In this episode of the Stewart Squared podcast, host Stewart Alsop speaks with his father Stewart Alsop II, who joins from Tokyo while Stewart broadcasts from Buenos Aires at 5 AM his time. The conversation covers NVIDIA's new Spark chip announcement and its partnership with Microsoft to bring ARM-based processors to Windows PCs, finally allowing Windows to compete with Apple's performance gains from five years ago when they switched to their own ARM-based M-series and A-series chips. They discuss the competitive dynamics between chip manufacturers, the token apocalypse affecting AI coding assistants like Claude and Codex, and how companies like Anthropic are struggling with inference costs while renting data center capacity from SpaceX's underutilized X AI facilities. The discussion also touches on the rise of small models for on-device AI, the dominance of Chinese models in developing markets, SoftBank's ownership of ARM and history of big bets, and how attention and access to insider deals have shaped the AI investment landscape. For more context on Microsoft's strategy, Stewart Alsop II references a Ben Thompson Stratechery interview with Microsoft CEO Satya Nadella that helped clarify how Windows now runs on ARM architecture. Timestamps 00:00 Stewart Alsop welcomes listeners, explains recording at 5 AM his time, 5 PM in Tokyo Japan, discusses NVIDIA's new announcement about processors and chips for Windows computers05:00 Discussion of ARM architecture versus Intel chips, Apple's competitive advantage using ARM-based M-series processors, how Windows has fallen behind Macintosh in performance capabilities10:00 NVIDIA positioning new chip as AI-focused but actually ARM-based architecture, Microsoft modifying Windows to run on ARM, multiple manufacturers producing laptops with NVIDIA chips instead of Intel15:00 Deep dive into ARM licensing model, SoftBank ownership of ARM, how NVIDIA's CPU competes with Intel while Microsoft adapts Windows for ARM architecture20:00 Intel's competitive position, Microsoft's alliance with NVIDIA, discussion of GPU versus CPU functions, how graphics processing naturally supports training large language models25:00 Token apocalypse experience with Claude and Codex, rate limiting issues, moving between coding assistants, quality regressions and improvements in different AI coding tools30:00 Anthropic efficiency improvements with Opus 4.8, competitive dynamics between Claude Code and Codex, strategy of using multiple subscriptions to avoid rate limiting35:00 Chinese models as workhorses for global users who cannot afford expensive subscriptions, frontier models limited to Google Anthropic and OpenAI, affordability challenges internationally40:00 Small models running on devices versus cloud-based large models, Apple's WWDC expectations for integrating models on iPhone, personal computing productivity shifts45:00 SoftBank history with Masayoshi Son making big bets, ARM acquisition rationale, attention-based access to insider deals, comparison to celebrity entrepreneurs gaining investment access50:00 Historical perspective on insider access to deals and IPOs, closing remarks about continuing conversation from Japan Key Insights 1. Microsoft and NVIDIA announced a new ARM-based processor called Spark that will run Windows, marking a significant shift in the PC market. This represents Microsoft finally moving away from its dependence on Intel chips, similar to what Apple did five years ago when it introduced its M-series chips for Macintosh computers and A-series for iPhones. The development is positioned as an AI chip for marketing purposes, but the real significance lies in the ARM architecture, which NVIDIA has licensed. This alliance between Microsoft and NVIDIA directly challenges Intel's dominance in the PC processor market and could make Windows machines more competitive with Apple's Macintosh in terms of performance and efficiency.2. The competitive landscape in AI coding assistants has dramatically shifted, with Anthropic's Claude Code releasing version 4.8 that significantly improved code quality and token efficiency. After experiencing severe rate limiting issues in May due to inference capacity constraints, Anthropic made their coding model much more efficient at the token level, allowing users to accomplish more within existing subscription tiers. Meanwhile, OpenAI responded with Codex to compete with Claude Code's success from last December. This competition has created a situation where programmers are now splitting subscriptions between multiple services, paying for both Codex and Claude Code while using Chinese open-source models as fallback options when they hit rate limits.3. The token apocalypse revealed fundamental business challenges for AI companies as they struggle to balance inference capacity with growing demand. Anthropic had to make difficult decisions to prioritize enterprise customers over individual users, causing noticeable degradation in their chatbot product quality. The company was spending enormous amounts of inference capacity on making conversations feel natural and philosophically relevant, which proved financially unsustainable. Companies like Uber reportedly burned through their entire token budgets in just three months, highlighting how the rush to maximize token usage became a poor metric for actual productivity, falling victim to Goodhart's Law where a measure that becomes a target ceases to be a good measure.4. The revenue growth projections for Anthropic demonstrate the explosive commercial potential of large language models. The company expected to end 2025 with 9 billion dollars in revenue, but by the second quarter had revised expectations to 50 billion dollars. This astonishing growth comes from companies paying substantial enterprise budgets for AI services. Meanwhile, SpaceX's X AI data center, built rapidly but underutilized due to poor adoption, has been rented out to both Anthropic and Google for approximately 2 billion dollars per month collectively, showing how infrastructure built for one purpose can be repurposed when the original business model fails to generate sufficient demand.5. SoftBank's strategic bet on ARM five years ago positioned the company at the center of the current processor revolution. Founded by Masayoshi Son, SoftBank has a history of making large, bold investments over four decades, including early deals with Microsoft for software distribution in Japan. The company took ARM private and then public again, with SoftBank retaining majority ownership. This investment proved prescient as ARM's licensing model became increasingly valuable, especially as Apple, NVIDIA, and others adopted ARM architecture for their processors, making it the de facto standard for CPU design across multiple device categories from smartphones to personal computers.6. The future of AI appears to be splitting between small models running on devices and large frontier models in the cloud. Apple is expected to announce at WWDC its integration of Google models on the iPhone, utilizing small models that can run locally on the device for personal productivity tasks like calendar and email management, while connecting to cloud-based large language models for more complex operations like programming. This hybrid approach addresses both privacy concerns and cost efficiency, as running everything through cloud-based large language models proves financially unsustainable for everyday personal computing tasks. The industry consensus currently recognizes only three companies as leaders in frontier models: Anthropic, OpenAI, and Google.7. Chinese AI models are emerging as the workhorses for global markets due to a...