“News Items is the first thing I read every morning”. — David Brooks.
1. From the Goldman Sachs China Report (via Paul Kedrosky):
China accounts for 77% of global token usage but only 22% of token spending. This is what widespread AI adoption at low prices looks like. And it is the channel through which Chinese competition will pressure prices elsewhere.
China shipped 97% of humanoid robots in the first half of 2026. Alongside China’s 75% share of EV production and 80% of battery production, China basically owns the physical side of AI and AI-adjacent technologies.
The industrial base behind the AI buildout is already concentrated in China: 60% of global transformer production capacity, 60% of PCB and optical-module production, 77% of new foundry capacity, and 90% of rare-earth refining. (Source: paulkedrosky.com)
2. Goldman Sachs:
The capital structures enabling defense, industrial AI, and data center buildout all eventually hit the same gating constraint: power.
AI training and inference are not the only sources of new load. As autonomous systems proliferate—electric and autonomous vehicles, industrial robots, and sensor-instrumented dark plants—each becomes a continuous generator of data that must be ingested, modeled, and acted on. That processing demand compounds the training and inference load already driving the buildout, and it sits largely outside projections calibrated to today’s language-model usage. Physical deployment is not only a consumer of AI; it is a fast-growing source of the compute—and the power—the next phase will require.
Today’s grid was built for a different demand profile. Utilities sized capacity around predictable residential and commercial load at roughly 60% average utilization; AI training is unpredictable and clustered in remote markets. A mature AI buildout pushes utilization toward 90%, where most utility operating assumptions cease to hold. The geographic mismatch compounds the problem—Abilene, Texas is now a market whose power demand equals that of San Francisco. Grid interconnection queues in the most important data center markets stretch eight to 12 years, two to three generations of GPU hardware. (Source: goldmansachs.com)
3. The Wall Street Journal:
