Anthropic S-1 Reveals Multi-Chip Compute Pool, Challenging Nvidia Pricing Power
Anthropic's S-1 filing reveals a unified compute resource pool integrating chips from Amazon, Google, Nvidia, and AMD. This strategy uses software scheduling to distribute tasks across platforms, aiming to reduce vendor dependency and strengthen bargaining power. While unlikely to disrupt Nvidia's training dominance in the short term, it may erode pricing power and accelerate compute commodification long-term.
Full Report
IT Home reported on October 28 that, according to a Business Insider article published early in the morning on the same day, Anthropic is simultaneously utilizing multiple types of chips for AI computing tasks to diversify risk. This strategy could potentially undermine NVIDIA's market dominance in the long term.
According to Anthropic’s S-1 filing released on October 26 (local time), the company aggregates computing power from multiple strategic partners into a unified resource pool, then uses software to allocate compute tasks to the most efficient chips.
As noted by IT Home, Anthropic employs four main categories of chips: Amazon Trainium, Google TPU, NVIDIA GPU, and AMD GPU. Achieving this cross-platform scheduling required years of effort and significant investment.
Analysts believe this capability may help large AI companies gradually reduce their dependence on NVIDIA, giving them greater leverage in price negotiations.
Patrick Moorhead, Founder and CEO of Moor Insights & Strategy, argues that Anthropic has not just started using multi-chip architectures but is now positioning this capability as a key advantage to attract investors. "What is truly at stake is NVIDIA's pricing power. Customers who can flexibly distribute compute tasks across four platforms negotiate from a completely different position than those who cannot."
Ashish Nadkarni, an analyst at IDC, pointed out that adopting multiple chip types also helps mitigate the impact of supply fluctuations in NVIDIA chips.
However, the threat this strategy poses to NVIDIA remains limited. Migrating tasks between different chip platforms is still difficult, particularly for model training, which is more challenging than inference. Only the largest AI companies have the resources to implement such cross-platform scheduling.
Anthropic continues to invest billions of dollars in NVIDIA computing infrastructure.
Gil Luria, Head of Technology Research at D.A. Davidson, suggests that major players like OpenAI are likely to adopt similar approaches. However, for smaller companies, investing heavily in time and capital for this purpose may not be cost-effective.
Brad Gastwirth, Global Head of Research and Market Intelligence at Circular Technology, noted that Anthropic’s strategy will not pose a significant short-term threat to NVIDIA. Having more chip options does not mean NVIDIA’s advantages in performance, networking interconnects, and software will disappear.
If AI computing power becomes increasingly commoditized, NVIDIA will face broader risks.
Ken Griffin, CEO of Citadel, warned in a Bloomberg interview last week that if computing supply exceeds demand, rental prices could drop: "Like airlines, GPUs have very low marginal operating costs. Once there is overcapacity, prices will plummet."