Ex-Googlers are building infrastructure to help companies understand their video data

Businesses are generating much video than ever. From years of broadcast archives to thousands of store cameras and countless hours of accumulation footage, astir of it conscionable sits unused connected servers, unwatched and unanalyzed. This is dark data: a massive, untapped assets that companies cod automatically but astir ne'er usage successful a meaningful way.

To tackle the problem, Aza Kai (CEO) and Hiraku Yanagita (COO), 2 erstwhile Googlers who spent astir a decennary moving unneurotic astatine Google Japan, decided to physique their ain solution. The duo co-founded InfiniMind, a Tokyo-based startup processing infrastructure that converts petabytes of unviewed video and audio into structured, queryable concern data.

“My co-founder, who spent a decennary starring marque and information solutions astatine Google Japan, and I saw this inflection constituent coming portion we were inactive astatine Google,” Kai said. By 2024, the exertion had matured, and the marketplace request had go wide capable that the co-founders felt compelled to physique the institution themselves, helium added.

Kai, who antecedently worked astatine Google Japan crossed cloud, instrumentality learning, advertisement systems, and video proposal models and aboriginal led information subject teams, explained that existent solutions unit a tradeoff. Earlier approaches could statement objects successful idiosyncratic frames, but they couldn’t way narratives, recognize causality, oregon reply analyzable questions astir video content. For clients with decades of broadcast archives and petabytes of footage, adjacent basal questions astir their contented often went unanswered.

What truly changed was the advancement successful vision-language models betwixt 2021 and 2023. That’s erstwhile video AI started moving beyond elemental entity tagging, Kai noted. Falling GPU costs and yearly show gains of astir 15–20% implicit the past decennary helped, but the bigger communicative was capableness until recently, models conscionable couldn’t bash the job, helium told TechCrunch.

InfiniMind precocious secured $5.8 cardinal successful effect funding, led by UTEC and joined by CX2, Headline Asia, Chiba Dojo, and an AI researcher astatine a16z Scout. The institution is relocating its office to the U.S., portion it continues to run an bureau successful Japan. Japan provided the cleanable testbed: beardown hardware, talented engineers, and a supportive startup ecosystem, allowing the squad to fine-tune its exertion with demanding customers earlier going global.

Its archetypal product, TV Pulse, launched successful Japan successful April 2025. The AI-powered level analyzes tv contented successful existent time, helping media and retail companies “track merchandise exposure, marque presence, lawsuit sentiment, and PR impact,” per the startup. After aviator programs with large broadcasters and agencies, it already has paying customers, including wholesalers and media companies.

Now, InfiniMind is acceptable for the planetary market. Its flagship product, DeepFrame, a long-form video quality level susceptible of processing 200 hours of footage to pinpoint circumstantial scenes, speakers, oregon events, is scheduled for a beta merchandise successful March, followed by a afloat motorboat successful April 2026, Kai said.

image credits: infinimind

The video investigation abstraction is highly fragmented. Companies specified arsenic TwelveLabs supply general-purpose video knowing APIs for a wide scope of users, including consumers, prosumers, and enterprises, Kai said, while InfiniMind focuses specifically connected endeavor usage cases, including monitoring, safety, security, and analyzing video contented for deeper insights.

“Our solution requires nary code; clients bring their data, and our strategy processes it, providing actionable insights,” Kai said. “We besides integrate audio, sound, and code understanding, not conscionable visuals. Our strategy tin grip unlimited video length, and outgo ratio is simply a large differentiator. Most existing solutions prioritize accuracy oregon circumstantial usage cases but don’t lick outgo challenges.”

The effect backing volition assistance the squad proceed processing the DeepFrame model, grow engineering infrastructure, prosecute much engineers, and scope further customers crossed Japan and the U.S.

“This is an breathtaking space, 1 of the paths toward AGI,” Kai said. “Understanding wide video quality is astir knowing reality. Industrial applications are important, but our eventual extremity is to propulsion the boundaries of exertion to amended recognize world and assistance humans marque amended decisions.”