Amazon Web Services (AWS) has considered the possibility of selling artificial intelligence chips, which it has so far mainly used in its own cloud infrastructure, to external customers. This step, which the company is considering, indicates a remarkable change in AWS’s long-held chip strategy. Statements on the subject were given to Bloomberg by Peter DeSantis, AWS Senior Vice President and one of the managers responsible for artificial intelligence activities. DeSantis confirmed that the company is in discussions about the possibility of selling its Trainium AI accelerators to other companies for use in data centers. However, no details were shared about who the potential customers might be.
According to information provided by company sources to TechCrunch, this plan is still in the early evaluation phase. The basis of the discussions lies in the statements made by Amazon CEO Andy Jassy in his annual shareholder letter published in April. Jassy stated that the demand for Amazon’s self-developed artificial intelligence chips has exceeded expectations and stated that it may be possible to sell chips or chip shelves directly to third-party customers in the future.
According to Amazon’s calculations, its annual revenue potential could reach nearly $50 billion when its chip operations alone are considered as a standalone company and sales to AWS and other customers are taken into account. Although this figure remains well below the revenue level reached by Nvidia today, it corresponds to a significant business volume in the global semiconductor industry.
Demand for Amazon Web Services Trainium chips strains production capacity
There are various reasons why AWS has stayed away from selling artificial intelligence chips directly to date. Chief among these is the additional revenue the company generates from cloud services. AWS offers artificial intelligence processing power to its customers by using Trainium chips in its own data centers. Thus, you not only generate revenue from transaction capacity; It also provides additional revenue from storage, network infrastructure, security, monitoring and other corporate services. This structure creates a more comprehensive revenue model for the company than chip sales.
In addition, it is known that AWS’s current production capacity has difficulty meeting demand. Andy Jassy stated in his shareholder letter that the capacity of the current generation Trainium chips was exhausted in a short time. In addition, it was stated that the capacity allocated for the new generation Trainium4, which has been available for more than a year, has been largely filled. These statements were made before AWS later added OpenAI models to its service portfolio. Therefore, it seems possible that demand has reached even higher levels today.
In such an environment, AWS starting to sell its chips to external customers could make capacity planning more complicated for existing cloud customers. In order to overcome this problem, the company needs to provide more capacity from its production partners. AWS’s chips are produced by TSMC, one of the world’s largest semiconductor manufacturers. However, due to the high demand on TSMC, obtaining additional capacity does not seem easy. The fact that Nvidia has recently increased its production volume significantly and the company has become one of TSMC’s largest customers further strengthens the competition.
AWS spokesman Doron Aronson also confirmed that the company may turn to direct chip sales in the future. Aronson said AWS has turned down such requests in the past, although selling chip racks to third-party companies is a possibility.
On the other hand, Nvidia does not aim to remain only in the GPU market. Jensen Huang, the company’s founder and CEO, recently announced that they see great opportunities in the artificial intelligence-focused CPU market. Thus, while Nvidia is focusing more strongly on the areas where Intel and AMD operate, Amazon is preparing to enter more directly into the artificial intelligence accelerator market dominated by Nvidia with its Trainium family. Under current conditions, it is not clear when the plans will be implemented. Despite this, the rapid increase in demand for artificial intelligence infrastructures paves the way for the competition between cloud providers and chip manufacturers to become more visible in the coming period.