AWS Trainium 3 Ramps Up, Taiwan Supply Chain H2 Orders Accelerate
Alina Collins
Amazon's in-house Trainium 3 AI chip servers began shipping in May, with volume accelerating in Q3. Taiwan thermal, rail, and assembly suppliers secured a high share of orders, making this one of the busiest non-Nvidia server ramps in H2 2026.
How fast is Trainium 3 shipping?
Board-level L6 orders began shipping in May 2026; rack-level L11 orders are expected to enter mass production in July.
Shipment volume accelerates in Q3, extending through Q4 and into Q1 2027.
This means → the ramp window spans at least three quarters — not a one-off burst.
Which Taiwan suppliers are in the chain?
Thermal: AVC, Microloops, and Cooler Master.
Rails: King Slide and Repon.
Assembly: Accton Technology handles L6 board-level; Wiwynn handles L11 rack-level.
This means → AWS allocated a high proportion of orders to Taiwan suppliers, giving the island's supply chain more upside — from mechanical components to full assembly — than Google's TPU chain offers.
What do thermal suppliers' numbers already show?
Microloops posted May revenue of NT$330 million (~US$10.45 million), up 54.96% month-on-month. The company said it is confident in H2 momentum.
Industry sources at AVC say ASIC servers — servers built around chips designed for specific tasks — contributed 20%–30% of AVC's revenue in 2025, with ASIC shipments expected to rise sharply from H2.
In plain terms = thermal suppliers' numbers are already moving. This is not a forward-looking story — orders are being filled now.
Why does the liquid-cooling switch matter on its own?
Trainium 3 servers currently use air cooling, but liquid cooling is expected to begin integration before year-end.
Liquid-cooling modules carry an average selling price several times that of air cooling.
This means → on the same server, switching to liquid cooling lifts per-unit revenue for thermal suppliers sharply — volume growth plus a unit-price upgrade stacked on top.
ASIC servers vs. GPU servers — which segment grows faster?
DIGITIMES Research estimates 2026 GPU server shipments grow 43.8% year-on-year; ASIC server shipments grow 64.2%.
Google TPU holds roughly 46% of the ASIC server market; AWS holds about 20.4%. Google's growth momentum is stronger.
But AWS orders carry more upside for Taiwan suppliers — Google TPU servers are primarily assembled by Canada's Celestica, not by Taiwan-based firms.
Put simply = the ASIC track overall outpaces GPU, but for Taiwan suppliers specifically, the AWS line delivers more revenue leverage than the Google line.
What does Wiwynn's CEO say about the AI bubble?
Wiwynn President and CEO Simon Lin stated: "In 2025 people talked about an AI bubble. I didn't see one then, and now I'm even more certain there isn't one. I see no bubble for the next four years."
He added that clients are pushing ahead at full speed on AI data-center builds — "we can only try to keep up."
This reflects extremely high order visibility across Taiwan's AI server supply chain. Whether Trainium 3 can sustain volume beyond Q1 2027 will be the key test of how deeply AWS's in-house chip strategy takes root.
Content is for reference only, not financial advice.