Institution SemiAnalysis Analyst Dylan Patel In-Depth Interview: Outlook on the AI Computing Power Industry Chain

Article: US Stock Investment Network

1. How has the popularization of AI fundamentally changed business logic?

Past = Ideas were very cheap + Implementation was very difficult.

Now = Ideas are everywhere and cheap + Implementation is extremely easy (through AI).

This means → Only truly high-quality ideas are worth investing computational power to execute. To put it simply, execution capability is no longer a moat; capital and teams must shift focus to “how to choose the right ideas” and “how to sell the results produced by AI.”

2. What is the driving force behind companies frantically purchasing AI Tokens? What are the consequences of not keeping up?

Core driving force = Extreme leverage of efficiency. For example, one person spending a few thousand dollars on Tokens can complete what used to take a hundred-person team a year in a few weeks (such as chip reverse engineering analysis, nationwide power grid modeling).

Result = Creation of “Phantom GDP,” where actual output increases significantly while costs plummet, leading to distortion in traditional GDP statistics.

If you don’t keep up → You will inevitably face dimensionality reduction. If you don’t spend more Tokens to create and capture excess value, you will become “the perpetual underlayer of the AI era,” quickly commoditized and eliminated by faster-moving competitors.

3. Where exactly is the bottleneck in current AI computing power supply?

On the surface = Nvidia GPUs are in high demand, and the lifespan of older cards has been greatly extended (from 5 years to 7-8 years), boosting cloud providers’ gross margins.

Deeper bottleneck 1 → Memory (DRAM): Capacity expansion is extremely slow, with new supply not expected until 2028, meaning memory prices could double or triple again.

Deeper bottleneck 2 → CPU: Reinforcement learning environments’ scoring mechanisms and the large amount of AI-generated code require massive CPU resources, leading to CPU shortages.

Deeper bottleneck 3 → TSMC and edge materials: TSMC’s capital expenditure may soar to $100 billion by 2028, and the entire supply chain for copper foil, PCB glass fiber, and other peripheral materials is already at full capacity, with the industry competing for “sky-high prepayments” to secure production.

4. What trends are emerging in the competition landscape and Token economics at the large model layer?

Current situation = Anthropic is temporarily leading with Opus 4.7 and its internal “Mythos” model, even controlling release pace to manage risk, with very high profit margins (>72%).

Computing power race = Anthropic is limited by total compute capacity, while OpenAI is attempting to leapfrog by massive fundraising and accumulating compute power (in partnership with Microsoft, Oracle, etc.).

Key conclusion → Token demand far exceeds infrastructure capacity. Even second- and third-tier large model developers will face “sold-out” issues due to insufficient top-tier compute power. Essentially, as long as you can produce high-quality Tokens, the market can absorb them entirely.

5. Why is there a prediction of large-scale anti-AI protests in the short term?

Cause = The massive business restructuring brought by AI will cause fear among ordinary people, who tend to blame long-standing social issues on AI.

Catalyst = Poor public communication strategies by AI giants (such as Sam Altman and Dario) + frequent sensational narratives about “AI changing the world/replacing jobs,” which heighten public anxiety.

Advice from US Stock Investment Network → The industry must stop exaggerating AI’s terrifying future capabilities and instead focus on demonstrating its positive current contributions. Otherwise, public anger could be weaponized by politicians or influencers, leading to large-scale resistance.

Dylan Patel Background:

Chief analyst at SemiAnalysis, a research organization (Dylan Patel), deeply dissects GPUs (especially Nvidia).

Analyzes AI compute supply and demand (who is short of chips, who is stockpiling).

Tracks industry chain (ASML equipment → TSMC → cloud providers → OpenAI, etc.).