Building Asia's First Authenticated Dataset Marketplace for Large Video Models

In every platform shift, there is a moment when the bottleneck moves. In Web2, it was scale. In Web3, it was infrastructure. In AI, that moment is now, and the bottleneck is data.

Large Language Models (LLMs) have shown us what's possible when you have vast, high-quality training sets. But as we shift into the era of Large Video Models (LVMs), systems like OpenAI's Sora, Google's Veo, Meta's VideoGen, the same data playbook doesn't apply. Video is fundamentally different: it's temporal, multimodal, and much harder to structure at scale. Scraped YouTube clips won't cut it.

The next wave of AI systems won't just need more video, they'll need licensed, annotated, domain-specific video. And that's where Asia has an opportunity.

The Model Bottleneck Has Moved

For the past five years, AI has been dominated by model innovation. Researchers squeezed more out of transformer architectures, scaled parameters into the billions, and stacked GPUs by the warehouse. But we're now seeing diminishing returns from scale alone.

Today's best models: Claude 3, GPT-4 Turbo, Gemini, aren't dramatically smarter than their predecessors. The big jumps are coming from new modalities (e.g. video), better data, and vertical fine-tuning.

OpenAI's Sora wowed people not just because of its visual quality, but because it demonstrated what happens when an LLM-like approach is applied to video. But Sora's real limitation isn't its model architecture, it's what it can learn. The training data behind the model is a black box. Most likely: scraped, compressed, and unstructured.

The risk for the next generation of LVMs isn't in the architecture. It's in the quality and provenance of the data. This connects directly to our analysis of the data bottleneck in large video models.

Why Asia is Uniquely Positioned

Asia is home to the largest, and most under-monetized video content ecosystems in the world. Consider:

• TikTok and YouTube Shorts creators in Indonesia, the Philippines, and India are producing billions of minutes of video content per day.

• E-commerce platforms like Shopee and Meesho host millions of product demonstration videos, rich in human-object interaction.

• Regional influencers, brand marketers, and beauty creators already produce the kind of structured content LVMs need: product-centric, gesture-rich, camera-aware.

But none of this data is in a usable format for training models. It's unstructured, copyright-uncertain, and fragmented across platforms. If Asia doesn't solve for this, others will, and they'll license that content back to us at a premium. Our perspective on Asia's role in national video datasets explores this dynamic further.

The Case for an Authenticated Dataset Layer

There are two core principles at play here.

First, AI models need better inputs. Not just more pixels, but more context. Video that comes with time-stamped speech, gesture annotations, object masks, and metadata about brand, lighting, interaction type. This kind of data is gold for LVMs building applications in fashion, retail, beauty, robotics, and autonomous movement.

Second, creators and data owners need to be part of the loop. As lawsuits like The New York Times v. OpenAI and Getty v. Stability show, models trained on unlicensed data are legally and ethically vulnerable. The future of AI needs to be permissioned by design. Which means dataset marketplaces will need to be authenticated, traceable, and compensatory. This relates to our analysis of the creator's paradox, where creators' work inadvertently trains machines that may replace them.

Clairva is Building This for Asia

A system where:

• Creators and content owners upload videos they own.

• Those videos are annotated, segmented, and enriched by an AI pipeline.

• Each usage is tracked and compensated via a blockchain-based licensing engine.

• AI companies can subscribe, search, and retrieve datasets for vertical-specific model training.

  
  

In effect, it becomes a data layer for LVMs, starting with structured video for fashion, beauty, and retail, and later expanding into robotics, education, and healthcare. For content creators, this creates new opportunities to monetize content through AI.

Why Now?

The timing is unusually clear.

Factor	Description
Model demand is surging	Every major foundation model company is racing to build LVMs. Most of them lack clean, structured video datasets.
Regulatory pressure is rising	From the EU AI Act to India's Digital Personal Data Protection Bill, ethical sourcing of training data is becoming law, not suggestion.
Creators are ready	They've spent the last decade building content. Now they're looking for monetization paths that don't rely on algorithmic reach.

The first dataset marketplaces that solve for licensing, structure, and scale will define the standard and capture the demand. This aligns with our research on how synthetic data is changing AI, highlighting the need for balanced approaches.

The Bigger Picture

When ImageNet launched in 2009, it changed computer vision forever. It provided a single, structured, labelled dataset that accelerated a decade of innovation. The next ImageNet won't be a spreadsheet of still images. It will be a living, evolving video corpus that powers the next generation of human-centric AI.

Asia has the content. Clairva is building the infrastructure. This work also connects to our commitment to ensuring diverse representation in AI, particularly in domains like fashion where representation matters deeply.