The Hyperscaler Moat: Why Distribution Beats Model Performance in AI Video
Evoken, the Chinese AI creative-platform founded by a former ByteDance executive, has closed nearly $300 million in Series B+ funding at a valuation exceeding $2 billion — but the headline number is less instructive than what it buys. The company reports annualized recurring revenue of approximately $300 million as of May 2026, a zero-to-300M sprint in three years Evoken raises nearly $300M. That ARR figure, paired with the company's ByteDance pedigree, reflects a market view that distribution, not model quality, now determines competitive advantage in generative media.
The same week, Kuaishou's Kling AI raised nearly $3 billion at an $18 billion valuation, with Tencent, Alibaba, and Baidu all joining the round despite each operating their own foundation-model labs. The financing will support Kling AI's transition to independent commercial operations, spinning out from parent Kuaishou Kling AI raises nearly $3B. When three hyperscaler competitors collectively back a single spinoff, the bet is not on the model — it is on the distribution surface that Kuaishou's 600 million+ monthly active users provide.
The Fastest ARR Ramp in Generative Media
Evoken's trajectory demands re-examination of how AI application-layer value captures. The company operates three products: LiblibAI (an image-generation community with 4 million monthly active users and over 500,000 original models), Xingliu (an AI design agent), and LibTV (an AI video-creation platform for short drama and advertising) LiblibAI community metrics. This is not a single-product story. Evoken has built a layered ecosystem where the image community feeds the design agent, which in turn feeds the video platform — each tier raising switching costs for creators invested in the workflow.
The ByteDance playbook is explicit here. ByteDance's CapCut and Dreamina operate as a three-stage pipeline: generate (Dreamina), edit (CapCut), distribute (TikTok/Douyin), with each stage informing the next ByteDance creative suite strategy. Evoken's founder Chen Mian previously led monetization for Jianying and CapCut, and the product architecture is structurally identical. The difference is that Evoken is executing this playbook outside ByteDance's corporate umbrella, raising the question of whether the strategy can survive without the parent's distribution feeds.
Midjourney, by contrast, has never disclosed venture capital and reported approximately $200 million ARR at the end of 2023 — a slower ramp despite being the market-defining product. The comparison exposes a structural divergence: Midjourney optimized for per-user revenue via subscription gating, while Evoken optimized for user acquisition via freemium community. LiblibAI's 4 million MAU and freemium credit model generate 5–10 million daily image generations, creating a creator ecosystem that produces its own training data and model improvements LiblibAI community metrics. That flywheel is harder to replicate than a better diffusion architecture.
Kling AI's $18 Billion Bet on Spinoff Distribution
Kling AI's $3 billion raise at an $18 billion valuation is the largest single funding event in generative video to date — and the investor composition is its most telling feature. Tencent, Alibaba Cloud, and Baidu all participated alongside CPE, Guofang Venture Capital, and dozens of other investors Kling AI raises nearly $3B. Each of these companies operates its own foundation model — Tencent's Hunyuan, Alibaba's Qwen, Baidu's Ernie — yet they are collectively funding a competitor's spinoff. This is not a bet on Kling's underlying technology. It is a bet that Kuaishou's short-video distribution gives Kling a go-to-market advantage in the AI video vertical that competing models lack.
Kling AI reported Q1 2026 revenue of 650 million yuan (approximately $90 million), up over 300% year-on-year, with over 60 million creators and 600 million videos generated Kling AI revenue breakdown. The customer base includes over 30,000 enterprise users across marketing, e-commerce, film, short dramas, and gaming. But the revenue mix between internal Kuaishou distribution and external API customers has not been publicly broken out, making it difficult to assess how much of the growth is organic demand versus internal ad spend. One source notes that internal ad spend on Kling-generated content exceeded 20 million RMB daily in Q1, a figure that blurs the line between product revenue and parent-company subsidy.
The structural question is whether Kling can replicate Kuaishou's distribution advantages after its spinoff. The independent commercial operations announced alongside the funding round will test whether the product can attract users who are not already inside Kuaishou's ecosystem. If the answer is yes, the $18 billion valuation may prove conservative. If the answer is no, the company becomes a single-platform dependency story with an inflated capital base.
Sand.ai's Counter-Consensus Bet
The same week, Sand.ai raised over $100 million for its video-generation model, committing to an open-source Mixture-of-Experts architecture that the company claims will achieve SOTA performance. The company plans to release a new MoE-architecture video generation model in Q3 2026 Sand.ai raises $100M+. Sand.ai's strategy is the mirror image of Evoken and Kling: build the best model, open-source it, and let distribution follow.
The model-first approach has a precedent in the generative media segment, but the precedent is cautionary. Stability AI's open-source flywheel ignited the consumer generative-image era in August 2022, but the company's governance crisis — cash burn exceeding $8 million per month against annualized revenue below $15 million, internal financials leaked to Forbes and Bloomberg, CEO Emad Mostaque's forced resignation in October 2023 — demonstrated that open-source leadership does not translate to commercial durability without distribution. Black Forest Labs, founded by ex-Stability researchers, executed a cleaner version of the template with immediate hyperscaler contracts (Meta's reported $140 million deal), but the lesson holds: model capability alone has never sustained a generative-media company.
Sand.ai's bet is that video generation requires a fundamentally different architecture — autoregressive next-frame prediction rather than diffusion — and that this technical divergence will create a window where a model-first company can capture distribution before the hyperscalers replicate the approach. The company's Magi-1 model has held the #1 spot on Google DeepMind's Physics IQ benchmark, and its music agent VidMuse reportedly reached $10 million in ARR within three months of launch Sand.ai raises $100M+. That distribution-from-model strategy is not yet disproven, but the capital requirements are structurally different from Evoken's community-driven flywheel.
The Distribution Moat Thesis Applied
ElevenLabs, the voice AI startup, provides a useful comparison. The company is in early-stage discussions for a secondary share sale that would value it at approximately $22 billion, a 20x increase from its $1.1 billion Series B in January 2023 ElevenLabs targets $22B valuation. ElevenLabs has no distribution moat comparable to Evoken or Kling — it does not own a creator community or a short-video platform. Its valuation premium rests entirely on the thesis that voice AI is a foundational interface layer, not a niche feature, and that its enterprise integrations (Disney Accelerator, Deloitte, Twilio) provide durable revenue. The comparison highlights a fork: distribution moats are easier to identify in China, where platform ecosystems are vertically integrated, but the valuation multiples are converging at the application layer globally.
The capital re-rating is already visible. Evoken's $300 million at $2 billion, Kling's $3 billion at $18 billion — these are not model-investor valuations. They are platform-investor valuations, applying social-media multiples to AI-native creative tools. The mechanism is straightforward: if a creator community generates 30 million cumulative users and 500,000 original models, the switching costs to a competing model that is 5% better at text-to-image generation are negligible. The switching costs to a competing community with no creators, no models, and no distribution are insurmountable.
The Counter-Signal: Single-Platform Dependency
The most specific risk in this week's reporting is Kling AI's unreported revenue mix. Kuaishou has not publicly disclosed the breakdown between internal distribution revenue and external API customers, and internal ad spend on Kling-generated content exceeded 20 million RMB daily in Q1, suggesting meaningful reliance on parent-company demand Kling AI revenue breakdown. If Kling's $90 million quarterly revenue is predominantly internal transfer pricing rather than external enterprise adoption, the spinoff thesis weakens significantly. An independent Kling AI would need to replicate distribution that the parent company currently provides for free — and no amount of $3 billion in funding guarantees that replication.
A second risk, surfaced by the Evoken story, is competition from the same hyperscalers whose playbook it is borrowing. ByteDance's Dreamina and Seedance operate inside CapCut's pipeline with direct access to TikTok/Douyin's distribution. If ByteDance decides to compete aggressively on community-building and workflow-nesting, Evoken's $2 billion valuation assumes that the talent inheritance (Chen Mian's product expertise) outweighs the parent's distribution inheritance. That assumption is testable within the next two product cycles.
The generative media segment is experiencing a structural bifurcation. Model-first companies (Sand.ai, Stability, Black Forest Labs) compete on architecture and open-source strategy, betting that video generation's technical frontier is wide enough to sustain multiple independent labs. Distribution-first companies (Evoken, Kling, potentially ElevenLabs) compete on workflow nesting, community moats, and platform integration, betting that creator switching costs are the only defensible metric in a segment where model quality converges within three quarters. The capital markets are placing larger bets on the second category. For now, that is the correct call.