DeepSeek's latest AI model, R1, is making waves-not just for its performance but for the existential questions it raises about AI chip demand.

The company claims its training costs were a mere $5.6 million, a fraction of what frontier foundation models demand.

Naturally, investors are wondering: if AI can be trained this efficiently, does that mean the industry's chip-buying frenzy is about to cool off?

Jevons Paradox Strikes Again

JPMorgan's Harlan Sur isn't hitting the panic button. Instead, he points to history-where efficiency gains in computing have paradoxically driven more demand, not less.

From x86 virtualization in the 2000s to ARM Holdings PLC's (ARM  ) dominance in mobile, every step forward in computing power has led to a proliferation of use cases that need more chips, not fewer. The same could happen here: DeepSeek's efficiency might not curb AI chip demand but rather accelerate the adoption of AI, pulling forward the need for high-performance semiconductors.

Custom Silicon Could Be the Winner

DeepSeek's low-cost efficiency doesn't just raise questions-it also opens opportunities. Sur believes that hyperscalers and cloud providers will keep pushing for greater AI capabilities, but they won't just rely on off-the-shelf GPUs. Custom-built ASICs-where companies like Broadcom Inc (AVGO  ) and Marvell Technology Inc (MRVL  ) thrive-could see an uptick as cost and power performance become critical differentiators.

Despite lingering uncertainties around DeepSeek's exact cost structure and reliance on open-source models, one thing remains clear: AI innovation never slows down-it only fuels further breakthroughs.

Sur reiterates his bullish stance on Broadcom, Marvell, and Nvidia Corp (NVDA  ), emphasizing that the race for AI dominance is far from over. If history is any guide, this is just the beginning of the next wave of semiconductor demand.