Hey friend. It's Tuesday, December 30, 2025
--- THE INDUSTRIALIZATION OF INTELLIGENCE ---
OpenAI faces a capex projection that matches the GDP of major nations.
China and Korea are rapidly shifting the global compute balance with new sovereign silicon.
Autonomous agents now recursively improve their own code and handle complex legal proceedings.
Market Pulse: Market Watch: [TSM]
TSMC uses its monopoly on advanced nodes to pass rising costs to AI giants through 2029. While these hikes pressure lab margins, they reveal an inelastic demand curve for frontier compute. Supply remains the primary constraint. We remain long-term bullish on semiconductor equipment providers because capacity expansion is the only priority.
Sentiment: Bullish
Retrieval Augmented Generation is hitting its performance ceiling; there's a need for neural file systems.
The Deep Dive
Citigroup projects OpenAI could spend $700B on infrastructure to maintain its lead. This scale is massive. It dwarfs traditional tech cycles.
Morgan Stanley identifies a 44GW power gap in US data centers. Energy is the new silicon. The Tactical Edge: Investors should pivot from model-layer speculation to picks and shovels like grid modernization and liquid cooling providers.
The Alpha: We are moving from the 'Model Era' to the 'Infrastructure Era' where the winner is determined by energy procurement and land rights rather than just algorithmic efficiency.
Rapid Fire
🏎️ The Compute Arms Race
Lisuan G100: This hardware provides a viable retail alternative to NVIDIA in China.
SK Telecom: The firm launched A.X K1, a 500B parameter model for Korea.
Softbank: The company acquired DigitalBridge for $4B to secure data center real estate.
Meta: The social giant acquired Manus AI for over $1B to integrate action-oriented AI.
Coinbase: Developers reduced agent build times from 12 weeks to 1 week using LangSmith.
Research Corner
SmartSnap: This paper introduces proactive self-verification where agents prove task completion via curated snapshots.
Use Case: Deploy this in high-stakes GUI automation to ensure agents do not hallucinate successful form submissions.
End-to-End Test-Time Training: Researchers achieved constant inference latency regardless of context length using a sliding-window Transformer.
Use Case: Use this for real-time analysis of massive legal or medical document sets without the typical latency penalty.
ManualVLA: This model allows robots to generate their own visual and text step-by-step guides for complex tasks.
Use Case: Scale warehouse automation by letting robots teach themselves new SKU handling procedures.
Generative DNA: AI designed synthetic regulatory DNA sequences for precision gene control.
Use Case: Accelerate drug discovery by using AI to design targeted gene therapies with fewer off-target effects.
Podcast of the Day
AI Engineer Podcast | Episode: Where does consumer AI stand at the end of 2025?
Brand loyalty is the primary moat protecting incumbents like Gemini from specialized startups.
Opinionated, high-value workflows outperform general-purpose assistants in user retention.
The 2026 market will reward builders who prioritize specialized design over raw distribution.
Have a tip? Send it our way.
Cheers, Teng Yan. See you tomorrow.