Goldman Sachs has formally declared that video generation is a separate competitive axis in artificial intelligence, distinct from general language model quality. On July 12, 2026, the investment bank published a detailed ranking of Chinese AI video-generation capabilities as part of a broader research note on China's AI value chain, marking the first time a major bulge-bracket bank has published an investable ranking of video-generation quality as its own category.

The ranking placed ByteDance's Seedance model at the top, ahead of competitors including Kuaishou's Kling AI, DeepSeek, Alibaba, Tencent, and Minimax. This wasn't a casual mention buried in footnotes. It was the headline finding of a roughly 50-page competitive deep-dive on China's AI large-model industry, led by Goldman analyst Ronald Keung.

Why Is Goldman Sachs Suddenly Focused on Video Generation?

The timing reveals something important about how institutional capital now views AI differentiation. Three separate capital events occurred within three weeks of Goldman's note. On July 2, Kuaishou's Kling AI confirmed a funding round exceeding $2 billion at an approximately $18 billion valuation, with Alibaba among the participants. In late June, within an 11-day window, both A24 and Lionsgate took equity stakes in AI video providers. Then came Goldman's formal ranking.

These weren't isolated moves. They reflected a shared conviction among sophisticated investors: real differentiation in AI now lives in video generation, not in general chat quality or foundational language model performance. Goldman's research framework evaluated video generation on time-to-market, Arena score (a benchmark measuring model quality), valuation, and pricing per token. This is the language of competitive advantage, not commodity technology.

The numbers behind Seedance's top ranking tell part of the story. According to Chinese financial outlets cited in Goldman's research, Seedance achieved an annualized revenue run rate exceeding $2 billion within roughly four and a half months of its Seedance 2 launch, at a gross margin near 70 percent. By comparison, Kuaishou's Kling took a full year to climb from $100 million to $500 million in annualized run rate, a slower ramp on a lower base.

What Does This Mean for the Broader AI Market?

Goldman's broader thesis on Chinese AI stocks provides context. The bank's Thematic Investing team published a note on July 9 titled "Investment Strategy: Long China's AI Value Chain," arguing that China's entire AI sector accounts for only about 10 percent of global AI market capitalization, despite China generating an estimated 16 percent of global AI-related revenue. Global mutual fund allocation to Chinese technology stood at just 1.2 percent as of January 2026.

The market reacted immediately. The Hang Seng China Enterprises Index jumped as much as 4.5 percent intraday on the day the note went out, its largest single-day gain since February 2025. That reaction suggested the thesis wasn't a surprise so much as a confirmation of something capital was already leaning toward.

On foundational text models, Goldman scored Zhipu and DeepSeek strongest, largely on pricing power and cost advantage. Chinese high-end models run at roughly $1 per million tokens against $4 to $8 for comparable US models, and Chinese architectures use 2 percent to 10 percent of the parameter counts of their US counterparts through mixture-of-experts routing, a technique that activates only the most relevant parts of a model for each task.

How Are Video Generation Models Being Evaluated Differently?

Goldman's scoring framework introduced a three-part evaluation system that separated video generation from other AI capabilities:

• Pricing Power: Measured by release speed, LMArena scores (a benchmark for model quality), and price per million tokens processed, reflecting how quickly models improve and how affordably they operate.
• Cost Advantage: Evaluated through throughput, cache-hit rate, parameter-activation ratio, and inference gross margin, capturing the efficiency of running these models at scale.
• Financial Strength: Assessed via cash reserves, net cash-to-assets ratio, and valuation multiples, indicating which companies have the resources to sustain development.

This framework matters because it treats video generation as a distinct business problem, not just a feature bolted onto a language model. Seedance and Kling are both text-to-video and image-to-video models at their core, and Goldman flagged both Kling and Minimax's Hailuo model positively for expected breakthroughs in the second half of 2026 on video-generation and LLM-integration capabilities.

Goldman initiated coverage on Zhipu, which trades as Knowledge Atlas Technology on the Hong Kong exchange, at a Neutral rating with a price target about 15 percent above where the stock closed when the coverage went live. The analyst note emphasized Zhipu's latest GLM-5.2 model, which reached near-frontier performance with significant adoption among domestic enterprises and global small-to-medium businesses.

What Practical Implications Does This Have for Developers?

For developers and companies evaluating video generation tools, understanding content policy is essential before committing engineering resources. Kling AI operates with strict content moderation that applies uniformly across all input types and subscription levels.

Kling AI's moderation system runs at three distinct layers on every generation request:

• Prompt Screening: Before any generation begins, text prompts are scanned against a prohibited content classifier, rejecting requests containing keywords, phrases, or semantic patterns associated with blocked categories before any computing resources are spent.
• Real-Time Generation Constraints: During the diffusion process itself, Kling applies policy-aware constraints that steer the model away from producing certain visual outputs even when prompts don't explicitly request them, which is why some neutral prompts may produce unexpected results.
• Output Review: The completed video passes through a final content classifier before being returned to the user, blocking any content that passed earlier layers but still produced flagged output.

Kling AI does not allow NSFW (not safe for work) content in any form in 2026. There is no adult mode, no toggle, and no API parameter to unlock explicit generation. This is a deliberate product decision embedded in the model architecture, not a limitation waiting to be patched.

The content restrictions that cause the most friction for legitimate creators fall into three specific areas. Realistic human skin exposure, including swimwear or lingerie, is frequently flagged. Violence in realistic contexts, such as action scenes involving realistic weapons fire or blood, is commonly rejected, though stylized or clearly animated violence is treated differently. Political and public figure content is the least predictable category because the model applies contextual judgment rather than simple keyword matching.

For developers using Kling through platforms like Atlas Cloud, the content policy applies identically whether accessing the model directly or via an intermediary platform. The restrictions apply uniformly across text-to-video, image-to-video, and reference-to-video modes, meaning reference images containing nudity, explicit material, graphic violence, or sensitive political imagery are rejected at the input stage before generation begins.

The convergence of Goldman Sachs' institutional ranking, major funding rounds, and Hollywood studio investments signals a maturation moment for video generation AI. The market is no longer treating video generation as a secondary feature of general-purpose AI models. Instead, it's pricing these tools as distinct competitive assets with their own business dynamics, revenue trajectories, and technical requirements. For companies building on top of these models, understanding both their capabilities and their content boundaries is now essential to production planning.