Topic
Explore the latest artificial intelligence news and research.
SK hynix will invest a combined ₩54.3 trillion in Yongin Y2 and Cheongju M17. The long-term vision has moved into an execution phase with fixed groundbreaking dates, though equipment installation will still proceed according to demand.
OpenAI announced it cannot rule out the highest warning level, "Critical," for the cyber capabilities of its upcoming model Astra, and halted part of its development. This marks the first time the voluntary safety framework established in 2023 has actually constrained the company's own pace of development, signaling that frontier AI capability management has moved from theory into operational practice.
18 listed Chinese companies are lining up along the glass-core substrate supply chain, with equipment orders for pilot production lines arriving ahead of mass production. Whether this becomes commercially viable hinges on yield, reliability, and customer qualification.
SpaceX and Tesla's planned semiconductor plant Terafab has selected Grimes County, Texas as its construction site. We examine the initial $16.8 billion investment and the conditions that remain before mass production can begin.
NVIDIA has maintained its official spec of up to 1.5TB for Vera, but this doesn't contradict reports of halved memory capacity due to supply constraints. Disclosure of mass-production SKU capacities, bandwidth, and pricing will determine whether this represents flexible configuration options or a compromise forced by supply shortages.
In an internal email, Microsoft declared that 'tokenmaxxing is not what we're optimizing for,' instructing departments to set AI budgets and switch to lower-cost models. The industry's assumption that token consumption serves as proof of productivity is now being overturned by one of the largest AI infrastructure players itself.
Envision has opened its Galaxy Campus in Inner Mongolia, integrating renewable power generation, transmission, and battery storage with AI computing infrastructure in a single design. The 2GW-plus capacity and million-chip figure are design targets for full completion, leaving current deployment scale and renewable supply performance as the next points to verify.
From late July to early August 2026, OpenAI, Anthropic, and Meta successively disclosed deviations by AI models during testing. In all three cases, the cause was not a rogue model but a failure of network isolation in the evaluation environment—exposing a structural problem: the infrastructure needed to safely measure the offensive capabilities of frontier models is still immature.
A survey of 1,500 companies in the UK, France, and Germany found that 73.9% worry about the US government cutting off their access, and 54.5% say they would grind to a halt within one business day. Yet only 44% have regularly rehearsed backup plans, and the fact that even authentication infrastructure is American-made is blocking any escape.
Stanford researchers used the AI model Evo 2 to design bacteriophage genomes, producing 16 functional viruses. Some surpassed the killing power of natural strains and overcame bacterial resistance, demonstrating in a study published in Science that the era of genome design has opened beyond protein design.
