@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260119

Episode Date: January 19, 2026

- OpenAI and Cerebras - Taiwan Pledges $250B US investment - TSMC in Arizona - Intel 18A fab and Apple - Inside an AI lab online movie is a 300-million-view hit [audio mp3="https://orionx.net/wp-content/uploads/2026/01/HPCNB_20260119.mp3"][/audio] The post HPC News Bytes – 20260119 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Hi, everyone. Welcome to HBC Newsbytes. I'm Doug Black of InsideHPC. And with me is Shaheen Khan of OrionX.net. Cerebrus, maker of the dinner plate-sized wafer scale GPU, introduced in 2019, has gradually moved from strange novelty, what some have compared to, a science. experiment, increasingly toward normality, in an HBC AI world-driven to find faster processing wherever and however it can be found. Shaheen, you and I have kept our eyes and ears open for

a breakout moment for Cerebrus, as customers, including the national labs, have reported eye-popping processing with the Waifer Scale engine, and that moment may have arrived. The Wall Street Journal reported last week that OpenAI agreed to a $10 billion deal, to purchase up to 750 megawatts of computing power over three years. This has been a good period, by the way, for the AI processor group that includes GROC, Sanbanova, along with Cerebrus. Intel reportedly is in advance talks to acquire Sanbenova for about $1.6 billion. And NVIDIA recently entered a non-exclusive licensing deal with GROC for roughly $20 billion.

It all shows the continued insatiable appetite for AI hardware and apparently also the ability of big players like OpenAI to continue to fund it. $10 billion is another nice big number. Assuming Cerebris has provisioned for capacity, the deal would also minimize parts shortages that are rampant in the industry. So back at envelope calculation says $10 billion of hardware in a 750 megawatt data center, comes out to about $500 million per year of power and cooling costs plus or minus, and depending on power costs, whether the system runs continuously, and power usage efficiency, a metric that goes by the wonderful initialism, maybe even an acronym P-U-E, which is expected to be very good for Cerebris,

since it has said that it is 2.2 to three times more power efficient than the NVIDIA B-200 blackwell-based systems in terms of performance per watt for AI training and inference. The data center would probably need some of their SwarmX networking, memory X disaggregated memory, and a bunch of SSD and HDD storage too. But even after all that, the power usage efficiency should be very good and below 1.2, maybe pretty close to 1.1. On the geopolitical front, Taiwan agreed to a trade deal with the Trump administration intended to increase U.S. production of semiconductors in exchange for a cut in tariffs. The deal calls for TSM to build several new factories to its cluster in Arizona as part of a $250 billion investment in the U.S.

According to the Commerce Department. In exchange, the U.S. is cutting tariffs on Taiwanese goods to 15 percent from 20 percent. and exempting Taiwanese chip companies like TSMC that are investing more in America. TSMC's expansion in the U.S. continues primarily focused on meeting internal U.S. demand. But high-end chips right now get exported back to Taiwan for advanced packaging, say for chiplet integration, and TSM's customers in the U.S., like Apple or NVIDIA or AMD, build full systems and sell them globally. So the chips will also get exported to other countries like, again, Taiwan or Korea or Europe. Over time, the capacity can serve other countries too.

At the moment, TSM has been producing 4 nanometer to 5 nanometer chips in its Fab 1 in Arizona since late 2004, early 2025. It's working to mass-produce 3 nanometer chips in its fab 2 by the second half of 2027, and is targeting 2028, 2030 for 2 nanometers A16 chips in its FAB 3, which it started building last year. And it expects to begin production in its advanced packaging FAP called AP1 in 20208. We talked last week about progress made by Intel with its advanced 18A FAB in Arizona, indicating 1.8 nanometer class technology.

Now there's another interesting piece of news regarding, Intel's Foundry Business. According to Mark Lapidus in the publication Semi-Ecosystem, Intel has won some foundry business from Apple. The loss of X86 chips for Max five years ago was of a piece with the drumbeat of negative news coming out of Intel over such developments as delays and shipments of CPUs canceled products, missing the opportunity in the GPU market and lagging behind NVIDIA and AMD on performance. But Cheyenne, whether the credit for the recent positive news for Intel Foundry Services can be credited to former CEO Pat Gelsinger, who vowed to revive Intel's chip fab business soon after returning to the company in 2021, or to Lipu-Tan, who replaced Gelsinger 10 months ago, or to both

of them, is hard to say. Well, regaining global leadership in chip manufacturing was viewed by many as simply impossible. So Pat Gelsinger's plan looked highly ambitious, but it looked like it was going to work a couple of years ago, and the progress has continued. It's a big effort. It's really a relay marathon that has to work over time and through many people at every level. And all indications are Lip Bhutan has continued the focus on engineering, but has also increased the company's valuation, which has benefited from deals with Nvidia, selling a stake to the U.S. government, but also continued cutting of projects and staff. Progress with the manufacturing business is the big geopolitically important part, though.

And winning a deal with Apple is very welcome news, obviously for Intel, but also for US-based manufacturing of high-end chips. The news quoted a research note from Key Bank Capital Markets, saying Intel would use the 18A FAP to manufacture low-end M-series processors for MacBooks, and iPads and the project would go into production in 2027. There's a new documentary out who's released last month that has gotten, if you can believe it, 300 million views. The movie called The Thinking Game, quote, is flattering to quote from a review in the Wall Street Journal, but it's also genuinely revealing and it's clearly reaching an audience that never expected

to be so enthralled by research about protein structure. By the end, to quote the reviewer, I had a better understanding of the people building their technology with the power to transform civilization than I did 84 minutes earlier. The documentary revolves around work on protein structures by Dennis Haseebis CEO of Google DeepMind and a Nobel Prize winner. Into the most replayed scene in the movie, Haseebis walks into a conference room to discuss the future of AlphaFold when a college. mentioned they can now easily predict all known protein sequences, Hasebus looks up from his phone. Why don't we just do that, he says. We should just run every protein in existence and then release that. So they did. The epiphany to fold everything led to opening the alpha fold system to the entire scientific research community and,

spoiler alert here, I'm not going to spoil the ending. Very nice. Well, there's even a Wall Street General article about this, which really is as much about the film crew and their history making documentaries exciting as it is about AI or the Google team. They reportedly started filming in 2018 and did it for six years and released the movie just a couple of months ago,

so the view count is pretty impressive, but it also shows the pent-up demand by everyone to better understand AI and the people behind such important advances. Google funded the movie and owns it, so as the article says, quote, some reviewers have dismissed the film as glorified advertising, end quote. I have to say that that is usually where I land, but you know this better than anyone, Doug, that the very act of covering a new technology is, well, publicizing it. And most often, journalists also need to properly attribute it to the people and companies that build a technology. So really, the key is to make it factual, informative, and maybe even a little entertaining, which it looks like this movie does.

All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of Orion X in association with Inside HPC. Shaheen Khan and Doug Black host the show. Every episode is featured on insidehpc.com and posted on Orionx.net. Thank you for listening.