The Shovel Sellers Are Winning

In the world of AI, money is returning up the supply chain. Samsung Electronics just posted its first-ever quarterly operating profit of 107.4 trillion won, a staggering 782% increase year-over-year. This record profit has analysts questioning whether the market's response was too timid given the company's impressive performance. Meanwhile, OpenAI's annualized revenue has been revised down from $70 billion to $50 billion, a $20 billion gap that highlights the growing demands of the AI trade.

The source of Samsung's success lies in high-bandwidth memory (HBM), the most sought-after component in AI servers. Samsung and SK Hynix dominate this market, with every NVIDIA GPU requiring a cut from these two companies. This shift in the market demonstrates that memory has transitioned from a cyclical commodity to a strategic resource, as buyers now sign multi-year supply agreements due to concerns over potential shortages.

Nvidia has also stepped up, announcing $1 billion over five years for U.S. superintelligence and quantum research. A chip company that built AI inference chips, Etched, has seen its valuation rise from $21 billion to $40 to $50 billion. This has been achieved by focusing on inference, the step where a model generates a response after receiving a prompt, delivering more tokens for the same dollar compared to competitors.

This marks a shift in AI chip competition from training to inference, with the focus shifting to driving down the cost of inference to gain pricing power.

As the demand for AI expands, the cost of electricity becomes a significant factor. The U.S. government is drafting a law requiring data centers to report their power use, and states like New York and New Jersey have already begun taking notice. The power industry is rapidly becoming a key component in the AI ecosystem, with policymakers recognizing the need to address electricity scarcity as the physical ceiling of AI expansion.