加速押注自研 AI 芯片！AWS 三季度 ASIC 服务器出货目标上调至多 30%

Amazon AWS is accelerating its bet on in-house chip strategy. Supply chain sources indicate that AWS has notified relevant suppliers to increase third-quarter ASIC server shipments by 20% to 30% compared to the original plan, demonstrating high optimism regarding the market prospects of the Trainium 3 chip.

On Monday, citing Taiwanese supply chain sources, Digitimes reported that motherboard-level components for servers equipped with Trainium 3 chips began shipping in May and are currently in a monthly ramp-up phase. The industry interprets AWS's increase in shipment plans as a signal of pulling in orders early to seize market share, rather than merely a routine upward revision of demand.

Multiple supply chain sources pointed out that Anthropic is one of the important drivers behind AWS's accelerated procurement—the AI company has stated that the computing power planned based on 10x growth benchmarks is still insufficient, requiring immediate supplementation of the gap.

Boosted by this news, several Taiwanese manufacturers in the AWS Trainium 3 server supply chain are expected to benefit, with room for upward revisions in their operational performance in the second half of the year. Meanwhile, the industry is also beginning to focus on whether AWS will launch the next-generation Trainium 4 ahead of schedule due to strong market momentum, further strengthening its competitiveness in the ASIC market.

Supply Chain Fully Enters Mass Production Ramp-Up Phase

The Taiwanese supply chain for AWS Trainium 3 servers has formed a relatively complete ecosystem. Thermal management involves AVC, Maike, and Sunon; rail manufacturers include King Slide Works and Nan Jun International; chassis manufacturers include AVC and MiTAC; and the L6 motherboard-level assembler is Accton.

Regarding the shipment rhythm, L6-related motherboard components have been shipping since May, with monthly volume increases; cabinet-level L11 orders and components such as rails will enter mass production in July. Currently, the overall status remains in the shipment ramp-up phase, with volume expected to surge in the third quarter. Following AWS's increase in procurement intensity, the shipment growth for related manufacturers will exceed previous estimates, leaving room for further upward revisions in operational performance in the second half of the year.

Strong Demand from Major Clients Like Anthropic and OpenAI

Supply chain sources believe that Anthropic is one of the main drivers behind AWS's early procurement. AWS's parent company, Amazon, is a significant investor in Anthropic. The two parties further strengthened their cooperation in April 2026 by signing a 10-year partnership agreement, under which Anthropic committed to expanding its procurement of AWS computing power. Anthropic has publicly stated that there is still a significant gap in the computing power planned based on current 10x growth benchmarks, which needs to be filled as soon as possible.

In addition to Anthropic, OpenAI and Uber are also important customers for AWS Trainium. Bedrock, the generative AI hosting platform launched by Amazon for enterprise clients, currently has 125,000 customers. Its inference workloads mainly rely on Trainium services, and the continuous expansion of enterprise-side demand constitutes important support.

Amazon CEO Andy Jassy stated in a recent earnings conference call that Trainium 2 is sold out, Trainium 3 is almost fully booked, and some customers have already started booking the next-generation Trainium 4. The company has simultaneously initiated the development process for Trainium 5.

Tightening Competitive Landscape, AWS Urgently Needs to Strengthen ASIC Scale

Competitive pressure in the cloud computing market is prompting AWS to accelerate its actions. AWS remains the leader in the public cloud market, followed closely by Microsoft and Google. In the ASIC chip sector, Google's TPU shipments rank first in the market, putting direct pressure on AWS and forcing it to accelerate the enhancement of Trainium's capabilities and scale.

Furthermore, Meta recently announced plans to open its proprietary AI computing power to the outside world, allowing enterprises and developers to pay for using its AI models or directly rent bare-metal GPU computing power. This move openly competes with AWS, Microsoft, and Google for enterprise clients, further intensifying market competition and increasing the urgency of AWS's expansion.

DIGITIMES Research analysis predicts that in 2026, among high-end AI server chips, the annual growth rate for GPU server shipments will be approximately 43.8%, while the annual growth rate for ASIC server shipments will be even more significant, expected to reach 64.2%. As both AWS Trainium and Google TPU servers enter the volume production phase in the second half of 2026, the overall ASIC server market is expected to see accelerated expansion.