A second, less visible race is unfolding beneath the headline-grabbing competition to generate video: making sense of the footage that already exists. While Kling AI, the video generation unit of Chinese platform Kuaishou, just secured roughly $2 billion at an $18 billion valuation, TwelveLabs, a startup focused on video understanding rather than generation, closed a $100 million Series B funding round on July 1, 2026. The contrast reveals a fundamental shift in how investors are thinking about artificial intelligence and video.

What's the Difference Between Video Generation and Video Understanding?

The two approaches tackle opposite problems. Video generation companies like Kling AI, Runway, and Higgsfield build models that create new footage from text prompts, images, or motion descriptions. Video understanding companies like TwelveLabs build models that watch video that already exists and convert it into something software can search, query, and reason over. Most large language models (LLMs), which are AI systems trained on vast amounts of text to understand and generate language, still process video by sampling it into a sequence of still images, which loses motion, timing, and continuity. TwelveLabs takes a different approach, building native video models instead of retrofitting language models to handle video.

The company's Marengo 3.0 embedding model, released in late 2025, converts sound, speech, and motion across time into a single machine-readable representation rather than a stack of still frames. On top of that sits Pegasus 1.5, which turns raw video into structured data: scene boundaries, entities, temporal segments, and semantic context that an LLM can reason over directly. TwelveLabs describes Pegasus as functioning like a domain-specific language for video, making raw footage parseable by any intelligent system built on top of it.

Why Is Amazon's Chip Deal More Important Than the Funding Amount?

The $100 million Series B, co-led by NEA and NAVER Ventures with participation from Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital, and Red Bull Ventures, is modest compared to generation-layer funding. The real signal lies in what Amazon committed alongside the check. AWS signed a multiyear deal to optimize and host TwelveLabs' video inference workloads specifically on its own Trainium chips, and TwelveLabs agreed that its new frontier models will launch on AWS first. Any company can buy cloud compute. Far fewer get a cloud provider to commit its own custom chips to their specific workload and agree to be the first launch surface for future models.

"Video is the data understanding has to answer to. Models commoditize. The intelligence layer that composes them does not," said Jae Lee, TwelveLabs co-founder and CEO.

Jae Lee, Co-founder and CEO at TwelveLabs

That distinction signals investor confidence in a category that is about to matter, not a small hedge. The chip deal ties TwelveLabs' compute roadmap to Amazon's own silicon rather than to generic rented graphics processing units (GPUs), which are specialized processors used for training and running AI models. This kind of commitment from a major cloud vendor typically arrives only when infrastructure capital is backing a second-order race that investors believe will outlast the first.

How Are Companies Using Video Understanding Today?

The use cases follow directly from TwelveLabs' architecture. The company points to four primary industries where video understanding creates immediate value:

• Sports Broadcasting: A sports broadcaster can pull every match-winning goal scored in the final minutes out of years of unlabeled game tape without manually reviewing footage.
• Security and Surveillance: A security team can search surveillance footage by description instead of scrubbing through timestamps by hand, dramatically reducing response time to incidents.
• Advertising and Marketing: Advertisers can index and search video assets by semantic content, making it easier to find footage that matches campaign requirements.
• Automotive and Enterprise Archives: Companies with massive accumulated video libraries can query and reason over footage that has accumulated faster than any human team can watch or index.

TwelveLabs also builds persistent memory into the system, so the more video an organization indexes, the more capable its own archive becomes to query, rather than starting from zero with every new clip. This compounds over time, creating a durability advantage that generation models do not have.

Why Is Generation Getting Cheaper While Understanding Stays Scarce?

The funding gap between generation and understanding reflects a fundamental economic reality. Generation is getting commoditized fast: cheaper, more capable models are shipping constantly, and the money chasing that layer is now a scale game between a handful of giants. OpenAI's decision to shut down Sora earlier in 2026 after the consumer app burned roughly $1 million a day against $2.1 million in lifetime revenue illustrates the economics. Generation is a high-compute, high-cost problem with limited revenue models. Understanding, by contrast, sits one layer over, in making all that newly abundant video legible to software in the first place.

TwelveLabs' funding trajectory reflects this shift. The company raised $5 million in a seed round in March 2022, then a $12 million seed extension later that year, bringing seed-stage funding to about $17 million. A $50 million Series A followed in mid-2024, co-led by NEA and NVIDIA's NVentures. The Series B brings total funding to roughly $167 million across three rounds since 2022. The company has grown from around 58 employees a year ago to roughly 178 as of June 2026, a headcount curve that tracks the funding curve.

What Does Tencent's Stake Reduction Signal About Kuaishou?

While TwelveLabs was closing its Series B, Kuaishou's parent company faced investor scrutiny. Tencent Holdings offloaded over a third of its stake in Kuaishou, the Chinese short-video platform operator, stripping the internet giant of its major shareholder status. Tencent sold 273 million shares in an off-market block trade, reducing its stake to about 9.4 percent from around 15.7 percent. Based on Kuaishou's intraday trading range, Tencent may have cashed out between $1.5 billion and $1.6 billion.

The timing is notable. The stake reduction comes only days after Kuaishou announced it will spin off Kling AI and raise as much as $3 billion for the artificial intelligence-based video-generation business at a roughly $18 billion valuation, bringing in external investors including Tencent, Alibaba Group Holding, and Baidu. In comments included in the filing, Tencent emphasized that it remains confident in Kuaishou's long-term prospects, adding that the pair will maintain mutually beneficial ties and ongoing strategic cooperation. Analysts suggest the decision to reduce the stake fits Tencent's broader strategy of concentrating resources on core business operations while gradually reducing exposure to non-core holdings, rather than signaling doubt about Kling AI itself.

The divergence in funding and investor positioning between generation and understanding companies reflects a maturing market. Generation captured headlines and billions in 2025 and early 2026, but understanding is where the infrastructure capital is now flowing. As video becomes cheaper to create and more abundant, the ability to make sense of it becomes the scarcer, more valuable capability. TwelveLabs' $100 million round and Amazon's chip commitment signal that investors are pricing that durability into their bets.