S&P 500 slips on report Fed’s Waller leading race to replace Powell; tech shines
Investing.com -- Concerns over declining compute demand that emerged following the emergence of DeepSeek’s R1 artificial intelligence model are misplaced, according to an AI expert.
Adam Beberg, founder of distributed computing software solutions company Mithral and former principal architect of distributed systems at Nvidia (NASDAQ:NVDA), dismissed the notion that R1 signals a shift toward lower compute requirements.
In a discussion hosted by investment firm Jefferies, Beberg stressed that "the demand for compute requirements is more likely going up, not down, as a result of the release of R1."
DeepSeek’s R1 model, like OpenAI’s o1, leverages Chain-of-Thought (CoT) reasoning, which requires significantly more inference compute time.
While some investors speculated that R1’s reported $5.6 million training cost indicated falling AI infrastructure needs, that number is widely questioned on Wall Street.
“Industry experts are in agreement that DeepSeek has considerably more powerful AI infrastructure than they’ve publicly reported,” Jefferies analysts wrote in a note.
The shift toward reasoning-based AI models is accelerating compute intensity, particularly on the inference side.
“So while pre-training may have been more efficient, inference is considerably less efficient,” they added. “As the market trends more in the direction of these reasoning models, demand for compute should accelerate.”
According to Beberg, the rising demand for inference computing is likely to benefit ASICs, as reasoning-based models like DeepSeek’s R1 require significantly more processing power. While R1 is recognized as a strong model, the expert argues that its efficiency improvements in pre-training do not translate to a meaningful drop in overall compute demand.
Beberg also highlights that the traditional scaling approach of increasing compute power is starting to show diminishing returns.
As a result, future breakthroughs are expected to come more from advancements in software rather than hardware. That said, he still believes that Nvidia remains firmly positioned as the technology leader, with higher compute capacity continuing to provide a competitive edge.