Daily AI Brief

At the Huawei Connect conference, the company unveiled its plan for the Ascend chip series: two variants of the Ascend 950 in 2026, followed by the Ascend 960 in 2027 and Ascend 970 in 2028. Alongside, Huawei is introducing powerful compute platforms "Atlas 950" and "Atlas 960" capable of supporting thousands of Ascend chips each, in "supernode" architecture. These systems claim, in some metrics, to outperform Nvidia's GB200 NVL72. Huawei also said it has developed its own high-bandwidth memory (HBM), reducing dependence on U.S. and South Korean suppliers.

NVIDIA and Intel have announced a collaboration to jointly develop multiple generations of custom data-centre and client CPUs, integrating NVIDIA's GPU/accelerated compute via NVLink with Intel's CPU strengths. As part of that, Nvidia is investing USD 5 billion in Intel common stock. The joint work includes x86 system-on-chips combining RTX GPU chiplets for personal computing, and custom CPUs for data centre applications.

A new WTO World Trade Report projects that AI technologies could increase global trade in goods and services by approximately 34-37% by 2040, with global GDP rising by 12-13%, assuming enabling conditions. But the report cautions that poorer nations and regions may lag unless trade policy, infrastructure, and digital readiness improve. Otherwise, there is a risk of widened economic divides.

OpenAI researchers have published work demonstrating that even with perfect training data, large language models (LLMs) will necessarily produce plausible but false outputs (hallucinations), due to inherent statistical and computational constraints. The admission frames a realistic boundary for what can be engineered away vs what is intrinsic to current model designs.

Irregular, a security lab working on frontier AI risks (including misuse, adversarial models, threats at scale), has raised USD 80 million. The lab aims to research, develop safeguards, and influence standards for AI safety in cutting-edge contexts. As AI systems scale, attention to security risks is becoming more urgent.

A new OECD report highlights that while many governments are piloting AI in core public functions, scaling up remains challenging due to issues like transparency, governance, ethical oversight, and risk of misuse. The report emphasises that responsible frameworks are essential to ensure trust and sustainable deployment.

A planned joint venture between SoftBank and OpenAI in Japan — intended to bring corporate AI services locally — is delayed. Originally expected to be up and running this summer, it is now unclear when it will launch, with updated progress expected by November.

Chinese AI developer DeepSeek disclosed that training its R1 model cost around USD 294,000. While this is still substantial, it is much lower than many comparable models of its class, raising questions about how cost efficiencies are being achieved and what that implies for competitive dynamics in AI model development.

A new academic study shows that the widespread availability of AI tools is changing team dynamics within organisations. Collaboration, boundaries of work roles, decision-making, and error checking are shifting. The study argues that organisations will need to rethink team structure, leadership, and responsibility as AI becomes a central part of workflows.