At today’s close, shares in U.S. chipmaker Nvidia Corp (NVDA) are off just under 17% after Chinese generative artificial intelligence (genAI) startup DeepSeek released an update to its large language model (or LLM, in shorthand the technology that provides the underpinnings of genAI capabilities). The update purports to radically reduce the cost of building and improving LLMs (“training”), while also slashing the cost to use the model (“inferencing”). Figures we read suggest a more than 90% reduction in both types of compute cost. The firm also published a paper explaining how its approach works and released the model under an open-source license, meaning most anyone can take advantage of its capabilities for free.
In fact, much of the broader genAI momentum trade went into reverse today, as DeepSeek’s achievements may well be shocking folks with the reminder that such progress is a core theme—perhaps the core theme—to the technology industry: that advances in technology initially are exorbitantly expensive but tend to become much less so over time. That is, investors should expect more disruption this to follow. The achievements may upend, in fact, a theme that had been emerging of late that genAI would break that lower-cost-over-time theme, a critique generally presented by genAI skeptics as rationale for the belief that the spending on infrastructure would forever surpass the potential revenue the infrastructure generates.
GenAI champions have been quick to demote DeepSeek’s advances. Even though the retorts provide some counter clarity to the “it’s over for the Silicon Valley bros!” side of the argument, it’s hard not to read many of their arguments as face-saving amidst a hindsight failure in the creativity department.
With this news, the dumb money of the Magnificent 7 and others seems increasingly dumb. CEO Mark Zuckerberg said just last Friday that his Meta (META) intends to spend $65 billion on genAI infrastructure this year. At the start of the year, Microsoft (MSFT) said it will spend $80 billion this year on genAI infrastructure and computing. And the White House will support a $500 billion-over-four-years project for a domestic infrastructure build. Lots of zeros there. Critically, we’ve learned of few advances touting the potential incremental revenue to be generated from the ongoing expansion in computing largesse.
The upshot of the DeepSeek news is that the genAI investment theme is vulnerable to disruption, precisely because so much of the pricing in the space was is based on a range of often overlapping winner-take-all arguments that ignore an observable value-oriented truth in both the technology and finance realms: expensive things tend to get cheaper over time.
