The Register

The European Commission has reached the preliminary position that Azure and AWS should be designated as gatekeepers under the Digital Markets Act (DMA). The gatekeeper designation would mean requirements imposed on the cloud giants, with fines of up to 10 percent of worldwide turnover if those requirements are not met. According to the Commission, AWS and Azure, "the largest and second largest cloud computing services in the EU respectively," are a gateway between businesses and their customers in the bloc. "They both have vast and entrenched user bases and appear to benefit from lock-in effects and high switching costs, in addition to a large ecosystem." Although the cloud giants did not meet the DMA's quantitative thresholds for designation (such as user numbers), their market positions have attracted scrutiny. Should the gatekeeper designations stick, obligations regarding interoperability, access to data, and competition would apply. The view is preliminary at this stage, and Amazon and Microsoft have the opportunity to respond before anything becomes final. A Microsoft spokesperson told The Register: "We continue to engage constructively with the Commission. The cloud sector in Europe is innovative, highly competitive and an accelerator for growth across the economy." The spokesperson added: "We remain concerned that ignoring the growing power of Google Cloud and Gemini will tilt the market in a harmful way." AWS also disagreed with the Commission's preliminary position. A spokesperson told The Register: "The Commission's preliminary findings disregard the breadth of cloud services available to European customers and risk deterring European investment and innovation. AWS faces healthy competition and customers across Europe have more choice, lower prices, and greater flexibility than ever before. "The EU already has comprehensive cloud regulation through the Data Act, and adding another heavy layer of overlapping regulation under the DMA undermines European competitiveness and access to cutting-edge information technology. We will continue to engage with the Commission to reach the right outcome for customers and Europe's digital future." Other parties responded more positively. A spokesperson for the Open Cloud Coalition told The Register: "Our members welcome the Commission's preliminary finding naming Microsoft and AWS as cloud gatekeepers. We particularly note the finding that existing customer lock-in may fuel enterprise AI, a development that mirrors long-standing market concerns over Microsoft's licensing and ecosystem practices. "Moving quickly to deliver remedies is now a priority to ensure choice and growth for European cloud customers." In 2024, Microsoft described the Open Cloud Coalition as a lobbying group for Google. Henna Virkkunen, executive vice-president for tech sovereignty, security and democracy, stated: "Cloud services have become a cornerstone of Europe’s economy – and a prerequisite for AI – with over half of EU businesses now relying on them, combined with record investment in public cloud infrastructure. "Given their central role in Europe's digital future, these services must operate in fair, open and competitive markets that foster trust and secure Europe's tech sovereignty." Should the preliminary findings be confirmed and Microsoft and Amazon be designated as gatekeepers for their cloud services (they already have gatekeeper status for other services), the pair will have six months to ensure compliance with the DMA's obligations. ®

Collabora has released version 26.04 of CODE, the company's web-based office suite. CODE stands for "Collabora Online Development Edition" – "Development" indicating that this is a preview of what will be coming soon in Collabora Online, its web-based function-rich office suite. The announcement is Collabora staking its claim to a slice of the increasingly competitive online office suite market. What's new Depressingly, the first bullet point in each section of the release announcement – for the Writer word-processor, Calc spreadsheet, and Impress presentation module – is improved AI integration. We asked the company about this, and Collabora assured us that it's off by default. "Basically, lots of partners and customers have been asking for the option for a while, so it was added, but we have ensured that people can decide whether or not to use it and how they use it. To be clear, AI assistance isn't present in the product unless you add it through File > Options or on the admin side through the server side settings panel." As an example, Collabora noted that the EU's proposed Cloud and AI Development Act (CADA) has AI all over it. There are other functional improvements here as well as optional Automatic Idiocy. For instance, there's Markdown support. If that sounds faintly familiar, it may be because that was one of the standout highlights of LibreOffice 26.2, which we covered back in March. Writer now has a multi-page view; the "Track Changes" functionality has been enhanced in multiple areas, including comment handling. Comments can now be exported to PDF along with the main content. The Navigator sidebar now has search, and the Formatting sidebar now has inline preview. The spreadsheet has per-user views, smarter formula error handling, calculated values in pivot tables – yes, the wider LibreOffice family does have pivot tables – as well as better formatting with bundled templates, colored tabs, better JSON import, lots of new, modern spreadsheet functions, smarter drop-down lists, and more. Impress presentations have been similarly revamped. The various components' ribbons have been reorganized, and the suite makes better use of limited screen space. Effort has gone into making the suite more accessible, with controls having more descriptive names, descriptions added to the tabs of dialog boxes, lots more keyboard controls, and more. These improvements not only help users with disabilities: more and better keyboard controls help power users navigate much quicker too. Family ties – and family feuds Some of the changes and improvements are things that mirror recent improvements in LibreOffice, because there is substantial overlap between Collabora's office suites and The Document Foundation's LibreOffice. There used to be a fairly clear split: the Document Foundation maintained two versions of LibreOffice – a current release and a slightly slower-moving stable release. LibreOffice is a completely local app, with versions for all the major desktop OSes. For organizations wanting support, the Document Foundation recommended multiple independent vendors, including Collabora, which did a lot of the development work, and offered its own version of the codebase, Collabora Online (COOL) – the same codebase, adapted to work as a web app. You install and run it on a web server in conjunction with an online cloud storage tool, such as Nextcloud. CODE is a preview of what's going into the next version of COOL. We last looked at CODE in 2022. The distinction between LibreOffice and Collabora Online became much less clear in November 2025 with the release of Collabora Office for Desktop, which is the COOL web suite bundled up with a local server and installer, so it can run on a standalone, non-networked machine (this is, incidentally, how ONLYOFFICE works: web apps, but running locally). As we covered back in March, the Document Foundation has responded by reviving its web-based version of the suite, LibreOffice Online. There's a fully local FOSS office suite (and, for completeness, the moribund OpenOffice), plus two online versions, one of which can also be run locally. The Document Foundation is in active competition with Collabora. There is blood in the water of the pool where all the descendants of StarOffice swim. Why here, why now? When we looked at CODE in 2022, cloud office suites were something of a niche interest. Now, though, this sector is rapidly hotting up – especially since the re-election of Donald Trump to the presidency of the USA. That has ignited great interest in European digital sovereignty, as we reported from the Open Source Policy Summit in February (oh, and Microsoft is going to ax Office Online too). There is money to be made here. Money from national and governmental organizations is even more tempting. Because those folks are the ones who print the money, there can be more of it: the public sector pocket is deep. The FOSS world is always short of money. Selling FOSS to company directors who only understand payments and contracts is tricky: it doesn't work by the only rules they understand, as Andrew Nesbitt's article Open Source vs the Invisible Hand neatly dissects. Selling FOSS to the public sector instead is an exciting new market, and it's resulting in a feeding frenzy. Suddenly, European organizations are struggling to replace fleets of Windows desktops linked into American cloud tools. Wholesale rip-and-replace is a big undertaking, and expensive: even if the software is free, the labor isn't. One tempting alternative is to switch to locally hosted online services. As we reported last October, Schleswig-Holstein did just that. In February, we reported on the French government opting for La Suite among other such changes. One of the next big moves came in April, when Nextcloud and Ionos announced Euro-Office, a fork of ONLYOFFICE. The Euro-Office announcement highlighted the new suite's visual resemblance to Microsoft's offerings, and its embrace of Microsoft file formats. It also criticized the dated default look and feel of LibreOffice. ONLYOFFICE was not happy, and suspended its partnership with Nextcloud. Collabora was not happy either. Even so, the Free Software Foundation endorsed the fork's legitimacy. ONLYOFFICE remained unhappy and in May released ONLYOFFICE 9.4 – with stricter license terms. As The Register reported when Euro-Office 1.0 shipped, The Document Foundation was not happy. Dangling juicy treats over the heads of squabbling hungry vendors was never going to result in harmony and increased cooperation. Even so, as long as all the tools work and can read one another's documents, then all the different bodies and organizations will be able to talk to each other, exchange files, and keep working. It is messy, and it will continue to be, but then, FOSS has always been messy, acrimonious, and reduplicative. Until the AI bubble implodes, this will continue. Then the hucksters will pivot to trying to sell us all quantum computing instead, and the government money will go toward climate remediation and geoengineering… and the bureaucrats will be left with whatever proves cheapest to deploy and run. ®

ZTE showcased at MWC Shanghai 2026 its comprehensively boost to TPS. Powered by multiple dimensional co-design, deep optimization, and acceleration - from chips, servers, clusters, and AIDC, to software algorithms and scheduling platforms - this innovation empowers customers to build TCO-Optimal AI factories providing robust support for the efficient development of the Token Economy. As large models enter the phase of scaled inference deployment, the "cost per Token" has emerged as the ultimate metric for measuring the commercial value of AI. ZTE proposes that a leap in Token generation efficiency can only be achieved through architectural-level innovation and system-level synergy. To that end, the OEX (Orthogonal Electrical eXchange) architecture based SuperPOD showcased at MWC Shanghai 2026 represents a milestone innovation designed to shatter computing power bottlenecks and maximize energy efficiency. Pioneering the OEX architecture to define the next-generation super-node standard ZTE pioneered the Orthogonal Architecture SuperPOD concept. Its OEX architecture features a midplane-free and zero-cable design to achieve physical decoupling and flexible replacement of core components such as GPUs, CPUs, and switch chips. By supporting mainstream high-speed interconnect protocols like CLink and SUE, it truly realizes "multi-chip synergy, open compatibility, and on-demand optimization". Compared with traditional architectures, OEX-based SuperPOD communication paths are shorter with lower signal loss, significantly improving overall interconnection efficiency, minimizing latency, and enhancing system reliability. ZTE's SuperPOD single rack achieves industry-leading ultra-high-density integration of 128 GPUs, and supports scale up to 16,000 GPUs to build an extra-large-scale cluster. It meets AI training and inference requirements ranging from thousand card to ten thousand card scale, providing a solid foundation for long context, high concurrency agent scenarios. Multi-dimensional optimization to achieve comprehensive improvements in inference efficiency and computing energy efficiency Hardware-software synergy unleashes ultimate efficiency per watt. By leveraging a PD disaggregation and integrating technologies such as network efficiency optimization, operator optimization, and multi level KV cache, performance bottlenecks are overcome and throughput is significantly increased. In close collaboration with multiple manufacturers, heterogeneous mixed inference and system level optimization are advanced on domestic chip platforms, resulting in a comprehensive enhancement of inference efficiency and a notable increase in TPS. Compute-storage-network synergy builds a large scale inference pool. ZTE offers the Full Series AI Server supporting high density deployment with 8 or 16 GPUs per server and 64 or 128 GPUs per rack, adapting to diverse scenarios. The AI-native KV cache is implemented through DPU hardware acceleration that enables direct GPU access to storage, achieving zero-copy data transfer, microsecond-level latency, and PB-scale scalability. Combined with intelligent prefetching and dynamic eviction mechanisms, the cache delivers a hit rate exceeding 70%, significantly boosting inference efficiency. Building an open and evolvable AI infrastructure with ecosystem partners ZTE emphasizes that AI computing power development must balance performance, cost, and sustainable evolution. To achieve this, ZTE's OEX-based SuperPOD adopts a "Pre Integration" model. Through Pre adaptation & Pre integration, the product adaptation and turning cycle is slashed from over one year to within six months, significantly accelerating ecosystem convergence and large-scale commercial rollout. Efficient computing power serves as the bedrock of the Token economy, and ZTE is dedicated to ensuring that every ounce of computing power translates into tangible AI productivity. This showcase fully underscores ZTE's deep technological heritage and innovative strength in intelligent computing infrastructure, while offering customers a definitive, future-proof blueprint for building highly efficient AI factories. Contributed by ZTE.

Elastic, whose products include Elasticsearch and Kibana, has announced an "approximately" 7 percent reduction in its workforce. In a blog post, CEO Ash Kulkarni thanked employees for their hard work as he announced the layoffs. He stated that the customer-facing sales team would continue to grow, but for others, "advances in AI and automation are letting us operate with leaner teams." Kulkarni also noted that engineering, "where the nature of the work is evolving fastest," would be split into three core areas, each led by a senior leader who will report directly to him. Last month, Elastic announced its Q4 FY 2026 figures, which included a total revenue of $451 million, an increase of 16 percent year-over-year. It's been an interesting few years for Elastic. In 2021, the company announced the adoption of the Server Side Public License (SSPL) in an effort to stop cloud providers from offering its software as a service. It had formerly used the more permissive Apache 2.0 license. The reaction was swift: Amazon Web Services (AWS) forked Elasticsearch and Kibana, calling it OpenSearch. OpenSearch was later transferred to the Linux Foundation. For those who haven't come across the company's wares, Elasticsearch is a distributed search and analytics engine, and Kibana is a data visualization tool. In 2024, shortly before the transfer of OpenSearch, Elastic announced that it was adding the GNU Affero General Public License v3 (AGPL) as an option. CTO Shay Banon said at the time: "I am so happy to be able to call Elasticsearch Open Source again." Kulkarni said: "The industry is changing. Advances in AI, automation, and technology are reshaping how work gets done, and we're changing with them." Part of the change, it appears, is laying off a chunk of the workforce. "We're shifting our pace of innovation," he stated, "simplifying how we operate, and investing in new skills. That's what this reorganization is for: a simpler structure, with fewer layers, less complexity, and less friction." In its last 10-K filing with the US Securities and Exchange Commission (SEC), the company stated it had a total of 4,019 employees. Slightly less than 300 Elasticians stand to be shown the door depending on what "approximately 7 percent" ends up being. In its SEC 8-K filing [PDF] last night, the company said in addition to the 7 percent workforce reduction, it "plans to continue hiring in key strategic areas and locations, including continuing to grow headcount in customer-facing go-to-market functions." It said that overall, it expects "total headcount to grow this fiscal year compared to last fiscal year." The Register asked Elastic for a breakdown of the figures, and will update this piece should the company provide more information. ®

ZTE announced that Cui Li, the company's Chief Development Officer, delivered a keynote speech titled "Unlocking Value and Embracing Uncertainty in the AI Era" at MWC Shanghai 2026. Cui Li noted that the world is undergoing a profound paradigm shift. AI is iterating at a breakneck pace and generating more customized demands, where the "one-size-fits-all" model is no longer applicable. We are now in an era where uncertainty is the only certainty. In this time of change, ZTE put forward the "All in AI, AI for All" strategy. Specifically, the company aims to unlock AI value to the fullest—deeply embedding AI-native capabilities into products and solutions to achieve a great leap in value delivery, and advancing the agile evolution toward a data-driven organization that features human-machine collaboration. Meanwhile, to address uncertainty, ZTE is committed to building a resilient AI system capable of agile actions and fast evolution from four key dimensions: openness and decoupling, flexible scaling, extreme synergy, and scenarios first. Looking into a future of human-AI symbiosis, ZTE will remain steadfast in its role as a value contributor in the ecosystem, and make continuous innovations and breakthroughs, to create a brighter future with global partners. Below is the full transcript of Cui Li's keynote speech: It's a great honor to meet you again here at MWC Shanghai. Last year at this venue, we explored how tech surge was reshaping the boundaries of industries under the title "Digital and Intelligent Evolution: Usher in a New Era of AI Civilization". While a year later, agentic AI is emerging as a dominant force driving industrial upgrading, redefining the business logic, and transforming how we collaborate. So today, I'll elaborate on how we can fully unlock value and embrace uncertainty in the AI era. Undoubtedly, we're now in an era where uncertainty is the only certainty. From a global perspective, the macro environment is undergoing profound transformation, and AI revolution and energy transition are redefining the underlying logic of industry competition. Against this backdrop, the global industries are moving from "efficiency first" to "efficiency + resilience"—where short-term volatility and long-term structural adjustments converge, reshaping the global economic and technological landscape. From a tech perspective, AI keeps iterating at a breakneck pace. As for software, large AI models evolve monthly or even weekly, with major capability leaps every six months. And we're seeing the profound transition from conversational AI toward agentic AI and embodied AI. Hardware upgrade is also accelerating. Companies like NVIDIA roll out new architectures every two years and update flagship chips annually, continuously expanding compute supply. Meanwhile, huge gains in inference efficiency are transforming cost structures, and the competition between open-source and closed-source models is advancing the democratization of AI. As we can see, compressed tech iteration cycles give rise to new application scenarios, while posing higher requirements on companies in choosing the right tech roadmap, setting an appropriate pace for asset updates, and aligning talent capabilities. In addition, markets are segmenting at an unprecedented pace. AI is generating more customized demands, and the "one-size-fits-all" model is no longer applicable. On the consumer side, the long-tail demand is exploding, driven by personalized recommendation algorithms, and niche brands are breaking through against all odds via content marketing in e-commerce platforms. In production, the "small order, fast supply" model becomes the norm powered by flexible manufacturing and AI quality inspection. Meanwhile, emotional economy and experience economy are rising fast, as evidenced by the exponential growth of segmented markets like AI pet, emotional healing, and silver hair technology. Under such trends, the traditional economies of scale are being disrupted, and a new business paradigm is taking shape, where niche is the new mainstream. Amidst such market uncertainty, innovation is the only way forward. In this time of change, we advocate for deeply embedding AI-native capabilities into our products and solutions to achieve a great leap in value delivery, while advancing the agile evolution toward a data-driven organization featuring human-machine collaboration. Also, we should fully take into account the exponential factors of AI iteration—namely, keep our architecture flexible and open, leave room for dynamic resource reallocation, and thus build a resilient system that is capable of addressing uncertainty. After over three years of AI sprints, model capabilities have continuously improved. Large AI models already perform at or beyond PhD level, especially good at info-heavy, super-uncertain tasks. Agents move from passive information processing and reasoning to active interaction and execution, steadily advancing toward continuous learning and self-evolution. This raises a fundamental question: how intelligent does a model really need to be? For most consumers, today's models are already overkill; yet for developers and enterprises, there is still a considerable gap in using AI well. ZTE's proposition is clear: unlock AI value to the fullest. We are upgrading products, solutions, and platforms, building full-stack, all-scenario intelligent computing solutions driven by end-to-end TCO optimization, and fully accelerating the company-wide intelligent transformation. In 2025, ZTE proposed the "All in AI, AI for All" strategy. At the core of it lies our commitment to fully unlocking AI value. To start with, with an AI-native mindset, we are redefining telecom networks and our offerings for enterprises, homes, and consumers. In connectivity, we bring intelligence to our products and solutions, such as AIR MAX and lossless WAN, while harnessing agentic AI and digital twins to accelerate the evolution toward L4 autonomous networks. For enterprises, our all-in-one gateway, integrating optical connectivity, video security, computing, and storage, delivers plug-and-play simplicity, making AI adoption effortless for small and micro-sized enterprises. For homes, we have partnered with operators to develop AI smart displays—an AI hub covering education, entertainment, eldercare, and home security. The cumulative shipments already exceed 3 million units. For consumers, our AI-native phones and AI Cloud PC WorkBuddy bring brand-new intelligent experiences. In computing, we focus on end-to-end TCO optimization and stay committed to building a homegrown, open AI ecosystem. At the core capability layer, we deliver full-range, system-level AI chip capabilities that enable optimal selection. This is empowered by our in-house switch chips for high-speed intra-HBD/cross-HBD interconnects, domain-specific CPUs, DPUs, and NICs, alongside the adaptation and optimization for various mainstream GPUs. As for infrastructure, we provide a full-stack intelligent computing solution that brings together computing, high-speed network, and green IDC—all designed to deliver the optimal TCO. Regarding platforms, we are building an AIOS and an agent hub, supporting elastic scheduling of intelligent computing resources, training and inference acceleration, and rapid agent development. At the application layer, based on the "1+N+X" model (one foundation model, N domain-specific models, and X scenario-based applications), we have delivered more than 1,000 AI projects across over 18 industries such as telecom, manufacturing, and power grid. At the terminal layer, we have launched innovative products including AI-native phones, AI cloud PCs, and AI smart displays, creating an intelligent home hub that integrates networks, computing, displays, and agents. This is how we promote scenario-based application and truly realize AI for all. Regarding organizational evolution, building on the digital transformation initiated in 2016, ZTE has achieved comprehensive intelligent upgrade across key areas, including R&D, marketing, operations, and supply chain. In R&D, we have successfully advanced from "Chip + Equipment" to "Chip + Equipment + Large AI Model", and are now accelerating toward "human-machine collaboration". In marketing, we have created "agile small teams within a large enterprise" to break down organizational and process barriers, and strengthen process reengineering and cross-enterprise IT system interconnection, thereby ensuring fast response and efficient execution. In operations, our in-house Co-Claw has been deployed across the company. As an enterprise-grade agent platform compared to OpenClaw, Co-Claw has been significantly enhanced in compliance, security, reliability, and integration with internal systems. Deeply embedded into our core systems, it can effectively execute high-frequency tasks, including document management, contract analysis, fault ticketing, event creation, and R&D collaboration. Over the past three years, innovations in AI algorithms and architectures have exploded, with landmark advancements such as MoE, MLA, quantization, distillation, attention mechanism optimizations, and ultra-long context windows. Different algorithms require tailored hardware to achieve optimal cost-performance, while system and hardware architectures typically follow a 2 to 3-year refresh cycle. This calls for keen foresight and built-in flexibility across solution development, architectural design, investment deployment, and operations management to secure medium- and long-term competitiveness. Drawing on our experience, ZTE advocates "embracing uncertainty". We are building a resilient system capable of agile actions and fast evolution from four key dimensions: openness and decoupling, flexible scaling, extreme synergy, and scenarios first. First, openness and decoupling. Given AI's rapid evolution, only openness and decoupling allow us to fully internalize the latest industry innovations and practices. That's why, in the intelligent computing field, ZTE remains committed to decoupling hardware from software, models from platforms, and training from inference. For hardware, our solutions support different GPU and CPU combos for optimal performance and compatibility, while delivering software-hardware co-optimization to maximize tokens per watt. Additionally, ZTE's platform is fully compatible with over 200 models, enabling efficient heterogeneous resource management and deep software-hardware synergy. It also accelerates the full lifecycle from model development to application with one-click deployment and seamless migration. For applications, our product packages support high-quality data synthesis, one-click rapid fine-tuning, intelligent generation of knowledge graphs, tacit knowledge mining, minute-level agent building, and autonomous tool evolution. Notably, ZTE's Co-Sight AI agent factory has been open-sourced. Second, flexible scaling. Take ZTE's AI server portfolio as an example. It spans standard 8-GPU/16-GPU nodes, SuperPOD, and modular solutions. The SuperPOD, built on an innovative OEX architecture, packs 64 or even 128 GPUs into a single rack. In cluster-level scaling, our modular design allows seamless expansion all the way to tens of thousands or even hundreds of thousands of GPUs. For our customers, this means they can scale incrementally rather than all at once. This approach significantly slows down hardware depreciation and maximizes the long-term ROI. Third, extreme synergy. For computing-network convergence, through scale-up, we boost computing density to up to 16,000 GPUs; through scale-out, we expand clusters beyond 100,000 GPUs via optical interconnects and electrical switching. This truly strengthens computing power through advanced networks. For computing and storage, our in-house switch chips guarantee high running efficiency, while our Dinghai DPU chips enable Pod-level shared memory, delivering microsecond latency and PB-scale expansion for KV cache. In computing-power coordination, operation strategies can be dynamically adjusted based on grid loads—enabling elastic power supply and intelligent energy saving. The rack power density scales flexibly from 8 to 120 kW. And our cold-plate liquid cooling solution reduces the PUE to 1.15, making green and sustainable development a tangible reality. Finally, scenarios first. We emphasize tailoring solutions to the unique requirements of each scenario and application—because the best fit delivers optimal cost effectiveness. To date, ZTE's intelligent computing solutions, including AI all-in-one machines, have empowered more than 18 vertical industries. Together with over 1,000 ecosystem partners, we have delivered over 100 best use cases, charting a path from pilot projects to industry-wide adoption. Here are some examples: At ZTE Nanjing Binjiang Base, our integrated intelligent operations system has shortened order scheduling cycles by 80% and improved per capita output by 81%. In the Yunnan Sunho Aluminum factory, ZTE's Co-Claw, as an intelligent hub, enables millisecond load regulation, with capacity utilization exceeding 99% and aluminum distribution efficiency up by 30%. At the Shaanxi Yanchang Petroleum Balasu Coal Mine, four digital employees have been deployed, accumulating more than 60 AI skills. This has accelerated major hazard identification to the minute level, reduced technical report generation to just 5 minutes, and lowered the proportion of repetitive operational and management tasks to below 15%, significantly improving production safety and operational efficiency. The future has already arrived. We are entering a new stage of civilization that is defined by "human-AI symbiosis". Through the profound synergy of computing, networks, energy, and algorithms, AI is seamlessly embedded into every facet of social operations—much like a vast, intelligent organism. In this living system, 6G serves as the "neural fibers", enabling millisecond information transmission; computing power is the "heart", pumping energy to sustain the entire system; energy functions as the "blood", which is intelligently scheduled and circulates in an endless cycle; and algorithms represent the "genes", which define the shape of the entire system and where it evolves. Looking ahead, research into continuous learning and self-evolution will further propel the system toward higher-level intelligence. For ZTE, we firmly believe that though the road is long, every step forward brings us closer to our destination. Our commitment is to fully unlock AI value and embrace uncertainty with agile iteration. We also believe that harmony lies in diversity and prosperity lies in symbiosis—and that only through open collaboration can we build an ecosystem that is stronger and truly thrives. ZTE will remain steadfast in its role as a value contributor in the ecosystem, and make continuous innovations and breakthroughs, to create a brighter future with global partners. Contributed by ZTE.

ZTE showcased the TCO-optimal AI factory, an AIOS-powered ecosystem spanning scenario-based applications and innovative terminals, and cutting-edge breakthroughs in AI-driven networks at MWC Shanghai 2026. Through the extreme synergy across computing, network, storage, energy, and software, as well as system-level architecture innovations, ZTE fully unlocked new momentum for token operations. As AI agents become widespread, industry competition focus is shifting from computing power scale to token efficiency. Leveraging its end-to-end, full-stack AI capabilities, ZTE is forging a complete value chain spanning token production, services, and circulation—to enable operators, enterprises, and industry partners to unlock AI value at scale. Building a TCO-optimal AI factory for more efficient, cost-effective token production Token efficiency is the core competitiveness in this era of inference. TCO-optimal AI factories are the key to boosting token efficiency and achieving "token freedom". At MWC Shanghai 2026, ZTE showcased solutions that enhanced token efficiency and minimized the cost-per-token through the deep synergy of various elements, including SuperPod, extremely cost-efficient inference, and AIDC with ultimate energy efficiency. Building on 41 years of accumulated expertise in the R&D and engineering of ultra-large-scale, complex systems, ZTE has launched its SuperPod solution. Leveraging multi-chip open collaboration design and the innovative Orthogonal Electrical eXchange (OEX) architecture, it enables plug-and-play deployment and rapid provisioning. A single rack supports 128 GPUs, with high scalability up to 16,000 GPUs, laying a foundation for high-efficiency computing in model training and inference. Through the synergy of network, storage, and computing, as well as hardware and software, ZTE achieves extremely cost-efficient inference. With non-blocking network and Global Server Load Balancing (GSLB), ZTE's solution enables efficient coordination across computing clusters ranging from thousands to tens of thousands of GPUs, further boosting overall inference efficiency. Moreover, ZTE has developed its AIDC solution with ultimate energy efficiency, which integrates 800V HVDC power supply, full-stack liquid cooling, and intelligent computing-electricity synergy, enabling highly efficient and low carbon operations. From computing cluster construction and inference efficiency optimization to green operations, ZTE is steering AI factories from being compute-centric to efficiency-driven, delivering a more cost-efficient and sustainable token production system. Developing an innovative AIOS technological foundation for smarter token scheduling and closed-loop services The AI factory addresses the challenge of low-cost token production, while AIOS plays a critical role in token scheduling, orchestration, and service-oriented output. ZTE has launched the NewStart AIOS and positioned it as the "technological foundation in the AI era". Built on AIOS, Co-Claw—an enterprise-level agent platform—is fully integrated into ZTE terminals, extending AI services to production and daily life. For enterprises, the focus is on enhancing office and R&D efficiency and enabling intelligent operations, to achieve self-evolving workflows and closed-loop decision-making. For homes, ZTE has launched innovative terminals with different screen sizes which serve as entry points to a fully connected smart home. They can be household AI assistant, AI entertainment companion, and AI security guardian—always on, always active for an intelligent lifestyle. For individuals, the company has introduced a new AI-native paradigm spanning AI phones, AI cloud PCs, AI smart displays, FreeScreen and more, enabling personalized recommendations, contextual responses, and immersive interactions across travel, life, and work. AI knows better and delivers enjoyable smart experiences anytime, anywhere. Leveraging the AIOS technological foundation, Co-Claw agent platform, and a full lineup of AI devices, ZTE transforms tokens into schedulable, serviceable, and monetizable resources. This allows operators and enterprises to develop a closedloop model for token-based services. AI + Network, co-building a future of intelligent, ubiquitous 6G connectivity 6G is more than a generational leap in connectivity. It also serves as the critical infrastructure that enables efficient token circulation in the AI era. As a key participant in global 6G research and standardization, ZTE centers around "AI+" and "SAGIN", speeding up the evolution of 6G from research to commercial use. Regarding SAGIN, ZTE has launched industry-leading LEO satellite communication payload solutions compatible with diverse application scenarios of SAGIN for 5G-A and 6G, fostering the critical connectivity foundation for a smooth evolution from 5G-A to 6G. As for AI-native communications, ZTE has demonstrated its GigaMIMO solution and the world's first 256 TR U6G prototype in the 6G Zone at this event. With GigaMIMO, a cornerstone technology for 6G, the solution leverages architecture innovation, computing synergy, and algorithmic breakthroughs to systematically overcome critical bottlenecks in future network capacity, coverage, and spectral efficiency, thus speeding up 6G deployment and industry evolution. Centered on AI-native capabilities, ZTE has driven the deep integration of AI with existing networks—through architecture upgrades and key scenario innovations—to build a next-gen communications network. The AIR MAX solution comprehensively upgrades the network architecture with AI at its core and delivers a three-tier capability system involving the AI-native infrastructure, autonomous operations, and monetization engine. Together, they empower operators in capability evolution, service transformation, operational shift, and ecosystem reshaping for the AI era. The AI HI-NET solution delivers three key capabilities—AI-native, ultra-high-speed and lossless, and intrinsically secure, laying a solid network foundation for ubiquitous intelligent computing access, lossless wide-area connectivity, and the integration of computing, network, security, and service. The 10G AI-Optical Network solution has been widely deployed across homes, campuses, small businesses, and other key scenarios, unlocking greater value of all-optical cities. Meanwhile, ZTE continues to promote key scenario applications through innovations such as new calling with AI assistant, "first-class" mobile network services, ISAC for low-altitude economy, and highly autonomous networks. These efforts are made to build a more intelligent, efficient, and trustworthy connectivity foundation for the AI era. From AI factory to AIOS, agent platform, and future networks, ZTE demonstrated an end-to-end capability system at MWC Shanghai 2026. The system covers token production, scheduling, and circulation, integrating computing, network, terminals, and agents to accelerate AI from technological breakthroughs to large-scale deployment. The AI industry is accelerating into the inference era, where tokens have become a key measure of AI-driven value creation. Looking to the future, ZTE will work with industry partners to build AI infrastructure and ecosystems that are more efficient, inclusive, and sustainable. Together, we aim to bring AI to various industries, unlocking new growth momentum for the AI era. Contributed by ZTE.

IBM has developed a sub-nanometer (nm) chip technology it says could be used to produce commercial chips within five years, and has mapped a path to 0.1 nm. Big Blue claims its new process node can cram nearly 100 billion transistors onto a silicon die the size of a fingernail, almost double the density of the 2 nm technology it unveiled back in 2021. The new process as disclosed is actually for 0.7 nm or 7 Angstroms (7A), compared with the cutting-edge manufacturing nodes now being prepared for production in 2028 by the likes of Intel and TSMC which are 1.4nm, or 14 Angstroms. Several structural and material innovations have gone into this latest manufacturing method, including a three-dimensional nanostack architecture that sees transistors stacked, with n-type and p-type field-effect transistors (FETs) arranged so that one is layered above the other. "We're announcing it's not just an incremental step, it's a meaningful leap forward, enabling up to 50 percent higher performance, or 70 percent greater efficiency [than 2nm], and pointing to a future where computing becomes significantly more powerful without a corresponding increase in energy," claimed director of IBM Research and IBM Fellow Jay Gambetta. And the firm sees a clear path to shrinking down to one-tenth of a nanometer over the next ten years, he added. "Nanostack is not one innovation. It is actually a device platform that can enable the future of scaling for another decade beyond nanosheet, as you can see from our technology roadmap all the way to 1 Angstrom." Although the firm touts nanostack as the industry's first three-dimensional, nanosheet-based design, Intel was talking about 3D stacking of transistors back in 2023 – though has not so far implemented it. Huawei has also come up with a similar concept in its LogicFolding architecture, using two separate wafers fused together. IBM's nanostack design also has a twist – the transistors in the upper layer are staggered, or offset, from those below. "Nanostack is nanosheet transistors stacking on top of each other. But it's not through a simple monolithic lithography and etch process," said Huiming Bu, VP of Silicon Technology Research & Development at IBM. "What happens here is we actually stack in vertical direction but also stagger, so the front side of each transistor and the backside of each transistor can be contacted independently for signal and power," he added. "Second, the stacking of this transistor is done by single dielectric bonding, which is a key innovation that we have developed. Through that technology, the channel materials, essentially the top FET and the bottom FET, can be optimized independently." IBM says the architecture could support multiple applications such as CPUs, GPUs, mobile chips and memory, such as SRAM. Gambetta hinted that the technology could be used in future AI accelerators. "This is why we were excited by the initial experiment that shows a 40 percent scaling in SRAM. There are many examples of AI chips that are using more SRAM to scale, but fundamentally, it comes down to: can we make transistors more efficient, less power, put more in there?" he said. But IBM no longer manufactures chips itself. When asked which foundry might adopt its sub-nanometer process, Huiming said the nanosheet architecture IBM invented is now used by all leading foundries at this point. "I'm not going to talk about a business model, but it's being adopted by all leading foundries. But today, we are focusing on helping Rapidus to be successful in bringing up 2 nm manufacturing capability in Japan," he stated. Rapidus is a government-backed semiconductor foundry set up to revitalize the nation's semiconductor industry. The nanostack transistor architecture is discussed in a paper, available for download from the IEEE. ®

The minutes of the government's recently announced digital ID advisory group will not be published, Cabinet Office minister James Frith has told a Conservative MP, while not answering his questions about its budget or how its members were selected. Andrew Snowden, MP for Fylde and an assistant whip, asked the Cabinet Office whether the minutes, recommendations, and advice of the digital ID advisory group announced earlier this month will be published. In separate parliamentary written questions, he also asked what budget the group had been allocated and what criteria were used to select its members, who include security expert David Rogers, Mumsnet founder Justine Roberts, and former New South Wales digital government minister Victor Dominello. "The running of the digital ID advisory group will be supported by the Cabinet Office's existing digital ID task force. The group is not a decision-making body and minutes will not be published," said Frith in identical replies to all three questions. "The answer was disappointing to say the least," said Snowden when asked by The Register if he was satisfied with Frith's response. "Digital ID was a deeply controversial policy when Keir Starmer announced it, causing one of the many U-turns that led to the situation we are in today. If the government are persisting with developing a system of digital ID then scrutiny of that policy by Members of Parliament is vital. To ignore key questions will not increase public support for digital ID." Snowden added that written parliamentary questions are an important way to hold the government to account and obtain further information. "When the government refuses to answer questions from Members it is not just us they are disrespecting. They are disrespecting Parliament itself and the people we represent." The digital ID scheme was announced last September at the Labour Party's annual conference by current prime minister Keir Starmer. The conference also heard Greater Manchester mayor Andy Burnham speaking against the plans. "I think there's a risk of an opportunity cost situation here, where something can consume a huge amount of time and actually doesn't come through," he said. Following his by-election victory last week and Starmer's announcement on Monday that he would resign, Burnham is widely expected to become the UK's next prime minister, which could leave him to decide the fate of an unpopular scheme introduced and championed by his predecessor. ®

It is half a century since the Soviet Union launched the final crewed Almaz space station, also known as Salyut 5, which was home to two crews, while a third mission failed to dock and nearly came to a watery end. The Almaz stations were launched for the Soviet military and are better known as Salyut 2, 3, and 5. There were additional stations under construction, but the crewed program was canceled after Salyut 5. While we hesitate to use the word "cursed," Salyut 5 was certainly an eventful program for its crews. The station was launched on June 22, 1976, atop a Proton-K rocket from Baikonur. Known internally as OPS spacecraft (Orbital Piloted Stations), the Almaz stations had a pair of solar arrays, reconnaissance equipment, and a cannon mounted at the base of each station, which was test-fired on Salyut 3 while uncrewed. The first crewed mission was expected to last more than 50 days, and three further missions were planned. The mission did not go according to plan. Soyuz 21, carrying cosmonauts Boris Volynov and Vitaly Zholobov, was launched to the station on July 6, 1976. At first, everything went swimmingly. The duo conducted experiments aboard the space station and undertook reconnaissance. After all, this was the Soviet answer to the US's Manned Orbiting Laboratory (MOL), which would have used astronauts to perform surveillance on targets from a space station. The MOL was canceled before any crews were flown. The Soyuz 21 crew began experiencing problems in August. David Harland's book, The Story of Space Station Mir, notes that Russian newspapers "reported on 18 August that the cosmonauts seemed to be suffering from 'sensory deprivation' and that psychologists monitoring their health had suggested that music be played to them over the voice uplink." A few days later, Radio Moscow took a different line, noting that a prolonged flight could be on the cards due to favorable levels of solar radiation. On August 24, the crew returned to Earth after 49 days, earlier than scheduled. It's unclear what happened, although most reports suggest an acrid odor developed aboard the station, possibly from chemicals used to develop photographs from the surveillance equipment or from fumes leaking from the Salyut's fuel tanks. The station's environmental systems were unable to address the problem, and officials, fearing for the cosmonauts' health, elected to bring them back early. A second mission to the station, Soyuz 23, was launched on October 14, 1976, but an equipment malfunction stopped the crew from docking with Salyut 5. The failed docking was only the beginning for the two cosmonauts on board, Vyacheslav Zudov and Valery Rozhdestvensky. The spacecraft splashed down in the freezing waters of Lake Tengiz, and the pair were stranded in the capsule until rescuers dragged it to shore the following morning. "It was fortunate that such an ordeal had not befallen Volynov and Zholobov in their weakened condition," Harland noted. Not to be deterred, the Soviet military tried again with Soyuz 24, crewed by cosmonauts Viktor Gorbatko and Yuri Glazkov. The mission launched from Baikonur on February 7, 1977, and successfully docked with the station. The pair had a shorter stay on board Salyut 5, departing after just over 16 days of docked operations, but did not suffer the problems of the first crew, and demonstrated an air-replacement technique that involved venting the station's atmosphere from one end while air was released into the station from the tanks of their Soyuz. Harland reported that the crew felt a "light breeze" during the operation. The Soyuz 24 backup crew was assigned to a planned fourth visit to the station, but, as described by Anatoly Zak of RussianSpaceWeb.com, the station's propellant was likely running low, and, with no way of refueling, Salyut 5 was deorbited on August 8, 1977. Salyut 5, or OPS-3, was the final crewed station of the Almaz program. Subsequent Salyut stations would have a civilian focus and additional docking ports for replenishment, but the project yielded useful operational lessons, even if there would be no more Almaz crews. The project does, however, have an intriguing coda. An unfinished Almaz station was acquired a few years ago by an Isle of Man-based company, Excalibur Almaz. The project came to naught, but the incomplete Almaz-206 can be viewed at the excellent Isle of Man Motor Museum. ®

Modern AI infrastructure, particularly in the realm of agentic AI, is often discussed through the lens of accelerators, model sizes, and training clusters. However, making large-scale agentic AI systems functional relies heavily on the CPU within these pipelines. As agentic AI deployments expand across cloud and enterprise environments, the CPU increasingly acts as the control plane for the entire system. It manages the coordinated movement of data between storage, memory and accelerators, ensures the secure isolation of workloads, and handles precise scheduling across distributed infrastructure. These are crucial tasks that underpin the efficiency and reliability of agentic AI operations. Without efficient and performant CPUs, even the most powerful accelerators cannot deliver their full value. This is one reason why Arm's role in the datacenter has expanded rapidly in recent years. Long associated with mobile devices, Arm has become the core architecture for cloud and AI infrastructure. The, built in close collaboration with Meta targets maximum rack-level density to scale up performance for agentic AI datacenters, it is built on the Arm Neoverse V3 platform and delivers high-performance cloud and AI deployments. The Arm AGI CPU also offers system efficiency, security features such as confidential computing, and the flexibility required for hyperscale environments. --- Hyperscalers and leaders in AI, including AWS, Google Cloud, and Microsoft Azure, have each introduced multiple generations of custom Arm-based processors for their platforms, reflecting a broader move toward purpose-built silicon optimized for modern workloads. Google's Axion processors are designed to support applications ranging from data analytics and microservices to AI inference. Testing has shown strong gains in performance and efficiency, while companies such as Spotify have reported significant improvements when evaluating the architecture for large-scale workloads. Similarly, AWS Graviton processors offer enhanced performance and cost-effectiveness, making them a competitive choice for various computational tasks. Microsoft is taking a similar approach with its Azure Cobalt processors, which are built on the Arm's Neoverse compute platform and designed specifically for cloud-native environments. Early deployments have demonstrated measurable improvements in performance and infrastructure efficiency for real production services. NVIDIA has used Arm Neoverse’s line of CPU designs in its Grace Hopper and Grace Blackwell chips, as well as the latest Vera Rubin NVL72 system.. These developments point to a broader architectural model sometimes described as the "converged AI data center." In this environment, compute, accelerators, networking, storage, and software are designed together as a unified system rather than as independent components. The CPU plays a central role in that model by coordinating how the rest of the stack operates. For architects planning the next generation of AI infrastructure, the lesson is increasingly clear: accelerators may drive model performance, but the CPU remains the engine that keeps the entire system running. To explore the architecture behind the Arm AGI CPU and Arm Neoverse in more detail, visit the Arm product page. Sponsored by Arm.

PWNED Welcome back to PWNED, the weekly column where we school ourselves on others' security failures. This week, we’ll learn about a school where the entire network was like an open-book test … and the IT department got a zero. Have a story about someone leaving a gaping hole in their network? Share it with us at pwned@sitpub.com. Anonymity is available upon request. Our tale of academic pwnage comes courtesy of a reader we’ll Regomize as Nathan. Nathan was 17 and attending sixth form at a UK school when he found a treasure trove of admin privileges and data at his fingertips. One day, our hero connected his laptop to his school’s Active Directory domain. There was no admin authentication required and Nathan was able to see domain controller tools in view mode, look at policy maps, and so on. Nathan then browsed the directory and located the domain administrator account. The password, “horse fence ditch,” was written right in the description field, where anyone with access to the network could view it. There were also backup accounts with passwords such as “bd” and “bigbaddog.” Once he had full God mode enabled, Nathan said, he could see student and staff data, gain Remote Desktop access to any server or domain controller, and even access LanSchool, a popular classroom management app. “I could've accessed sensitive leadership docs, reset passwords, deleted accounts, wiped the whole network, etc,” Nathan told The Register. Moreover, the entire system was synced with Google Workspace, so Nathan had access to user mailboxes as well. He even found firewall settings, security policies he could change, and keystroke histories. Because Nathan was a student and did not want to get in trouble at school, he didn’t actually use any of these privileges. He kept his head down and graduated from school without incident, but also without reporting the vulns, which might still be in place today for all we know. So what can we learn from this tale of academic malpractice? First, as we learned a few weeks ago, do not store passwords in description fields for Active Directory. In fact, do not store passwords in cleartext anywhere without serious controls! Second, Nathan should not have been able to see Active Directory domain controller tools. And it might also have helped if Google Workspace had different admin credentials. Imagine the restraint required not to change people's grades, take over their computers, or delete data. Would you have been able to exercise the same level of discipline as a 17-year-old? ®

Infosys chairman Nandan M. Nilekani has predicted AI – even AI that does the kind of coding work his company does for many clients – will be good for services companies. Nilekani made his prediction in a speech delivered at the Indian services giant’s annual general meeting on Tuesday. “The industry is going through a major technology transition and whenever there is such a transition, questions are asked about our relevance, leadership or ability to maintain growth and margins,” he said. “Given that AI is a much larger and disruptive technology transition than ever before, the questions are louder and the doubts are more insistent. Moreover, the existential question that is asked of us is, if coding becomes automated, then why are we needed at all?” Nilekani said Infosys “will embrace the best coding tools and improve our productivity” – and also pointed out that there’s more to writing software than just writing software. “Enterprise context is paramount,” he said. “Solutions must complement existing investments. They demand rigorous testing, resilient architecture, and foundational cybersecurity.” The chair thinks AI can’t do that, but Infosys can. He also thinks AI will create more demand for services that only humans can handle. “The AI revolution has made legacy modernization urgent in a way nothing else has, and clients are moving to retire the technical debt accumulated over decades,” he said, presumably referring to the idea that AI coding tools make it possible to rewrite old code more quickly and efficiently than humans can do the job. Nilekani thinks that organizations using AI to modernize software will prefer to build custom replacements instead of buying packaged programs. “All this creates even larger opportunities for us,” he said. “The defining opportunity lies in integrating intelligent AI systems with mission-critical enterprise platforms. The greatest value will come from combining the world of models and agents with traditional transaction systems that continue to underpin enterprise operations. That convergence is where the next wave of opportunities will emerge.” Infosys, naturally, has done the work needed to make itself an ideal provider of the necessary services. “More than three years after GenAI’s launch, Infosys is more relevant than ever and well-positioned for the decade ahead,” the chairman said. And of course he would say that – although he made no mention of “AI deflation”, the decline in services revenue that Infosys and India’s other tech services giants recently said is likely because AI coding tools mean there’s less work for them to do. ®

Australia’s Security and Intelligence Organisation (ASIO) has established dedicated teams to counter nation-state attacks on critical infrastructure, the org’s director general Mike Burgess revealed yesterday. “We discovered nation-state hackers had compromised the network of an Australian critical infrastructure provider,” Burgess said yesterday in remarks accompanying the release of ASIO’s annual threat assessment, a task it performs in its role as Australia’s equivalent to the FBI and MI5. “ASIO assessed the hackers were preparing for sabotage. They weren’t planting ‘digital dynamite’ as such; they were mapping out the network and maintaining access so they could cripple it at a time of their choosing.” “In this case, a state-sponsored group didn’t just achieve access to the Australian critical infrastructure provider, it successfully acquired credentials – login details and passwords – for active users of the networks, including the IT professionals guarding it,” he added. Burgess said ASIO “identified, tracked and attributed the hack, and worked with the victim company and our security partners to remediate the compromise – work which is ongoing.” “The scale of this activity – led by one nation-state in particular – is difficult to overstate,” he added, before saying Australia is not alone in facing such attacks. “We struggle to find a single country in our region that has not been compromised by this state’s cyber apparatus.” He described cyber sabotage as “an evolving threat. I have established dedicated teams to counter it.” Burgess also shared an example of espionage targeting Australia’s military to gain information about the AUKUS pact – the US/UK/Australia defense collaboration that will see The Land Down Under acquire nuclear submarines, and which also includes collaborations around information technology capability, and intelligence activities. “A spy from a foreign intelligence service approached an Australian security clearance holder online, pretending to be from a consulting company,” Burgess revealed. “The spy paid the official to write two reports on Australia’s relationship with our Pacific neighbours, and then, thinking he’d been hooked, offered money for inside information on AUKUS.” The Australian official became suspicious, reported the incident and conducted interviews with ASIO during which Burgess said the spy agency “gained valuable insights into the foreign service’s information gaps and tradecraft.” The Australian official even handed the money they were paid by the foreign spy to ASIO. “In effect, ASIO disrupted the foreign intelligence service’s operation and made them pay for it,” Burgess crowed. ASIO then scored another win. “My officers borrowed the phone from the official and rang the so-called consultant in her home country. Thinking it was her target, the spy picked up and got a very unwelcome surprise when she realised she was speaking to ASIO,” Burgess said. “We demonstrated we knew exactly who she was, demanded she cease targeting Australian citizens, stated we have zero tolerance for spying on AUKUS, provided a quick overview of Australia’s espionage laws and pointed out the Director-General reserves the right to speak publicly about these matters. At that point the spy hung up.” ASIO officers later mentioned this incident to members of the foreign intelligence service that ran the op. Burgess seems to think that officers at that foreign agency may not have told their superiors about the op failing. “In case they did not report it up – I’m confirming it now,” he said. Burgess also pointed to abuse of online spaces continuing to represent a threat to Australia. “Instead of being radicalised by associates in the real world, individuals are often being radicalised by strangers online,” he said. “Instead of being radicalised over months and years, individuals are increasingly being radicalised in weeks. Instead of being radicalised as adults, individuals are all too often being radicalised as minors. Instead of gathering in prayer halls or backyards, radicalised individuals are frequently gathering in encrypted chat rooms.” “And, instead of spending time and resources planning sophisticated attacks, radicalised individuals are moving to low-capability attacks with little or no warning,” he said. “Traditional groups such as Islamic State and al-Qa’ida and their affiliates are growing their capability to conduct and inspire attacks, enabled both by permissive geographic and online spaces.” Burgess revealed ASIO has “resolved” 14 “significant-terror related cases” since the December 2025 terror attack at Sydney’s Bondi beach, and 31 “major terrorism plots” since 2014. He said ASIO is now “aggressively adopting new tools and techniques – including artificial intelligence – to navigate our security environment,” and invited Australians to work for the agency, perhaps as offensive hackers. “All ASIO’s teams contribute to our mission and every ASIO officer makes a difference, whether you collect the dots or connect the dots, run cables or run sources, code networks or penetrate networks,” he said. ®

Memory-maker Micron has found a way to keep prices for its products sky-high for another five years, by signing 16 “strategic customer agreements” (SCAs) that include a floor price the company says comes with “a very robust gross margin for Micron, well above our peak quarterly margins in any past cycle.” Micron CEO, president and chairman Sanjay Mehrotra explained the SCAs in prepared remarks delivered during the company’s Q3 earnings call. He explained that Micron has signed 16 SCAs, most of them covering 2026 to 2030, and that they involve a commitment to buy a certain quantity of product and pay for it in a pricing band that has a floor and a ceiling price. The floor price covers the historically high gross margins mentioned above, and the ceiling price means those who commit to an SCA are insulated if memory prices go even higher. The CEO said 16 customers have signed SCAs and then explained why it’s worth locking into the deals even though they bake in such high margins. “Our customers are recognizing that supply shortages in memory and storage will take considerable time to improve,” he said. “Even as we expect industry supply to improve gradually in 2028, we currently do not have line of sight as to when memory supply will be able to catch up with increasing demand.” Even massive efforts to build new chip fabs aren’t much help, he said, because the increasing complexity of new memory types means it takes longer to build factories – and when they come online there still won’t be enough capacity to build both the high-bandwidth memory needed for AI and other types of NAND and DRAM. “Supply is structurally constrained in its growth and ability to meet industry demand, despite our comprehensive efforts to increase supply,” he said. Don’t assume that SCAs mean your suppliers get price certainty, because Mehrotra said the deals will account for 40 percent of Micron revenue – meaning the company is reserving most of its inventory to sell at prices it can negotiate. The CEO did have a little good news in the form of predictions that Micron’s DRAM output in 2026 will “grow in the low- to mid-20s percentage range, slightly above our prior outlook.” He also revealed that the SCAs see customers pay up front, which helps Micron to fund its fab expansions. Q3 results were also exceptional. Revenue landed at $41.5 billion, a fifth consecutive quarterly revenue record and 346 percent year-over-year jump. DRAM revenue alone reached a record $31.3 billion, up 343 percent year-over-year. NAND revenue rose 361 percent year-over-year, to $9.9 billion. Net income reached $28.9 billion, and consolidated gross margin came in at 84.9 percent. Execs predicted even better results, offering Q4 guidance of $50 billion in revenue, gross margins hitting “approximately 86 percent”, and telling investors that the increased complexity of memory means future products will cost even more than Micron’s current kit. But while Micron thrives, IT pros will toil to make workloads perform with less memory: Mehrotra said Micron expects sales of conventional servers will grow in the mid-teens percentages during calendar 2026 but predicted “modest reduction in average server DRAM content growth as customers focus on maximizing unit shipments amid a very tight allocation of memory.” Investors liked what they heard: Micron’s share price popped 15 percent in after-hours trading. ®

AI vendors have been pushing organizations to board the AI hype train as it races by at full speed. But many of the companies doing so, unable to move quite that fast, have stumbled along the way. According to a survey of 406 IT decision makers, 93 percent of organizations have experienced AI-caused infrastructure incidents, but a mere 19 percent had the necessary governance to respond. The survey, conducted in April by Panterra Group at the behest of Spacelift, forms the basis of the orchestration platform's 2026 State of Infrastructure Automation report [PDF]. It posits an "AI Readiness Gap," meaning that companies are adopting AI before they're ready to do so and are paying the price. "The findings are unambiguous: organizations are using AI to generate infrastructure code at a rate their governance frameworks were never designed to handle,” said Paweł Hytry, co-founder and CEO of Spacelift, in a statement. The consequences of these incidents, respondents say, consist of reworking AI-generated changes (37 percent), security misconfigurations that reached production (36 percent), compliance violations (36 percent), infrastructure drift attributable to AI changes (35 percent), and incidents caused by agentic systems (33 percent). The report characterizes 24 percent of organizations as "exposed." "Exposed organizations are using AI, but without the governance or frameworks to support it safely," the report says. "What they are doing diverges significantly from what they have in place to manage it." And then there are the "fragmented" entities, 32 percent of respondents, that use AI sometimes, unevenly, and have some governance, but no coherent plan. The two remaining categories, "outpacing" and "pioneer," at 25 percent and 19 percent respectively, describe heavy AI adoption that's ahead of business controls, and AI use in conjunction with structural discipline, respectively. In terms of AI-caused infrastructure incidents, 97 percent of "exposed" organizations reported at least one such snafu. Meanwhile, among "pioneer" entities, 17 percent said they had no AI-related infrastructure incidents. Spacelift, an infrastructure-as-code (IaC) platform, contends that automated validation accounts for the difference here because it outperforms manual code review. Across the board, respondents report greater use of AI for generating code – 82 percent say between 25 percent and 74 percent of their code was created with help from AI. This has a downstream effect on the infrastructure teams that deploy said code: 40 percent of respondents say security vulnerabilities are showing up more frequently, 40 percent say governance has become more challenging, 37 percent cited higher change volume, 35 percent see strains on the development pipeline, and 35 percent report infrastructure drift. Spacelift's report calls out the cognitive dissonance – a blameless formulation of "self-delusion" – among organizations adopting AI: 86 percent say they can govern it, while only 30 percent actually have a formal AI governance policy in place. The report advises organizations to start paying attention to AI-oriented metrics that few organizations bother to track, specifically the volume of AI-generated IaC in deployment pipelines, error rates due to AI-generated changes, and infrastructure drift attributable to AI changes. It also stumps for greater automation through IaC, for building governance to cover that automation, getting AI-generated code into governed IaC orchestration workflows, and planning for the governance of AI agents. ®

It’s looking like another tough week (month? year?) for Switchzilla amid reports of new serious vulnerabilities under attack. First up is a server-side request forgery bug in its Unified Communications Manager tracked as CVE-2026-20230. Cisco disclosed and patched this flaw in early June. The comms control platform doesn’t properly validate some HTTP requests, and an attacker could exploit this bug to gain root privileges on a compromised device. At the time, Cisco said that a proof-of-concept exploit was available – and now it seems unknown miscreants are putting that exploit code to use, with threat intel company Defused warning that it observed miscreants exploiting CVE-2026-20230 over the weekend. “The observed chain abuses the WebDialer SSRF to deploy a rogue Apache Axis service, uses that service to write a first-stage JSP file-writer, then drops a second-stage command-execution shell under /platform-services/axis2-web/,” the firm noted on LinkedIn. Cisco Catalyst SD-WAN zero day Then, a Mandiant advisory on Wednesday warned that a Cisco SD-WAN zero-day tracked as CVE-2026-20245 was exploited much earlier than initially disclosed, including at a communications service provider where the attacker elevated a compromised admin account to full root-level access. While the Google-owned threat hunting biz said it can't assess the full scope of the intruders' post-compromise activity, this SD-WAN device compromise could have been dire, potentially giving the attacker total visibility across an entire corporation's internet traffic. This is what makes SD-WAN zero-days such a hot target for government-sponsored spies looking to set up shop for long-term snooping activities. It also explains the rash of attackers battering Cisco SD-WAN devices since the start of the year. Cisco had issued an advisory for CVE-2026-20245 in early June, admitting that attackers had a head start on abusing this security hole. “In June 2026, the Cisco PSIRT became aware of exploitation of this vulnerability,” the vendor said at the time. In a Wednesday report, however, Google’s Mandiant incident response and consulting biz reported that exploitation of this bug – Cisco’s sixth SD-WAN vulnerability listed as under attack since the start of the year, and the second zero-day in two months – began much earlier. “In early 2026, Mandiant identified a threat actor targeting SD-WAN infrastructure at a service provider,” Mandiant threat hunters Chester Sng, Pete Boonyakarn, and Logeswaran Nadarajan wrote. “After gaining initial access, the threat actor exploited a zero-day vulnerability (CVE-2026-20245) in Cisco Catalyst SD-WAN to escalate privileges from a compromised administrative account to root-level access.” The attacker gained initial access via an unauthorized peering connection, abusing the SD-WAN fabric to authenticate between network components and facilitate Secure Shell (SSH) access. In this case, they authenticated to the SD-WAN manager device via SSH using the vmanage-admin account on the same victim devices. Then, they changed the default password on the admin account, authenticated directly to the SD-WAN Manager web application interface using the admin account, and exfiltrated SD-WAN fabric configurations. Likely in an effort to cover their tracks and not get caught, the attacker changed the password of the admin account back to its original one before terminating their active session. Neither the vmanage-admin nor the admin accounts on Cisco Catalyst SD-WAN controllers possess root shell access, however. To gain root access, the attacker exploited CVE-2026-20245, which allows an authenticated, local attacker to execute arbitrary commands as root by supplying a crafted file to the vulnerable system. The attacker uploaded a file named evil_tenant.csv that contained the exploit payload. Upon execution, the digital intruder created a user account named troot with full root privileges. Mandiant says it later observed the miscreant accessing this new troot account from the admin account using the substitute user command. The Register reached out to Cisco about the reported exploitation of CVE-2026-20230, plus Mandiant’s investigation into CVE-2026-20245, and did not receive any response. We will update this story when we hear back from the networking vendor. ®

We knew it was coming, but now it's official: Qualcomm is making a major push into the datacenter market. And though it is late to the game, the mobile-chip giant believes it can make an impact by delivering a lower total cost of ownership and better performance per watt than rival platforms. It has to go somewhere, and the company is already dominating the chip space elsewhere, Qualcomm's datacenter EVP and GM Tony Pialis said during the company’s Investor Day presentation on Wednesday. Pialis said that Qualcomm is already a winner in mobile, PC, and automotive, but will now be able to play in a market with well established competitors such as Nvidia. He started by addressing the obvious question of "are you too late" to the market. "When the company turns its attention to solve a new problem, we revolutionize the solution and push our way to the forefront," he said. "And folks I'm here to tell you today that is what we will and are doing in datacenter." Pialis’ portion of the presentation included the formal unveiling of the Dragonfly compute platform we’ve been hearing about for weeks, the … ahem … core of which is the C1000 CPU he showed off. Pialis claimed that the new datacenter-grade chips offer 2x better performance per watt and 30 percent more speed than competitors' processors. The CPU cores, Pialis explained, are based on Qualcomm's custom Oryon architecture and will operate at more than 5 GHz in a chiplet-based design featuring more than 250 cores. Most interestingly, Pialis argued that the design of the C1000 chips addresses the memory bottleneck facing AI datacenters with what Qualcomm is calling "High-Bandwidth Compute" (HBC) technology. Pialis described HBC as combining compute and memory more closely by integrating an XPU beneath a DRAM stack, claiming it delivers SRAM-like performance advantages inside a high-bandwidth memory package while reducing data movement and improving performance per watt. Those Dragonfly C1000 processors are expected to enter production in the second half of 2028. Qualcomm plans to offer multiple C1000 variants targeting agentic AI, general-purpose computing, and AI head-node workloads. But CPUs alone do not a datacenter pivot make, and Pialis said that there are three other datacenter segments Qualcomm is targeting along with new CPUs: connectivity, custom silicon designed for individual customers, and AI accelerators. There are also the aforementioned AI accelerators, which Qualcomm says will use its HBC technology to address memory bottlenecks in AI workloads. Microsoft CEO Satya Nadella and Meta CEO Mark Zuckerberg made guest appearances during Investor Day. Qualcomm said Microsoft plans to use its HBC-based AI accelerators, while Meta separately announced plans to deploy Dragonfly C1000 CPUs under a multi-generation agreement. Pialis also detailed new connectivity technologies that form part of the Dragonfly portfolio. According to Pialis, Qualcomm wants to enable new distances in cluster-to-cluster optical connectivity up to 20 kilometers with its new QAM16 coherent-lite optical modules. “We have everything you need to scale from millimeter technology to tens of kilometers,” Pialis said. As for custom silicon, that’ll involve making bespoke chips for what Pialis said will be Qualcomm’s “highest tier of customer” who needs someone to design and fabricate AI and cloud DC CPUs from end to end. What Qualcomm is bringing on the hardware side will be supplemented by Modular, a company that develops AI software stacks. Qualcomm announced on Wednesday that it had reached an agreement to acquire Modular in order to flesh out the software side of its Dragonfly endeavors, which the company said will give it access to hundreds of billions in new market space. ®

Writing prompts is so … 2025! AI influencers and industry luminaries have declared that prompts are out and loops are in, and maddeningly this has become blog fodder and grist for the news cycle. Never mind that AI agents, which are models using tools in a loop, have involved loops since people started yammering about them last year. Never mind that programming has always had constructs for repetition, even before the do-loop appeared in Fortran. The word from the AI-pilled is that if you're writing prompts and checking each response, as opposed to having AI agents address multi-step tasks with minimal input, you're doing it wrong. The recent focus on loops can be attributed in part to Peter Steinberger, the creator of OpenClaw who went on to join OpenAI, even though the likes of Andrej Karpathy were writing about and implementing AI loops months earlier. "Here’s your monthly reminder that you shouldn’t be prompting coding agents anymore," Steinberger wrote in a social media post on June 7. "You should be designing loops that prompt your agents." Ed Zitron, the harbinger of AI doom, landed the first body blow with his reply: "Does OpenAI bill itself for its token spend?" That's primarily what this is about – celebrating and evangelizing autonomous token consumption, spending that OpenAI and its peers would very much like to stimulate and capture. Imagine a for-profit but heavily indebted utility advising customers to remember to leave their lights and appliances on all night. That's about where we are. To underscore the spending assumptions baked into loops über alles, Zitron in a recent post skewered Anthropic's Boris Cherny for his advocacy of loops as the successor to prompts: "Pretty convenient for a guy who’s allowed to burn upwards of $130,000 a month in tokens by Anthropic." But back to the X thread: Gautham Pai, founder of corporate learning biz Jnaapti, answered Steinberger, "Oh god, LinkedIn will now start a new fad, 'Loop Engineering'. Harness Engineering is so last year. Loop Engineering is what you should be doing." In response, Steinberger quipped, "Don’t worry it’ll take 3 months until it’s there. We’ll be talking about fleets that design your loops then." A blog post titled "Loop Engineering" appeared that very day, not on LinkedIn, but on the blog of developer Addy Osmani. And unlike the humble-brag, jargon-strewn river of scintillating affirmation that is LinkedIn, Osmani's take is worthwhile if you happen to be actually trying to implement an AI agent loop. His conclusion is particularly worthwhile, though it rather undercuts the whole idea of loops: "The loop changes the work, it does not delete you from it." Another way to put that is, "Automate at your peril." If you set an inherently non-deterministic AI model on a task and expect flawless operation, you deserve to clean up the inevitable mess. The major AI companies would love to see set-and-forget service consumption. Your job is to be the human in the loop. Your livelihood may depend on that. ®

OpenAI and Broadcom have teamed up – with a little help from some of the former’s AI models – to develop the frontier model lab’s very first inference chip, dubbed Jalapeño, the companies announced in a press release on Wednesday. Details of the spicily named silicon are scarce in the announcement, with the company admitting that it’s running engineering samples of Jalapeño in its lab “at target frequency and power,” but noting that it won’t have any technical details to share until a report on its performance is released in the coming months. But that doesn't stop it from claiming that "early testing shows that Jalapeño will deliver performance per watt substantially better than current state-of-the-art." Alright then! OpenAI said that it designed Jalapeño itself, with Broadcom serving as its implementation and integration partner (i.e., they made the darn thing), but it wasn’t humans alone who helped come up with Jalapeño’s made-for-inference ASIC architecture: AI helped, natch, and the result is what OpenAI says is the fastest ASIC development cycle ever in the high-performance advanced semiconductor space. “Jalapeño was co-developed from initial design to manufacturing tape-out in just nine months,” OpenAI claimed. “That speed reflects deep software-hardware co-development with OpenAI’s engineering teams, Broadcom’s silicon implementation expertise, and the use of OpenAI models to accelerate parts of the design and optimization process.” In other words, AI is now helping design the chips it’ll run on. Here’s hoping they ironed out the hallucinations before heading to production. OpenAI explained in the announcement that Jalapeño is just the first of its AI accelerators, with the chip serving to define its “vision for the future of LLM inference,” and one that will involve OpenAI controlling the entire stack behind its models and products. According to the release, OpenAI envisions a future where it doesn’t just own the frontier models and the products built on top of them, but the infrastructure underneath as well. “Chip architecture, kernels, memory systems, networking, scheduling, deployment systems, and product experience” are all part of OpenAI’s full-stack vision, which it said will enable it to make models “faster, more reliable, and more affordable.” And more locked in, one would assume, like a proverbial walled garden. The Apple of the AI world, if you will. OpenAI is far from alone in developing its own silicon to help power AI – most of the giants in the space, including Amazon, Google, Meta, and Microsoft, have been building and using their own silicon for AI for several generations now, and OpenAI arch-rival Anthropic is reportedly considering a similar move. No telling, either, how OpenAI is intending to continue funding this capital-intensive initiative, given that it ran an operating loss of over $20 billion last year, according to leaked financials reported by Ed Zitron, and has apparently committed massive amounts ($600 billion? $1.4 trillion?) to infrastructure spending over the next few years. But hey – if we questioned AI economics, nothing would ever get built, would it? ®

Microsoft, its friends, and international law enforcement - with an AI assist - disrupted two widely used pieces of malware and their infrastructure, in what Redmond describes as a novel approach to cybercrime disruption that targets the cyberattack supply chain instead of a single tool or service. “What’s new is how we’re combining AI analysis with an expanded use of that law,” Steven Masada, assistant general counsel for Microsoft’s Digital Crimes Unit, said in a Wednesday blog, referring to the Racketeer Influenced and Corrupt Organizations Act (RICO). Typically Microsoft uses RICO and other US laws to take legal action against a single cybercrime service or infrastructure. The disruption involved the takedown, suspension, and blocking of more than 200 domains and command-and-control (C2) servers that formed the backbone of StealC and Amadey infrastructure. Multiple security companies, including ESET, BitSight, Mitsui Bussan Secure Directions (MBSD), IBM X-Force, and Proofpoint, also played a role in dismantling the alleged operations. Combined with the earlier SocGholish disruption announced last week, a Europol-led law enforcement coalition flagged and restricted cryptocurrency assets valued at more than $47 million and recovered about 27 million stolen credentials. StealC and Amadey are two separate malwares developed by different criminal crews, but they used the same infrastructure and were operating in concert. StealC collects multiple browser credentials and cookies, cryptocurrency wallets, chats from messaging apps, and other sensitive data, and exfiltrates the stolen goods to a C2 server. It also works as a secondary loader, allowing criminals who rent the stealer to download additional malware on compromised devices. Amadey is a malware-as-a-service used to deliver StealC and other stealers, plus other types of malware including remote access trojans, cryptominers, and ransomware. In just the first two weeks of May, Amadey and StealC were linked to more than 140,000 infected computers globally, according to Microsoft. “It’s no longer enough to go after threats one by one,” said Masada. “We need to interrupt how the attacks are put together.” In this case, Redmond’s investigators used Copilot and other AI tools to analyze both malwares and their infrastructure, “asking questions in plain English instead of manually combing through complex code,” Masada wrote. “That helped surface key details, uncover hidden data, and test findings in a fraction of the time, turning what would have taken hours or days into minutes and enabling the team to spot connections faster.” One of these key details: both Amadey and StealC used the same infrastructure. This allowed Redmond’s legal team to treat both malwares as part of a single conspiracy under RICO and bring civil claims against five defendants allegedly involved across both operations. “Defendants comprise a group of cybercriminals operating a Malware as a Service enterprise that leverages malicious software commonly known as the Amadey Malware Suite and StealC Malware Suite (the "MaaS Enterprise"),” the court documents say. “Through the Maas Enterprise, Defendants and their accomplices have victimized hundreds of thousands of innocent computer users, including many users of Microsoft's software and services.” ®