The AI Gold Rush Enters the "Mining Equipment" Stage: Whoever Controls Energy and Capacity Controls the Future

The Event

The U.S. government is converting federal subsidies into Intel equity stakes and actively promoting collaboration between Intel, Apple, and Nvidia. Meanwhile, TSMC's quarterly revenue hit a new high of 1.27 trillion Taiwan dollars, representing a 36% year-over-year increase. News coverage focuses on "capacity competition" and "how geopolitics is reshaping the semiconductor supply chain."

Why This Is More Than Industry News

In the first two years of the AI revolution (2023-2024), the market debated "whose model is smarter." Now (2025-2026), the focus has shifted:

1. Large model architectures have become transparent: open-source, papers published, rapidly replicated 2. Software-level competition is fading: computational power = intelligence capacity, marginal costs are known 3. The competitive frontier is retreating: from "algorithmic innovation" → "hardware manufacturing" → "electricity, land, capital"

This is a classic case of resource scarcity reshaping competitive dynamics. When an industry transitions from innovation-driven to maturity-driven, the technical moat of first-movers gets eroded by "resource integration capability."

Three Layers of Resource Competition

1. Semiconductor Manufacturing Capacity

• TSMC: monopolizes >90% of global advanced process capacity; quarterly revenue at all-time high
• Intel: accumulated losses of $10.4 billion over four quarters; yet the U.S. government is "forcefully injecting" equity subsidies and demanding it catch up in advanced packaging
• The logic: America has accepted that catching up in process technology is hopeless, but advanced packaging can narrow the gap in the short term

The implicit observation here is: Once the technological roadmap becomes clear, the U.S. government abandoned the fantasy of making Intel "a second TSMC," instead aiming to make Intel "domestically manufactured"—lower quality, politics first.

2. Electricity and Infrastructure

The news about "data centers competing for land and water/electricity" is now common knowledge by 2026:

• A single hyperscale data center requires 100-500 MW of continuous power supply
• OpenAI, Google, and Meta are signing long-term electricity contracts globally to deploy LLM inference and training
• Electricity in Taiwan, Arizona, and the Middle East has become a strategic asset

AI capability is coupled with electricity supply: no matter how sophisticated the algorithm, it cannot run without cheap, stable power.

3. Capital and Policy Support

The Trump administration converting subsidies into equity and promoting Apple-Intel collaboration is not a market choice—it's policy-driven capital allocation.

The logic chain: - Acknowledge that Intel cannot catch up to TSMC in process technology - But use government capital to forcibly maintain "U.S. semiconductor self-sufficiency" - The cost: customers are forced to use suboptimal chips and pay a premium - The benefit: decentralizing the semiconductor supply chain (reducing dependence on Taiwan)

Historical Analogy

This mirrors the Cold War logic of Soviet-American military industry:

• In the 1950s-1980s, the Soviet Union could not catch up to the U.S. in chip manufacturing
• But the Soviets used administrative directives and capital to forcibly maintain an "independent military-industrial electronics sector"
• Result: Soviet weapons systems had overall inferior performance, but solved the fear of "being strangled by America"

Trump's policy toward Intel is essentially replicating Cold War logic—not seeking the optimal, but seeking autonomy.

Implications for Industry

What About TSMC?

TSMC's moat is gradually shifting from "only we can manufacture the most advanced chips" to "we can manufacture sufficiently advanced chips at the lowest cost."

Political risks are rising: - The Taiwan government may be forced to restrict TSMC exports to China - The U.S. government may demand TSMC build factories in America (already operating Arizona and New Mexico facilities) - Long term, TSMC becomes a "U.S.-Taiwan co-managed asset" rather than a "Taiwanese company"

Opportunity in Open-Source Chip Design (RISC-V)

When manufacturing capacity is divided by geopolitics, open-source chip design becomes more attractive. If different countries each manufacture their own designs, they need not depend on a single foundry.

But this requires 5-10 years to form a viable alternative.

Electricity and Cooling Costs

The marginal cost of AI training is no longer "computing costs," but "electricity costs + cooling costs."

Future competition will center on: - Who can build data centers near nuclear power plants - Who can sign 20-year electricity contracts with governments - Whose cooling technology is most advanced

Applicable Principles

1. Mining industry evolution: from "prospecting (innovation)" → "mining (scaling)" → "competing for mining rights and transportation routes (resource wars)" 2. Institutional economics: when technology becomes replicable, the role of institutions and policy rises 3. Geopolitical economics: when an industry becomes a strategic asset, markets fail and politics replaces markets

---

Unanswered Questions

• Can Intel, with government subsidies and customer support, really achieve acceptable manufacturing yields within 5 years?
• When electricity becomes a bottleneck, will "regional AI hubs" become locked in place?
• Can open-source chip design pose a real threat to TSMC's dominance within 10 years?