The recent Dwarkesh pod with Jeff Dean touched on the need for multi-data center training runs as hyperscalers run into energy bottlenecks

According to SemiAnalysis, Google and OpenAI are likely already constructing campuses in concentrated areas for distributed training runs (1/6)
The centralized data center paradigm isn't stopping any time soon, but it is running into constraints:
- Power infra: transformers, generators, substations all have long lead times
- Grid constraints: upgrading transmission infra and grid interconnection has been very slow in the US, especially for clean energy sources like nuclear
- Power density needs: centers need massive amounts of specialized cooling equipment and new power supply units; some older centers need to be retrofitted
- Sustainability pressures: companies like Meta with net-zero emission goals have an additional constraint on how and where they will build large data centers
- Supply chain issues: every piece of the buildout (GPUs, interconnects, cooling, power, etc) can be subject to supply chain disruptions (2/6)

1) The Fat Protocol Thesis Reborn
The market senses the immense potential for agents to be both better consumer interfaces and autonomous economic actors. As L1 blockspace had difficult-to-value, but infinite potential last cycle, so does agent infra this cycle.
Lets dig in: 2) Agent adoption IRL is sky-rocketing
- The percentage of LLM-app developers on LangSmith using agents jumped from 7 to 43% this year
- Agentic workflows are becoming more complex (more steps and tool-calling)
- Automation and simplification are primary goals for enterprises: this will likely manifest in customer service, IT, marketing, and software dev. workflows

Tether's record-breaking profitability last quarter puts it in the ranks of tradfi giants. But that $5.2 billion number also puts it in the crosshairs of new competitors who want a piece of the pie (1/9): My latest piece dives into the rapidly evolving world of stablecoins - in both the centralized and decentralized space. We break down the vertical into several major categories (2/9):