The AI video market is no longer a single race to build the best video generator. A new $100 million funding round reveals that a parallel category is maturing fast: tools that understand and search existing video rather than create new footage. This split matters because it changes which vendors matter for different jobs, and it shows how cloud providers are reshaping AI startup economics.

What Is Video Understanding, and Why Does It Matter?

TwelveLabs, a San Francisco and Seoul-based startup, just closed a Series B funding round co-led by NEA and NAVER Ventures on July 1, 2026, bringing its total funding to roughly $150 million. But the company is not trying to compete with OpenAI's Sora or Google's Veo, which generate new video from text prompts. Instead, TwelveLabs builds models that index and reason over existing footage, treating video understanding as a distinct product category.

The company's two core models work differently than generative tools. Marengo converts raw footage, including speech, sound, and motion, into searchable representations. Pegasus, a video-language model, reasons over that data across up to two hours of continuous footage rather than sampling isolated frames. For enterprises managing surveillance archives, sports libraries, or broadcast tape, this capability solves a different problem than video generation does.

"The goal is durable intelligence layer value as underlying models commoditize," said CEO Jae Lee, who previously worked as a data scientist for South Korea's Ministry of National Defence.

Jae Lee, CEO and co-founder of TwelveLabs

That framing is important. Lee is betting that as the underlying video models themselves become cheaper and more available, the real value will shift to the intelligence layer that sits on top, helping organizations make sense of video they already have.

How Is Cloud Provider Investment Reshaping AI Startups?

The funding round itself is solid, but the structure reveals a bigger pattern in AI infrastructure deals. Amazon, already a repeat investor in TwelveLabs, used this Series B to formalize AWS as the company's preferred cloud provider. New models optimized for AWS Trainium chips will launch there first. This is not just a partnership; it is a strategic lock-in that mirrors similar arrangements Amazon has made with other AI video startups earlier in 2026.

For teams evaluating video AI vendors, this dynamic matters as much as the funding total itself. Cloud providers are increasingly trading investment dollars for locked-in compute commitments and roadmap influence rather than pursuing pure equity upside. That means the startups with the deepest cloud partnerships may have advantages in hardware access and pricing that competitors cannot match.

Steps to Navigate the Shifting Video AI Landscape

• Identify your actual need: Do you need to generate new video content, or do you need to search, index, and understand video you already have? The answer determines which category of tool matters for your workflow.
• Evaluate cloud dependencies: Check whether your chosen vendor has preferential access to accelerator hardware through cloud partnerships. Startups with deep cloud ties may have cost or latency advantages that affect long-term feasibility.
• Plan for API sunset windows: If you are considering tools like Sora 2, note that OpenAI's Sora API is scheduled to discontinue on September 24, 2026, so any new product dependency on that API is not a stable long-term bet.

What Does This Mean for Sora and Other Generative Tools?

While TwelveLabs is scaling video understanding, the generative video landscape is contracting. OpenAI's Sora consumer app and web experience ended on April 26, 2026, and the Sora API is scheduled for discontinuation on September 24, 2026. For developers still using Sora 2 through the API, the remaining window is a sunset period, not a stable platform for new product work.

The official Sora 2 API pricing during this window is $0.10 per second for standard 720p generation and $0.30 per second for the higher-fidelity Pro model at the same resolution, with batch pricing at half those rates. But pricing alone is not the decision driver anymore. The removal date is. Any team building a new long-term product on Sora 2 is planning past the documented removal date, which makes migration planning more important than squeezing cost savings from batch pricing.

The practical guidance for Sora 2 users is narrow: use the API only for bounded, current work that can finish before September 24, 2026. Set a hard final generation date on your side to prevent unfinished jobs from becoming stranded work. Cap the dollar budget by model and resolution, because sora-2-pro at 1080p can turn short videos into double-digit costs quickly. Export every output and store metadata, because after the sunset window, your local archive may matter more than the API job history.

Why the Split Between Generation and Understanding Matters

The divergence between generative and understanding-focused tools is reshaping how enterprises think about video AI. TwelveLabs has grown to roughly 178 employees as of June 2026, up from about 58 a year earlier, and industry estimates put the AI video search market at roughly $3.2 billion by 2028. That growth trajectory suggests the understanding category is not a niche; it is a maturing product market with real enterprise demand.

For teams evaluating video AI vendors, the widening split between generation-focused and retrieval or reasoning-focused tools is now as important as the funding total or the model capability. The question is no longer just "which video AI tool is best?" It is "which category of tool solves my actual problem, and which vendor has the cloud partnerships and stability to support it long-term?"

The TwelveLabs funding round signals that video AI is no longer a single market. It is two markets, with different technical approaches, different customer needs, and increasingly different infrastructure dependencies. Understanding that split is the first step to making the right vendor choice.