Article · AI translation
IT Home reported on October 28 that, according to a Business Insider report published early this morning (the 28th), Anthropic uses several types of chips to handle AI computing tasks in order to spread risk. Over the long term, this strategy could undermine Nvidia's market dominance.
On the 26th local time, Anthropic's public listing filing (Form S-1) showed that the company combines computing capacity provided by several strategic partners into a unified pool of resources, then uses software to assign computing tasks to the chips best suited to them.
IT Home learned from the report that Anthropic uses four main categories of chips: Amazon Trainium, Google TPUs, Nvidia GPUs and AMD GPUs. The company has spent years and substantial sums of money developing this cross-platform scheduling capability.
Analysts believe this capability could help large AI companies gradually reduce their reliance on Nvidia and gain more leverage in price negotiations.
Patrick Moorhead, founder and CEO of Moor Insights & Strategy, believes Anthropic has not just begun using multiple types of chips; rather, it has begun presenting that capability as a major advantage to attract investors. “What could really take a hit is Nvidia's pricing power. A customer that can flexibly allocate computing tasks across four platforms is in a completely different negotiating position from one that cannot.”
IDC analyst Ashish Nadkarni noted that using multiple types of chips also has the benefit of reducing the impact of fluctuations in Nvidia chip supply.
However, this strategy poses only a limited threat to Nvidia. Moving tasks between different chip platforms remains difficult, especially for model training, which is harder to move than model inference. Only the largest AI companies have the resources to schedule tasks across platforms.
Anthropic also continues to invest billions of dollars in Nvidia computing capacity.
Gil Luria, head of technology research at D.A. Davidson, believes large companies such as OpenAI are also likely to use similar approaches, but for smaller companies, the substantial investment of money and effort required may not be worthwhile.
Brad Gastwirth, head of global research and market intelligence at Circular Technology, said Anthropic's strategy will not pose a major threat to Nvidia in the short term. Having more chip options does not mean Nvidia's advantages in performance, networking and software will disappear.
If AI computing capacity gradually becomes commoditized, Nvidia will face broader risks.
Ken Griffin, CEO of Citadel, warned in an interview with Bloomberg last week that rental prices could fall if the supply of computing capacity exceeds demand: “GPUs are like airlines: their marginal cost of operation is very low. Once there is excess capacity, prices will plunge.”