Last week’s piece argued that the US-China AI question has become an architecture question, and that marketing organisations need a sovereignty-tiered routing model to operate in a bifurcated stack. The sharpest practical test of that argument is happening in AI video, and the picture became measurably clearer over four weeks this spring.
The sequence: in December 2025, Disney announced a $1 billion investment in OpenAI’s Sora, with licensing rights to over 200 characters across Disney, Marvel, Pixar, and Star Wars. In February 2026, ByteDance pre-released Seedance 2.0 and was hit within days with cease-and-desist letters from Disney, Netflix, Warner Bros, Paramount, Sony, and Universal. On 24 March 2026, OpenAI shut down Sora and Disney walked from the deal without a dollar changing hands. Four weeks. The same architecture decision now sits on every CMO’s desk in compressed form.
Focus On: The economics that broke
Sora was burning $15 million per day in inference costs against $2.1 million in lifetime revenue, per Forbes reporting on Cantor Fitzgerald estimates. Each ten-second clip cost roughly $1.30 to generate. Annualised burn: $5.4 billion on a product whose own head had described the economics as “completely unsustainable” the previous October. Downloads collapsed 66% in three months. Disney was reportedly informed less than an hour before the announcement.
The lesson that travels: the consumer-subscription model failed at the price-per-clip Western frontier labs are able to deliver. The cost gap with Chinese models is structural, not cyclical. ByteDance’s Seedance 2.0 Fast generates 1080p video at $0.022 per second. Alibaba’s Wan 2.6 sits at $0.07. Kuaishou’s Kling 3.0 at roughly $0.10. Sora 2 and Kling Video O3 at $0.15. A $10 budget on Seedance Fast produces ninety five-second 1080p clips; the same $10 on Sora 2 produces thirteen. Practitioners report 70-90% production cost reductions versus traditional shoots when routing is done properly.
Why Chinese labs lead this category
The structural advantage rests on three reinforcing pillars. Training-data asymmetry comes first: ByteDance owns Douyin and CapCut, Kuaishou operates China’s second-largest short-video platform, and the corpus available to Chinese video labs is the corpus they themselves run at scale. Vertical integration follows — model, editor, and distribution surface sit inside one company, with a feedback loop from creative output to engagement data to model improvement that’s shorter than anyone else’s. Hardware independence is now real. DeepSeek V4 trains and runs on Huawei Ascend 950 clusters and Cambricon silicon, no Nvidia. Domestic Chinese chips reached 41% of China’s AI chip market in 2025. That lowers the cost floor for Chinese inference while US frontier models remain tethered to Nvidia’s pricing power.
The IP counterweight
The same training-data asymmetry that produces the price gap creates the legal exposure. In February 2026, Netflix called Seedance 2.0 “a high-speed piracy engine”, accusing ByteDance of pervasive unauthorised reproduction of Stranger Things, Bridgerton, Squid Game, and K-Pop Demon Hunters. Disney, Warner Bros, Paramount, Sony, and Universal each followed with cease-and-desist letters. ByteDance halted Seedance’s international rollout and committed to safeguards under litigation pressure, not policy.
The asymmetry that matters for marketing organisations: guardrails on Western frontier models are commercial and contractual; guardrails on Chinese models are reactive. Google’s Nano Banana refuses content that could harm third-party rightsholders’ interests. OpenAI signed a Disney licensing deal in December 2025. Add the March 2026 US Supreme Court decision declining to hear Thaler v. Perlmutter — AI-generated content remains uncopyrightable in the United States — and the picture sharpens. Hero content built purely on AI has no defensible IP moat, regardless of jurisdiction.
Brand creative in three tiers
The “AI ads are coming” debate is settled. Kalshi aired a fully AI-generated spot during Game 3 of the 2025 NBA Finals, made for $2,000 in generation costs by an experienced AI filmmaker. Amaysim ran a fully AI-generated ad nationally across Australian broadcast and digital. Agencies running serious AI video pipelines route by use case rather than standardising on a single model: Veo for product fidelity, Seedance for camera tests and volume, Kling for dialogue and motion, Wan for cinematic narrative. A documented working split allocates Seedance Fast to roughly 50% of budget on concepts and drafts, Kling to 30% on polished deliverables, and a frontier model to 20% on premium hero content. Reported savings versus a single-premium-model approach: 30-50%.
The architecture maps directly to the sovereignty-tiered routing model from last week. The volume tier — performance creative, social variants, multi-market localisation, A/B test assets — runs on the cheapest production-quality model that meets the workload’s quality bar. The frontier tier — brand films, flagship campaigns, anything where licensing must be unimpeachable — runs on licensed-data Western models with paper trails. The IP red line cuts across both: any creative that uses, evokes, or could be confused with copyrighted IP, talent likeness, or branded environments must run through licensed models, regardless of cost. The Hollywood litigation against ByteDance and MiniMax is a forward indicator. Brand-side litigation against advertisers using infringing AI tools is a twelve-to-eighteen-month risk worth pricing in now.
The procurement question for this quarter is not which AI video tool to standardise on. It is whether the routing policy lives somewhere finance, legal, and brand can all sign, and whether the engineering infrastructure exists to enforce it across vendor boundaries. The marketing organisations that build that layer this year will run an order of magnitude more creative tests at materially lower cost. The pattern is already visible in agency pipelines and across the e-commerce category. By Q4, it will be visible in ad-creative budgets.
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programs, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.