@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260706

Episode Date: July 6, 2026

- Jim Keller's fab for fabs - First Arm AGI servers spotted at Computex - 3D-printed Small modular Rector module? - Fuel Cell Technology - Data Center Water Use [audio mp3="https://orionx.net/wp-content/uploads/2026/07/HPCNB_20260706.mp3"][/audio] The post HPC News Bytes – 20260706 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies. Hi, everyone. Welcome to HBC Newsbytes. I'm Doug Black, and with me, of course, is Shaheen Khan. Chipmaking is an enormously complex process, as we all know, that requires on-site integration and tuning and constant supervision. There's been no product that just stamps out chips. Well, not surprisingly, people have been thinking about that, and Fab 2, a semiconductor startup co-founded by Chip architect Jim Keller, wants to address a part of that need.

Originally formed four years ago as Atomic Semi in San Francisco, Keller's startup just expanded into Texas and rebranded itself as Fab 2, as in Fab Fab, as in a Fab, whose product is the production of a complete fab. It sounds like a great idea because we all know the world needs more chips and that means more fabs, but the complexity of a fab is a function of how high end the fab is. For something like EU V lithography, the complexity demands the 18-wheeler-sized EUV equipment that only ASML produces. Clearly, that's not the market that Fab 2 and Jim Keller are aiming at. Yes, generating EUV light, super flat mirrors,

specialty optics, alignment, cleaning, et cetera, will keep it complex for a while. But if you set your sights way lower, say 300 nanometer chips versus three nanometer chips, and if you don't need high volume production, then you can pursue it. Fab2 uses electron beam lithography, which is easier and slower, and it targets a small batch and low-end chips, say for prototyping, education, special needs, Now, high-end chips also need super advanced packaging before they can be used. Interposers and substrates, hybrid bonding, power delivery and cooling, maybe HBM HBM integration, etc., which need other specialized tool chains from other companies. Chips coming out of a FAP2-style FAB would also need work before they can be used,

but FAP2 is aiming to make the chip-making part of the process more turnkey than it was before. but they're kicking off distributed low-volume chip making. And what they are doing was also super complex at some point in the past, so who knows where that leads. At the recent Competext 2026 show, the publication Serve the Home spotted a server based on arms upcoming AGI CPU. The system is by ASROC, the Taiwanese system company, that had revenues of about $1.5 billion last year and operates in some 90 countries. Arm announced earlier this year that it was going to expand beyond licensing and start offering its own chips. Provocatively named AGI chip as an artificial general

intelligence is its first in-house server CPU with meta reported as an early major customer. The ASRoc RAC system is a 1U chassis with one AGI processor. The chip is a TSM 3 nanometer CPU with 136 Arm Neuverse V3 cores, 12-channel DDR5 memory, 96 lanes of PCIE, Gen 6, and CXL3.0 support. It's a nice system. The big news continues to be how the dynamics will work out as Arm directly enters the server CPU market. It strengthens arm offerings, but it also creates tension with armed licensees who might see new competition from upstream. It's a big long-term realignment of the market. The AGI design seems optimized for dense Rackenstack-and-stack cloud-scale deployments. These days, that means running Agentic AI inference workloads, where CPUs are emerging as having a more, well, central role again. The Register reports that a

Florida startup called Ampira is working on a 3D-printed nuclear microreactor module. It would be in the 15 to 30 megawatt range and may have the potential to run for up to 30 years without refueling. It is described as a 3D-printed solid-state thorium-based reactor with a silicon carbide core and pressure vessel. The unit would be subcritical, which means the core does not sustain the nuclear chain reaction by itself. An external neutron source or neutron driver would be used to control the reaction. Modular nuclear reactors have been part of proposed methods of generating energy for AI data centers, military applications, and off-grid locations. Several approaches are making progress. This one is a nice milestone, but of course far from a licensed commercial reactor actually

producing electricity. At this stage, the claims are pretty ambitious. Thorium, subcritical operation, silicon carbide 3D printing, no refueling for decades, and data center deployment. Every piece needs validation, licensing, safety assurance, and customer acceptance. But the direction towards smaller, denser, more deployable power system is very appealing. Then we have news from a clean energy source, hydrogen fuel cell technology, which has not been a big part of the conversation about finding sustainable sources of electricity for energy-strapped data centers. News came out last week that Bloom Energy, a leader in the fuel cell, field won a $5 billion deal with the Brookfield investment company expanding their partnership

to finance power projects for AI infrastructure. Roiders reported that Bloom has already deployed its fuel cell technology to data centers via partnerships with American Electric Power, Equinix, and Oracle. While fuel cell power today constitutes a small portion of total power delivered on the grid, fuel cell technology is getting cheaper to manufacture. So perhaps we'll see fuel cells becoming more a part of the data center power solution as well. A major related topic is water usage by AI data centers, especially as the intended and announced combined investment in AI infrastructure by big cloud providers and hyperscalers approaches a trillion dollar scale wave.

The part of water usage they control directly to cool their data. centers is called the direct use. Then there is indirect use, which occurs upstream at power stations during electricity generation. The Wall Street General reports that among the biggest industry players only meta reports this indirect water footprint. For energy production, the highest water consumers are coal and nuclear power, followed by natural gas. Renewables do significantly better on the waterfront, with solar requiring almost none. Of course, right now, operators must secure energy any way they can to meet immediate demands while actively working to increase the share of lower cost renewables.

The scale of indirect water use depends heavily on the energy mix, but across the U.S., it has historically averaged about 12 times the amount of direct on-site water use, according to an analysis by the Lawrence Berkeley National Lab. That ratio, of course, also depends on how efficient data center cooling is. Now, you might ask why the water is not recycled, and you'd be right. Most of today's data center cooling, unfortunately, relies on evaporative systems. They are efficient because they use the cooling effect of water turning into vapor, but unless that vapor is captured, the water is gone.

That makes the system efficient, but also water heavy. Closed loop cooling systems are coming, with vendors showcasing different approaches and governments increasingly pushing in that direction. But that mostly addresses the direct use side. The indirect use problem has to be solved through the energy system itself. More low-water renewables, more efficient grids, and more careful planning around where these giant AI facilities get built. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of OrionX. Shaheen Khan and Doug Black host the show.

Every episode is posted on OrionX.net. If you like the show, please rate and review it. Thank you for listening.