TR41: Mandate of Heaven (Nov 2025 Edition)

Deep dive into re-ranking AI labs after Gemini 3 Pro and Opus 4.5 releases. The team debates moving Google DeepMind to S-tier and downgrading OpenAI to A-tier with negative outlook due to lack of recent pre-training runs and competitive pressure.

• •Gemini 3 Pro and Opus 4.5 are outperforming OpenAI's models on benchmarks, raising questions about OpenAI's leadership
• •OpenAI hasn't done a major pre-training run in months while competitors have shipped new models
• •Google's challenge is product execution and distribution, not model quality - 'killed by google.com' problem
• •Meta's Llama dropped from InferenceMax repository for being irrelevant, Chinese models taking over open source
• •OpenAI's competitive response could take 6+ months due to training run timelines

Analysis of how different AI labs allocate compute resources between serving users and training new models. Discussion of OpenAI's massive compute commitments and whether infrastructure delays could actually help their cash flow situation.

• •OpenAI serves nearly 1 billion users, creating massive compute demands that limit training capacity
• •Google and Anthropic have fewer serving demands, allowing more compute for pre-training
• •OpenAI committed to $1.4 trillion in compute purchases, but delays may prevent bankruptcy by pushing costs to 2035
• •Users switching to Gemini while still paying for ChatGPT improves OpenAI's gross margins
• •Data center build-outs are significantly delayed, reducing near-term capital pressure

Evaluation of Chinese AI labs including Moonshot (Kimi), GLM, and MiniMax. Discussion of their technical capabilities, particularly in creative writing, and the challenge of monetization in China's market.

• •Kimi (Moonshot) produces superior English writing and lyrics compared to ChatGPT, Claude, and Gemini
• •Chinese labs likely to go public before OpenAI, providing first look at AI company financials
• •China's market is 'allergic to paying for software' - cultural resistance to SaaS pricing
• •Alibaba and Tencent CapEx dropping after H20 stockpiling, unclear future investment plans
• •DeepSeek downgraded to 'solidly mid' tier as talent has been poached by other companies

Dylan recounts presenting Semi Analysis's AI infrastructure forecasts to TSMC executives, including explaining free monetization models and the need for 65 gigawatts of capacity. Highlights the security measures and organizational silos within TSMC.

• •Presented to multiple TSMC teams for 7 hours, including one-on-one with exec reporting to CEO CC Wei
• •TSMC has extreme security: phones confiscated, laptop ports taped over, no external devices allowed
• •Different TSMC teams don't share information - had to repeat same presentation 8 times to different groups
• •TSMC exec referenced Sam Altman claiming 'AGI is next year' but found Semi Analysis forecast more credible
• •Presented need for 65 gigawatts just to match current OpenAI capabilities across industry

Discussion of AI company business models, financial disclosure issues, and comparisons to historical railroad financing. Debate over OpenAI's revenue accuracy and the lack of audited financials for private AI companies.

• •ChatGPT launching shopping features and ad tier to diversify revenue beyond subscriptions
• •Source claims OpenAI may be overstating revenue by 2x, though cash flow makes this questionable
• •Private investors often don't get information rights - relying on 'vibes' rather than audited financials
• •GPU useful life debates miss the point - companies can collateralize and refinance infrastructure
• •20% of UK GDP at peak was railroad construction, similar scale to current AI infrastructure build

Detailed comparison between 19th-century railroad speculation and current AI bubble. Covers how Moody's, modern banking, annual reports, and insider trading regulations were all created to support railroad financing.

• •Moody's was invented specifically to categorize railroad securities for investors
• •J.P. Morgan financed the Civil War through railroad bonds, then bought collapsed railroads post-crash
• •Railroads owned ~15-20% of US land through government grants, selling it to finance construction
• •Railroad executives owned construction companies, paying themselves 200% annual dividends while issuing bonds
• •Over half of railroad funding came from European investors - similar to Chinese internet bubble dynamics
• •Annual reports were invented for railroads but were mostly fraudulent, leading to SEC creation
• •Insider trading was legal - congressmen received railroad shares in exchange for granting charters

Analysis of OpenAI's CFO Sarah Friar's background at Square/Block and Nextdoor. Discussion of the 'glass ceiling' phenomenon where women executives are brought in to lead struggling companies.

• •Sarah Friar was 'the only grown-up in the room' at Square, managing Jack Dorsey's chaotic leadership
• •Nextdoor stock dropped 90% under her CEO tenure before joining OpenAI
• •Pattern of CFOs taking CEO shots at companies in crisis, often failing
• •Lisa Su at AMD cited as successful example of turnaround leadership when company was near bankruptcy
• •Friar's temperament may be good match for Sam Altman's 'crazy founder' style

Discussion of Ari Emanuel's proposed America vs China robot UFC battle. Analysis of whether US or Chinese robotics would win based on compute hardware, AI models, and teleoperation capabilities.

• •Chinese UniTree robots use almost no American chips, limiting compute for AI models
• •American robots with NVIDIA Orin chips could run superior models with more compute
• •All robot demos are teleoperated - real competition might be best human operators
• •Outcome depends on whether AGI is achieved before competition - 'if bubble pops before AGI, China wins'
• •Need to 'keep the bubble growing' to fund continued AI development and maintain competitiveness