Dylan Patel is the founder and CEO of Semi Analysis, where he tracks the semiconductor supply chain and AI infrastructure buildout. Patrick O'Shaughnessy is the CEO of Positive Sum and host of Invest Like the Best.

The conversation explores the explosive growth in token demand and supply constraints across the AI infrastructure stack. Dylan shares how his firm's AI spending skyrocketed from tens of thousands to $7 million annually, representing 25% of their salary costs.

They discuss Anthropic's Mythos model and its restricted release, the bottlenecks in memory and semiconductor manufacturing, and the concentration of AI capabilities among fewer companies with enterprise access.

The discussion covers supply chain constraints from memory to fab equipment, the role of CPUs in reinforcement learning, and predictions about public backlash against AI companies in the coming months.

Semi Analysis Token Spending Explosion

Semi Analysis went from tens of thousands in AI spend last year to a $7 million annual run rate, representing over 25% of their $25 million salary expense.

"If this trajectory continues, then we'll spend more than 100% by the end of the year, which is a bit terrifying" - Dylan on their accelerating AI costs.

Non-technical employees are now spending thousands of dollars daily on Claude, with one person building a GPU-accelerated chip analysis application for a few thousand dollars that previously required an entire Intel team.

Malcolm, an economist, single-handedly built economic analysis tools and AI benchmarks that would have taken 200 economists a year to complete at his previous bank.

Anthropic's Revenue Surge and Margin Expansion

Anthropic's revenue exploded from $9 billion to $40+ billion ARR while compute didn't grow proportionally, resulting in gross margins of at least 72%.

"Whoever can pay for them, Anthropic has the same problem... those tokens are generating way more than $40 billion in value" - Dylan on token demand exceeding supply.

Enterprise contracts and rate limit increases have become more valuable than subscription pricing as token scarcity drives premium access.

Mythos Model Capabilities and Restricted Access

Mythos represents potentially the biggest capability jump in two years, achieving L6 engineer performance compared to the L4 target for Opus 4.6.

"We knew it was supposed to be really good... Mythos is potentially the biggest step up in model capabilities two years" - Dylan on the model's breakthrough performance.

Anthropic only released Mythos to select companies for cybersecurity use, despite it being available internally since February, demonstrating increasing model access restrictions.

Model release cadence has compressed from six months to two months as implementation becomes easier while ideas remain cheap.

Memory and Semiconductor Supply Bottlenecks

DRAM prices will double or triple again as memory capacity can only grow 20-30% annually while demand explodes exponentially.

"Even if they wanted to as fast as possible, it doesn't come till 28, late 27 at best" - Dylan on when incremental memory capacity will arrive.

TSMC may spend $100 billion on CapEx by 2028, up from $57 billion currently, creating massive downstream supply chain bottlenecks.

H100 prices are skyrocketing and useful life is extending to 7-8 years instead of the predicted 5 years as demand outstrips supply.

CPU Demand and Reinforcement Learning Infrastructure

CPUs are completely sold out due to reinforcement learning environments and deployment of AI-generated code requiring massive CPU resources.

Reinforcement learning requires CPUs to run complex environments that score model outputs, from simple text validation to complex physics simulations.

Next-generation AI racks require 120 FPGAs each, creating bottlenecks beyond just GPU availability across the entire infrastructure stack.

Token Economics and Value Measurement Challenges

"The hardest area for us and for everyone is understanding tokenomics, the economics of tokens" - Dylan on the difficulty of modeling AI adoption.

Traditional GDP metrics fail to capture the economic value created by AI tokens, creating a "phantom GDP" that's difficult to quantify.

Token costs for equivalent capabilities fall 100x over time, but demand shifts to frontier models that enable entirely new use cases.

Predictions for AI Industry Backlash

"AI is less popular than ICE, less popular than politicians" with large-scale protests against Anthropic and OpenAI expected within three months.

"Sam Altman and Dario have to stop getting on interviews. They're so uncharismatic" - Dylan on leadership communication failures.

AI companies need to stop discussing future capabilities and focus on present uplifting applications to counter growing public resentment.

Resources Mentioned

Compute!'s Second Book of Amiga

grown to the same degree. And if you do the calculations and you assume they didn't decrease their research and development compute, they clearly didn't. They have Mythos. They have Opus 4.7. So the

Cost-Effective Data Pipelines Balancing Trade-Offs When Developing Pipelines in the Cloud

, they clearly didn't. They have Mythos. They have Opus 4.7. So they clearly didn't decrease their research compute spend. So ultimately, what they've done, even if you assume all incremental compute

Human Development A Life-Span View (MindTap Course List)

s are at a floor of 72%. In reality, some of that incremental compute they've got probably went to research and development. It may be higher than 72% gross margins. To be clear, at the start of the

compute that the labs are spending is actually turning into

etter. Along the whole way, we're also getting these compute efficiency wins, which are as all this research compute that the labs are spending is actually turning into, if I want X capability tier mo

or Applied Materials or ASML or their further downstream supply chains like MKSI and all these other companies

can't fathom that. But what does that mean for their downstream supply chains? Companies like Lamb Research or Applied Materials or ASML or their further downstream supply chains like MKSI and all th