Amazon is getting deeper into the AI arms race with the launch of Trainium 3, a chip designed to rival Nvidia’s dominant GPU hardware.
The new chips, available through Amazon Web Services (AWS), promise to quadruple the training speed compared to the previous version, while maintaining the same energy footprint. The move will put the tech giant in competition with Google and Nvidia as the fight over infrastructure intensifies.
Each cluster of Amazon’s new “UltraServers” can run up to 144 Trainium 3 chips, positioning them to handle large-scale language model training and other heavy computing tasks. The launch is part of Amazon’s broader push to expand its AI infrastructure and reduce reliance on others.
Amazon’s push, coupled with Google’s dominance in the AI model race, where it now has an 87% chance of landing the best one by the end of the year, has OpenAI’s Sam Altman declaring a “code red.”
AI and cryptography
However, building more AI servers creates a problem that few tech giants can solve alone: finding enough power and space. That’s where crypto miners, who already have large data centers operational, are stepping in, using some of their hardware to enter the AI arms race and profit from it.
Amid the arms race and following the Bitcoin halving in 2024, which halved block rewards, several large mining companies began repurposing their energy-intensive operations into AI-ready facilities. Companies like Core Scientific, CleanSpark, and Bitfarms are now seen less as bitcoin bets and more as utility providers for hyperscalers.
Bitcoin mining company turned neocloud IREN (IREN) soared last month after signing a $9.7 billion AI cloud deal with Microsoft (MSFT). Similarly, TeraWulf (WULF) signed a $9.5 billion AI infrastructure joint venture with Google-backed Fluidstack.
These companies control gigawatts of power capacity, with existing infrastructure ready for AI clusters requiring advanced cooling and stable network connections.
Bubble risk?
Still, the shift carries risks.
Miners are borrowing heavily to retrofit sites for AI workloads, and as investors become more wary of the pace and scale of costs behind “AI trading,” correlated risk assets (such as tech stocks and cryptocurrencies) are under pressure.
bitcoin is down more than 17% over the past 30 days, while the broader CoinDesk 20 index (CD20) lost 19.3% of its value over the same period. The tech-heavy NASDAQ 100 index is down about 1.5% over the past month, having recently recovered from a more than 7% drop in the period.
Analysts have warned that the AI infrastructure boom resembles past bubbles. OpenAI, for example, has committed to spending billions on infrastructure, funds for which it still needs to raise.
Much of the capital that goes into the AI arms race is recycled through the actors themselves, selling AI chips or cloud services. If demand for AI slows, Bain & Co. has forecast a shortfall of up to $800 billion for these companies, which would need $2 trillion in combined annual revenue by 2030 to fund the computing power needed for projected demand.
If demand for AI computing slows, these hybrid operations could face the same liquidity crisis that hit the crypto sector in 2022. Such a hit would likely hit the broader market, sending risk assets down sharply.
For now, however, miners are betting the future of their business on a new kind of gold rush powered by GPUs, not ASICs.
