2025

The mood shifted at the start of the year. In January, China's DeepSeek released R1. Read as high performance at low cost, it sent US semiconductor stocks down sharply by the end of the month. That same month, OpenAI announced the Stargate Project with SoftBank and Oracle, a $500 billion AI infrastructure venture.

Compute contracts changed scale. OpenAI raised $40 billion in a SoftBank-led round in March, agreed 10 gigawatts with NVIDIA in September and 6 gigawatts with AMD in October. Anthropic expanded to as many as a million Google Cloud TPUs and said in November it would put $50 billion into US data centres.

The labs took turns in front. Gemini 2.5 Pro in March, Claude Opus 4 in May, GPT-5 in August, Gemini 3 and Claude Opus 4.5 in November. Claude Code, out in February, became generally available in May and moved into daily work through the writing of code. In August OpenAI released gpt-oss with open weights.

In March, image generation spilled into the street. When OpenAI put GPT-4o's native image generation into ChatGPT, redrawing your own photographs in the style of Studio Ghibli became a worldwide pastime, and Sam Altman changed his own profile picture on X to one of them. The argument about how far a style may be imitated ran on from there.

It was also the year AI's output itself got named. "AI slop" — mass-produced generative content — swept the dictionaries' word-of-the-year picks in English and entered Japanese as-is. In the summer, "clanker," borrowed from Star Wars, spread on social media as a slur for AI and robots, and Harvard Business Review coined "workslop" for shoddy AI-made work product. The long-standing "dead internet theory" — most of the web is already machine-made — was cited anew as the slop piled up. On the approving side, Karpathy's February coinage "vibe coding" became a buzzword of its own, and in November "Chappie," the Japanese nickname for ChatGPT, made the shortlist for Japan's buzzword awards. The insults and the pet names grew up in the same year.

The map of alliances was redrawn. Microsoft released MAI-1, its first in-house foundation model, in August. In October OpenAI completed its move to a for-profit structure and signed a new agreement with Microsoft; in November Microsoft and NVIDIA invested in Anthropic and Claude arrived on Azure.

On copyright, a federal court held in June that training on books was fair use while keeping a pirated library was a separate matter. In September, Anthropic settled with the authors for $1.5 billion.