Microsoft announced Maia 300, its next-generation custom AI inference chip, and plans to begin multi-gigawatt Azure deployments in 2027. Microsoft says a rack configuration with 96 chips is planned; one account says it will use about 167 kW of power, while another says the configuration promises seven times the token throughput of its predecessor. The company plans to make Maia 300 its primary AI inference system for Microsoft products and services, with production workloads scheduled to begin in 2027.
WASHINGTON and SAN FRANCISCO, Oct. 9, 2026 — The Department of Energy’s (DOE) Office of Science, the National Institutes of Health (NIH), and Biohub signed a Memorandum of Understanding (MOU) to advance the frontier of Artificial Intelligence (AI) in biology, deepen our…
IonQ has agreed to acquire Elisity, a Zero Trust security company, in an all-stock transaction valued at up to $1.3 billion. The deal would add Elisity’s least-privilege access technology to IonQ’s security portfolio. IonQ says the acquisition combines its post-quantum cryptography and quantum key distribution capabilities with Elisity’s enterprise security technology.
AMD introduced Helion, an open rack-scale reference design for agentic AI infrastructure, at the OCP Global Summit. Its two configurations include a CPU-only rack with up to 32,768 Zen 6 cores and an enterprise design combining EPYC processors with up to 16 Instinct MI350P GPUs. AMD says the blueprint builds on its Helios rack-scale platform and is based on an Open Compute Project design. Partners are expected to offer CPU-only and mixed CPU/GPU configurations in November.
Hewlett Packard Enterprise announced that its Spaceborne Computer-4 will run AI experiments and what the company describes as the first high-performance computing benchmarks on the Moon. The compute module is slated to travel aboard Astrolab’s FLIP rover, with Voyager Technologies’ Griffin-1 lander delivering it near the lunar south pole. The mission is planned for late 2026 under NASA’s Commercial Lunar Payload Services initiative. Tests include supercomputing benchmarks, language-model retrieval-augmented generation, vision inference and LiDAR processing.
Meta published details about Prometheus, its New Albany, Ohio, supercluster used to train the Muse family of AI models. The system combines Nvidia Grace Blackwell-based clusters, with a Vera Rubin-based cluster expected to join; Meta says Prometheus will eventually scale to a gigawatt of capacity and about a million accelerators. Meta also described measures to keep training running through hardware failures, including a pool of redundant GPUs: more than 11 percent of the GPUs are dead or idle as replacements, one account says.
Socionext is expanding its use of agentic AI into physical design for large-scale system-on-chip projects using advanced process technologies, working with Cadence on the effort. The company says it already applies AI agents to front-end design, including design and verification tasks, engineering analysis and decision-making. The broader initiative aims to improve engineering productivity and design quality while allowing engineers to focus on higher-value design decisions.
The Barcelona Supercomputing Center’s AI Factory Accelerator Programme has moved into full operation after a pilot, giving startups access to sovereign computing resources, technical support and training. NVIDIA says some BSC AI Factory services are now accessible through its Inception portal, which also offers training, business mentoring and venture-capital connections. The programme supports companies developing technologies for healthcare providers, utilities and other sectors.
A feature article argues that Ethernet will play a growing role in AI training as large GPU clusters depend on reliable communication. Even small amounts of packet loss can degrade performance, and synchronized operations may be held up by a delayed GPU or data transfer. The piece describes training jobs that distribute work across GPUs, which exchange gradients, parameters or activations before proceeding. It frames network performance as a constraint on the efficiency of increasingly large AI systems.
AMD and Intel are combining their x86 CPU optimizations for PyTorch in oneDNN, an open-source deep-learning library governed by the Unified Acceleration Foundation. Following a recommendation from the x86 Ecosystem Advisory Group, the companies aligned on oneDNN as PyTorch’s default open-source library for x86 acceleration. AMD is integrating its ZenDNN optimizations into the library and has joined the UXL Foundation as a Steering Member. The shared path is intended to make PyTorch workloads easier to build, maintain and deploy across AMD and Intel processors.
The EOSC Federation and EOSC Node Finland entered production at the annual EOSC Symposium in Florence, held Oct. 14–16, 2026. Finland is among the first nodes to go live, as the federation’s first 14 production nodes arrive 11 months after its November 2025 launch. Fourteen additional nodes launched in 2026, bringing the network to 28. Finland has developed federation services including access and authorization infrastructure, service catalogues and a helpdesk, while working to onboard national research resources.
The U.S. Department of Energy selected 75 PhD students from 58 universities for its 2026 Office of Science Graduate Student Research program. The researchers represent 30 states and Puerto Rico and will receive training and access to facilities, expertise and resources at DOE national laboratories. The program supports doctoral research connected to the Office of Science’s mission and prepares participants for careers advancing scientific and energy research.
EuroHPC JU and Greece’s GRNET inaugurated the DAEDALUS supercomputer in Lavrion, Greece, at a ceremony on Oct. 19. DAEDALUS is the most powerful supercomputer ranked in Greece; in its first appearance on the June 2026 lists, it placed 31st in the TOP500 and 23rd in the Green500. The inauguration was attended by Greek Prime Minister Kyriakos Mitsotakis and European Commission and Greek government officials.
NVIDIA ran a 104-qubit simulation of a fault-tolerant quantum processor on JUPITER, Europe’s first exascale supercomputer at the Jülich Supercomputing Centre. The computation used 8,192 GH200 GPUs and modeled magic state cultivation under realistic hardware noise, including coherent phase errors. NVIDIA calls it the largest comprehensive simulation of a fault-tolerant quantum processor to date and says the method requires 90% fewer qubits than previous approaches. The results are intended to help researchers assess future quantum-computer designs before building them.
Zapata Quantum appointed mathematician Peter Shor and three other scientists to its Scientific Advisory Board on Oct. 20, 2026. The new members are Shor, Paler, Devitt and Izmaylov, and will advise on the company’s scientific and technical direction, including development of its Quantum Pilot platform. Shor said he sees a need to develop algorithms as quantum hardware advances toward fault tolerance.
Accenture Federal Services and Google Public Sector announced Gemini launchpads for federal agencies, a service intended to help move agent experiments into operational use. The centralized hubs will provide Forward Deployed Engineers to work alongside agencies on building, deploying and scaling agents based on Gemini. Agencies can use the launchpads to connect Gemini securely to mission datasets, prioritize use cases, deploy agents and train staff, while leadership tracks performance and mission impact.
Alice & Bob reported results from the first phase of France’s PROQCIMA quantum-computing program, including three operational cat-qubit demonstrators built and tested on real hardware. The French government program aims to develop two universal quantum-computer prototypes, each supporting 1,000 logical qubits, for industrialization by 2032. Five companies entered the 10-year competition, which will narrow to three candidates after four years and ultimately select two beyond year eight.
Researchers at the U.S. Department of Energy’s Argonne National Laboratory completed two of the largest cosmological simulations ever performed using the Aurora exascale supercomputer. Each modeled the evolution of 13.4 trillion dark matter particles across a region 28.6 billion light-years wide. The gravity-only simulations are intended to help scientists investigate cosmic inflation and analyze observations from DESI, SPHEREx and other major sky surveys.
Quantinuum says an error-correction result from its Helios quantum computer strengthens its confidence in a 2029 target for Apollo, its planned fault-tolerant system. The company reported 99.9995% logical memory fidelity, or roughly five failures per million cycles, and an average twelvefold reduction in logical memory errors as color-code distance increased. It estimates the approach could support 100 logical qubits with about 30,000 physical qubits, and says the results clarify a route toward scalable error correction.
Great Sky’s superconducting neural-network chip passed tests processing data from a fusion reactor, advancing work on hardware intended for fusion applications. The chip uses superconducting computing and optical connections to run an analog neural network, rather than emulate one digitally. The Genesis Mission project aims to build a larger chip for further fusion testing next year. Great Sky says its approach could process data at higher rates than conventional chips, including GPUs, for the same energy.
EuroHPC JU has elected Sinéad Ryan as Chair and re-elected Maria Girone as Vice-Chair of the EuroHPC User Forum Coordination Group. The elections followed the completion of the previous leaders’ two-year mandates and took place during EuroHPC User Days 2026. Ryan and Girone will serve for the next two years as co-chairs, continuing the group’s work to foster dialogue and collaboration across Europe’s HPC user community.
Great Sky, Brookhaven National Lab, Princeton Plasma Physics Laboratory and the DIII-D National Fusion Facility have partnered to develop a superconducting AI chip for fusion-plasma control. Their chip uses analog neural networks built with superconducting circuitry and optical communication, rather than conventional silicon transistors. The team demonstrated three tasks relevant to tokamak operation, including classifying anomalies and predicting disruptions using DIII-D data sampled at one million times per second. The tests showed the chip could keep up with that data rate.
An article describes applications of El Capitan, the supercomputer at Lawrence Livermore National Laboratory, in public safety, medicine and energy. It says scientists used the system to create a tsunami “digital twin” that combines ocean sensors with physics-based models to estimate localized coastal risks in under 0.2 seconds. The article contrasts this with traditional emergency models, which it says can take minutes or hours, and presents the faster calculations as a way to improve warning and evacuation decisions.
New York University has deployed Torch, a supercomputer designed to expand computing capacity for its academic research community. It delivers 10.79 petaflops on the LINPACK benchmark, five times the performance of NYU’s previous Greene cluster, and is described as New York State’s highest-performing academic computing system. Developed with HPE and NVIDIA, Torch brings computing resources on campus for faculty, researchers and students.