In this issue
The short version
- Nvidia's DSX POD puts seven Vera Rubin racks in a factory-built container and claims deployment in as little as 14 days at a prepared site, down from up to six months.
- AMD and Intel will make oneDNN the default library for x86 CPU acceleration in PyTorch, folding in AMD's ZenDNN, so developers get one optimisation path for both makers' processors.
- China's new guidelines on fusing technological and industrial innovation push AI+ across industry and name frontier fields from quantum technology to embodied AI and 6G.
- OpenAI bought Orchestra Bio, whose agents coordinate drug programmes, and will fold it into its Rosalind life-sciences suite.
October 22 was a day for infrastructure and specialised AI. Nvidia showed how to build an AI data centre in a fortnight, AMD and Intel agreed on a common software path for AI on ordinary processors, China set out how it wants research and industry to work together, and OpenAI pushed further into drug development. Each section explains what was announced and how to use it.
01
Nvidia DSX POD: an AI factory in a container
What Nvidia announced
At GTC in Berlin on October 21, Nvidia introduced DSX POD, a reference design for modular AI data centres built in factories. Each container holds seven liquid-cooled Vera Rubin NVL72 racks, 504 GPUs in all, assembled and tested before shipping so the unit can be connected on arrival. Nvidia says this cuts deployment at a prepared site from up to six months to as little as 14 days; the clock starts only once power, cooling and networking are already in place. Shipping is expected to begin in early 2027 through partners such as Dell, HPE, Lenovo (联想), Schneider Electric and Vertiv, and pricing has not been announced.
Nvidia also introduced an OpenShell reference architecture for governing AI agents, which runs each agent in its own sandbox with limited access, monitoring and policy enforcement.
What it means for teams that rent AI computing
Faster construction mainly helps cloud providers and large companies add capacity sooner, which should ease shortages of the newest chips over 2027. For most teams the practical point is timing: if you plan a large AI project, ask your cloud provider when it expects Vera Rubin capacity in your region and on what terms, and avoid long commitments to older hardware at today's prices.
The OpenShell approach is worth copying now, whatever hardware you use. Run each agent with only the access its task needs, log what it does and set rules for what it may not do. That applies to a single coding agent on a laptop as much as to a data centre.
Get HAI Daily in your inbox
The day's AI news, what it changes, and which tools to try.
02
AMD and Intel unify PyTorch on x86 processors
What was agreed
AMD and Intel announced on October 20 that they will focus on oneDNN, the open-source library, as the default for accelerating PyTorch on x86 processors. AMD will fold its ZenDNN optimisations into oneDNN and join the UXL Foundation, which oversees it, as a steering member. The two companies are also preparing AI Compute Extensions (ACE), new x86 instructions for faster matrix calculations, which will reach PyTorch through future oneDNN updates.
The recommendation came from the x86 Ecosystem Advisory Group the companies formed in 2024. Neither has named the PyTorch release that will include the combined code, given a timeline or published performance figures. The change concerns CPUs only, not AMD's GPUs.
What developers should do
Many AI workloads still run on ordinary server processors: small models, data preparation, embeddings and inference where GPUs are not worth the cost. Today, getting the best speed on both AMD and Intel CPUs can mean maintaining separate settings. Once the unified path lands, a standard PyTorch install should perform well on both.
For now, record your current CPU inference speed on a few representative jobs so you can measure the change when new PyTorch and oneDNN versions arrive. Avoid building on vendor-specific extensions that may be retired. If you want a free place to experiment with models and notebooks, Kaggle provides both. Our open-source ML workflow shows a full setup.
03
China's guidelines on fusing science and industry
What the guidelines say
The General Offices of the Communist Party Central Committee and the State Council issued guidelines on October 21 on deeply integrating technological and industrial innovation, with a company-led system linking business, universities and research institutes. They call for fully implementing the AI+ initiative and orienting national science programmes toward industrial use. They set out three routes: breakthroughs in key technologies to strengthen supply chains; original research to create new industries, including quantum technology, biomanufacturing, hydrogen and fusion energy, brain-computer interfaces, embodied AI and 6G; and advanced technology to upgrade traditional industries.
What it means for companies working with Chinese partners
The guidelines indicate where Chinese public funding and procurement will concentrate. If your business supplies or partners with Chinese manufacturers, expect more demand for AI applied to engineering and production data, and for joint projects with universities. Prepare short case studies showing measurable results, such as defect rates or energy use before and after, because that is the evidence these programmes ask for. Tools in the Data Analysis category help build those comparisons.
04
OpenAI buys Orchestra Bio for drug development
What OpenAI announced
OpenAI acquired Orchestra Bio, a start-up founded in 2022, and will add its team and technology to Rosalind, OpenAI's life-sciences model, and Rosalind Workbench, the tool that connects it with laboratory software and data. Orchestra does not design molecules or find drug targets. Its AI agents track a drug programme's progress, risks and hand-offs from the lab to clinical trials, and help coordinate drug companies with the contract research and manufacturing organisations they rely on. Terms were not disclosed.
The deal extends OpenAI's life-sciences work from scientific reasoning into running drug programmes.
Where AI helps in research operations
Much of the delay in research comes from coordination rather than science: waiting for data from partners, missed hand-offs and status reports assembled by hand. Even without specialised tools, you can apply AI here. Keep project updates in one structured document, and ask an assistant such as ChatGPT each week to list overdue items, risks and decisions needed. Keep confidential data out of general assistants unless your organisation has approved them.
For the science itself, tools such as Google's AI Co-Scientist and Elicit help generate and check hypotheses against the literature. The Healthcare AI category lists more, and our AI academic research workflow shows how to keep sources traceable.
What to try next
Put one change into practice
Give every agent you run only the access its task needs and keep a log of its actions. Record your current CPU inference speed so you can measure the coming PyTorch changes. If you run research projects, try a weekly AI-generated risk and status summary.
Explore Code Generation, Healthcare AI and Data Analysis to build your shortlist.
How we put this issue together
The H.AI Tools Team checks each story against the official sources linked in the text and removes claims we could not confirm. Product links lead to H.AI Tools directory profiles. Unless an issue says otherwise, product notes describe vendor documentation, not our own benchmark tests.