News related to
developers

Rotating chairman David Wang said at Huawei Connect 2026 in Shanghai on Thursday that the company plans to release the 960DT in the first quarter of 2027 and the Ascend 960PR in the third quarter. The roadmap places Huawei’s next generation of accelerators at the centre of its effort to build domestic alternatives for training and running increasingly demanding AI models.

Huawei also unveiled the Ascend 960 SuperPoD and an upgraded version of its UnifiedBus interconnect technology, designed to link large numbers of processors so they can function as one computing system. The company said a combination of UnifiedBus and its Hi-ONE optical system could connect as many as 4,000 processors within a SuperPoD, extending high-speed optical links deeper into the machine.

The strategy reflects a broader shift towards system-level performance rather than relying solely on the capability of individual chips. Advanced AI workloads require thousands of processors to exchange data rapidly, making networking bandwidth, memory access and cluster efficiency increasingly important alongside raw chip performance.

Wang said Huawei has developed 11 semiconductors based on UnifiedBus technology for use across its large computing systems. Its biggest planned superclusters will be able to support up to one million AI processors, he said, while the company has already shipped more than 1,000 smaller supernodes to more than 370 customers.

Huawei did not identify those customers or give a detailed breakdown of shipments. It also did not provide independently verified benchmarks directly comparing the new 960-series chips with Nvidia’s leading accelerators, leaving questions over single-chip performance, software compatibility and manufacturing scale.

The company had previously said the Ascend 960 would offer twice the computing power, memory capacity, memory bandwidth and interconnect ports of the Ascend 950 generation. Huawei has also been developing the Atlas 960 SuperPoD and SuperCluster, with the larger system intended to combine more than one million neural processing units when it reaches the market.

Huawei’s current product roadmap also includes the Ascend 950DT, due in the fourth quarter of 2026. The company has said that processor will use its HiZQ 2.0 high-bandwidth memory, providing 144 GB of capacity and 4 terabytes per second of bandwidth. Those specifications are intended to support both model training and the decode stage of inference workloads.

US restrictions have limited China’s access to cutting-edge Nvidia processors and advanced semiconductor manufacturing equipment, giving domestic suppliers greater room to expand. Huawei has emerged as the most prominent of those suppliers, but Nvidia retains a major advantage through its CUDA software ecosystem and a broad global developer base.

Chinese AI developers have increasingly turned to domestic accelerators for inference, where trained models generate responses, even as many frontier laboratories continue to depend on Nvidia hardware for the most demanding training jobs. That split has made Huawei’s ability to improve both hardware and software a critical test of whether China can reduce its dependence on foreign computing infrastructure.

Huawei is also contending with shortages of high-bandwidth memory, a key component in AI accelerators. The company has designed memory for some newer Ascend products, but supply constraints across the industry have increased costs and complicated efforts to scale production quickly enough to meet demand.

The group is seeking to offset those limitations through architectural changes, advanced packaging, optical networking and tighter integration between chips, systems and software. Its SuperPoD approach is intended to compensate for weaker performance from individual processors by allowing much larger numbers of chips to work together with lower latency.

Bae said South Korea was still trying to close the gap with leading AI powers while creating new markets, making its position different from countries or companies at the technological frontier. “We cannot afford to slow down AI right now,” he wrote on social media, describing the government’s approach as one of “speed responsibility” rather than speed control.

His remarks place Seoul firmly on the acceleration side of a widening debate over whether developers should moderate the pace of capable AI systems. Anthropic chief executive Dario Amodei has called for a slowdown in frontier model development to allow more time for safety work, while OpenAI chief executive Sam Altman and other technology leaders have voiced support for stronger precautions as capabilities advance.

Bae said South Korea should continue expanding research, industrial adoption and home-grown model development while strengthening protections against deepfakes, cyber threats and privacy violations. Faster progress, he said, required safety and trust measures to move at the same pace rather than follow after deployment.

The minister pointed to a broader technology agenda that includes sovereign foundation models, systems capable of processing vision and language, and models able to interact with physical environments. He said the country also needed to prepare for AI becoming deeply embedded across industries and everyday life.

The government has made domestic foundation-model development a central part of its ambition to place South Korea among the world’s top three AI powers. The Ministry of Science and ICT has been running a sovereign AI foundation model programme intended to improve domestic capabilities and reduce dependence on foreign technology.

Three teams led by Upstage, SK Telecom and LG AI Research advanced after the programme’s second-stage evaluation this year. The ministry has said models developed under the initiative were built with comparatively limited resources but were released as open source as part of an effort to expand the domestic AI ecosystem.

Bae indicated that further details on the next phase of the foundation-model programme would be announced after ongoing reviews. His comments also referenced government work on security-focused AI models and frontier systems, alongside an “AI for All” initiative intended to spread adoption across the economy.

The policy push is being developed alongside South Korea’s AI Basic Act, which establishes rules covering both the development of artificial intelligence and the creation of a framework for trust. The law’s stated purpose includes protecting rights and human dignity while supporting quality of life and national competitiveness.

That legal structure gives Seoul a basis for pursuing faster adoption without presenting safety as a separate objective. Bae said competitiveness and sustainability were linked, arguing that development speed and safeguards should advance together.

The global dispute has intensified as leading executives and policymakers weigh the economic benefits of faster AI deployment against concerns over misuse and loss of control. Amodei has proposed independent safety evaluators, coordination among major laboratories and greater international cooperation, while Nvidia chief executive Jensen Huang has argued against slowing technological progress and favours advancing safety alongside innovation.

European Commission President Ursula von der Leyen on Wednesday backed calls for greater caution and said frontier AI laboratories would be invited for talks on managing risks. Meta chief executive Mark Zuckerberg, by contrast, said AI companies already had strong incentives to build systems safely and did not require an industry-wide slowdown.

Bae’s intervention highlights the competitive pressures facing countries outside the United States and China, where most frontier AI investment and model development remains concentrated. South Korea has strong semiconductor and technology industries, but its government has acknowledged the need to strengthen domestic models, computing capacity and commercial adoption to narrow the gap.

Rahul Chari, PhonePe’s founder and chief product and technology officer, said the company is concentrating on AI deployments that produce measurable gains in operating, product or revenue metrics rather than adding the technology indiscriminately across its platform.

Chari told the Global Fintech Fest in Mumbai that the company’s approach starts with identifying operational bottlenecks where artificial intelligence can change outcomes. “Not everything that is an experience today gets genuinely better using AI,” he said. The emphasis, he added, is on a baseline shift in performance rather than the number of AI-labelled features. That approach is shaping investment across customer interactions, back-office workflows and engineering, while preserving conventional systems where automation offers little advantage.

The fintech processes close to 12 billion transactions a month and operates customer support with roughly 500 agents. Automation has increased the share of support cases handled without human intervention from about 90 per cent to 94 per cent, and PhonePe is targeting 96 per cent.

Chari said the remaining cases are harder because they often involve unusual circumstances, scattered information and the need for rapid responses. Large language models are being used to address those complex interactions, while human staff remain focused on cases that require judgement or escalation.

Fraud investigation is another area where automation has advanced. About 80 per cent of investigations can now be handled through automated systems, allowing specialist teams to concentrate on complicated incidents and new fraud patterns, according to the company’s account of its AI programme.

PhonePe is also extending artificial intelligence into consumer-facing services. It has been developing natural-language search that allows users to navigate the app, initiate tasks and reach relevant services using voice or text commands. The company is exploring broader conversational capabilities for personal finance, insurance and investment products.

Its search system, launched earlier this year using Microsoft Foundry, can interpret requests such as making a payment, recharging a FASTag or locating transaction information. PhonePe has said the interface is designed to reduce dependence on conventional menu navigation by directing users according to their stated intent.

Merchant products are another focus. PhonePe launched an AI-powered integration layer in May aimed at helping online businesses connect to its payment gateway through conversational instructions, reducing a process that can otherwise require developers to work through documentation and technical configuration.

The company followed in June with SmartPages, an AI-powered payment-page creator embedded in its payment gateway dashboard. Merchants can describe what they are selling in ordinary language, including vernacular languages, and the system generates product copy, selects relevant customer information fields and structures a mobile-oriented payment page.

PhonePe is considering further agentic functions for its business app, including answering merchant and customer queries and helping businesses create and distribute offers or coupons. The company has also examined tools that could connect commerce activity with services such as WhatsApp and Instagram.

Artificial intelligence is being incorporated throughout PhonePe’s software development process, including solution design, coding, code reviews, testing, deployment, troubleshooting and debugging. Internal tools include TraceMine for crash analysis, Lens for initial root-cause analysis using logs and deployment data, and KadAI for code orchestration and agent-based development.

Corporate functions are being brought into the same framework through an internal Agent Hub designed to let non-technology teams automate workflows. Legal, compliance, finance and human resources teams are among those using the platform.

PhonePe says the system was built for a regulated financial-services environment, with controls covering data access, leakage and localisation. Its architecture includes an in-house large language model router, open-source models hosted on company premises, and internal data and tools.

One target is the company’s new-product introduction process, where reviews involving compliance, privacy, legal, trust and safety can take about 30 days. PhonePe is exploring whether AI-enabled workflows can compress that process to less than a day.

The regulator said the work will focus on creating clearer rules for datasets used by embodied AI systems, which combine artificial intelligence with physical machines such as humanoid robots. The effort is intended to support more consistent data collection, processing, evaluation and use as companies race to improve machines that must perceive, reason and act in real-world environments.

China has already moved to build a broader standards framework around the sector. National standards published on August 27 cover the quality of real-world embodied-intelligence data and technical requirements for data-generation platforms, while several additional specifications remain under development.

Among projects overseen by the National Data Administration are proposed standards covering the sources and constituent elements of high-quality embodied-intelligence datasets, simulated synthetic-data generation and processing, and specifications for data collection and model training at training bases. The projects are being developed under the National Data Standardisation Technical Committee.

One draft programme, registered in April, sets a 12-month timetable for a standard on data collection and model training at embodied-intelligence training bases. Its drafting group includes the China Electronics Standardization Institute, Beijing Institute of Technology, the Institute of Software at the Chinese Academy of Sciences and robotics companies.

A separate draft addresses synthetic data, which is becoming increasingly important because collecting enough physical-world robot interactions can be expensive, slow and difficult to reproduce. China’s policy framework encourages the use of simulation and synthetic techniques to expand supply where genuine operating data are scarce or costly to obtain.

The National Data Administration has made high-quality datasets a central part of its wider artificial-intelligence strategy. An implementation plan issued in June called for faster construction of datasets in strategic and emerging fields including embodied intelligence, intelligent driving and the low-altitude economy. It also called for data covering physical interaction, environmental perception and motion control in key scenarios.

The plan directs authorities and industry participants to improve data cleaning, enhancement, labelling, alignment and quality inspection, while developing national standards covering formats, categories, annotation and quality assessment. The regulator said such work should produce datasets that are structurally complete, diverse, accurately labelled and suitable for model development.

China’s push reflects the distinct data demands of robots operating in the physical world. Unlike text-based systems trained largely on digital information, embodied AI requires large volumes of multimodal material that can include video, images, sensor readings, point clouds, movement trajectories and feedback from physical interaction.

Data supply has therefore emerged as a competitive constraint for developers seeking to train machines to perform long sequences of actions reliably across changing environments. The National Data Development Research Institute has said embodied intelligence and world models are driving sharply higher demand for three-dimensional, video and other multimodal data, while also raising requirements for collection and annotation quality.

The government has also been expanding the underlying pool of AI-ready datasets. At a national data industry event in Guiyang in August, the National Data Administration said more than 126,000 high-quality datasets had been built across China, with a combined volume exceeding 1,815 petabytes.

Officials described a system combining a national dataset management platform, an open-source community and a data-labelling network with sector-specific pilot programmes. The framework includes 106 leading organisations and 140 pilot tasks, with regional authorities encouraged to develop projects suited to local industrial strengths.

Embodied intelligence has featured in those pilots alongside other emerging fields. The administration has promoted cooperation among robotics companies, universities, research institutes and public bodies to increase the quantity and usefulness of machine-training data while testing standards before wider application.