@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20260831

Episode Date: August 31, 2026

- Brazil Plays Both Sides of the AI Stack - Nvidia's AI Economics Expand Up the Stack - Nvidia + HuggingFace? - IBM Puts Arm Inside the Mainframe - Hot Chips roundup - Quantum Computing roundup [audio mp3="https://orionx.net/wp-content/uploads/2026/08/HPCNB_20260831.mp3"][/audio] The post HPC News Bytes – 20260831 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 is Shaheen Khan. We've been seeing more news from Brazil about that country building out its advanced technology and AI capacity. The country is putting about $44 million into its national AI infrastructure, and the interesting thing is how it is dividing its AI resources, if you will. One major project in Rio de Janeiro involves China's Huawei and I Fly Tech, while a new Brazilian AI supercomputer is expected to use Nvidia technology.

Brazil is also pursuing greater national cloud capacity and Risk 5 semiconductor development. The government is explicit about its strategy, build domestic air, AI capability without becoming dependent on one country or technology supplier. So this is bigger than buying a couple of supercomputers. Brazil is building an AI stack while deliberately maintaining relationships with competing U.S. and Chinese technology ecosystems. Let's also note that Brazil ranks seventh in the world for population and tenth for its economy.

For most countries, sovereign AI may increasingly mean how. they manage dependencies rather than eliminating them. And this is very much the type of development that we predicted a few episodes ago. What Brazil is doing is an example of technological non-alignment as technology driven by AI becomes an imperative for economic competitiveness. Some countries outside of the United States-China rivalry may want to be or at least look neutral if they can manage it. They want access to technology, but also bargaining power and some control over their own infrastructure. But that option has risks too, because it creates silos and geopolitical pressure can make mixed technology strategies hard to sustain. Invitya keeps reporting better than

expected quarterly results, and surprisingly, this keeps surprising the investor community. But I suppose if you keep saying Nvidia can't do it again, eventually, you'll be right. The company reported another extraordinary quarter, $96.2 billion in revenue, including $89 billion from Data Center, with overall revenue more than doubling year over year. For the record, the stock had declined to $1.90 in late July, but last week on earnings announcement day, shares returned nearly to their record high of $2.35. At almost the same time as the earnings announcement came reports that NVIDIA maybe is in negotiations to a $2.00,000, acquire Hugging Face for around $13 billion. The reporting isn't consistent yet about whether an agreement has actually been signed, and neither company has confirmed the transaction. So we're treating this as a reported deal rather than a completed one. Hugging Face has become an important

piece of the AI infrastructure for sharing, finding, and deploying open models. Put the two developments together, and Nvidia's reach increasingly extends from accelerators and systems into software models and developer ecosystem. The rapid and consistent growth in AI hardware begs the persistent question, like you said, of when it will all slow down or stop. And yet, quarter after quarter, AI continues to be seen as the imperative I just mentioned by more and more governments and companies around the world, and NVIDIA continues to produce the receipts, amassing more financial power and technological influence in the process. Now, for NVIDIA, paying $12,000,000,000. $12.9 billion for Hugging Face is like a person spending two-week salary on something.

It's roughly two weeks of Nvidia revenue. But buying HuggingFace suggests one way that that power can be used. And it's a nice move as Nvidia strategically goes up the stack, morphing from a chip company into a systems, software, and AI ecosystem company. Buying Hugging Face is especially valuable because it is the leading repository and distribution platform for AI development, including Open AI models, weights, data sets, tokenizers, and related tools. So this is kind of like when Microsoft bought GitHub for traditional code. It also prevents anyone else from buying it and possibly steering it away from Nvidia's interests. The acquisition, if it happens, also could augment Hugging Face with serious compute resources so developers can

test and deploy models right away and potentially add AI engineering services around it, like a GitHub co-pilot sort of a thing. Nvidia has its own very good foundation models, particularly around physical AI, but Hugging Face is a good hedge against a handful of close-source model providers controlling a very key layer. It helps drive self-hosted and distributed AI developments too, which is an important and growing market for Nvidia. At the same time, large AI companies like OpenAI in Anthropic and Google and Microsoft

are increasingly designing their own AI chips and accelerators. And while Envideo has investments in many of these companies, it's probably healthy to make sure the industry continues to have a lot of options that run on Invideo Harbor. IBM used last week's Hot Chips Conference in California to disclose an unusual processor for future Z mainframes and Linux 1. systems. Each processor core is designed to execute both IBM's Z instruction set and Arm-Arc 64 natively. These aren't separate Arm and Z cores packaged together. The same cores can execute either architecture,

allowing Arm-native Linux environments to run alongside Z operating system and Linux on Z. The processor is built on a 2-nometer process running at 5.7 gigahertz and also incorporates AI and I.O. acceleration. The immediate objective is software access, bring the very large arm software ecosystem into the mainframe environment while retaining the reliability, security, transaction processing, and large memory characteristics that differentiate these systems. IBM mainframe chips have historically been very strong, and this is another very cool technology in the many decades of evolution of this architecture. Mainframes have always been highly reliable and integrated environments, but it's been clear that most new applications are developed on merchant architectures, like X86 or ARM, and gradually also Risk 5. Those are the main instruction sets.

So for the mainframe to natively run software that's built for ARRRRM, Linux opens up a wide range of applications and reignites the software catalog for the mainframe. And the chip is fully bilingual, able to run arm and legacy mainframe workloads natively on the same cores switching between the two instruction sets within nanoseconds, according to IBM. If there is a broader architectural point here, it is the new flexibility of mixing instruction sets, which might reduce the need to port or recompile some applications. and make it easier to bring new software into established systems. Speaking of hot chips, the conference gave us a useful snapshot of where processor design is going.

Open AI showed measured results from Halapeno, its custom inference accelerator. Microsoft disclosed more of Maya 200, its second generation internal AI accelerator. Sci-5 introduced Big Sky, a rackable Risk 5 development server intended to let data center developers port and tune real software. Intel showed its next generation Zeon architecture alongside a GPU aimed at inference, and arm detailed AGI moving beyond licensing CPU intellectual property towards supplying a complete data center processor. These are very different products, but the common theme is specialization. Cloud providers, AI companies, processor vendors, and IP suppliers are all moving deeper into systems architecture rather than assuming a small number of general purpose processors will serve every workload.

Yeah, there were a few shifts that stood out for me. First, as we said last week, Custom Silicon is becoming normal. Google, Amazon, Microsoft, Meadow. Tesla, and now Open AI and Anthropic and others are all designing their own chips. Second, Risk 5 continues its march, now pushing into enterprise class hardware, while Arm itself is moving from supplying IP towards providing finished server silicon. Third, the chip is increasingly the wrong unit of analysis. Memory, interconnects, networking, packaging, power, cooling, and software all impact how designs manage

the trade-offs that they must make. And AI makes all of the above especially visible. So we have more kinds of processors, but each one is increasingly designed as part of a tightly coordinated system. Four quantum developments fit together unusually well. The National Science Foundation is putting more than $290 million into eight Quantum Leap Challenge Institutes,

covering areas including error correction, networking, hardware, algorithms, and sensing. Separately, researchers demonstrated hybrid workflows on the production supermuck-NG supercomputer, treating an attached quantum processor as a scheduler-managed accelerator. Treasury launched a public-private quantum readiness task force, focused on moving the financial sector towards quantum safe cryptography, and IBM completed its acquisition of HRL laboratories, adding silicon spin-cubit expertise,

alongside its superconducting approach, together with capabilities in cryogenics, packaging, and control electronics. Research capacity, operational integration, defensive preparation, and industrial development are all advancing at the same time. The quantum ecosystem is starting to build the structures and surrounding infrastructure that it needs while waiting for useful quantum systems.

And there's good progress towards that too. The number of modalities remains large, photonics, neutral atoms, trapped ions, superconducting, silicon spin and quantum dot, topological, diamond-based, and quantum annealing. But hardware companies are investing in manufacturing, packaging, and controls, and some of the roadmaps are moving noticeably forward. Qera, for example, is now talking about a fault-tolerance system in 2008, followed by more than a thousand logical qubits around 2028 or 2029. Microsoft has also accelerated its roadmap and is targeting a scalable, practical system by

29. It's also gratifying to see HPC emerge as an important, if not the first, stop for quantum computers. Supercomputing centers need schedulers and workflows that understand quantum processors, and the research you mentioned points to that direction. And as those roadmaps move forward, banks and governments need quantum safe cryptography migration plans earlier than once envisioned. So the progression remains, microcontroller units or MCUs at one end, moving to CPUs, GPUs, and ultimately QPUs for quantum computing at the other end.

And the bookends will be the site of a lot of cool technology action. By the way, we haven't talked much about the diamond-based approach. It uses atomic defects in diamond called color centers, which are optically addressable and whose electronic spin states can act as cubits. The surrounding diamond crystal provides a very stable environment that helps protect those quantum states. And one of the attractions is that some of these systems can operate at room temperature. 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.