HPC News Bytes – 20260803

**Doug Black** (0:05)
Welcome to HPC News Bytes, a weekly show about important news in the world of supercomputing, AI, quantum computing, and other advanced technologies. Hi, everyone. Welcome to HPC News Bytes. I'm Doug Black, and with me is Shahin Khan.
Amidst a welter of recent AI security news, NVIDIA and a group of technology, cyber security, cloud, enterprise software, and open source organizations have launched the Open Secure AI Alliance. Participants include the Linux Foundation, Microsoft, IBM, Hugging Face, Cisco, CrowdStrike, CloudFlare, Red Hat, Palo Alto Networks, and HPE, among others. Early contributions address several layers of the agent stack.
Cryptographic workloads, identity, safer model formats, digitally signed software patches, multi-agent vulnerability scanning, secure coding workflows, and tools for tracing and governing agent behavior. NVIDIA is contributing models, weights, data, and its open source NOOA research framework. That acronym stands for the NVIDIA Labs Object Oriented Agent.

**Shahin Khan** (1:20)
Well, setting aside the question of why these capabilities are not already present and what everybody provides, the Alliance is about security, but it's also very much about the open versus closed AI debate. It basically says that cybersecurity defense is a community project. It requires transparency. It needs systems that defenders can inspect and modify and self-host. And since AI needs cybersecurity, it would need the open model. Hosted models can be opaque and may refuse legitimate requests, because as we've discussed here before, malicious commands and genuine debugging requests can look identical.
Open models give security teams greater control, although they also give attackers access to capabilities that might be used in malicious ways. But historically, they have been proven to be more secure. The more difficult task is securing the complete agent stack. Model weights are only one layer. There's also the usual requirements of identity, permissions, isolation, provenance, logging, and the entire software supply chain, all the way to applications and users. And in this case, the users are increasingly AI agents.
We've joked about, quote, AI agents being people too, kind of a thing. But it's a pretty good way to point to the complexity as AI agents are granted or gain more autonomy. Anyway, shared standards in those areas could become foundational for AI agents and are important. Another consideration is how a coalition of so many large companies can work together productively on technologies, and in this case, also to act as a lobbying body to influence policy.

**Doug Black** (3:08)
The Euro HPC Organization has launched its formal call for consortia to build and operate European AI gigafactories. These facilities are intended to sit above the existing AI factory program and scale, combining very large accelerator installations, data infrastructure, networking, software, and access mechanisms for European model developers and industrial users. The call asks prospective consortia to address financing, construction power, operations technology sourcing, and long-term commercial viability. It also reflects Europe's preference for public-private infrastructure rather than relying entirely on foreign cloud providers. The program remains a procurement and policy framework rather than an operating system as of today.
Many questions remain for this effort, including how much it will cost, accelerator supply, power availability, delivery schedules, and how many installations will be justified by European demand.

**Shahin Khan** (4:13)
Europe is marching on with its unique take on AI. It has worked on regulating data and then AI, built large-scale systems around the European Union, and is now moving towards financing the infrastructure that AI data centers need. The Gigafactory model recognizes that frontier-scale AI requires hard infrastructure, like land, power, cooling, and system infrastructure, like systems and clusters, and that all of it needs capital at one end and customers at the other. As AI development has turned into a race, everyone's challenge has been execution speed, either that or an inability to explain why a slower path will be okay.
While it has its defenders, Europe is typically seen as lacking speed, even within Europe. Public-private partnerships have emerged as a way to bring in more capital and efficiency into the process. These structures can aggregate resources and support strategic autonomy, but they can also create fragmented governance and slow procurement. The European program will be strategically useful, even if Europe does not match the scale and scope of American investment. It can support regional models, scientific computing, defense applications, industrial AI, and negotiating leverage with external suppliers.
Now, much of the system infrastructure comes from non-European companies and that will continue. So a deeper question is what sovereignty means, and sovereignty cannot mean complete technological independence. The US comes closest to having that, but it's not 100 percent. Having your own models, weights, jurisdiction, operational control, data governance, and assured access are arguably more important considerations.

**Doug Black** (6:04)
Speaking of sovereignty, our next story involves the question of whether sovereign AI is not just a matter of policy, but also of architecture. A new archive paper presents a German-English open-source AI foundation model designed around European sovereignty and efficient deployment. It uses a mixture of experts' architecture with roughly 30 billion total parameters. It also combines transformer and Mamba-style components, which the authors say keeps the inference cache comparatively stable as context length grows.