@Asianometry & Dylan Patel — How the semiconductor industry actually works

Semiconductor Supply Chain & Geopolitics

• China's Centralization Potential: If China centralizes compute resources, it could train a larger model than any Western lab by next year, despite current sanctions, due to their superior ability to build 10-gigawatt data centers quickly (e.g., near the Three Gorges Dam).
• Export Control Efficacy: Sanctions are currently "flipped" in effect; China can manufacture domestic chips (SMIC 7nm/5nm) that outperform the restricted chips (NVIDIA H20) allowed for export to them.
• Espionage vs. Engineering: China's semiconductor progress is attributed to a mix of espionage (hacking ASML, poaching TSMC talent) and a unique cultural "struggle" mentality that drives extreme work ethic (e.g., Huawei's 996 culture).
• Talent Poaching History: Former TSMC R&D head Liang Mengsong was described as a "nut" who does not care about people but cares only about taking business to the limit; he recruited a "conga line" of talent to Samsung to force them to the leading edge, eventually moving to SMIC.
• SMIC Yield Challenges: SMIC produces approximately 50–80 good chips per wafer at 7nm due to bad yields caused by the lack of EUV lithography tools, though they have ~60,000 wafers/month capacity if not constrained by location.
• Knowledge Silos: Semiconductor knowledge is highly stratified and siloed; no single entity knows the entire stack, with critical "apprentice-master" intuition lost in the US but preserved in Taiwan via university networks (NTU, NTHU).

AI Scale, Costs, and Future Trajectory

• Capital Requirements: Building the scale of clusters planned for OpenAI next year requires raising $50–100 billion, which the speaker believes OpenAI will secure through partnerships with sovereign wealth funds (Saudi Arabia, UAE) and tech giants.
• Compute Efficiency Gains: Even if process nodes stopped shrinking, architectural improvements (optimizing data movement, memory integration) could still yield 100x gains in compute efficiency.
• Architectural Divergence: AI architectures are expected to diverge between the US and China due to hardware constraints; Chinese models may prioritize memory-centric or sparsity-focused designs (e.g., State Space Models) to overcome limited FLOPs per chip.
• Future Compute Scale: By 2028–2029, the industry could see clusters delivering $10^{30}$ FLOPs (across pre-training, synthetic data, and inference), representing a 100,000x increase over current capabilities.
• Market Dynamics: The AI hardware market is currently a buyer's market for short-term GPU leases (prices dropping from ~$3.00 to ~$1.70/hour for H100s), but demand for large-scale clusters (32k–100k GPUs) remains unmet.

Industry Infrastructure & Bottlenecks

• Power & Infrastructure: The US faces a "non-existent" power supply chain compared to China; data center bottlenecks are shifting from chips to transformers, substations, and grid capacity, with US power generation growing negligibly while demand triples.
• Data Center Locations: New massive data centers (gigawatt-scale) are being planned in Malaysia, the Middle East (UAE/Kuwait), and the US, with OpenAI and Microsoft already securing deals with CoreWeave and Crusoe to utilize existing crypto mining sites.
• Moore's Law Viability: Two-nanometer nodes are economically unjustifiable without the immense demand from AI; the AI sector is effectively subsidizing the cost of next-generation nodes for the entire industry.
• Cluster Efficiency Loss: Connecting multi-site clusters introduces a 20–50% efficiency loss compared to single-site clusters, making single-site gigawatt centers (like the planned 2026 Microsoft site) highly desirable.

Founders' Backgrounds & Perspectives

• Dylan Patel's Path: Transitioned from a "shit poster" on gaming and stock forums to running SemiAnalysis by attending 40+ technical conferences annually, networking directly with industry experts, and publishing deep-dive technical/economic reports.
• John (Asianometry): Started as a travel/hobbyist vlogger before pivoting to semiconductor and geopolitical analysis; works with an intense "Huawei-like" work ethic, often producing two videos a week alongside a former textile business.
• Investment Thesis: The current AI boom is viewed as a "delusional" bubble akin to the Dot-com era but potentially larger; major tech CEOs are making a "Pascal's Wager" that under-investing is riskier than over-investing, driving massive CapEx despite lagging revenue.

Risks & Future Scenarios

• Taiwan Contingency: An invasion or earthquake in Taiwan would cause a catastrophic "tech reset," halting production of everything from cars (2,000+ chips each) to consumer electronics, as Taiwan dominates all process nodes.
• Revenue Lag: Revenue generation will lag significantly behind the current multi-billion dollar CapEx spend; companies are betting that GPT-5 (or similar future models) will be transformative enough to justify the current debt-financed build-out.
• Innovation Catalyst: The crisis of AI scaling is expected to revitalise traditional industrial sectors (power, manufacturing) faster than anticipated, as the need for massive compute drives innovation in grid infrastructure and chip packaging.