AI infrastructure

The 2025 WAIC features Huawei’s debut of the Atlas 900 A3 SuperPoD, built on the Ascend 384 Super Node. This super-node utilizes advanced bus technology for high-bandwidth, low-latency interconnection between 384 NPUs, addressing communication bottlenecks in large AI clusters. Huawei’s CloudMatrix 384 (CM384) AI cluster, built around Ascend chips, delivers 300 PFLOPs of dense BF16 compute power, reportedly surpassing NVIDIA’s GB200 NVL72. Analysts suggest Huawei’s scaled solution surpasses current market offerings from NVIDIA and AMD.

OpenAI is significantly expanding its AI infrastructure through a massive deal with Oracle to build new data centers across the US as part of its Stargate initiative. These centers, requiring 4.5 gigawatts of power, will house over two million chips and support wider access to advanced AI. This supports OpenAI’s pledge to invest heavily in US AI infrastructure, potentially exceeding $500 billion, and is projected to create over 100,000 jobs. Collaborations also include SoftBank and Microsoft, highlighting the extensive industrial effort behind AI development.

CoreWeave (Nasdaq: CRWV) announced a private offering of $1.5 billion in senior notes due 2031, guaranteed by its subsidiaries. The AI hyperscaler intends to use the proceeds to repay debt and cover expenses, bolstering its position in the expanding AI infrastructure market. The offering targets qualified institutional buyers and non-U.S. persons, adhering to securities regulations. CoreWeave provides cloud solutions for accelerated computing and cautions investors about forward-looking statements, advising due diligence and consulting SEC filings.

Nvidia CEO Jensen Huang, despite the company’s $4 trillion market cap, feels constant pressure, stating Nvidia is “30 days from going out of business.” He attributes this to the rapid pace of technological obsolescence in the chip and AI industries. Nvidia faced near-bankruptcy twice: once due to the Dreamcast chipset failure and again from a graphics card design flaw leading to recalls. Huang’s commitment to innovation fuels his drive to guide Nvidia far into the future.

Microsoft announced a second major round of job cuts this year, potentially impacting up to 9,000 roles across departments, geographies, and levels. This move aims to streamline operations and reduce costs amidst increasing AI infrastructure expenses and broader industry trends of workforce adjustments. These cuts represent less than 4% of Microsoft’s global workforce.

IREN Limited has appointed Anthony Lewis as its new Chief Capital Officer to bolster its capital markets strategy and fuel aggressive growth in the AI infrastructure sector. Lewis brings over two decades of financial markets experience, including a significant tenure at Macquarie Group, to shape IREN’s capital structure and financing initiatives.

DataCanvas launched Alaya NeW Cloud 2.0, a next-gen intelligent computing platform featuring serverless architecture and reinforcement learning. It offers on-demand AI infrastructure, targeting compute-intensive applications with a user-friendly toolchain. By offering a “pay-as-you-go” model and streamlining AI development, the platform aims to reduce costs, provide efficient performance, and make AI computing accessible to developers worldwide. The platform also introduces the world’s first reinforcement learning intelligent computing service, AgentiCTRL, for improved efficiency.

Broadcom unveiled Tomahawk 6, the world’s first 102.4Tbps data center switch chip. Designed for AI, it doubles existing switch performance, supporting up to 100,000 GPUs and boosting GPU cluster utilization. The chip offers flexible architecture with features such as 100G/200G SerDes interfaces and CPO support. Tomahawk 6 promises to reduce AI training costs, with further energy efficiency improvements planned by the end of 2025.

At Computex 2025, NVIDIA CEO Jensen Huang repositioned the company as an AI infrastructure architect, unveiling advancements in consumer GPUs, data center solutions, industrial metaverse tools, and robotics. Key highlights included the Blackwell architecture for AI training/inference and Isaac GR00T N1.5, a foundation model for humanoid robots. Huang emphasized physical AI as the next industrial revolution, citing partnerships with firms like Boston Dynamics and Morgan Stanley’s projection of a $60T humanoid market by 2050. NVIDIA aims to dominate through its full-stack ecosystem, bridging AI models, autonomous systems, and simulation tools to power smart factories and cities. (100 words)

The Kunpeng & Ascend Developer Conference 2025 in Beijing showcased Huawei’s AI infrastructure innovations under the theme “Innovation Through Collaboration.” Key releases included Kunpeng’s modular AI+ Solution Suite, open-source platform openFuyao for hybrid computing, and Ascend’s SuperNode Architecture, boosting cluster training efficiency. Upgraded tools like the MindSpeed RL Suite and CATLASS Operator Library aim to accelerate AI development. Huawei reported 665,000 global developers and 8,800 partners, with real-world deployments like RAG Solution 1.0 yielding 40% latency reductions. Emphasizing ecosystem growth and open standards, Huawei positions itself as a foundational force in industrial AI transformation through scalable tooling and infrastructure. (99 words)