This video offers a qualitative, "vibes-based" analysis of the AI boom in Silicon Valley, reflecting on observations from a recent visit and the Hot Chips conference. It notes a shift in AI chip development, with a growing focus on photonics driven by the extreme power demands of AI. The author delves into the unique strategy of the Japanese foundry Rapidus, the concept of "peak software" due to AI's impact on developer productivity, and the subsequent redirection of investment towards harder, more physical problems. These include AI-accelerated materials discovery, novel computing paradigms (photonic quantum and superconductor-based), and advancements in physical AI/robotics. Despite concerns about an "AI bubble" and the profitability of large-scale AI investments, the video concludes with an optimistic outlook, suggesting that the unexpected success of ventures like OpenAI has reignited a willingness to fund and commercialize ambitious, difficult technologies over traditional software-as-a-service models.
Exploring the AI Boom: Hot Chips, Hard Problems, and the Future of Compute
0:00 The AI Boom Enters Year Three: The author reflects on his third year observing the AI boom in the Bay Area, noting a sense of "locking in" for certain assumptions and a renewed excitement driven by a new generation of builders.
0:53 Hot Chips Shifts Focus: Unlike last year's Hot Chips conference, which featured a plethora of new AI accelerator teams, this year indicated a closing window for new AI chip designs, with Google's Ironwood TPU being a notable exception.
1:45 Photonics Emerges as a Key AI Solution: A significant event at Hot Chips was the back-to-back presentations by four photonics companies (Celestial AI, Labs, Light Matter, Nvidia), highlighting how the extreme power demands of AI chips are revitalizing the photonics industry.
2:33 Concerns for Silicon Photonics: The author notes that silicon photonics is a challenging manufacturing problem with a historically small market, raising concerns about potential fallout in the event of an "AI winter."
2:53 Rapidus's Novel Foundry Strategy: Rapidus, a Japanese foundry experiment, is developing a 2nm process licensed from IBM with a unique focus on achieving the world's fastest possible turnaround times (as little as 15 days for single wafers) rather than high-volume manufacturing. The author questions the scalability of this "fine art wafers" approach.
4:37 AI and "Peak Software": The video discusses the rise of AI-fueled software development, characterizing it as "peak software" where AI makes programmers highly productive. This leads to concerns about rapid replication of successful software concepts ("vibe coding threat") and challenges for startups without strong proprietary advantages.
7:00 Investment Shifts to AI-Accelerated Materials Discovery: As software becomes "cheapened," investors are increasingly funding "harder problems," such as AI-accelerated materials discovery. High-profile examples like Periodic Labs (co-founded by ChatGPT co-creator Liam Fedus) are building automated wet labs to revolutionize material science.
9:02 Exploring Alternate Forms of Compute: The author highlights visits to companies developing novel compute architectures:
9:09 SciQuantum: Focused on light-based quantum computing, the company has raised $1 billion and is converting a former semiconductor fab into an assembly facility for its photonic quantum chips, operating at superconducting temperatures.
10:35 Snowcap Compute: Commercializing a CMOS-compatible, superconductor-based approach for low-energy computing, evolving from earlier Josephson Junction technologies.
11:24 Progress in Physical AI and Robotics: The author observes progress in robotics, revisiting an earlier assessment about Chinese companies' dominance.
11:47 One X Technologies: Aims to introduce household robots, designed with a unique cable and pulley system for safety. Their strategy involves an expensive initial model, leveraging remote teleoperation to generate training data, with autonomous capabilities for simple tasks. They aim to ship by 2026.
13:01 Robotics Adoption Compared to Cars: The robotic future is predicted to unfold more like the gradual, localized adoption of cars over many years, rather than the rapid, widespread integration seen with consumer software products like ChatGPT.
14:03 Coding: AI's First Killer App: The author identifies coding assistants (e.g., Anyphere/Cursor, Replet) as AI's clear "killer app," foreseeing a significant impact on software engineers' wages globally.
14:34 Search for a Second AI Killer App: Discussions in San Francisco reveal that a definitive "second killer app" for AI beyond coding has not yet emerged, with existing agentic AI products often being "brittle" or limited in broad applicability.
15:55 Growing Concerns About an AI Bubble: Outside of San Francisco, there's increasing acceptance of an "AI bubble," with concerns about potential economic challenges. Quantitative analyses suggest unfavorable returns on hyperscaler AI capital expenditures, and the growing involvement of third-party investors signals a potential shift from pure cash-flow funding.
16:51 Breakthroughs Could Reignite Hype: The author suggests that if AI solves a "millennium problem" (e.g., Navier-Stokes), it could significantly alter the dynamics and reignite hype.
17:17 Optimism for Hard Technologies: Despite the speculative nature of many AI-fueled ventures, the author concludes with optimism, drawing a parallel to OpenAI's unexpected success. This has fostered a renewed belief in commercializing hard technologies, making the current investment landscape "way better than funding another Uber for X or another SAS."