OpenAI's Internal Share Sale Makes Around 75 People Billionaires

A $6.6 billion internal share sale conducted by OpenAI in October 2025 allowed roughly 75 current and former employees to cash out up to $30 million each (the cap), it has emerged. According to The Decoder, more than 600 current and former employees took part in the sale, and President Greg Brockman has stated that he holds shares worth approximately $30 billion.

OpenAI's soaring valuation has become a symbolic event illustrating the concentration of capital in the AI industry. At the same time, it has sparked debate over wealth disparities among employees at private companies and the sustainability of the AI investment bubble.

The "Not-Hardware" Moat Built by NVIDIA's CUDA

WIRED has analyzed that NVIDIA's competitive advantage lies not in the chips themselves but in the CUDA ecosystem. CUDA is a software foundation built up over many years, and the switching costs formed by its developer community, libraries, and toolchains are not something rival chipmakers can easily overcome.

While AMD, Google TPU, and others are following suit in the AI semiconductor market, the analysis highlights the reality that CUDA's ecosystem lock-in functions as a substantial barrier to entry. The argument that NVIDIA is less a hardware company and more of a software platform player offers an important perspective for considering the direction of AI infrastructure investment.

Japan's AI Adoption Rising at Three Times the Global Average Pace

According to a Microsoft report, AI adoption in Japan is rising at three times the pace of the global average. The story, reported by ITmedia AI+, is believed to be underpinned by the dramatic improvement in LLMs' Japanese-language performance.

Because Japanese differs greatly in structure from English, LLM support for Japanese had long been limited. However, analysts attribute the acceleration in adoption to performance gains in multilingual models such as the Qwen series and reinforced Japanese-specific training data, which have brought the technology to a practical level. Business use cases are also expanding rapidly.

Sony CEO Hiroki Totomi Reveals AI Adoption in Game Development

Sony Group CEO Hiroki Totomi shared progress on AI use in creative fields at the company's business strategy briefing on May 8. The company is also introducing AI into game development, positioning it as a tool that supports creators' creativity.

As AI adoption advances across the entertainment industry, Sony emphasized its policy that "AI should support creativity rather than replace creators." Meanwhile, as WIRED has reported, there is also the reality of Hollywood screenwriters being pushed into AI gig work, revealing a divide within the industry over how AI should be used.

NTT Aims for GAFAM-Level Presence with Full-Stack AI Offerings

The NTT Group has unveiled its AI-related service initiatives. Under a business structure delivering a full stack from infrastructure to applications, the question is whether the company can achieve a presence on par with GAFAM.

It represents the largest-scale AI investment by a domestic telecom carrier, with edge AI and private AI environments leveraging its own network infrastructure expected to form the core of the offering. How to differentiate itself from GAFAM, which has already established a dominant position in the cloud market, remains a challenge.

Unlocking the Collapse Phenomenon in Ultra-Deep Diffusion Transformers

A study titled "Mean Mode Screaming," attracting attention on Hugging Face's Daily Papers, analyzes a peculiar collapse phenomenon in ultra-deep Diffusion Transformers. The researchers identified the cause of the phenomenon — dubbed "Mean Mode Screaming," in which training that proceeds stably for thousands of steps suddenly diverges — as the geometric asymmetry of residual connections.

The research team proposed a solution called "Mean-Variance Split (MV-Split) residual" and demonstrated its effectiveness on 400-layer and 1000-layer DiTs. It resolves the problem of existing methods compressing even the signal components by separating the mean path from the centered path. It is a noteworthy achievement that opens a new path toward stable training of ultra-deep models.