The archive · Developer & Business Tools · Product decision · 2024–2026
InfiniMind bets enterprises will mine unviewed video; $5.8M seed from UTEC
Ex-Google Japan duo build no-code infrastructure that turns PB-scale 'dark' video into queryable data; TV Pulse is live, DeepFrame ships 2026.
InfiniMind
What the business is
InfiniMind makes enterprise video-intelligence infrastructure: TV Pulse analyzes broadcast TV in real time for media and retail clients, and DeepFrame turns up to 200 hours of footage into a searchable knowledge base with visual, audio and speech understanding, deployable in a VPC or on-premises.
How it started
Aza Kai and Hiraku Yanagita spent nearly a decade together at Google Japan, Kai in cloud/ML/data science and Yanagita in brand and data solutions. At Google they watched clients sit on petabytes of broadcast and camera footage and fail to answer even basic questions about it — object tagging existed, but nobody could track narrative, causality or long-form content. Kai says 2021–2023 advances in vision-language models changed what was possible, and by 2024 the market demand was clear enough that the two left to found InfiniMind (formerly SDio) in Tokyo, joining programs like AWS GAIA 2025, METI's GENIAC and NVIDIA Inception along the way.
What happened
InfiniMind launched its first product, TV Pulse, in Japan in April 2025: it analyzes broadcast TV in real time so media and retail companies can track product exposure, brand presence, sentiment and PR impact. After pilots with major broadcasters and agencies it signed paying customers among wholesalers and media companies, and DeepTech reported the platform had analyzed more than 100,000 hours of content. In February 2026 the startup announced a $5.8M seed round led by UTEC, with CX2, Headline Asia, Chiba Dojo and an a16z Scout-linked AI researcher participating; it said it would relocate its headquarters to the US while keeping its Tokyo engineering office, hire more engineers, and ship its flagship DeepFrame — long-form video intelligence over up to 200 hours of footage, with audio and speech understanding — in beta in March 2026 and full release in April 2026.
No ending yet — it is still running.
Background
InfiniMind is a Tokyo-founded startup, built by two ex-Google Japan employees, that sells infrastructure for making unviewed video useful. Businesses generate video constantly — broadcast archives, store cameras, production footage — and most of it sits unanalyzed. InfiniMind's products, TV Pulse and DeepFrame, convert that so-called dark data into structured, queryable information that enterprises can search with natural language.
The bet is a timing call. Aza Kai (CEO) says video AI spent years doing little more than labeling objects in individual frames, which could not track narrative, causality or long-form content. When vision-language models improved sharply between 2021 and 2023, and GPU costs kept falling, he and COO Hiraku Yanagita judged the technology finally capable — and founded the company in 2024. Their wedge is enterprise-grade infrastructure, not a general-purpose consumer API: no-code workflows, unlimited video length, and VPC or on-premises deployment for companies with data-sovereignty requirements.
Proof points so far are Japanese. TV Pulse, launched in April 2025, analyzes broadcast television in real time for media and retail clients tracking product exposure, brand presence, sentiment and PR impact; by early 2026 DeepTech reported 100,000+ hours analyzed and paying customers among wholesalers and media companies after pilots with major broadcasters and agencies. In February 2026 UTEC led a $5.8M seed round joined by CX2, Headline Asia, Chiba Dojo and an a16z Scout-affiliated AI researcher.
InfiniMind plans to use the funding to move its headquarters to the US, keep its Tokyo engineering office, hire more engineers and launch DeepFrame, its 200-hour long-form video intelligence platform, in beta in March 2026 and fully in April 2026. The strategic claim is that video understanding is a path toward AGI — that indexing real-world events like a database is worth building even beyond the industrial use cases that pay the bills today.
What has to be true
- The supply of unanalyzed video already exists in every archive and camera feed, so the unlock is analysis cost and capability, not new data capture.
- The 2021–2023 vision-language model jump is what made founding in 2024 rational — the same pitch would not have worked two years earlier.
- Enterprise constraints — privacy, data sovereignty, no-code workflows, long-form queries — are a moat against general-purpose video APIs aimed at consumers and prosumers.
- Japan's broadcasters and agencies made a compact testbed where demanding customers could validate the technology before a US push.
- Cost control through adaptive sampling and hybrid indexing targets the per-minute economics that keep long-form video analysis out of reach today.
What can be applied
Founding should wait for the capability curve: video AI only tagged objects for years, so Kai waited until vision-language models could follow narrative — timing made 2024 credible to UTEC.
Aftermath
As of 2026-09-02 InfiniMind is live and seed-funded: TV Pulse has paying Japanese media and wholesale customers and 100,000+ hours analyzed, and the $5.8M round funds DeepFrame development, hires and the US HQ move. DeepFrame's beta was slated for March 2026 and full release for April 2026; no later milestone was verified for this entry. With 10+ Tokyo employees and University of Tokyo collaborators, it argues that long-form, causal, multimodal understanding — not clip-level tagging — is the defensible enterprise niche.
Sources
- Ex-Googlers are building infrastructure to help companies understand their video data
- 前谷歌团队创办公司，把闲置视频变成结构化数据
- 东京初创InfiniMind以AI破解企业视频暗数据难题，获580万美元种子轮融资推动全球扩张
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card