Dylan Patel - The Infinite Demand for Tokens, Claude Mythos, and Supply Constraints - [Invest Like the Best, EP.469]

Episode Summary

Executive Summary: Dylan Patel argues that AI token demand is exploding because frontier models have become dramatically more capable and useful, making implementation cheap and prompting users to spend aggressively. He says this demand is colliding with real supply bottlenecks across GPUs, memory, wafers, CPUs, and the semiconductor equipment stack, creating a prolonged period of shortages, rising margins, and concentration of value among the best-capitalized firms.

Main Topics: Explosive token demand and personal usage surge (Priority: 5/5): Patel describes SemiAnalysis’s own AI spend rising from tens of thousands to a $7M annual run rate, driven by broader adoption of Claude Code and frontier-model workflows across technical and non-technical staff. Frontier models are creating a step-change in usefulness (Priority: 5/5): He argues that the newest frontier models are the only ones people really want because they unlock materially better economic output, and users are rapidly moving to the latest release even when older models still worked fine. AI makes implementation cheap and reorganizes business advantage (Priority: 5/5): The conversation emphasizes that ideas are now cheap while implementation is increasingly easy, so competitive advantage shifts toward choosing the right ideas, moving fast, and capturing value before commoditization. Supply-side bottlenecks across the semiconductor stack (Priority: 5/5): Patel walks through shortages and pricing pressure in GPUs, memory, logic/fabs, wafer equipment, copper foil, optics, and CPUs, arguing that supply cannot scale quickly enough to match demand. Token economics, margins, and concentration (Priority: 4/5): He highlights that labs like Anthropic may already have very high gross margins because token demand outstrips compute supply, and suggests tokens will increasingly accrue to fewer companies and connected customers. Robotics and the next demand wave (Priority: 3/5): Patel says robotics will become a major new token and compute demand center as software-only gains spill into physical automation, though real breakthroughs likely require better pre-trained robot models. Public backlash and AI messaging risk (Priority: 3/5): He predicts rising protest and political backlash against AI, arguing labs need better public communication focused on immediate benefits rather than abstract future risk.

Key Arguments: Frontier-model capability is improving fast enough that many users feel compelled to pay for the latest version immediately, even at high cost. SemiAnalysis’s AI spend rose from tens of thousands of dollars to a $7M annual run rate, showing that practical usage is exploding. A single person using Claude Code can now build tools that previously would have required entire teams, such as chip reverse-engineering visualization and a US grid mapping system. AI is commoditizing information and analysis businesses, so firms must constantly improve or be displaced by competitors using the same tools. Token demand is not just from consumers; it is driven by new, high-value use cases that generate more value than their token cost. Supply constraints are structural because compute, memory, fab capacity, and equipment cannot expand at the same pace as token demand. Memory capacity can only expand modestly each year, so DRAM and related components may still face major price increases and prolonged shortages. CPUs are becoming a major bottleneck because reinforcement-learning environments and deployed AI applications depend heavily on CPU infrastructure. The economic value created by tokens is poorly measured by GDP, so current macro statistics understate the scale of AI’s impact. As frontier models become scarce and more tightly controlled, access may concentrate among the best-connected and best-capitalized users. Robotics could become a second major wave of demand once software capabilities spill into the physical world and sample-efficient robot learning improves. Public fear of AI may grow if labs continue emphasizing capability progress without clearly communicating practical benefits.

Data Points: SemiAnalysis AI spend last year: tens of thousands of dollars - Patel says the firm’s AI usage was modest in the prior year before exploding in 2025. SemiAnalysis AI spend run rate: $7 million per year - Current annualized Claude Code spend at SemiAnalysis. SemiAnalysis salary expense: about $25 million - Used to compare AI spend as a share of payroll. AI spend as % of salary expense: north of 25% - Patel says AI spend is already more than a quarter of salary expense. Anthropic revenue: from $9B to $35B-$45B ARR (claimed estimate range) - Patel uses this to argue that demand is outpacing compute growth. Anthropic gross margins: at least 72% - Estimated floor based on revenue growth versus compute growth. Earlier leaked gross margins: 30-something percent - Referenced as an older leaked estimate from funding documents. BLS task benchmark: 2,000 tasks - Malcolm’s benchmark for determining which bureaucratic tasks can be done by AI. Tasks doable by AI now: about 3% - From the BLS task evaluation project described in the transcript. Energy data services market: about $900 million - Used to illustrate the size of the addressable market for SemiAnalysis in energy intelligence. TSMC CapEx: $56B to $57.4B - Patel notes TSMC has already raised capex estimates and may increase them further. Possible future TSMC CapEx: $100B in 2028 - Patel says this is a real possibility and would cascade through suppliers. Memory capacity growth: roughly 20%-30% per year - He says DRAM/NAND supply cannot grow much faster than this. GPU cluster useful life: 3-4 year renewals, possibly 7-8 year useful life - He argues claims of sub-5-year life are wrong because old clusters are being re-signed. Frontier model access price sensitivity: 5-10x token cost for selective cyber release - He references a selective model release being priced much higher for specialized customers.

Pivotal Quotes: "If you don't use more tokens, you'll never escape the permanent underclass." — Dylan Patel: Patel frames token usage as central to economic mobility and competitive advantage. "What used to matter a lot was execution was very, very fucking difficult and ideas were cheap. Now ideas are cheap and plentiful, but execution is very easy." — Dylan Patel: He explains how AI changes the economics of building products and businesses. "The demand is so high. They're able to cut back on usage limits, rate limits, all these things." — Dylan Patel: He describes why frontier labs can maintain very high margins despite capacity limits.

Implications: AI adoption is moving from experimentation to operational dependency, raising spend, compressing time-to-build, and intensifying bottlenecks across hardware and infrastructure. Firms that move fastest and secure model access may gain outsized advantages, while slower firms risk commoditization and disintermediation.