Dylan Patel: Supply & Demand of AI Tokens
Category: Pricing & Economics · Duration: 46 min · ▶ Watch
Speakers: Dylan Patel · Patrick O’Shaughnessy
Segments (12)
- 00:00:41 · The Explosion in AI Token Spend
- Dylan Patel describes how his firm’s spending on AI tokens skyrocketed from tens of thousands of dollars to a $7 million annual run rate, driven by non-technical employees using AI for coding.
- 00:02:33 · Use Cases: From Chip Reverse Engineering to Macroeconomics
- Patel details two powerful examples of AI’s impact: automating chip reverse engineering, a task that previously required a full team, and enabling a single economist to build a macro analysis tool that would have taken a 200-person department a year.
- 00:04:56 · The Business Owner’s Dilemma: Spend or Be Commoditized
- The conversation shifts to the strategic necessity for businesses to invest heavily in the latest AI, as failing to do so risks rapid commoditization and being outcompeted.
- 00:10:14 · Tokenomics: The Economics of AI Compute
- An analysis of the demand for AI tokens, which is so intense that it’s driving massive revenue growth and margin expansion for model providers like Anthropic, creating a new form of arbitrage.
- 00:11:56 · The Frontier Model ‘Mythos’ and the Capability Leap
- Patel reveals the existence of a significantly more powerful, unreleased Anthropic model codenamed ‘Mythos’, highlighting the rapid, accelerating pace of AI model improvement.
- 00:15:02 · Supply Chain Bottlenecks: From GPUs to Copper Foil
- A deep dive into the supply side, where every component of the AI infrastructure stack, from GPUs and memory to niche materials like copper foil, is facing extreme shortages and long lead times.
- 00:17:17 · The Unseen Demand for CPUs and FPGAs
- Beyond GPUs, Patel explains that demand for CPUs is also exploding due to their critical role in reinforcement learning environments and in running the vast amounts of code generated by AI.
- 00:19:12 · The ‘Permanent Underclass’ and the Value of Tokens
- Patel argues that individuals and companies who fail to aggressively use AI tokens to generate outsized economic value will be left behind in a new ‘permanent underclass’.
- 00:22:40 · The Coming Wave of Robotics
- With software implementation becoming trivial, the next major breakthrough will be in robotics, with few-shot learning enabling a new wave of physical automation in the next 6-18 months.
- 00:26:10 · The AI Lab Horse Race: Anthropic vs. OpenAI
- Despite Anthropic’s current lead with its Opus models, the race is far from over, as the primary constraint is compute capacity, which OpenAI is aggressively acquiring.
- 00:29:03 · The Challenge of Measuring ‘Phantom GDP’
- The true economic value created by AI is difficult to measure with traditional metrics like GDP, as it manifests in productivity gains and deflationary effects that are not easily captured.
- 00:30:28 · Prediction: Large-Scale Protests Against AI
- Patel predicts that growing public fear and misunderstanding of AI, coupled with its increasing economic disruption, will lead to large-scale protests against AI companies within three months.
Specific Prices (14)
| Timestamp | Item | Value | Context |
|---|---|---|---|
| 00:00:51 | Firm’s AI spend last year | tens of thousands of dollars | Describing the initial, relatively low level of AI spending at his firm. |
| 00:01:28 | Firm’s AI spend run rate (initial mention) | $5 million | The annualized spending rate on AI tokens mentioned to the host a week prior. |
| 00:01:30 | Firm’s AI spend run rate (current) | $7 million | The updated, current annualized spending rate on AI tokens, up from $5M in one week. |
| 00:01:42 | Individual daily spend on Claude Code | thousands of dollars | The amount some non-technical employees are spending per day to use AI for coding. |
| 00:01:53 | Firm’s total salary expense | $25 million | The firm’s annual salary expense, used as a benchmark to compare against AI spend. |
| 00:02:55 | Cost to build chip analysis app | a couple thousand dollars of Claude tokens | The cost for one person to build a complex GPU-accelerated application for chip reverse engineering. |
| 00:06:26 | Energy data services market size | $900 million | The size of the market his firm is now able to compete in thanks to AI. |
| 00:06:48 | Individual daily spend on energy model | $6,000 a day | The daily AI token spend by one employee to build a comprehensive model of the US energy grid. |
| 00:10:20 | Anthropic ARR (initial) | $9 billion | The annualized revenue run rate of Anthropic at a recent point in time. |
| 00:10:22 | Anthropic ARR (current) | $35-40 billion | The estimated current annualized revenue run rate of Anthropic, showing massive growth. |
| 00:12:34 | Mythos token cost | 5 or 10x the token cost | The pricing for the unreleased, more powerful ‘Mythos’ model during its selective release for cyber applications. |
| 00:13:33 | DeepSeek cost vs GPT-4 | 1/600th the cost | Illustrating the rapid cost decline for achieving a certain level of AI capability. |
| 00:16:18 | TSMC 2028 Capex Prediction | $100 billion | A prediction for TSMC’s annual capital expenditures by 2028, highlighting the massive scale of investment required. |
| 00:17:25 | FPGAs per AI rack | 120 | The number of FPGAs required in a single next-generation AI server rack. |
Predictions (4)
- [00:23:45, 6-18 months] There will be real breakthroughs in robotics enabling few-shot learning.
- [00:30:28, 3 months] There will be large-scale protests against AI companies like Anthropic and OpenAI.
- [00:16:31, By 2028] TSMC will spend $100 billion on capex in a single year.
- [00:14:58, Unspecified] DRAM prices will double or triple again from current levels.
Companies Mentioned (28)
Vanta · Ridgeline · WorkOS · Rogo · Ramp · OpenAI · Anthropic · CoreWeave · Intel · FRED (Federal Reserve Economic Data) · Shopify · Stripe · Cursor · Perplexity · Vercel · Jane Street · Citadel · DeepSeek · TSMC · Lam Research · Applied Materials · ASML · MKS Instruments · Siemens · Nvidia · Oracle · SoftBank · Amazon (AWS)
Notable Quotes (6)
If this trajectory continues, then you know, we’ll spend more than 100% by the end of the year. — Dylan Patel @ 00:01:55
If this person can do the work of 5 to 10 to 15 people using Claude Code, then all of a sudden, I should probably cut people. — Dylan Patel @ 00:02:21
Dude, this would have taken the team of 200 economists a year. — Dylan Patel @ 00:04:52
If I don’t move up the bar, then I will be commoditized. — Dylan Patel @ 00:05:48
If you don’t use more tokens, you’ll never escape the permanent underclass. — Dylan Patel @ 00:19:12
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. So really only the good ideas are the ones that can justify the spend on super cheap implementation. — Dylan Patel @ 00:18:56
Key Topics
AI Token Economics · AI Compute Cost · GPU Supply Chain · Semiconductor Capex · AI Model Pricing · Enterprise AI Adoption · AI Return on Investment (ROI) · AI-driven Productivity · Model Training vs. Inference · Hardware Bottlenecks · AI Commoditization · Scaling Laws · Robotics and AI · CPU Demand in AI
Takeaways
- The demand for AI tokens is exploding, with some firms’ spending increasing by orders of magnitude in months, now representing a significant percentage of their total operational costs.
- AI is dramatically lowering the cost and difficulty of execution, shifting the primary source of value from the ability to build to the ability to conceive of high-value ideas for AI to implement.
- Non-technical employees are becoming major consumers of AI compute, using tools like Claude Code to build applications that previously required entire teams of specialized engineers, leading to massive productivity gains.
- The entire AI infrastructure supply chain is severely constrained, from leading-edge GPUs and memory to niche components like copper foil, leading to skyrocketing prices, extended hardware lifecycles, and expanding margins for suppliers.
- For businesses, aggressively adopting and spending on the latest AI is becoming an existential necessity; failing to do so risks rapid commoditization by more nimble competitors.
- The immense value generated by frontier AI models is growing faster than the infrastructure can be built to serve it, creating a feedback loop of increasing demand, compute shortages, and rising costs.
- The next wave of AI-driven disruption will be in the physical world, as breakthroughs in software and model efficiency will soon unlock few-shot learning for robotics, leading to an explosion in physical automation.
- The intense, concentrated demand for frontier AI models is creating a new form of economic stratification, where access to the best models (and the capital to pay for them) determines who can generate outsized value, potentially creating a ‘permanent underclass’ of those left behind.