Amazon.com said Tuesday that its newest artificial-intelligence chips are now available for customers, offering a challenge to industry leader Nvidia and custom chips from Google.
Amazon said “UltraServers” powered by up to 144 of its in-house Trainium 3 chips are now available, offering more than four times the computing performance and a similar level of improved energy efficiency from its previous generation. The announcement came amid Amazon’s AWS re: Invent event taking place in Las Vegas this week.
“Trainium 3 UltraServers enable organizations of all sizes to train larger AI models faster and serve more users at lower cost—democratizing access to the compute power needed for tomorrow’s most ambitious AI projects,” Amazon said in a statement.
In-house chips have been an important project at Amazon as it seeks to make sure it doesn’t become too dependent on costly hardware from Nvidia. The big question for Trainium 3 is how many major external clients will adopt the hardware, especially with Alphabet’s Google making inroads into the AI processor market with its Tensor Processing Units.
“Trainium is used by a limited cohort of large AWS [Amazon Web Services] customers, but we expect AWS to expand access to a broader base of customers,” wrote J.P. Morgan analyst Doug Anmuth in a research note written before Tuesday’s announcement.
One important question is the balance of chips that will be used by AI startup Anthropic, which said in October that it had struck a deal to use up to a million of Google’s TPU chips, diversifying its supply beyond Amazon and Nvidia.
Amazon stock closed hight 0.2% at $234.42 on Tueaday, while Alphabet rose 0.3% to $315.81, and Nvidia advanced 0.9% to $181.46.
Amazon has committed up to $8 billion of investment to Anthropic, founded by former OpenAI engineers in 2021. Google has invested about $3 billion in Anthropic, according to Bloomberg.
Anthropic has said that Amazon remains its “primary training partner and cloud provider.” It has said it expects to be using more than one million Trainium 2 chips by the end of the year, including its use of the Project Rainier supercomputer with nearly 500,000 Trainium processors.
Amazon said thousands of its UltraServers can be connected together, to form clusters containing up to one million Trainium chips.
“Project Rainier was built exclusively for Anthropic, but we believe AWS could replicate the architecture for Trn3 to serve other large customers if it can gain adequate access to power,” J.P. Morgan’s Anmuth wrote.
Amazon declined to provide direct benchmark comparisons for the Trainium 3 servers against the latest hardware from Nvidia and Google, or their power requirements. Each chip integrates 144 gigabytes of high-bandwidth memory, compared with 192 GB for Google’s latest Ironwood TPU and up to 288 GB for Nvidia’s latest Blackwell GB300 chips.
The Trainium 3 chips will be manufactured by Taiwan Semiconductor Manufacturing, but Amazon didn’t disclose any additional partners. Chip designer Marvell Technology has worked on previous generations of Trainium but the stock has been dogged byworriesthat it might lose out on future versions.
Amazon said Tuesday it is already working on Trainium 4.
