【競争するな！】GoogleやOpenAIには真っ向勝負しない。日本発スタートアップが数人でユニコーンを目指せるAIネイティブな開発AIdeaLab冨平×田所 後編

AIdeaLab founder Tomihira argues anime-specialized video generation is Japan’s defensible wedge against OpenAI Sora and Google, backed by a METI national AI project.

• AIdeaLab was selected for GENIAC, a Japanese government (METI) national AI project providing GPU subsidies to build a domestic video-generation foundation model.
• Their model, IdeaLab Video JP, applies a MoE (Mixture of Experts) architecture to video generation — they claim this is a first in the video-gen domain.
• AnimeGen, their text-to-anime service, targets anime studios and individual creators; they claim anime output quality exceeds Chinese competitor Kling.
• Sora 2 infringed on Ghibli, Pokémon, and Dragon Ball IP; Tomihira frames Japan’s deep anime IP catalogue as a structural moat that Western foundation model labs cannot easily exploit.
• AI token costs fell roughly 2,000–3,000x over three years; vibe coding slashed dev costs further — building to MVP now costs ~¥300K vs. millions previously.
• Key moats have shifted: feature depth and fast iteration no longer differentiate; proprietary datasets, network effects, and brand trust are now the defensible layers.
• Peter Thiel’s ‘don’t compete’ maxim is central: target regulated or reputationally risky spaces where large incumbents cannot move, not head-on races against Google or OpenAI.
• AI-native startups will be defined by AI-led development and review with minimal headcount — Tomihira expects unicorns to emerge from very small teams.

2026-01-28 · Watch on YouTube