Dylan Patel Explains the AI War While Cooking | In-Context Cooking
Dylan Patel of SemiAnalysis discusses the AI arms race, highlighting the massive $200B+ hyperscaler capex, the true semiconductor bottlenecks shifting from data centers back to fabs, the geopolitical chess game surrounding Taiwan and TSMC, and Nvidia's strategic battle against vertical integration.
The AI Arms Race and The Hyperscaler Innovator’s Dilemma
The AI industry is in a period of explosive, unprecedented growth, marked by a massive increase in infrastructure spending. Hyperscalers are making enormous capital expenditure (capex) commitments, with Google planning around $180 billion and Amazon $200 billion. This represents a fourfold increase from just a few years ago and is driven by the clear, tangible demand they are witnessing. The party can stop at some point, but for now, the demand for compute is insatiable, as evidenced by AI labs like Anthropic adding billions in revenue monthly, which directly translates into billions in required compute.
This aggressive spending has created an interesting conundrum. The financial markets are reacting negatively to these announcements, punishing the stocks of companies that are sacrificing short-term profitability. This is the new innovator’s dilemma for Big Tech:
- Pascal’s Wager: They are caught in a game-theoretic trap. If one hyperscaler doesn’t spend aggressively while others do, they risk being left behind permanently. The only viable strategy is to keep spending to stay in the race.
- Profit vs. Dominance: These historically profitable companies are signaling a pivot where they no longer care about profit in the traditional sense. The belief is that by 2027, a company like Google might have zero free cash flow, reinvesting every dollar into building AI infrastructure and models—what the market might perceive as “building pixie dust.”
- A Tale of Two Cities: This will create a sharp contrast between the perception in Silicon Valley, which sees skyrocketing revenue and adoption, and the view from financial centers like New York and London, which will see the world’s most profitable companies destroying their business models for uncertain future returns.
The Real Bottleneck in AI Progress is Shifting
The primary constraint on the speed of AI development is not static; it evolves year by year.
- In 2023, the bottleneck was in the semiconductor supply chain, specifically advanced packaging like CoWoS (Chip-on-Wafer-on-Substrate).
- In 2024-2025, the bottleneck has shifted to data centers and obtaining sufficient power. However, the industry is reacting quickly, with creative solutions like on-site power generation using industrial gas engines, bypassing traditional grid limitations.
- Looking ahead to 2026-2028, the ultimate bottleneck will circle back to the most fundamental layer: semiconductor fabrication capacity. Fabs are the most complex structures humans build, requiring multi-year timelines and intricate supply chains. The industry has simply not built enough fabs to meet the exponential demand for chips, and this will be the primary limiting factor for the rest of the decade.
Geopolitics, Export Controls, and Taiwan
The geopolitical situation surrounding Taiwan and TSMC remains a critical variable. There are several endgame scenarios, with the status quo being the most favorable. A “galaxy brain” scenario for the U.S. would be for the more pro-China KMT party to win in Taiwan. This would likely placate Beijing and reduce the risk of an invasion, while TSMC would still be forced to comply with U.S. export controls due to its deep integration with American equipment and banking systems.
The debate over export controls on AI chips and technology to China is complex.
- One argument is to sell China the chips but not the most advanced models. This keeps them reliant on the U.S. ecosystem and gives them less incentive to invade Taiwan, as an invasion would stop the flow of chips entirely.
- The counterargument is that the current regime of controls is not stringent enough, as China remains only slightly behind the U.S. in AI capabilities. Given the potentially world-changing impact of AI, proponents of this view argue for a much stricter lockdown to maintain a decisive lead, even if it risks antagonizing China.
Nvidia’s Strategy: Paranoia and Heterogeneity
Nvidia’s position at the top is not guaranteed, and the company operates with a healthy dose of paranoia, embodying the “only the paranoid survive” ethos of its founder, Jensen Huang. He is known for being extremely agile, updating his priors, and changing course quickly based on market signals, such as the move to acquire Groq-like technology after learning of a potential OpenAI-Cerebras deal.
To combat the innovator’s dilemma, Nvidia’s hardware roadmap is embracing heterogeneity. The upcoming “Ruben” platform is a move away from a one-size-fits-all GPU. It will feature a diverse lineup of chips optimized for different tasks:
- CPX chips for context processing and prefill.
- The mainline GPU for general-purpose training and inference.
- Groq-like chips for extremely low-latency applications.
Nvidia’s primary challenge is a “business model deficit.” Vertically integrated competitors like Google and Amazon can always save money by designing their own chips. To justify its high margins, Nvidia must innovate relentlessly to ensure its products are not just slightly better, but massively better than in-house alternatives. In this environment, with hundreds of billions of dollars being invested, traditional moats are “as shallow as they’ve ever been.”
AI Adoption and the Coming Backlash
While financial markets are wary, AI adoption is accelerating at an incredible rate in the real economy. The use of AI for coding, for example, is exploding, with tools like Anthropic’s Claude accounting for a rapidly growing percentage of commits on GitHub. This signifies a shift from consumer novelties to tools that can generate trillions of dollars in economic value.
However, this rapid progress is on a collision course with public opinion. The general public largely dislikes or fears AI, and this sentiment is likely to fuel a significant political backlash. As AI begins to impact jobs, from software development to transportation, anti-AI platforms could become a central issue in upcoming elections, creating a powerful headwind for the industry.