Amazon is intensifying its involvement in the AI competition with the introduction of Trainium 3, a chip engineered to compete with Nvidia’s leading GPU technology.
The newly released chips, accessible via Amazon Web Services (AWS), claim to enhance training speed by four times compared to the previous version, all while keeping energy consumption consistent. This strategic move positions Amazon in direct competition with Google and Nvidia as the race for infrastructure intensifies.
Each cluster of Amazon’s innovative “UltraServers” can accommodate up to 144 Trainium 3 chips, making them well-suited for extensive language model training and demanding computational tasks. This launch is a component of Amazon’s larger strategy to broaden its AI infrastructure and lessen reliance on external sources.
Amazon’s initiative, alongside Google’s leading position in the AI model sector—where it currently holds an 87% likelihood of achieving the best model by year-end—has reportedly prompted OpenAI’s Sam Altman to issue a “code red.”
AI and crypto
Nevertheless, the surge in AI server construction presents a challenge few tech giants can independently tackle: securing adequate power and space. This is where crypto miners come in, as they already possess large operational data centers and are utilizing some of their hardware to join the AI competition and capitalize on it.
In the midst of this arms race, following the 2024 Bitcoin halving that reduced block rewards by half, several prominent mining companies have started transforming their energy-demanding operations into AI-ready facilities. Firms like Core Scientific, CleanSpark, and Bitfarms are increasingly viewed as utilities for hyperscalers rather than mere bitcoin investments.
The Bitcoin mining company turned neocloud firm IREN (IREN) experienced a significant rise last month after securing a $9.7 billion AI cloud partnership with Microsoft (MSFT). Likewise, TeraWulf (WULF) has established a $9.5 billion joint venture for AI infrastructure development with Fluidstack, supported by Google.
These companies have control over gigawatts of power capacity and already possess infrastructure tailored for AI clusters that demand advanced cooling and stable grid connections.
Bubble risk?
However, this transition carries inherent risks.
Miners are incurring substantial debt to retrofit their sites for AI operations, and as investors become increasingly cautious regarding the rapid pace and magnitude of expenses involved in the “AI trade,” correlated risk assets (including tech stocks and cryptocurrencies) face pressure.
Bitcoin has decreased by over 17% in the last 30 days, while the broader CoinDesk 20 (CD20) index has lost 19.3% within the same timeframe. The tech-heavy NASDAQ 100 index has seen a slight decline of approximately 1.5% over the past month, recovering from a more significant drop of over 7% during that period.
Experts have issued warnings that the burgeoning AI infrastructure market resembles historical bubbles. For instance, OpenAI has pledged to invest trillions in infrastructure, funds for which it is still in the process of raising.
A significant portion of the funding directed towards the AI race is being cycled through the same key players who are selling AI chips or cloud services. Should the demand for AI diminish, Bain & Co. forecasts a potential deficit of up to $800 billion for these firms, which would require $2 trillion in combined annual revenue by 2030 to sustain the computing power necessary for anticipated demand.
If the demand for AI computing dwindles, these hybrid operations might encounter the same liquidity crises that afflicted the cryptocurrency sector in 2022. Such a downturn could adversely impact the broader market, resulting in significant declines in risk assets.
For the time being, however, miners are betting the future of their operations on a novel gold rush driven by GPUs rather than ASICs.
