Taiwan's Industrial Technology Research Institute (ITRI) has published work on water immersion cooling as an answer to rising thermal loads in AI chips. The announcement itself is thin on numbers. That absence matters more than the claim.

The institute frames direct liquid contact as the next step beyond air cooling and cold plates. No capacity figures, no partner fabs and no deployment dates appear in the material. Readers should treat this as a research direction, not a qualified manufacturing process.

Why heat now defines the AI roadmap

AI accelerators keep drawing more power per package. Air cooling struggles once package power climbs into the hundreds of watts. Cold plates help, but they still leave thermal resistance between the die and the fluid.

Immersion cooling removes that barrier. The chip sits directly in a dielectric or treated water medium. ITRI's contribution centers on water-based immersion, which carries more heat per liter than oil-based dielectric fluids.

Water near live electronics raises obvious concerns. The institute says its approach addresses those risks. It does not publish leak-rate data, corrosion results or long-run reliability figures alongside the claim. Without those numbers, the method remains a laboratory proposition.

What the announcement does not say

Several questions stay open:

• No named chip customers or accelerator vendors appear in the material.
• No pilot line, fab location or volume timeline is disclosed.
• No thermal resistance or watts-per-square-centimeter target is stated.
• No cost comparison against cold-plate liquid cooling is offered.

Each gap is normal for early research output. Each gap also blocks any commercial assessment. Hyperscalers already deploy two-phase immersion and direct-to-chip liquid cooling at scale, so ITRI enters a crowded field with established competitors.

The competitive and supply context

Taiwan has a structural reason to push cooling research. The island hosts TSMC and much of the global AI silicon supply chain. If thermal limits cap accelerator performance, they cap the value of that supply chain. Cooling research is therefore strategic, not merely academic.

Nvidia-class accelerators and their successors set the thermal agenda for the industry. Any cooling method must prove compatibility with advanced packaging, including CoWoS-class substrates and high-bandwidth memory stacks. The ITRI material does not state which package types it has tested.

What to watch next

The credible milestones are measurable ones: a named partner, a published watts-per-package figure, and reliability data over thousands of hours. Until ITRI discloses at least one of these, the work sits in the long queue of promising thermal ideas awaiting qualification.

By Sophie Lindqvist, Chip Dispatch