@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260105

Episode Date: January 5, 2026

- China's access to H200 GPUs - Meta acquires Manus - Sandia's new photonics chip [audio mp3="https://orionx.net/wp-content/uploads/2026/01/HPCNB_20260105.mp3"][/audio] The post HPC News Bytes – 20260105 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 Dog Black of Inside HPC, and with me is Shaheen Khan of OrionX.net. In early December, the Trump administration once again changed policy on AI chip exports, allowing Nvidia to see. sell its second most powerful GPU, the H-200, to China. It was a surprising policy shift, given that earlier this year, Trump first allowed and then banned exports of the NVIDIA H-20 chip, which the H-200 exceeds in power by 6x. Jensen Wong reportedly argued that

selling powerful AI technology to China will inhibit the PRC from developing a strong AI stack of its own. Now a writer named Noah Smith, who used to be an economics columnist for Bloomberg opinion, has come out with a piece slamming that decision, saying it sabotages America's effort to stand up to Chinese power. He argues that it's unrealistic to assume that giving China access to powerful AI chips, even if they're a generation behind, Nvidia's most powerful processors, will keep them dependent on U.S. and Taiwanese chip technology. China, while accepting H-200 imports is sticking to its strategy of developing an indigenous AI capability, and according to Smith, Huawei is not expected to catch up to the H-200 until the end of next year

at the earliest. Well, it all depends, and there are a lot of parameters it depends on, how you see the world, how you assess the competitive landscape, how you manage risks over what period of time, and on and on. And you'll either be right or surprised. It's what military theory would call, quote, situational awareness, end quote, which is what it takes to have effective decision making in the face of complexity, complete with a structured model for, well, awareness, which includes perception, comprehension, et cetera. And in this case, it comes down to what you think you're enabling or preventing and what gain or pain that causes for who. It is not easy, and software and hardware advances can change the dynamics. As you indicated, China will continue

to pursue independence to address what it sees as supply chain risks, and that means building internal capabilities. For example, the PRC announced last week that Chinese chip makers are required to use at least 50% domestically made equipment for adding new capacity. Smith seems to focus on available compute capacity as the one important metric. He asserts that if there were, quote, no AI chip exports to China and no smuggling, the U.S. would hold a 21 to 49x advantage in 2006 produced AI compute, depending on whether FP4 or FP8 performance is used for Blackwell chips. This advantage would translate into a much greater American capacity to train frontier models, support more and better resourced AI and cloud companies, and run more

powerful inference workloads for more capable AI models and agents. Unrestricted H-200 exports would shrink disadvantage to between 6.7x and 1.2x, end quote. There were several hundred AI-related acquisitions in 2025. One report counted 381 transactions in Q1 of last year alone. And there is widespread expectation that the frenzy will continue in 2026. One notable deal that happened just before the end of the year was Meta acquiring Singapore-based startup Manus for over $2 billion. Manus launched its chatbot agent in March of 2025 to great acclaim. It performed better than OpenAI's deep research on the general AI assistance or GAIA benchmark. GAA poses real-world questions to test the AI's ability to reason, plan, and use multimodal understanding and search.

It's a test of progress toward artificial general intelligence. Madness also leads on another benchmark called the Remote Labor Index benchmark, which similarly aims to measure how AI agents handle complex remote jobs. So unsurprisingly, one month after launch in April 2025, it raised a 75 million Series A with a valuation of 500 million, led by U.S.-based benchmark capital with Tencent and other Chinese funds joining in. The 100-person Manus team, including its CEO, will join Mehta and report to the COO. This is Meadow's third largest acquisition, but it's a massive increase in valuation for Manus

compared to just a year ago. Its launch did not become a, quote, deep-seek moment, end quote, like some people expected, but Manus reportedly became profitable quickly with a nice growth curve to 100 million plus ARR in just eight months. AARR stands for annual recurring revenue, a metric that is commonly used for subscription-based businesses. While a 4 or 5x or even 10x increase in valuation within a year for an AI company is not so notable these days and Manus's financial metrics are impressive, what is not.

notable is Manus' trajectory, starting out as a Chinese company and its subsequent, what's the word designization? Anyway, six months ago, Manus closed its China operations, severed all ties with the mainland, and moved it all to Singapore. That is what enabled suitors like Meda to come in and enable the acquisition. So now that path is looking like a model for other Chinese entrepreneurs, presumably as long as Chinese authorities allow it, set it up in China, and grow it, and then move it outside to sell it. The journal Nature Communications has run an article about a new quantum microchip developed by researchers at Sandia National Labs that allows photons to be generated, transmitted, and manipulated on a small silicon chip. And it could

enable the development of portable quantum computers. The chip is about 100 times smaller than a human hair and is produced using standard semiconductor fabrication processors, which means it could be scaled up for commercial production. The key to the chip is enabling stable control over light at exceptionally precise frequencies, helping quantum machines to control millions of qubits at once. This could allow quantum systems to stabilize multiple particles in superpositions, a key challenge on the road to practical quantum computing. Photonics show up across the tech world, so you see photonics research in many places.

Photonic quantum computing is also one of the major modalities in that industry. Companies like Zanidu in Canada, Psy Quantum in the U.S. and Australia, Kondela in France, Quicks Quantum in the Netherlands, and Orca computing in the UK, all pursue photonic quantum computers. Photons are fast and can operate at room temperature, but they also are hard to control and are fleeting. They can turn into energy and disappear just like that. According to the article, the new Sandio technology is its ability to produce very stable

laser frequencies. This is done with, quote, optical phase modulators with microwave vibrations that oscillate trillions of times per second, end quote. And do it all in silicon, which reduces the size to a microchip and the required electricity by something like 80 times. So as usual, great progress, though a lot more is left to be done. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of OrionX in association with InsideHPC. Shaheen Khan and Doug Black host the show. Every episode is featured on Insidehpc.com and posted on Orionx.net. Thank you for listening.