Frore Systems’ LiquidJet is a direct-to-chip liquid cooling coldplate aimed at the thermal limits of modern AI accelerators, and it was first unveiled at OCP in October 2025. Instead of treating a GPU package as one uniform heat source, it treats it as a power map, so the coldplate can be tuned for die, HBM stacks, and any localized hot spot that dominates junction temperature. https://www.froresystems.com/products/liquidjet-dlc-coldplate
A lot of today’s data-center coldplates are built with skived 2D microchannels: coolant enters, runs a long path, warms up, and you get a noticeable temperature gradient across the plate. LiquidJet flips that by using short-loop “jet-channel” microstructures and multi-level manifolding, so cooling can be uniform when you want it, or intentionally biased toward a 10×11 mm hot region on a larger die area. Feature sizes can get down to ~75 µm, which is why the manufacturing approach matters a lot.
The manufacturing angle is borrowing semiconductor-style fabrication, but on metal wafers: etch the microstructure rather than machining long channels, then build the stack as a precise, repeatable flow network. In this CES Las Vegas 2026 walkthrough, Frore shows hotspot demos and explains how the approach can support very high heat flux, up to about 600 W/cm² with liquid-metal TIM, versus roughly 300 W/cm² in many conventional plates, while holding a tighter temperature field on the plate.
On current high-power GPU platforms, Frore quotes results like ~7.7°C lower GPU temperature, ~75% higher heat removed per unit flow (kW/LPM), and ~50% lower coldplate mass using copper where it counts and a lighter top construction. Lower required flow (they cite about 1.0–1.4 LPM per kW) can also reduce pumping load and pressure stress on rack plumbing, which matters when a CDU is feeding many parallel loops at scale, with production targeting around June.
Looking ahead, the discussion links cooling needs to packaging trends: single-reticle versus multi-reticle dies, rising total module power toward multi-kilowatt designs, and the need to cool both compute and adjacent HBM without overbuilding the whole loop. If hyperscalers can hold junction temperature down with less flow and less pressure drop, they can often sustain higher clocks and improve throughput-per-watt at the rack level, which is the real prize in an AI factory rack.
I’m publishing about 100+ videos from CES 2026, I upload about 4 videos per day at 5AM/11AM/5PM/11PM CET/EST. Check out all my CES 2026 videos in my playlist here: https://www.youtube.com/playlist?list=PL7xXqJFxvYvjaMwKMgLb6ja_yZuano19e
This video was filmed using the DJI Pocket 3 ($669 at https://amzn.to/4aMpKIC using the dual wireless DJI Mic 2 microphones with the DJI lapel microphone https://amzn.to/3XIj3l8 ), watch all my DJI Pocket 3 videos here https://www.youtube.com/playlist?list=PL7xXqJFxvYvhDlWIAxm_pR9dp7ArSkhKK
Click the “Super Thanks” button below the video to send a highlighted comment under the video! Brands I film are welcome to support my work in this way 😁
Check out my video with Daylight Computer about their revolutionary Sunlight Readable Transflective LCD Display for Healthy Learning: https://www.youtube.com/watch?v=U98RuxkFDYY
