News related to
compatibility

The Hangzhou-based AI developer released Ascend versions of several key components on September 30, including its DeepGEMM matrix-multiplication library and DeepEP communications software. The projects are designed to work with Huawei’s Ascend processors while keeping public interfaces aligned with versions built around Nvidia hardware, reducing the amount of code developers may need to rewrite when moving workloads.

That portability targets a critical part of Nvidia’s competitive advantage. Nvidia’s CUDA software platform has become deeply embedded in AI development, giving programmers mature libraries, tools and optimisation techniques built over years. Competing chips therefore have to overcome not only differences in raw computing performance but also the cost and complexity of shifting software away from CUDA.

DeepSeek’s DeepJIT project takes a further step by providing a shared runtime interface for Nvidia CUDA graphics processors and Huawei Ascend neural-processing units. It handles compilation, caching, loading and launching of specialised AI kernels while retaining backend-specific code where necessary. The approach does not make the two architectures interchangeable, but it can reduce friction for developers maintaining software for both.

DeepGEMM-Ascend is fully API-compatible with DeepSeek’s original DeepGEMM package and supports BF16, FP8 and FP4 calculations used in demanding AI workloads. DeepSeek says the library uses Ascend-specific optimisation methods to approach the hardware’s performance limits while shielding programmers from lower-level complexities including memory layouts, alignment constraints and address calculations.

Its DeepEP-Ascend library similarly aligns its public buffer interfaces with the Nvidia version while using Huawei’s communications technologies underneath. DeepSeek has also added Ascend support to DeepSelect, a specialised component used for sparse-attention and sampling operations, and released Ascend sparse-attention kernels for its model infrastructure.

Huawei, meanwhile, has been opening more of the software surrounding Ascend. Its CANN platform — the company’s counterpart to the programming infrastructure required to exploit AI accelerators — supports frameworks including PyTorch and works with open-source projects such as Triton, vLLM and veRL.

Huawei said last month that CANN had moved to sustained community-driven open-source development, with external programmers accounting for 61 per cent of its CANN developers. The company said the community had more than 5,200 monthly active developers, while more than 40 models had been natively pre-trained using Ascend and CANN.

The combination matters because China’s challenge to Nvidia has increasingly shifted from producing an alternative accelerator to assembling a usable full computing stack. Hardware can win orders only if model developers can train and run workloads efficiently without sacrificing too much engineering time or performance.

DeepSeek’s software also gives Huawei an influential independent developer helping test and improve that stack. The AI laboratory became prominent for building capable models with an emphasis on computational efficiency, and its infrastructure code is closely watched by developers seeking lower-cost approaches to training and inference.

The effort comes as US export controls continue to restrict China’s access to Nvidia’s most advanced processors, creating stronger incentives for technology groups to adopt domestic alternatives. Those restrictions have helped expand the strategic importance of Huawei’s Ascend line, even though China still faces constraints in advanced semiconductor manufacturing.

Nvidia nevertheless retains formidable advantages. CUDA has a vast installed developer base and an extensive collection of mature libraries, while Nvidia’s leading accelerators remain central to major AI training systems outside China. Compatibility layers and open-source libraries also require sustained testing, documentation and optimisation before they can match the reliability of an established platform across diverse workloads.

The Bank of Korea said the Bank of Korea Won International Wire Network started trial operations at 9am local time with KB Kookmin Bank, Woori Bank, Hana Bank and Shinhan Bank participating. Two of the banks completed transactions worth about 1.4 billion won during the initial test, according to the central bank.

Full operation is scheduled for January 2027, when foreign financial institutions are expected to gain wider access to the network. The system is designed to allow won payments and settlement during the business hours of major overseas financial centres rather than limiting activity to South Korea’s domestic banking day.

Under the planned framework, offshore investors will be able to trade the won with other foreign participants using accounts opened with Registered Foreign Institutions for KRW Business, known as RFI-Ks, in their home markets. This is intended to reduce the need for investors to open and operate separate accounts directly with financial institutions in South Korea.

The network will function continuously from 9am on each business day until 9am the following day, effectively providing round-the-clock settlement on weekdays. It will remain closed on weekends and public holidays. Foreign banks are also due to participate in the trial, although the Bank of Korea has not specified when they will enter the testing phase.

The central bank said it would monitor the pilot, refine operating arrangements and expand participation before the system moves into full service. It expects the new infrastructure to improve foreign investors’ access to won settlement and support wider international use of the currency.

The trial follows South Korea’s move in July to extend onshore dollar-won trading to a 24-hour schedule. That change removed the previous gap between Seoul trading hours and active periods in major centres such as London and New York, allowing transactions to continue through the global trading day.

Authorities have paired the longer trading window with changes to settlement arrangements because extended dealing hours alone do not remove operational barriers if investors cannot complete payments outside domestic banking hours. The new network is intended to address that mismatch by keeping settlement infrastructure available while overseas markets are active.

The initiative forms part of a broader programme to open South Korea’s foreign-exchange market and make local assets easier for global institutions to access. Seoul has also been seeking improvements in market infrastructure as it pursues developed-market classification for its equities from MSCI, whose accessibility assessments examine areas including foreign-exchange convertibility, investment flows and operational arrangements.

South Korea has already introduced the RFI framework to permit eligible overseas financial institutions to participate directly in the domestic foreign-exchange market. The settlement network extends that approach by enabling those institutions to hold and transfer won through designated accounts linked to the central bank’s payment infrastructure.

The Bank of Korea has also been modernising its core settlement systems. Its 2025 payment and settlement systems report, released last week, said operating hours for BOK-Wire+ were extended in March and that work continued on the international won network and the adoption of the ISO 20022 financial messaging standard.

The central bank has said the international network will use real-time gross settlement, a structure in which transactions are settled individually and continuously rather than being netted for later settlement. That approach is intended to limit credit risk during the payment process while supporting cross-border compatibility.

Preparations have included linked-system testing and business-continuity planning before the scheduled commercial launch. The authorities have also revised foreign-exchange regulations to accommodate the wider use of overseas won accounts and settlement channels.

The assessment evaluated multiple leading AI computing resource management platform providers in China and noted Phancy’s industry-leading technological maturity across heterogeneous accelerator compatibility and adaptation, large-scale cluster operations, and delivery in hybrid cloud environments. Drawing on its accumulated capabilities in heterogeneous computing resource management, fine-grained GPU virtualization and agent-oriented computing services, Phancy received full scores and outperformed the industry benchmark in four out of six evaluated dimensions[1].

According to the IDC report, as AI moves into large-scale deployment, the growing adoption of AI agents is reshaping cloud infrastructure. In the meantime, AI infrastructure is evolving beyond basic computing capacity to become a core part of the enterprise technology stack, supporting digital and intelligent transformation.

Dr. Dai Wenyuan, Founder and Chief Executive Officer of Phancy, said: “The AI industry is moving beyond the ‘model race’ and entering a new phase of large-scale deployment. For enterprises, the real challenge is no longer whether they can access large language models, but about how to translate agentic AI capabilities into sustained productivity while keeping costs under control and operations stable. IDC’s assessment reflects the growing importance of unified management, flexible scheduling and fine-grained operations for heterogeneous computing resources. Phancy will continue to advance its capabilities in heterogeneous computing resource scheduling and agent infrastructure, helping more enterprises unlock the value of AI with greater efficiency, cost-effectiveness and reliability.”

Five Core Capabilities Behind Phancy’s Vendor-Neutral Unified Management Platform

At the core of Phancy’s offering is vendor-neutral unified management software that supports large-scale training and high-concurrency inference for mainstream models across chips from over ten leading vendors[2]. Its core technological capabilities include:

• Code-compatible multi-chip virtualization: Phancy supports the partitioning of a single AI accelerator across multiple workloads, enabling mainstream frameworks and inference engines, including PyTorch, TensorFlow, vLLM and SGLang, to run without changes to application code. Resource quotas help maintain stability when multiple workloads run concurrently.
• Three-tier resource pools and intelligent workload scheduling: Phancy’s platform provides three types of resource pools: dedicated, shared and overcommitted pool. Resources can be intelligently scheduled based on four dimensions (workload priority, hardware topology, real-time load and resource requirements), enabling large-scale training and high-concurrency inference workloads to run on the same infrastructure.
• AI gateway and dual-track metering: The AI gateway provides unified management of model routing, API calls and audit activities. Usage is measured through both accelerator hours and token consumption, helping enterprises track costs and resource utilization from the perspectives of both resource consumption and model usage.
• Agent-oriented computing services: Phancy’s platform turns heterogeneous computing resources into standardized services that AI agents can call on demand. These services support partitioning, reuse and elastic scaling. Combined with intelligent routing through the AI gateway, this accommodates the frequent calls and fluctuating workloads associated with AI agents, while token quotas and circuit-breaking mechanisms also help enterprises manage usage risks and costs when predefined limits are exceeded.
• Operations and maintenance for heterogeneous computing resources: Phancy’s platform provides end-to-end observability spanning nodes, accelerators, workloads and model services. It supports fault isolation, workload rescheduling and private deployment in environments without public internet access, helping enterprises improve operational efficiency across complex heterogeneous computing environments.

The report highlighted a use case in which a Chinese policy bank significant improved resource utilization by using Phancy’s three-tier resource pool architecture to manage computing resources across three data centers in two locations, while also enabling a smooth business migration.

Looking ahead, Phancy will continue to advance its AI computing resource management platform as AI infrastructure shifts from traditional “cloud + AI” architectures toward greater cloud-native, agent-oriented integration. The Group will continue to help enterprises manage computing resources more efficiently, deploy AI across a wider range of production environments, and turn AI into sustainable business productivity at scale.

Hashtag: #Phancy

The issuer is solely responsible for the content of this announcement.

About Phancy Group

Phancy Group (6682.HK) is a leading full-stack AI cloud services platform, providing comprehensive solutions for the AI 2.0 era. Our offerings include SageAIOS, HAMi vGPU and ModelHub XC, delivering efficient and scalable AI infrastructure with end-to-end capabilities. We provide a complete solution from heterogeneous compute resource management and optimization to the deployment of intelligent agent models. These solutions empower digital transformation across a wide range of industries, supporting our vision of building a large-scale and efficient “Token Factory.”

Guided by the mission of “AI for Everyone” and positioned as the “Navigator of AI,” Phancy Group is committed to becoming a global leader in Artificial General Intelligence.

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.

Next-generation built-in oven features 350°C pizza cooking, an 81-liter capacity and smart connectivity ahead of its planned global launch in 2027.

BERLIN, GERMANY – Media OutReach Newswire – 2 September 2026 – Midea will showcase its new SpaceMaster Series built-in oven at IFA 2026 in Berlin, offering an early look at its next generation of smart kitchen appliances ahead of a planned global market launch in 2027.

At the core of the SpaceMaster Series is Midea’s Graphene Turbo Hot Air Technology, designed to deliver rapid and even heat distribution. Leveraging graphene’s high thermal conductivity, the heating system responds in as little as 0.2 seconds, enabling precise temperature control while reducing preheating time.

The technology also supports a dedicated 350°C pizza program. Combined with the included refractory pizza stone, the high-temperature cooking mode is designed to produce crisp, stone-baked-style pizza at home.

Key features of the Midea SpaceMaster Series include:

• Graphene Turbo Hot Air Technology: Rapid 0.2-second heating response for precise and energy-efficient cooking.
• 350°C Pizza Program: High-temperature cooking with a dedicated pizza stone for crisp, pizzeria-style results.
• 81-Liter Capacity: A spacious cavity supports simultaneous cooking across three levels with independent temperature control.
• Smart and Easy-Care Design: Dual self-cleaning, A++ energy efficiency, Midea SmartHome App connectivity, and compatibility with Google Home and Amazon Alexa.

The 81-liter oven is designed for households preparing multiple dishes at once, while its multi-level cooking system enables different foods to be cooked simultaneously with individual temperature settings. A dual self-cleaning system simplifies maintenance, while connected controls provide additional flexibility for everyday use.

“Cooking at home has become more than a routine task for many consumers; it is increasingly part of how people experience quality, creativity and enjoyment in everyday life,” said a Midea spokesperson. “With the SpaceMaster Series, we are combining advanced thermal technology with intuitive operation to make professional cooking capabilities, including 350°C pizza baking, more accessible in the home.”

Hashtag: #Midea

The issuer is solely responsible for the content of this announcement.

About Midea & Midea MOA

Midea is a global technology group ranked No. 246 on the 2025 Fortune Global 500 and No. 39 among the world’s most valuable technology brands in the Brand Finance Top 100 2026. Midea has ranked No. 1 globally in microwave appliances for 10 consecutive years (Euromonitor, 2016–2025). Its Microwave & Oven Appliance (MOA) division generated USD 3.05 billion in revenue in 2025.

Midea operates seven global R&D centers, holds more than 5,500 granted patents and serves markets in more than 160 countries and regions.

The Malaysian Palm Oil Board has developed Sawit EcoTherm, a non-conductive dielectric fluid made largely from palm-derived materials that allows servers and other computing components to be immersed directly in coolant. The technology eliminates the need for conventional evaporative cooling towers and is being positioned for AI and hyperscale facilities where rising processor densities are producing substantially more heat.

Sawit EcoTherm remains at the pre-commercial stage, with pilot projects, technical validation and partnerships with data centre operators among the steps required before widespread deployment. Researchers say the fluid could reduce cooling-related energy consumption by about 30% to 50%, largely by removing server fans and transferring heat directly from electronic components.

MPOB estimates that conventional water-based cooling across Malaysia’s expanding data centre industry could eventually require as much as 876 million litres of water a day. Other projections based on currently identified developments are lower, illustrating the uncertainty surrounding future capacity and cooling configurations. Data centres planned across Selangor, Johor and Negeri Sembilan have been estimated to require more than 445 million litres daily if projected demand materialises.

The pressure is particularly acute in Johor, which has become one of Asia’s fastest-growing data centre clusters because of its proximity to Singapore, available industrial land and established power and connectivity infrastructure. Water availability and grid capacity have increasingly become constraints as developers seek to commission large AI facilities.

Malaysia’s data centre and cloud computing investment approvals reached RM144.4 billion between 2021 and the middle of 2025. Global technology and infrastructure groups including Microsoft, Amazon, Google, ByteDance, Equinix and AirTrunk have committed billions of dollars to expanding computing capacity in the country.

AirTrunk announced plans this year to invest about $3 billion in two additional hyperscale campuses in Johor, adding more than 280 megawatts of capacity. Equinix separately unveiled an investment exceeding $190 million for another Kuala Lumpur facility designed to support more than 2,200 cabinets and equipped for advanced liquid cooling.

Those developments are accelerating interest in alternatives to traditional air and water cooling. AI servers can generate several times more heat than conventional computing equipment, making direct liquid cooling and immersion systems increasingly attractive as rack power densities rise.

Sawit EcoTherm uses palm-derived ester formulations engineered to provide electrical insulation while absorbing heat from processors, memory, power supplies and graphics processing units. MPOB has developed two formulations, EcoTherm 5s and EcoTherm 8s, both with dielectric breakdown strength exceeding 75 kilovolts.

The fluids are designed for single-phase immersion systems, where coolant remains liquid while circulating through a closed loop. The 8s formulation has a minimum flash point of 200 degrees Celsius, while the 5s version has a minimum of 160 degrees. Their characteristics are intended to combine thermal performance with relatively low pumping requirements.

Palm-derived material can also offer a cost advantage. Commercial synthetic and petroleum-based immersion fluids can cost about RM50 to RM80 per kilogram, while the raw-material cost for producing the palm formulation has been estimated at roughly RM20 to RM25 per kilogram. Final commercial pricing has not been announced.

MPOB says up to 95% of the material used in Sawit EcoTherm can come from renewable bio-based sources. The formulations are biodegradable and free from PFAS, a class of persistent chemicals that has attracted growing regulatory scrutiny because of its resistance to environmental degradation.

The technology nevertheless carries broader sustainability questions because expansion of oil-palm cultivation has historically been associated with deforestation, biodiversity loss and pressure on wildlife habitats. The environmental case for palm-based coolant will therefore depend partly on traceable sourcing, lifecycle emissions and whether demand can be met without encouraging additional land conversion.

Immersion cooling itself is not new. Vegetable-derived, mineral-oil and synthetic dielectric fluids are already available, while many operators developing AI facilities favour direct-to-chip liquid cooling because it can be integrated into server architectures without fully submerging equipment. Commercial success for Sawit EcoTherm will depend on compatibility with hardware, long-term fluid stability, maintenance requirements and acceptance by server manufacturers.