Nvidia's AI Factories Struggle with Diminishing Returns

As of August 30th, 2025, Nvidia's data center business continues to thrive, with the company's "AI factories" producing an impressive 864,000 GPUs per quarter. However, the company is grappling with the paradox of diminishing returns, as its own efficiency gains drive more demand, creating a treadmill effect.

Nvidia's chief financial officer, Colette Kress, has framed the company's scale in industrial terms, likening its production to a "steel mill" that cranks out "tokens per watt." The efficiency pitch is seductive, with Kress claiming that a $3 million investment in Nvidia hardware can yield $30 million in token revenue, a 10x return. But this same economic model is now proving to be a double-edged sword, as falling inference prices invite more usage, which drives more spend, which lowers costs again.

Adding to Nvidia's challenges is the slowdown of Moore's Law, with chips no longer doubling their performance on schedule. This means that customers are "sprinting harder just to stay in place," as CEO Jensen Huang puts it. Nvidia is responding by accelerating its product cycle, aiming to deliver more compute power without necessarily more returns per dollar.

Geopolitical tensions also loom large, with Nvidia recording zero H20 chip sales to Chinese customers in the second quarter of 2025. The company is working with the U.S. administration to secure licenses, but analysts are treating China as a "ghost market" for now.

Despite these headwinds, Nvidia remains the crown jewel of the chip trade, with Wall Street analysts maintaining a bullish outlook on the company. However, the market's reaction to Nvidia's latest earnings report suggests that the thrill of "more" is starting to sound a lot like "enough" as the AI revolution enters a new phase of maturity.