Who Auras is
Auras Technology (3017.TW) sells thermal and mechanical kits as one package: fans and heatsinks still ship, but the growth is cold plates, manifolds, quick disconnects, chassis and racks. It is not on NVIDIA's published 800 VDC list. It sits on the other half of the same cabinet — the 45°C liquid loop NVIDIA specifies next to the new power architecture.
Business mix
By the first half of 2026 the mix had already flipped toward AI infrastructure:
- Thermal products — about 63% of sales, including liquid-cooling modules.
- Chassis — about 20%, growing faster than thermal as rack density rises.
- Servers and networking as an end market — about 66% of sales, up from the high 40s a year earlier.
- A listed hinge subsidiary remains a smaller slice.
Management's operating rule is profit first, not revenue scale. Q2 2026 revenue was roughly flat sequentially while gross margin jumped.
Where it sits in the NVIDIA 800 VDC stack
The liquid-cooling block on NVIDIA's diagram: CDU, sidecar, 45°C loop, liquid-cooled busbar.
Auras ships the loop hardware that sits on the tray and in the rack — cold plate, manifold, QD, chassis — and says it can also build CDU internals or a complete water-to-water / water-to-air CDU if the customer defines it as a system. In practice it is investing more in components than in whole CDUs.
It is not a named 800V vendor. The 2026 calls never said 800 V. The 800V read is that Vera Rubin, Blackwell and the CSP ASIC cabinets are going fully liquid, and Auras claims a core-supplier seat on both the GPU and ASIC sides. Capacity for liquid-cooling modules is being lifted from 200,000 to 1 million sets a month, mostly in Vietnam.
For ASIC customers it says it designs the thermal architecture from scratch, not just builds NVIDIA's drawing, and that those programmes carry a higher gross margin than NVIDIA work. 2026 is framed as the ASIC first year, with ASIC revenue given a chance to overtake GPU in 2027.
Coverage note
Two notes, Q1 and Q2 2026. The August call is the one with the ASIC-year framing and the Google TPU question; the May call is the one with the five-times capacity target and the NVIDIA-versus-ASIC design distinction.