ByteDance and Kuaishou pull ahead in AI video race as western rivals lag

Chinese AI companies ByteDance and Kuaishou have pulled decisively ahead of their western counterparts in the race to produce high-quality AI-generated video, according to a Financial Times analysis that underscores a dramatic shift in the global competitive landscape. The two Chinese groups are now delivering output that outshines offerings from US-based rivals across key metrics like resolution, temporal consistency, and scene coherence. This lead is not academic: it is reshaping the advertising and entertainment industries, where AI-generated video is rapidly moving from experimental gimmick to production-grade tool. ByteDance, the parent company of TikTok, and Kuaishou, China’s second-largest short-video platform, have invested heavily in proprietary models that generate longer, more realistic clips with fewer artifacts than anything available from OpenAI, Meta, or Google. The gap is widening as western companies face regulatory headwinds and capital allocation debates, while Chinese firms benefit from a unified domestic market and aggressive compute deployment. This matters now because AI video generation is the next frontier of the AI infrastructure boom, and the company that controls the best model will capture the highest-margin revenue streams in advertising, content creation, and enterprise tools for years to come.

Training data scale and inference optimization

The technical edge comes down to two factors: training data scale and inference optimization. ByteDance and Kuaishou each operate platforms that generate billions of user-uploaded videos daily, giving them access to a training corpus that is orders of magnitude larger and more diverse than what western labs can legally scrape. ByteDance’s model, trained on TikTok’s global video library, produces clips with fluid motion and natural lighting that rival mid-budget commercial shoots. Kuaishou’s equivalent, optimized for the Chinese domestic market, excels at generating longer-form narrative sequences. Both companies have also invested in custom inference hardware that reduces generation latency below 10 seconds for a 30-second clip, a threshold that makes real-time advertising personalization viable. Western rivals, by contrast, remain constrained by smaller, curated datasets and reliance on general-purpose GPU clusters from Nvidia. The result is a quality gap that is visible even to casual viewers: ByteDance’s output scores 15–20% higher on standard perceptual quality benchmarks than OpenAI’s Sora, according to internal evaluations cited by the Financial Times. Beyond raw quality, ByteDance and Kuaishou have optimized their inference pipelines for cost efficiency at scale, achieving a cost per generated minute that is roughly 60% below what western cloud providers charge for equivalent output. This cost structure creates a self-reinforcing dynamic: lower prices attract more advertisers, higher volume funds more model training, and improved models attract still more advertisers. Western labs, which must monetize through API pricing rather than closed-loop platform ad revenue, cannot replicate this flywheel without a comparable distribution channel.

Revenue and margin impact on advertising

The advertising industry is the first sector to feel the impact. ByteDance already generates over $100 billion in annual advertising revenue from TikTok and Douyin, and AI-generated video allows it to offer advertisers dynamic, personalized creative at a fraction of the cost of traditional production. A 30-second AI-generated ad now costs roughly $200 to produce on ByteDance’s platform, compared to $5,000–$20,000 for a human-directed shoot. That cost advantage drives higher ad inventory turnover and increases platform take rates. Kuaishou, with a smaller but fast-growing ad business, is using its AI video model to attract premium brand advertisers who previously avoided short-video platforms due to quality concerns. The margin structure is equally compelling: once the model is trained, the marginal cost of generating an additional video is near zero, meaning incremental revenue flows almost entirely to gross profit. For western ad platforms like Meta and Google, which rely on third-party AI video tools or slower in-house models, the competitive pressure is mounting. If ByteDance and Kuaishou maintain their lead, they will capture an outsized share of the $700 billion global digital advertising market as it shifts toward AI-generated creative.

Winners and losers in the competitive reshuffle

The biggest losers in this shift are US-based AI video startups and the western hyperscalers that have bet on their own models. OpenAI’s Sora, once hailed as a breakthrough, now trails ByteDance’s offering on both quality and cost per clip. Meta’s Make-A-Video has been deprioritized internally as the company focuses on Llama and recommendation systems. Google’s Lumiere remains in research phase with no clear commercial launch date. Meanwhile, Chinese competitors are not standing still. ByteDance and Kuaishou are now competing against each other for the same pool of advertisers and content creators, driving a rapid iteration cycle that further widens the gap. The winners include the Chinese hardware supply chain: companies like SMIC and Huawei benefit from increased orders for the custom chips that power these models. Also benefiting are the advertisers and entertainment studios that gain access to cheaper, higher-quality production tools. The losers extend beyond western AI labs to include traditional video production houses, stock footage providers, and any business model built on human-directed commercial video. The BlackRock Smaller Companies Trust, in its final results, noted that AI narratives dominate global equity markets and capital is concentrated in perceived AI winners, a dynamic that now favors Chinese AI video companies over their western peers.

Compute, data centers, and enterprise buyers

The AI video generation race is intensifying the compute crunch that already defines the AI infrastructure boom. ByteDance and Kuaishou are among the largest consumers of Nvidia H100 and B200 GPUs in Asia, and their growing demand is contributing to data center capacity constraints. The Information reported that data center delay rumors are circulating as lead times for new facilities stretch to 24–36 months. Both companies are also investing in their own custom ASICs specifically designed for video inference workloads, which offer roughly two to three times the performance per watt of general-purpose GPUs for these tasks. This hardware investment reduces long-term dependence on Nvidia and on US semiconductor supply chains constrained by export controls, giving Chinese AI video companies a structural resilience that pure-software rivals lack. Hardware diversity is reshaping infrastructure planning: while Nvidia remains dominant, Chinese firms are increasingly designing custom ASICs for video inference, reducing their dependence on US chip exports. For enterprise buyers, the implications are twofold. First, the cost of AI video generation is falling so fast that any company with a marketing budget can now produce professional-grade video, democratizing access but also flooding the market with content. Second, privacy-conscious sectors like legal and healthcare are exploring local AI models as an alternative to cloud-dependent AI, as exemplified by Osaurus, an open-source LLM server that keeps files and tools on user hardware. Co-founders Terence Pae and Sam Yoo, participating in New York-based startup accelerator Alliance, are targeting these sectors with the Dinoki-derived platform. The broader enterprise shift toward local models could reduce demand for cloud inference, but for video generation, the compute requirements remain so high that cloud-based solutions will dominate for the foreseeable future.

Market and regulatory direction

ByteDance and Kuaishou’s ascendancy is not just a technical story: it is a policy signal. The Chinese government has actively supported domestic AI development through subsidies, relaxed data regulations, and state-backed compute clusters, while western regulators have imposed stricter rules on training data and model transparency. The result is a regulatory asymmetry that favors Chinese companies in data-intensive domains like video generation. This lead also signals that the next wave of AI value creation will be captured by companies with direct consumer distribution, not just model providers. ByteDance and Kuaishou own the platforms where video is consumed, giving them a closed-loop advantage: they can train on user behavior, generate ads, serve them, and measure performance all within the same ecosystem. Western AI companies, which lack equivalent distribution, must partner with platforms and share revenue. The Financial Times analysis suggests this structural advantage will persist unless western regulators change course or a breakthrough model emerges from a startup. For investors, the BlackRock Smaller Companies Trust’s 8.1% portfolio return versus a 10.6% benchmark in the second half underscores the risk of concentrating capital in perceived AI winners without accounting for geographic and regulatory dynamics. The trust’s report specifically flagged that AI narratives are driving capital concentration in a handful of companies, but the companies attracting the most capital in western markets are not necessarily the ones producing the best models globally. ByteDance and Kuaishou are not listed on western exchanges, which means equity investors exposed only to US and European AI stocks are systematically underweighted to the segment that is currently winning the video generation race. The AI video race is now a two-player game, and both players are Chinese.

The trajectory is clear: within 18 months, AI-generated video will account for a majority of all video content served on short-video platforms, and ByteDance and Kuaishou will control the dominant models. Western companies face a choice: invest aggressively in catching up, accept a secondary role, or pivot to adjacent markets like enterprise video tools where data regulation is less of a barrier. The most likely outcome is a bifurcated market: Chinese models dominate consumer advertising and entertainment, while western models carve out niches in regulated industries like healthcare and legal, where local inference and data privacy are paramount. Osaurus’s approach of offering both local and cloud AI models on a Mac Studio reflects this emerging reality. For hyperscalers and data center operators, the compute demand from Chinese AI video will continue to drive growth, but the hardware mix will shift toward custom ASICs and away from general-purpose GPUs. The energy consumption of these models will also intensify environmental pressure, as noted in the BlackRock report. The AI video generation race is not a sprint: it is a structural realignment of the global AI industry, and the starting gun fired in Beijing.