← Back to list
Colossus (Invest Like the Best / Business Breakdowns)Podcast23 Apr 2026Source: colossus.comHost: Patrick O'Shaughnessy

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

In plain words

This interview says AI token demand is exploding, and only the most advanced models are wanted, with people willing to pay almost anything. Dylan Patel thinks supply bottlenecks (like memory and chip factories) will keep widening the gap, boosting profits for model labs like Anthropic. Key holdings: Anthropic (its new model Mythos is a huge leap, revenue jumped from $9B to $35-40B); TSMC (chip maker, capex may hit $100B by 2028); DRAM makers (memory prices could double or triple due to slow capacity growth).

AI SummaryAI-generated · may contain errors · verify against the original

Dylan Patel discussed the explosive growth in the supply and demand of AI tokens during the program. On the demand side, frontier models are the only ones widely needed, with virtually no upper limit on willingness to pay. His company SemiAnalysis has seen AI spending surge from tens of thousands of

~11 min full read · 7 sections
Deep Analysis

This Issue at a Glance

Dylan Patel (Founder & CEO of SemiAnalysis, analyst of semiconductor supply chain and AI infrastructure) discusses with host Patrick O'Shaughnessy the explosive growth in AI token supply and demand. Dylan Patel's core judgment is that AI token demand has "completely exploded," frontier models are the only widely needed product, and willingness to pay is nearly unlimited. However, supply-side bottlenecks (memory, logic chips, fab equipment) will determine the pace of scaling — and this supply-demand gap will continue to widen, driving up profit margins for model labs.


Theme 1: Token Demand "Completely Exploding" — A Microcosm from SemiAnalysis's Own Experience

Dylan Patel believes that demand for AI tokens is experiencing unprecedented explosive growth, with his own company's experience serving as a microcosm.

  • Data chain: SemiAnalysis's AI spending has surged from tens of thousands of dollars last year to a run rate of $7 million this year (and is still accelerating). Its payroll is approximately $25 million, and AI spending has already exceeded 25% of payroll. Patel judges: "If this trend continues, AI spending will surpass 100% of payroll by year-end."
  • Mechanism breakdown: Patel points out that non-technical employees (such as company president Doug O'Loughlin) were the first to extensively use Claude Code for coding, which then "infected" the entire team. Token consumption per user varies wildly — some spend thousands of dollars a day, others hundreds — but the overall trend is a rocket-like ascent.
  • Specific cases: A former Intel engineer used a few thousand dollars' worth of Claude tokens to build a GPU-accelerated application for chip reverse engineering (automatically identifying the location of various materials in chips), "a task that would have required a full team." Another former bank economist single-handedly built an AI capability evaluation benchmark containing 2,000 tasks and created the "Phantom GDP" indicator — "a task that would have taken a 200-person economist team a year."
  • Extrapolation: Patel argues that if he does not actively adopt AI, he will be replaced by competitors. "AI is commoditizing everything. Those who don't move will lose."

Theme 2: Frontier Models Are the Only Ones Needed — "Mythos" Represents the Biggest Capability Leap in Two Years

Dylan Patel emphasizes that the market only wants the most advanced models, and Anthropic's "Mythos" represents the biggest capability leap in two years.

  • Historical Context: Anthropic's revenue surged from $9 billion ARR to $35–40 billion (at the time of recording), while compute resources did not grow proportionally. Patel calculates that even assuming all incremental compute is used for inference, Anthropic's gross margin is at least 72% — whereas leaked documents earlier this year showed a gross margin of just over 30%. "How can a business improve its margins so quickly?"
  • Mechanism Breakdown: Patel explains that frontier models are sought after because they are more "efficient" — while a single token is more expensive, the number of tokens required to complete a task drops significantly, making the actual cost lower. "Mythos is more expensive than Opus 4.6, but it is actually cheaper on most tasks because it is smarter."
  • Unique Insight: Patel reveals that Anthropic achieved a model with L4-level software engineer capabilities in early 2025, and two months later, Mythos had reached L6 level. "The pace of model improvement is accelerating, not slowing down." He himself "begged on his knees in front of Anthropic's co-founder for access to Mythos."
  • Contrary to Market Consensus: Patel notes that Anthropic deliberately refrained from broadly releasing Mythos, instead offering it only to select clients (e.g., top-tier banks) for cybersecurity. "They are worried about the impact on the world. They released a worse version (Opus 4.7) to us." This means models will increasingly be deployed in a concentrated manner to a small number of clients.

Theme 3: Supply Bottlenecks – From Memory to Fab Equipment, the Entire Chain Is "Sold Out"

Dylan Patel argues that supply-side bottlenecks are the key factor determining the pace of AI expansion, and these bottlenecks are more persistent than the market expects.

  • Memory Bottleneck: DRAM production capacity can only grow by 20-30% annually. Even if memory companies react immediately, incremental capacity will not arrive until late 2027 or 2028. "DRAM prices will double or triple from here." Because the only way to "steal" capacity is through demand destruction driven by higher prices.
  • Logic Chip Bottleneck: TSMC's capital expenditure for 2025 is $56-57.4 billion, but Patel predicts it could reach $100 billion by 2028. "People cannot imagine this." However, this implies a massive "bullwhip effect" for the downstream supply chain (Lamb Research, Applied Materials, ASML, etc.)—demand is amplified at each stage.
  • Other Bottlenecks: FPGAs (120 units required per next-generation AI rack), CPUs (for reinforcement learning environments and application deployment), copper foil, fiberglass, lasers, etc.—"Anything with a pulse that is sold out, people are scrambling for incremental supply."
  • Comparison Data:
Supply Side Current Status Key Data
DRAM Capacity growth constrained 20-30% annual growth, incremental capacity arrives in 2028
TSMC CapEx Continuously raised ~$57 billion in 2025, potentially $100 billion by 2028
GPU (H100) Prices rising, lifespan extended 3-4 year old clusters renewed for 3-4 years, lifespan may reach 7-8 years
Copper foil/PCB materials Completely sold out Prepayment required to secure supply
  • Extrapolation: Patel believes that because supply cannot respond quickly, profit margins for model labs will continue to widen—"until the people in the hardware supply chain realize, 'Why don't I raise my profit margins?'"

Theme 4: The Public Perception Crisis of AI — "AI Is Less Popular Than ICE"

Dylan Patel predicts large-scale anti-AI protests within the next three months, and AI industry leaders are exacerbating the problem.

  • Current Situation: Patel notes that AI's public approval rating is lower than that of ICE (Immigration and Customs Enforcement) and politicians. Sam Altman's home was firebombed twice in two weeks, and social media comments were "cheering."
  • Mechanism Breakdown: Patel argues that the public demeanor of AI leaders (Sam Altman, Dario Amodei) is "utterly charmless," fueling public fear. "Every interview they do makes ordinary people hate them more." At the same time, their constant talk about future capability shifts leaves ordinary people feeling fearful and alienated.
  • Recommendations: Patel proposes three areas for improvement: 1) Stop discussing the future and only talk about how AI improves lives today; 2) Showcase positive use cases of AI; 3) Stop "scaring people" in interviews.
  • Falsification Condition: If the AI industry can effectively reshape public narrative, the scale of protests may be smaller than expected. However, Patel believes that because ordinary people "don't know anyone at Anthropic or OpenAI," they view these companies as "a secret cabal of 5,000 people that will automate all jobs and destroy society."

Mentioned Positions

Position Guest Sentiment Key Data
Anthropic Bullish (leading position) Revenue from $9B → $35-40B ARR; gross margin at least 72%; Mythos model capability is the biggest leap in two years
OpenAI Neutral (lagging but still has advantages) Has massive computing resources (Oracle, CoreWeave, SoftBank, Microsoft, Amazon); "Even second-tier labs will sell out their tokens"
NVIDIA Bullish (hardware layer) Gross margin 75%+; H100 price rising, useful life extended to 7-8 years
TSMC Bullish (but not fully capturing value) CapEx ~$57B in 2025, potentially $100B by 2028; only single-digit price increases rather than triple-digit
DRAM Manufacturers Bullish (prices will double or triple) Capacity grows 20-30% annually, incremental capacity not arriving until 2028
ASML Bullish Completely sold out, needs Carl Zeiss to expand capacity
Lam Research / Applied Materials Bullish (downstream beneficiaries) TSMC CapEx growth will create a "bullwhip effect"
FPGA Manufacturers Bullish Each next-gen AI rack requires 120 FPGAs
CPU Manufacturers Bullish Reinforcement learning environments and deployment applications have led to CPUs being completely sold out

Judgments Worth Remembering

1. Dylan Patel: Demand for AI tokens is "completely exploding," and frontier models are the only ones needed. Evidence: SemiAnalysis’s own AI spending surged from tens of thousands of dollars to $7 million (exceeding 25% of payroll); Anthropic’s gross margin jumped from 30%+ to 72%+, proving willingness to pay is nearly unlimited.

2. Dylan Patel: Mythos represents the largest model capability leap in two years. Evidence: Anthropic advanced from an L4-level software engineer to L6-level within two months; the model is deliberately not widely released, offered only to specific clients—"they are concerned about the impact on the world."

3. Dylan Patel: DRAM prices will double or triple from here. Evidence: Capacity grows only 20-30% annually, with incremental capacity not arriving until 2028; the only way to "steal" capacity is through demand destruction via higher prices.

4. Dylan Patel: TSMC’s CapEx could reach $100 billion by 2028. Evidence: Current CapEx has been raised from $56 billion to $57.4 billion; the downstream supply chain will experience a "bullwhip effect"—demand is amplified at each stage.

5. Dylan Patel: GPU service life is severely underestimated. Evidence: 3-4 year old Hopper clusters are being renewed for another 3-4 years; A100 clusters are being renewed for several years; service life may reach 7-8 years rather than the market’s assumed 5 years.

6. Dylan Patel: Large-scale anti-AI protests will emerge in the next three months. Evidence: AI is more unpopular than ICE and politicians; a Molotov cocktail was thrown at Sam Altman’s home while social media cheered; the public behavior of AI leaders has intensified public fear.

7. Dylan Patel: If you do not use more tokens and create value from them, you will never escape the "permanent underclass." Evidence: There are three distinct problems—using more tokens, creating value from tokens, and capturing value from the created value; "AI is still not table stakes, so act now."

8. Dylan Patel: Model lab profit margins will continue to expand until the hardware supply chain realizes it can raise prices. Evidence: The economic value of frontier models is growing faster than the supply capacity of infrastructure; the supply-demand gap continues to widen, giving labs pricing power.