Blab AI
Amazon Signals Plans to Sell Trainium AI Chips to External Data Centers
Amazon is breaking out of its cloud‑centric mold. During a Paris interview, senior vice‑president Peter DeSantis revealed the company is negotiating to sell its Trainium chips to firms that operate their own data centers. While no customers have been announced, the move is still under exploration.
The Trainium family, born in Annapurna Labs, Amazon’s fabless semiconductor arm, has only ever been sold through AWS. The second‑generation chip, Trainium 2, debuted in 2024 and offered up to four times the training speed of its predecessor. Trainium 3, which began shipping earlier this year, has already sold out, Bloomberg reports. A fourth generation is slated for next‑year release and has already attracted early interest.
Amazon’s chief executive, Andy Jassy, first hinted at external sales in his April shareholder letter. He noted that the company’s annualized chip revenue—comprising Trainium, Graviton, and Nitro products—was roughly $20 billion, and that opening the market could lift that figure toward $50 billion. The letter also cited more than $225 billion in revenue commitments for the Trainium line.
DeSantis said the shift is driven in part by a surge in demand for “locally controlled computing resources,” especially in Europe, where regulatory and political pressure has amplified calls for sovereign AI infrastructure. He added that the way businesses build AI infrastructure is “on the cusp of rapid, wide‑ranging change” and that Amazon intends to play a role.
The strategy mirrors Google’s own announcement in April that its Tensor Processing Units (TPUs) would be made available to external data centers through a limited program. Both Amazon and Google position their silicon as cheaper alternatives to Nvidia’s GPUs. DeSantis emphasized that Amazon’s end‑to‑end control—from design to manufacturing to shipping—places it in direct competition with Nvidia.
Despite the prospect of new revenue streams, DeSantis dismissed worries that selling Trainium chips could erode AWS’s cloud business. He said there is “so much underconsumption in AI” and that the company is not worried about cannibalization.
The announcement comes amid a broader shift in the AI hardware market. Demand for high‑performance training and inference accelerators has outpaced supply, prompting companies to seek alternatives to Nvidia’s dominant GPUs. Amazon’s Trainium line, built on TSMC’s 4 nm and 3 nm processes, delivers high throughput and lower power consumption for training workloads.
At present, no external customers have been named, and no launch timeline has been disclosed. Nevertheless, early interest in the forthcoming fourth‑generation chip and the company’s stated intent to broaden its reach suggests that Amazon is preparing to enter the data‑center silicon market.
In the coming months, observers will look for concrete agreements, pricing, and deployment plans. The move could reshape the competitive landscape for AI infrastructure, especially for organizations seeking sovereign or region‑specific hardware solutions.
Today, Amazon is exploring external sales of its Trainium chips, with no customers identified yet. The next steps will likely involve formalizing agreements, setting pricing, and integrating the chips into third‑party data‑center environments. The broader AI industry will watch how this development affects supply dynamics, competition with Nvidia, and the availability of affordable AI training hardware.
The Trainium family, born in Annapurna Labs, Amazon’s fabless semiconductor arm, has only ever been sold through AWS. The second‑generation chip, Trainium 2, debuted in 2024 and offered up to four times the training speed of its predecessor. Trainium 3, which began shipping earlier this year, has already sold out, Bloomberg reports. A fourth generation is slated for next‑year release and has already attracted early interest.
Amazon’s chief executive, Andy Jassy, first hinted at external sales in his April shareholder letter. He noted that the company’s annualized chip revenue—comprising Trainium, Graviton, and Nitro products—was roughly $20 billion, and that opening the market could lift that figure toward $50 billion. The letter also cited more than $225 billion in revenue commitments for the Trainium line.
DeSantis said the shift is driven in part by a surge in demand for “locally controlled computing resources,” especially in Europe, where regulatory and political pressure has amplified calls for sovereign AI infrastructure. He added that the way businesses build AI infrastructure is “on the cusp of rapid, wide‑ranging change” and that Amazon intends to play a role.
The strategy mirrors Google’s own announcement in April that its Tensor Processing Units (TPUs) would be made available to external data centers through a limited program. Both Amazon and Google position their silicon as cheaper alternatives to Nvidia’s GPUs. DeSantis emphasized that Amazon’s end‑to‑end control—from design to manufacturing to shipping—places it in direct competition with Nvidia.
Despite the prospect of new revenue streams, DeSantis dismissed worries that selling Trainium chips could erode AWS’s cloud business. He said there is “so much underconsumption in AI” and that the company is not worried about cannibalization.
The announcement comes amid a broader shift in the AI hardware market. Demand for high‑performance training and inference accelerators has outpaced supply, prompting companies to seek alternatives to Nvidia’s dominant GPUs. Amazon’s Trainium line, built on TSMC’s 4 nm and 3 nm processes, delivers high throughput and lower power consumption for training workloads.
At present, no external customers have been named, and no launch timeline has been disclosed. Nevertheless, early interest in the forthcoming fourth‑generation chip and the company’s stated intent to broaden its reach suggests that Amazon is preparing to enter the data‑center silicon market.
In the coming months, observers will look for concrete agreements, pricing, and deployment plans. The move could reshape the competitive landscape for AI infrastructure, especially for organizations seeking sovereign or region‑specific hardware solutions.
Today, Amazon is exploring external sales of its Trainium chips, with no customers identified yet. The next steps will likely involve formalizing agreements, setting pricing, and integrating the chips into third‑party data‑center environments. The broader AI industry will watch how this development affects supply dynamics, competition with Nvidia, and the availability of affordable AI training hardware.
