Invest Like the Best
The Supply and Demand of AI Tokens | Dylan Patel Interview
Episode Summary
AI-generated · Apr 2026AI-generated summary — may contain inaccuracies. Not a substitute for the full episode or professional advice.
Dylan Patel, Chief Analyst at SemiAnalysis, reveals a profound and rapid reordering of economic dynamics driven by the explosive demand for AI tokens and the profound ease of implementation. He shares his firm's experience, where AI token spend has skyrocketed from tens of thousands to an annualized $7 million—over 25% of their salary expense—allowing them to grow faster and perform tasks that previously required entire teams, such as reverse-engineering chips and building complex economic models, with just a few thousand dollars worth of tokens. Patel argues that while ideas were once cheap and execution difficult, now ideas are plentiful and execution is "very easy," shifting value to those with "good ideas" and the capital to leverage AI's capabilities.
The conversation delves into the insatiable demand for frontier AI models, highlighting how Anthropic's revenue has surged to an estimated $40-45 billion ARR with gross margins at a floor of 72%, despite users being rate-limited due to compute constraints. Patel notes the visceral user demand for the latest models like Anthropic's Mythos, which he describes as an unprecedented leap in capability (from an L4 to an L6 engineer in two months), even at 5-10x the token cost. This creates a scenario where access to the best models and the ability to leverage them for high-value tasks becomes critical, potentially leading to a concentration of resources among those with capital and connections.
Patel also provides a deep dive into the constrained supply side, detailing bottlenecks across the entire AI infrastructure stack. He highlights memory (DRAM) as a critical bottleneck where true incremental supply won't arrive until 2027-2028, predicting prices will double or triple. Logic (TSMC), wafer fabrication equipment (ASML, Lam Research), and even obscure components like copper foil for PCBs are all sold out, with lead times extending for years. Furthermore, CPUs emerge as an underappreciated bottleneck, essential for reinforcement learning environments and for running the "slop code" generated by AI models in deployed applications.
Ultimately, the episode presents a future where the economic value delivered by the best AI models is growing faster than the infrastructure's ability to serve those tokens, leading to expanding margins for AI labs and an increasing scramble for access. Listeners will gain a stark understanding of AI's deflationary impact, the concept of "phantom GDP," and a challenging outlook on societal reform, with Patel even predicting "large scale protests" against AI due to public fear and misunderstanding.
👤 Who Should Listen
- Business leaders and entrepreneurs making strategic decisions about AI adoption and investment.
- Investors seeking to understand the economic impact of AI on technology supply chains and market dynamics.
- Economists, policymakers, and academics interested in the measurement of GDP and the broader societal implications of AI's rapid advancement.
- Technology and infrastructure managers facing increasing compute demand and supply chain constraints.
- Anyone concerned about the ethical considerations, power concentration, and potential social unrest driven by AI's development.
- Software engineers and developers evaluating the shift from execution-centric roles to idea- and value-centric roles in an AI-driven economy.
🔑 Key Takeaways
- 1.SemiAnalysis's AI token spend skyrocketed from tens of thousands to $7 million annually in one year, now representing over 25% of their salary expense, enabling faster growth and reduced hiring needs.
- 2.The ability to implement ideas has become 'super cheap,' shifting economic value to those with 'good ideas' and the capital to leverage frontier AI models.
- 3.AI is causing 'phantom GDP' where output goes up due to massively reduced costs, but traditional GDP metrics may paradoxically shrink because the cost of producing that output falls so dramatically.
- 4.Frontier AI models like Anthropic's Mythos represent significant capability jumps (e.g., L4 to L6 engineer in 2 months) and are driving insatiable demand despite higher per-token costs.
- 5.The supply chain for AI infrastructure—including DRAM, logic, wafer fabrication equipment, and CPUs—faces significant bottlenecks, with true incremental supply for memory not arriving until 2027-2028.
- 6.The economic value that the best AI models can deliver is growing faster than the infrastructure's ability to serve those tokens, leading to expanding margins for AI labs and continued capacity problems.
- 7.Failing to use more tokens, generate value, and capture that value from AI risks placing individuals in a 'permanent underclass' as AI capabilities skyrocket and resources concentrate.
- 8.The public's negative perception of AI is growing, with potential for 'large scale protests' against AI labs in the near future, fueled by fear and misrepresentation.
💡 Key Concepts Explained
AI Psychosis
This describes the visceral, almost addictive demand for the latest, most capable AI models. The episode highlights how users, including the guest's own firm, become immediately discontent with older versions (e.g., Claude 4.6) once a new, superior model (e.g., Opus 4.7, Mythos) is released, driving dramatic increases in token spend despite higher costs and rate limits.
Phantom GDP
Coined by an economist on the guest's team, this framework explains how AI-driven cost reductions can lead to increased economic output while paradoxically causing traditional GDP metrics to shrink. Output rises because tasks are cheaper to perform, but the value measured in GDP decreases because the cost of producing that output falls so dramatically, creating significant unquantified value.
Software Only Singularity
This concept posits a future where AI reaches a singularity in the software domain, making 'implementation super cheap' and rapidly advancing digital capabilities. However, it suggests the physical world (robotics) lags behind, with breakthroughs in 'few-shot learning' for robot models—where robots learn from limited examples—needed to bridge the gap and accelerate physical goods and services.
Permanent Underclass
This refers to the potential future societal division where individuals or businesses who do not actively use, generate value from, and capture that value from AI tokens will be left behind. As AI capabilities skyrocket and access to resources concentrates, failing to leverage AI effectively could lead to a significant economic disadvantage.
⚡ Actionable Takeaways
- →Prioritize securing access to the newest, most capable AI models, even if they are more expensive per token, as they are more efficient and enable economically valuable use cases.
- →Seek enterprise contracts with AI labs to gain better rate limits and potentially earlier access to frontier models, as general access becomes increasingly constrained.
- →Shift focus from execution skills, which are becoming commoditized by AI, to cultivating 'good ideas' and strategies for capturing value from AI-driven implementation.
- →Explore novel and complex use cases for AI beyond existing tasks, as these applications (e.g., chip reverse engineering, economic modeling) are generating the most significant value.
- →Prepare for potential societal backlash and protests against AI by focusing on uplifting present-day applications and improving public communication, rather than emphasizing future capabilities.
- →Investigate the entire AI supply chain, from memory and logic to obscure components like copper foil, for potential bottlenecks and opportunities due to skyrocketing demand.
- →Develop strategies to arbitrage AI tokens by identifying high-value tasks and deploying them effectively, as this could become a primary business model in the near future.
⏱ Timeline Breakdown
💬 Notable Quotes
“What used to matter a lot was execution was very very difficult and ideas were cheap. Now ideas are cheap and plentiful but execution is very easy.”
“If I don't adopt AI, someone else will and they will beat me.”
“Whoever you are, if you have enough capital, you should get a freaking enterprise cloud uh enterprise anthropic subscription where you pay per token... and then you must you need to figure out how to leverage those tokens to the highest value task um and make money off of it.”
“If you don't use more tokens and generate the value from them and capture that value... you'll never escape the permanent underclass.”
More from this guest
Dylan Patel
Listen to Full Episode
📬 Get weekly summaries like this one
No spam. Unsubscribe anytime. By subscribing you agree to our Privacy Policy.
Continue Exploring
