2026 AI Token Revenue Rankings: Why Pricing Power in Video and Coding Beats Raw Model Capability

Overview of the 2026 Token Revenue Ranking

In September 2026, 爱分析 (iResearch) released the China AI Vendor Token Revenue Ranking . The data shows a stark disparity: ByteDance leads with 30 billion RMB annualized token revenue , Alibaba ranks second at 16 billion RMB, and Zhipu third at 11 billion RMB. The gap between the top and bottom exceeds tenfold, with three clear tiers (10B+, 5B+, 3B+). The ranking reveals a critical inflection point: token revenue leaders are not necessarily the strongest in language model capability, but those who have established "pricing power" in specific scenarios .

ByteDance: Video Model Pricing Power

ByteDance's 30 billion RMB revenue is almost entirely built on its absolute leadership in video models (Seedance series). This pricing power manifests in two dimensions:

Price premium : Seedance video generation prices are at least 50% higher than comparable models . Seedance 2.0 720P generation costs ~1 RMB/second, while domestic peers range 0.5–0.8 RMB/second.

Discount discipline : Channel discounts are minimal — "rarely below 90%" — and even internal business lines like Jianying and Huoshan receive no discounts.

From a techno-economic perspective, video generation is a compute-intensive task with lower storage bandwidth requirements than language models, enabling cheaper domestic chips for inference and supporting high margins. More importantly, demand-side willingness to pay drives "value pricing" over "cost pricing": video generation directly replaces traditional per-second film production costs (hundreds of RMB/second), so even at a premium over rival models, the migration benefit far exceeds migration cost. This allows ByteDance to capture the highest token revenue despite lagging in language models overseas.

Alibaba: Full-Stack Closed-Loop Financial Model

Alibaba's 16 billion RMB revenue stems from a fundamentally different path: it is "the only vendor to monetize token revenue after connecting chips, cloud, models, and applications across four layers." The core advantage is financial model health — the report discloses token business gross margin exceeds 50% , with the T-Head GPU entering mass production as a key milestone. Alibaba even states "AI compute expected to break even in 2–3 years." A significant portion of token revenue is effectively vertically integrated profit from chips and cloud infrastructure.

Data shows Alibaba's token ARR nearly doubled in three months (May–August 2026) , indicating strong growth momentum. However, in the open-source model arena, Qwen's position has been overtaken by later entrants like Kimi K3 and Zhipu GLM . Alibaba's token revenue relies more on ecosystem enterprise customers and cloud service bundling than on absolute model technical moats.

Zhipu: Coding Niche Validates "Volume-Price Dual Rise"

Zhipu's 11 billion RMB makes it the most notable "efficiency sample" — API average price rose 101% while token call volume grew 40x since the beginning of the year. The technical logic lies in AI coding's rigid demand characteristics : unlike general chat, programming tasks exhibit high professionalism, high frequency, and strong payment willingness . OpenRouter data shows >70% of token consumption comes from mid-to-large enterprises and professional developers in production environments, with coding being a top scenario.

By betting on the coding vertical, Zhipu avoided a head-on war of attrition in general language models against ByteDance and Alibaba, instead establishing bargaining power in a high-value, high-barrier niche . Zhipu officially projects year-end ARR to rise another 50% to 16 billion RMB.

Call Volume ≠ Revenue: Structural Mismatch

Juxtaposing token revenue with call volume data reveals a striking mismatch. On OpenRouter's call volume leaderboard, MiniMax M2.5, Kimi K2.5, DeepSeek V3.2 consistently rank top , and Chinese models surpassed US models in monthly token call share for the first time in early 2026. Yet on the revenue ranking, MiniMax (4.8B), DeepSeek (6B), Kimi (5.8B) sit in the second tier , an order of magnitude behind ByteDance's 30B.

The root cause is uneven distribution of pricing power . Video models command tens of times higher per-token value than language models — video per-second pricing can be dozens of times language model per-token pricing. ByteDance uses video pricing power to offset relatively lower language model call volume; meanwhile DeepSeek, MiniMax, etc., despite call volume advantages, struggle to convert scale into comparable revenue due to fierce language model price wars.

DeepSeek exemplifies this tension: API gross margin reaches 82.9% , but 2025 AI infra spend was only 1.2 billion RMB — "not even one month of top-tier vendor investment" — causing supply to lag demand and widespread API service anomalies. The tension between high margin and low investment limits the ability to translate technical advantage into revenue scale.

Conclusion: Pricing Power Resides in Scenario Irreplaceability

In the token economy competitive landscape, call volume ≠ revenue, technical capability ≠ pricing power . ByteDance's video pricing power rests on customer value logic where "migration benefit far exceeds migration cost"; Alibaba's full-stack loop provides structural cost advantages; Zhipu's coding focus sidesteps general-market price erosion.

For second- and third-tier vendors, the real challenge is not improving model capability per se, but finding a scenario cut that supports "value pricing" over "cost pricing." Video models' high unit prices prove that when model capability becomes irreplaceable in a specific scenario, pricing power becomes the ultimate determinant of token revenue stratification .