Summary
Highlights
The AI boom is in its third year, feeling more cemented than previous tech cycles. Silicon Valley is reinvigorated with a new generation of builders, fueled by the Open AI dream. Last year's focus on AI accelerator chips has shifted, with the window for new entrants closing, leaving Nvidia largely dominant, with Google's Ironwood TPU as a notable exception.
The power demands of modern AI accelerators are driving a resurgence in the photonics industry. Presentations from four photonics companies at Hot Chips signal AI's insatiable need for scale, pushing advancements in semiconductor supply chains despite the challenges of silicon photonics manufacturing. Concerns remain about potential fallout in an 'AI winter' due to the niche market.
Rapidus, a Japanese foundry, is working on a 2nm process node licensed from IBM. Their unique approach focuses on achieving the fastest possible turnaround time for single wafers (as little as 15 days), rather than high-volume manufacturing. The author expresses skepticism about the target market for this 'fine art wafers' strategy, questioning its scalability for fabless companies like Nvidia that require massive volumes.
The SAS (Software as a Service) model, once highly valued for its recurring revenue, faces new challenges with AI. While AI makes software teams more productive, it also creates a 'vibe coding threat' where competitors can quickly replicate successful software products. AI is seen as 'peak software,' making it difficult for new startups to gain an advantage without proprietary data or established access, leading investors to seek harder problems to solve.
Investors are increasingly funding efforts to solve complex problems, such as AI-accelerated materials discovery. Startups like Periodic Labs, co-founded by a ChatGPT co-creator, are using AI models to accelerate the discovery of new materials, aiming to automate a traditionally trial-and-error process. The author questions whether AI can truly predict behaviors for entirely new material branches.
The video highlights two companies developing alternative compute solutions: PsiQuantum, focused on light-based quantum computing, is investing heavily in building an assembly facility for their photonic quantum computer. Snowcap Compute is commercializing a superconductor-based approach for low-energy compute, which is designed to be CMOS compatible and more manufacturable than previous attempts.
Progress in physical AI and robotics is becoming evident. One X Technologies is developing humanoid robots for household use, employing a cable and pulley system for safety. Their go-to-market strategy involves initial teleoperation to gather data for training, with autonomous capabilities evolving over time. The author emphasizes the importance of 'shipping' products frequently to get feedback, but notes that the robotics rollout will be slow and localized, unlike consumer software.
San Francisco remains the epicenter of AI, with ongoing discussions about scaling, power demands, and efficiency. Coding is confirmed as one of AI's killer apps, demonstrated by the rise of coding assistants. However, beyond coding, a second killer app for AI has yet to emerge. Agentic AI products are currently deemed 'brittle' and useful mainly for the business process outsourcing industry, which isn't considered big enough to redefine the market.
Outside of San Francisco, there's growing acceptance of an AI bubble and concerns about its eventual pop. Questions arise about the returns hyperscalers are making on AI capital expenditures, with reports of joint ventures and debt offerings suggesting financial challenges. The author contrasts this with the success of OpenAI, a company that commercialized a 'hard technology' against early skepticism, fostering a new belief in tackling complex problems in Silicon Valley.