- JetCool is working on a new silicon-level liquid cooling system called SmartSilicon
- The technology is several years from coming to market, but testing has shown it can handle more than 3,000-watts
- JetCool's work could signal ultra-high-power chips are on the way, Dell'Oro's Lucas Beran said
Nvidia already has the 1,200-watt Blackwell GPU. Intel is reportedly working on a 1,500-watt chip. But there’s a hint that bigger things could be on the horizon. Much bigger. In an exclusive interview with Fierce Network, JetCool CEO Bernie Malouin revealed the company is working on a new direct liquid cooling system capable of handling chip power levels of 3,000 watts and beyond.
JetCool’s single-phase direct-to-chip cooling technology is already unique in that its cold plates house hundreds of micro-nozzles which spray chilled water directly above the hot spots on a chip. The new system it’s working on — called SmartSilicon — replicates that same technology but builds the system into another layer of silicon that gets bonded to the chip.
“When you do this, you can do wildly high powers,” Malouin told Fierce. “Let’s just say that you can comfortably do things that are 3, 4, 5,000 watts in a device with that kind of solution.”
That’s a big deal, given questions have swirled about the viability of single-phase direct liquid cooling beyond 2,000 watts or so. That’s because the current system of designing the cold plate system separate from the processor is “always going to limit performance and efficiency,” Dell’Oro Group Research Director Lucas Beran told Fierce.
Thus, beyond 2,000 watts or so, the focus tends to shift to two-phase systems like those offered by Zutacore and Accelsius, Beran added.
SmartSilicon is a work in progress and still “a couple years out,” Malouin said. But “as we look toward the future roadmaps of power, if power continues to climb…with that kind of technology the sky is the limit.”
According to Malouin, the system has drawn keen interest from those on the cutting edge of tech — think hyperscalers and chipmakers. And JetCool just banked $17 million at the end of last year to fuel development of its future cooling solutions. But the tight integration SmartSilicon requires means JetCool’s role in the ecosystem has to evolve a bit.
JetCool has tested its technology. The biggest challenge is that it requires "a very tight level of integration," Malouin said. “Now we’re not just the cooling partner." JetCool needs to also be the design partner, and partner with the foundry. "And it takes time to build out all of those different interrelated parts, and then of course any time you’re going into a foundry there’s a lot of process validation.”
According to Beran, JetCool’s work could be an indicator that ultra-high power chips are coming down the pipeline. "Maybe it’s not the next generation or even the one after that but in three years’ time we certainly could be talking about 3,000 watts," he said.
Nvidia earlier this year unveiled its three-year roadmap, which includes the release of a GPU called “Rubin Ultra” in 2027, though details beyond the name are sparse.
Beran said he doesn’t expect 3,000-watt chips to go mainstream. Malouin agreed, noting that he expects there will be a “plateau” in power levels around 1,500 watts.
Why? Well, because in several years’ time, the artificial intelligence (AI) traffic that’s spurring high-power chip adoption will shift from high-intensity training workloads to lower-intensity but higher volume inferencing ones. As Beran put it, the question becomes who will need such strong chips at a time when we’ll already have robust foundation models in place.
Though it’s hard to see where the future of AI will go, Beran speculated that perhaps ultra-high-power chips could be used to rapidly train custom models for niche use cases.