Daily AI Digest

NVIDIA plans a $1 billion NAVER investment as Korea's DSX AI factory targets 200 MW

NAVER, NVIDIA and Brookfield propose expanding the NVIDIA DSX AI factory at NAVER's GAK Sejong data center from 55 MW to 200 MW by 2028, with a longer term path to 1 GW. NVIDIA plans to invest $1 billion in NAVER, Brookfield has signed a nonbinding term sheet for up to $9 billion, and NAVER will fund the balance. NVIDIA's investment is subject to customary closing conditions and to NAVER securing at least $9 billion in committed financing separate from NVIDIA's investment, so none of the headline capital should be treated as closed. The multi-tenant system is expected to use Blackwell and Vera Rubin infrastructure and serve AI innovators in Korea and the United States.

Verdict: The structure binds a chip vendor, a cloud operator and infrastructure capital to one capacity plan, showing that sovereign AI projects are becoming integrated financing, power and operations programs rather than simple hardware purchases.

Thibault Sottiaux of OpenAI's Codex and ChatGPT team says ChatGPT Work has officially overtaken Codex in active users, though the post gives no counts or definition of an active user. In a second post, he jokes that people have long been able to speak to computers but got little back until now, pointing to ChatGPT Voice on desktop, where users speak their thoughts and the model chooses how to present its answer. The product signal is clear even if the adoption claim is self-reported: reducing input friction can make delegated work easier to start.

Nan Yu, Head of Product at Linear, jokes that the real reason there is no true software factory is that nobody willingly chooses to vibe code in Java. The post is a one-line riff, not a definition of a software factory, but the joke points to a practical constraint: agentic development follows the languages and ecosystems builders actually want to use.

Zara Zhang, a builder, says AI-native companies have a culture akin to an open-source community. In another post, she asks what people do while waiting for AI output; an earlier post in the thread says waits of one to five minutes are too short for deep work but long enough to drive her back to X. Together, the posts suggest that transparent collaboration still needs deliberate handoffs and latency design, or agent work can become a source of constant distraction.

Terence Tao says mathematics is moving from proof scarcity to proof abundance, shifting the bottleneck to verification, exposition and acceptance

In his ICM 2026 public lecture, Terence Tao asks the mathematical community to assume that AI will soon handle a meaningful share of research-level tasks, then reconsider what progress should mean. In the second controlled First Proof batch, at least one team produced a publication-quality solution for seven of ten new problems, at compute costs ranging from $10 to $1,000 per problem. Tao argues that proof generation and formal verification are only the start: an AI-written proof can be correct yet poorly understood, while excessive polishing can hide where the real difficulty lies. The likely bottlenecks are expert review, clear exposition, community digestion, acceptance and eventual incorporation into canonical theory.

Verdict: Teams building research agents should treat explainability and expert review as first-class deliverables, rather than equating formal verification with a result's acceptance into the discipline.

Token relays advertise model access at discounts as deep as 97.8%, exposing a potential fraud supply chain

Vectoral researcher Matt Lenhard maps a Chinese-language market of relay services that proxy requests into systems from OpenAI, Anthropic, Google and other providers. One comparison site offered what it described as $3,333 of Anthropic credit for RMB 425, while the cheapest relay in Vectoral's sample advertised a 97.8% discount to official pricing. Based on public forum posts and the author's previous anti-abuse work, the reported supply methods can include free-trial abuse, chargebacks, stolen cards, leaked keys and pooled accounts. This is threat research, not a law-enforcement investigation, so it does not establish that every relay uses the same source, but the price gap alone is a serious supply-chain warning.