze Along the Memory Cascade: Not just HBM supply, but all layers are squeezed—leading to exotic tech for scavenging old DRAM (see CXL).
  • “They’re going to take every bit of spare memory they can find and attach them via CXL...” (104:48)
• Emerging CPU Shortage: Underinvestment in CPUs due to AI capex and refresh cycles means even basic web services are degrading.
  • “We might actually be seeing a CPU shortage partially because of this refresh cycle but also... RL and coding agents generate crazy demand.” (112:56)
• New Computation Paradigms: Discussion of burning model weights directly into chips (e.g., Thales) to obviate memory bottlenecks—potential future bifurcation of compute hardware for inference vs. training.

5. Tooling: Claude Code, Codex, Agent Swarms, Design Patterns

(26:20–34:07, 72:46–80:19)

• AI Coding Agents Horse Race: Codex 5.3 catching up to Claude Code, but each has different RL training biases (Claude more generalized, Codex hyper-focused on code).
  • “Codex is pretty good... I think Codex is back.” (75:59)
  • “Having them (Claude & Codex) yell at each other is really great.” (79:19)
• Agent Teams & Swarms: Subagents and agent swarms are promising but experimental; simpler subagent task delegation outperforms complex RL-prompted swarms for now.
  • “Subagents are okay... The agent team is actually really bad. Clearly not RL done.” (32:17)
  • “Kimmy 2.5 swarm is actually good… Agent team makes it worse.” (33:20)
• Hooks and Functional Patterns: Doug underutilizes hooks but sees potential for executing repeat, structured analyses; advises “less is more” and clarity of context in prompts.

6. Career, Writing, and Personal Philosophy

(114:34–123:50)

• Writing Discipline: Doug is a voracious reader; kept a streak of weekly newsletter writing for years, advocates writing more for non-fiction quality.
  • “I wrote every week since October 2021… All the people who write about writing say, just write more.” (116:32)
  • “Pre-write; think about it, go to sleep, and write with a fresh context window.” (118:48)
• Hiking and Life Fulfillment: Doug completed the Continental Divide Trail (CDT)—6 months, ~2,800 miles—as a personal sabbatical, stressing the importance of knowing one’s limits and seeking “raw experience.”
  • “I wouldn’t give up that experience for anything... The raw experience of life is so meaningful and you don't get to experience it while doing it that way.” (121:17)

Notable Quotes & Memorable Moments

• On AI Work Product:
  • "It makes mistakes all the time... I think of it once again as like a junior analyst." — Doug (00:00)
  • "It’s all a skill issue now." — Doug (21:36)
• On the Future of Tools:
  • "Excel is the IDE for analysts. Bloomberg is the IDE for analysts. All of them are dead." — Doug (70:10)
• On the Pace of Change:
  • "All of knowledge work is going to change in a decade, and we’re watching it from the front row.” — Doug (60:13)
• On the Global Memory Shortage:
  • “The most obvious logical leg of the AI trade is investing in memory capacity.” (102:50)
  • “We could see DRAM prices go up 100% again… that’s just crazy.” (101:33)
• On Microsoft & Incumbents:
  • "Microsoft is the horizontal software company... I cannot paint a bigger target." — Doug (81:49)
  • "Claude for Excel, Claude for Powerpoint... Microsoft should have built it." (85:02)
• On Personal Growth:
  • "Self-mastery is your most important tool... The journey is the destination.” (123:43)

Key Timestamps

• 00:00–12:02 — Doug's career journey, being early to the semis trade
• 12:02–26:20 — Breakout moment: Claude code “one-shots” complex research tasks, benchmarked vs. actual analyst work
• 19:54–31:53 — Analyst workflows, the waning of Excel/Bloomberg, information work disruption
• 32:14–34:07 — Agent swarms, subagents, and future of agentic automation
• 45:54–48:19 — The limits of AI: meta-learning, human-in-the-loop, AI as "junior analyst"
• 61:19–62:48 — Historic perspective: economic cycles, new productivity layers
• 62:56–111:13 — Deep dive: global memory shortage, GPU/TPU rivalry, semicon supply chains, CPU shortage, hardware innovation
• 114:34–123:44 — Writing philosophy, life outside work, hiking the CDT

Tone & Style

Casual, direct, often irreverent but with deep technical and market insight. Doug and John frequently use vivid metaphors (AI as an amplifying force for analysts, the global AI buildout as “railroads”/“steam engines”), and acknowledge the limits as well as the promise of emerging AI tools. Both speakers are candid about their learning process, willingness to change opinions, and working through complex “meta” challenges. The conversation is dense but lively, rich with practical takeaways for practitioners and observers of both finance and AI tech.

For Listeners

• For AI Engineers/Builders: Detailed on-the-ground validation of Claude Code & other AI agents, with discussion of agentic design patterns and pitfalls.
• For Investors/Finance Professionals: Maps concrete productivity gains to capital flows, market structure, and the emerging hardware shortage story.
• For Tech Observers: Vivid “current thing” synthesis of how foundational models are changing real research work—what’s hype, what’s reality, what’s next.

Memorable Catchphrase:
"It's all a skill issue now." — Doug (21:36)

Further Reading / Links

Full show notes and references at: https://latent.space