
One Short Video on Douyin Pushes Recommendation Engine Computational Costs to New Heights|抖音一条短视频,将推荐引擎算力账单推至全新高度
English Version First|Complete Chinese Translation Below
English Version
On July 27, 2026, Douyin Group announced a full upgrade to the core engine of its minors’ mode — the age-adaptive recommendation algorithm — with official deployment of Multimodal Large Language Model (MLLM) technology to empower intelligent content review and personalized distribution.
The upgraded technical framework follows a rigorous and standardized logic. The multimodal large model conducts cross-modal semantic understanding on massive video resources, comprehensively analyzing text, image and audio information to evaluate content cognitive difficulty and value orientation. After preliminary AI intelligent screening, professional human experts carry out secondary review to screen high-quality compliant content into an exclusive content pool. The age-adaptive recognition model further performs in-depth comprehension and classification to match content with accurate age groups.
Supported by the multi-layer protection mechanism of manual screening + LLM classification + age-adaptive recommendation + personalized preference matching, the platform adds approximately 6,000 high-quality contents daily, with the total volume of the exclusive minor-oriented content pool reaching 7.7 million pieces.
On the surface, this is a remarkable technological upgrade for safer and finer youth content governance. However, the entire AI industry shares a clear consensus: every content recognition and semantic inference performed by multimodal large models comes at the cost of massive token consumption.
Technically, processing a single 10-second 1080p video consumes around 350,000 tokens. For Douyin’s 7.7 million reserved videos, even a one-time full-scale inference requires a staggering 2.7 trillion tokens. Most importantly, this enormous computational overhead only belongs to a single functional module of one mobile application, reflecting the soaring cost pressure of modern AI recommendation engines.
Runaway Token Inflation in Recommendation Engines
Industry-wide data further reveals an unprecedented token inflation crisis sweeping China’s AI sector.
China’s average daily token consumption has skyrocketed from 100 billion at the beginning of 2024 to 30 trillion by mid-2025, and climbed to 180 trillion in February 2026. According to official data from the National Data Administration, the average daily token invocation volume exceeded 140 trillion by the end of March 2026, surging more than 1,000 times in just over two years.
Liu Liehong, Director of the National Data Administration, stated explicitly at a State Council Information Office press conference that the daily token volume had increased by over 40% in merely three months, compared with 100 trillion at the end of 2025.
The current token economy is experiencing structural inflation: it is not that individual tokens have become more expensive, but that token consumption per unit time and per user is growing exponentially. This explosive growth is primarily driven by the maturity of multimodal technology and the large-scale commercial implementation of Agent applications. Morgan Stanley forecasts that China’s AI inference token consumption will jump from approximately 10 trillion in 2025 to 390 trillion in 2030, a 37-fold increase within five years.
Recommendation engines have become the hardest-hit area of token inflation. State-based hybrid recommendation systems such as RecGPT-V3 require continuous long-context inference on massive user behavior sequences. Every single recommendation refresh relies on real-time computation of millions of tokens, resulting in sustained and massive computational consumption.
From Hardware Leasing to Token Monetization: The Rise of Computing Power Trading Hubs
While global token demand continues to expand rapidly, the token supply chain has exposed fatal structural flaws that restrict industrial development.
The traditional computing power procurement model relies on fixed hardware units, including individual servers, GPUs, and fixed-term leased computing resources. This model suffers from severe resource imbalance: computing power is extremely insufficient during business peak hours, while massive hardware resources remain idle during off-peak hours. Statistics show that the average cabinet utilization rate of domestic intelligent computing centers only ranges from 20% to 30%, with some enterprise-level centers even below 10%.
For small and medium-sized enterprises, self-built computing infrastructure means locking massive working capital — which should have been invested in core R&D — into fixed assets, bringing heavy operational burdens.
Token factories have completely overturned the traditional computing business model. Enterprises can directly purchase standardized token outputs generated by AI models, adopting on-demand usage, volume-based billing and real-time settlement to eliminate redundant hardware costs.
Beijing Data Group’s JingSuan Token Factory delivers standardized token and API capabilities, freeing enterprises from hardware procurement, operation and maintenance expenditures. National computing platforms have incorporated token products from three major telecom operators, supporting flexible billing modes including time-based leasing, quota-based leasing and pure token metering. Guangzhou has launched the country’s first comprehensive computing service platform based on token-level scheduling, realizing diversified flexible billing systems such as volume-based and cycle-based charging.
At present, computing power has evolved into the fundamental public infrastructure of the digital economy, equivalent to water, electricity and coal, supporting all AI scenario operations.
Unexhub: Building a True Global Token Trading Marketplace
Token factories solve the basic demand for on-demand billing, yet the global token market remains highly fragmented and isolated. Most vendors’ token resources are exclusive to their self-developed models and cannot be universally shared across the industry. Closed vendor ecosystems, opaque API pricing and drastic price fluctuations have created a harsh operating environment for small and medium-sized suppliers, who face prepayment pressure for computing costs while suffering from long deferred revenue settlement cycles.
Unexhub precisely targets this core industrial pain point and builds a new-type token aggregation and distribution computing power trading hub.
Unexhub does not independently produce tokens. Instead, it integrates scattered global token supply resources and performs accurate on-demand distribution for worldwide demanders. Its business logic mirrors a public power trading center: it does not generate energy, but efficiently connects production and consumption ends to realize optimal resource allocation.
Instant settlement eliminates full-cycle capital pressure
Traditional token trading adopts monthly or quarterly deferred settlement, causing long-term capital occupation for suppliers. Unexhub’s full real-time settlement mechanism allows suppliers to withdraw funds instantly based on actual token supply volume, completely zeroing out cash flow risks and operational pressure.
Zero-commission policy gathers global high-quality industrial resources
Early cooperative settled suppliers enjoy a permanent zero-commission privilege on the platform. By sacrificing short-term service revenue, Unexhub rapidly aggregates massive global token providers. Large-scale resource convergence enables the platform to become the industry’s authoritative token pricing center, forming fair and unified market trading standards.
Dynamic API bidding realizes authentic market price discovery
The innovative API dynamic bidding model enables real-time floating pricing linked to market supply and demand. During demand peaks, moderate price increases boost supplier profit margins; during off-peak idle periods, automatic price reductions lower invocation costs for individual developers and enterprises.
This fundamental upgrade transforms tokens from unilaterally vendor-priced API services into standardized tradable assets. Token prices are no longer controlled by individual manufacturers but determined by global market supply and demand, realizing real market value discovery.
The New Era of On-Demand Paid Recommendation Engines
Douyin’s minors’ mode upgrade is just a microcosm of industrial token consumption. The 7.7 million-scale content pool, 6,000 daily updated premium contents, and continuous multimodal inference only represent a single functional module’s computational consumption.
As recommendation engines fully embrace multimodal large models, token consumption has shifted from one-time fixed investment to high-frequency, 24/7 continuous inference costs. Industrial tidal computing pricing models are gradually popularizing: Tongyi Qianwen’s token plan launches 10% off for daytime usage and 2% off for nighttime idle usage. Alibaba Cloud MaxCompute’s model computing services support out-of-the-box volume-based paid inference, enabling flexible cost control.
Tokens are evolving from auxiliary byproducts of AI models into the core measurement unit of the global AI economy. In the next round of recommendation engine competition, the dominant players will be those that provide cheaper, more flexible and more transparent token acquisition channels.
Unexhub does not develop AI models, build computing clusters or launch end-user applications. It focuses exclusively on building efficient, transparent and standardized global token trading systems, just like stock market transactions. Instant settlement solves supplier cash flow dilemmas, zero commission gathers global high-quality supply, and dynamic bidding realizes fair market pricing.
Every token consumed by Douyin’s multimodal content moderation and recommendation system, whether from ByteDance’s self-operated computing power or third-party global supply, will eventually complete matching and circulation through professional computing power trading hubs — this is the most critical underlying logic of the current AI computational inflation era.
Currently, Unexhub has gathered a full spectrum of global AI suppliers, covering top mainstream models including ChatGPT, Claude, Gemini, DeepSeek and GLM, as well as high-quality customized model resources from individual developers and SMEs. Supported by self-developed millisecond-level intelligent routing technology, it achieves ultra-fast and accurate supply-demand matching.
For developers and enterprise users, the platform supports pure token-based real-time payment, with the lowest discount reaching 30% of the official market price, greatly reducing AI inference and business operational costs. For suppliers, one-click platform access is available with real-time settlement of invocation revenue shares, realizing zero-delay profit realization.
中文译文 | Chinese Translation
2026年7月27日,抖音集团全面升级未成年人模式核心引擎——适龄推荐算法,正式落地多模态大语言模型(MLLM)技术,实现内容审核、智能推荐全链路智能化升级。
本次升级技术逻辑清晰且严谨:多模态大模型对海量视频开展跨模态语义解析,综合研判文本、画面、音频多维信息,精准评估内容认知难度与价值导向;经AI智能初筛后,由人工专家二次复审,筛选优质合规内容纳入专属内容池;再通过适龄识别模型完成内容深度理解与分层分类,精准匹配对应年龄段的适配内容。
依托「人工筛选+大模型分类+适龄推荐+内容偏好匹配」四重防护机制,平台目前日均新增优质适龄内容约6000条,未成年人专属内容池总储量已达770万条。
表面来看,这是一次优化青少年内容生态、强化合规能力的优质技术升级,但所有AI从业者都清楚背后的成本代价:多模态大模型的每一次内容识别、每一轮语义推理,都在海量消耗Token算力资源。
据行业算力测算,AI模型处理一条10秒1080P短视频,需消耗约35万Token算力。仅抖音770万条存量合规视频,即便完成一次性全量推理识别,Token消耗总量就高达2.7万亿。而这,仅仅是抖音单款APP的单一功能模块算力开销,足以窥见当下推荐引擎的算力成本压力。
推荐引擎Token通胀彻底失控
放眼全行业,算力消耗数据更为惊人,一场席卷AI行业的Token通胀危机已然爆发。
国内日均Token消耗量持续暴涨:2024年初仅为1000亿,2025年年中突破30万亿,2026年2月主流大模型日均总消耗量攀升至180万亿级别。国家数据局数据显示,截至2026年3月末,全网日均Token调用量突破140万亿,短短两年多时间涨幅超1000倍。
国家数据局局长刘烈宏在国新办发布会上明确表示,相较于2025年底100万亿的日均用量,仅三个月时间,全网Token消耗量就再度增长40%以上,行业算力需求增速持续突破预期。
当下行业正经历结构性Token通胀:并非单枚Token单价上涨,而是单位时间、单用户的算力消耗呈指数级飙升。本轮通胀的核心驱动力,来自多模态技术的全面成熟与智能体(Agent)应用的大规模商业化落地。摩根士丹利预测,2025年国内AI推理Token消耗量约10万亿,2030年将增至390万亿,五年涨幅高达37倍。
各类AI推荐引擎成为通胀重灾区。以RecGPT-V3为代表的状态混合推荐系统,需要对海量用户行为序列进行超长上下文持续推理,每一次页面推荐刷新,背后都是数百万Token的实时算力计算,形成高频、持续的巨额算力开销。
从硬件租售到Token交易:算力交易中心颠覆行业格局
在全球Token需求持续爆发式增长的同时,Token供应链的结构性致命裂痕彻底暴露,成为制约行业规模化发展的核心瓶颈。
传统算力采购模式以固定硬件单元为核心,企业通过采购单台服务器、独立显卡或固定租期算力获取算力资源。该模式存在严重的资源错配问题:业务高峰期算力严重不足,低谷期海量硬件资源闲置浪费。目前国内多数智算中心机柜平均利用率仅维持在20%-30%,部分企业级算力中心利用率甚至不足10%,资源浪费现象极其严重。
对中小科技企业而言,自建算力需要投入巨额流动资金购置硬件设备、搭建运维体系,本该用于核心研发、市场拓展的资金被牢牢锁定在固定资产中,极大加重企业经营负担、制约发展速度。
Token工厂彻底重构行业算力交易规则,企业无需绑定硬件资产,可直接采购AI模型量产的标准化Token算力成果,实现按需取用、按量计费、据实结算,彻底剥离硬件采购、运维、折旧等冗余成本。
北京数据集团推出的「京算Token工厂」,可直接交付标准化Token与API即用能力,让企业零成本搭建算力基建。全国算力平台全面接入三大运营商Token产品,支持卡时租赁、核时租赁、纯Token计费三种灵活模式。广州更是上线全国首个词元级调度综合算力服务平台,搭建按量计费、周期计费等多元化柔性结算体系,推动算力服务轻量化、普惠化。
如今,算力已然成为数字经济的底层公共基础设施,如同水、电、煤一般,支撑所有AI场景的常态化运转。
Unexhub:打通Token孤岛,打造全球化标准化算力交易市场
Token工厂解决了企业按需付费的基础算力需求,但行业依旧存在严重的资源碎片化问题。各大厂商Token资源相互隔离、无法通用,API定价不透明、波动剧烈,行业缺乏统一交易标准。中小算力供应商深陷困境:算力采购需要预先垫资,而平台收益结算周期冗长,长期资金积压导致经营困难。
Unexhub精准捕捉行业核心痛点,搭建专业化Token聚合分发算力交易枢纽,补齐产业链中游短板。
平台不自主生产Token算力,核心整合全球分散、碎片化的Token供给资源,根据企业、开发者的实时需求智能调度、按需精准分发。其模式对标国家级电力交易中心:不生产能源,却打通供需壁垒,让算力资源高效流转、最优配置。
实时即时结算:彻底清零供应商资金周转压力
传统Token行业普遍采用按月、按季度延迟结算,供应商长期面临资金积压难题。Unexhub落地全程实时结算机制,供应商可依据实际Token供给量即时回款,零账期、零占用,彻底消除现金流断裂风险。
早期零佣金扶持:汇聚全球优质算力资源
平台对早期入驻、合规结算的合作供应商实行永久零佣金政策,以短期让利换取规模化资源集聚。海量优质算力与模型供应商集中入驻后,平台自然成为全球Token市场的权威定价中心,主导行业公平交易规则。
API动态竞价机制:实现市场化公允定价
依托供需联动的动态竞价模型,平台实现API价格智能浮动:业务需求高峰期价格适度上浮,助力供应商提升盈利空间;算力闲置低谷期价格自动下调,大幅降低开发者与中小企业的AI调用成本。
这一模式的核心变革,是让Token彻底摆脱厂商单方面定价的弊端,从封闭的API服务升级为市场化交易标的,价格由全球供需关系共同决定,实现真实、透明的价值定价。
按需付费,开启推荐引擎全新时代
抖音未成年人模式的算力消耗,只是行业Token通胀的缩影。770万条存量内容、日均6000条新增内容、多模态模型全天候持续推理,仅仅是单款APP单一功能的算力开销。
随着推荐引擎全面迭代升级、深度拥抱多模态大模型,AI算力消耗彻底告别一次性固定投入模式,转变为全天候、高频次、持续性的推理成本支出。行业潮汐算力定价模式逐步普及:通义千问Token计划推出「白天1折、夜间0.2折」的错峰低价策略,阿里云MaxCompute模型服务实现开箱即用、按量付费的轻量化推理,助力企业灵活控本。
Token正在从AI模型的附属产物,升级为AI数字经济的核心计量单位。在新一轮推荐引擎技术军备竞赛中,谁能为行业提供更低成本、更灵活、更透明的Token获取渠道,谁就能抢占产业发展先机。
Unexhub不研发大模型、不自建算力集群、不触碰终端应用,专注深耕Token流通交易赛道,让全球Token交易如同股票交易般高效、透明、标准化。即时结算解决供应商资金痛点,零佣金聚合全球优质供给,动态竞价实现市场化公允定价。
抖音升级未成年人模式的背后,多模态模型审核、推荐所消耗的每一枚Token,无论源自字节自有算力还是第三方全球供给,最终都将通过算力交易中心完成撮合流通。这是当下AI算力行业最隐蔽、最核心的产业发展暗线。
目前,Unexhub已汇聚全层级全球AI供应商资源,涵盖ChatGPT、Claude、Gemini、DeepSeek、GLM等一线主流大模型厂商,同时吸纳海量个人开发者、中小企业的优质定制化模型资源。依托自研毫秒级智能路由技术,实现供需极速匹配、算力高效调度。
对企业与开发者而言,平台支持纯Token实付计费,最低可享市面市场价3折极致低价,大幅压降AI开发、推理与运营成本;对算力供应商而言,支持一键入驻对接全渠道流量,调用分成实时结算、收益零滞后兑现,高效盘活算力资产。