AWS Raises Trainium 3 Shipment Plans, Taiwan Supply Chain Sees Upward Revision in H2 Demand
N.R. Finch
AWS has told its server supply chain to lift Q3 Trainium 3 shipments by 20%–30%, boosting Taiwan thermal, chassis, and assembly suppliers — while ASIC server shipments are forecast to grow 64.2% this year, outpacing GPUs.
Why is AWS pulling orders forward?
Trainium 3 servers only began shipping in May, yet AWS has already told suppliers to raise Q3 volumes 20%–30% above the original plan.
This means → AWS is not ramping gradually — it is pulling inventory forward and compressing lead times to capture demand before the window closes.
The biggest driver is Anthropic, which says its compute roadmap targets tenfold growth but still faces a shortfall. AWS, Anthropic's lead investor, signed a ten-year partnership in April, under which Anthropic will expand its purchases of AWS compute.
How strong is Trainium demand really?
On the latest earnings call, CEO Andy Jassy said Trainium 2 is sold out, Trainium 3 is nearly fully booked, and customers are already reserving the next-generation Trainium 4.
In plain terms = three generations are simultaneously supply-constrained, and Amazon has already begun developing Trainium 5.
Beyond Anthropic, OpenAI and Uber are major Trainium customers. Amazon's managed-AI platform Bedrock serves 125,000 enterprise clients, with inference workloads running primarily on Trainium.
Which Taiwan suppliers are in the order flow?
Taiwan firms span multiple tiers of the Trainium 3 server: thermal (AVC, Microloops, Cooler Master), rack slides (King Slide, Repon), chassis (AVC, Chenbro Micom), L6 assembly (Accton), and L11 assembly (Wiwynn).
L6 motherboard components — including thermal and chassis modules — have been ramping since May, with monthly volumes rising steadily. L11 rack orders and slide-rail components are expected to enter mass production in July.
This means → both product lines are accelerating now; Q3 shipments are on track to beat earlier forecasts, and the H2 earnings-upgrade thesis is already in motion.
Why are ASIC servers growing faster than GPUs?
DIGITIMES analyst Hsiao Ya-wen forecasts 2026 GPU server shipments will grow 43.8% year-on-year, while ASIC servers will grow even faster at 64.2%.
This reflects a structural shift: hyperscalers are moving from buying general-purpose GPUs to designing custom ASICs to grab share — AWS's Trainium and Google's TPU (chips purpose-built for AI workloads) will both ramp aggressively in H2.
Google's TPU already leads in ASIC shipment volume, pressuring AWS. Meta recently announced it will open its AI compute to outside developers, emerging as a new competitor alongside AWS, Microsoft, and Google.
What still needs to fall into place?
Whether Taiwan suppliers' H2 earnings actually upgrade ultimately depends on AWS customers' real procurement plans holding firm.
In plain terms = the signal from AWS is strong, but between "notifying the supply chain" and "customers placing binding orders" there is still a gap — if Anthropic or other large buyers slow their compute expansion, the uplift could be trimmed.
The near-term link with the highest certainty is L6 components and thermal modules already ramping; L11 rack orders entering mass production in July are the next checkpoint.
Content is for reference only, not financial advice.