从抖音未成年模式看多模态大模型内容安全:算力交易中心如何支撑AI审核规模化落地

The upgraded technical logic is rigorous and complete. The multimodal model first conducts cross-modal semantic analysis on massive video resources, comprehensively evaluating text subtitles, visual imagery and audio tracks to judge each content’s cognitive difficulty and value orientation.


A TikTok Algorithm Upgrade Exposes the Computational Black Hole of AI Content Moderation|TikTok算法升级,揭开AI内容审核的算力黑洞

English First, Chinese Translation Below|英文在前,中文完整译文在后

English Version

On July 27, 2026, TikTok Group officially launched a full upgrade to the core engine of its minors’ mode — the age-adaptive recommendation algorithm — fully integrating Multimodal Large Language Model (MLLM) technology to reshape its youth content review and distribution system.

The upgraded technical logic is rigorous and complete. The multimodal model first conducts cross-modal semantic analysis on massive video resources, comprehensively evaluating text subtitles, visual imagery and audio tracks to judge each content’s cognitive difficulty and value orientation. After initial intelligent screening by the AI model, professional human experts conduct secondary manual review to filter compliant and high-quality content into an exclusive minor-oriented content pool. Finally, the age-adaptive recognition model performs in-depth classification and age matching to deliver precise age-appropriate content.

Supported by the multi-layer protection mechanism of human screening + large-model classification + age-adaptive recommendation + personalized preference matching, the platform adds approximately 6,000 pieces of premium youth-oriented content every day, with the total volume of the exclusive content pool reaching 7.7 million pieces.

On the surface, this is a comprehensive technological upgrade for safer and finer content governance. Yet for every AI industry practitioner, there is an unavoidable hidden cost behind this sophisticated mechanism: every full-scan content recognition and inference by multimodal large models consumes massive token resources at an extremely fast speed, forming a huge computational black hole.

Industry technical data shows that processing a single 10-second 1080p video via multimodal model inference consumes around 350,000 tokens. For TikTok’s 7.7 million reserved minor-protection videos, even a one-time full traversal inference generates an astronomical token consumption volume. Most importantly, this enormous computational overhead only comes from a single functional module of one application, foreshadowing the industry-wide token inflation crisis.

Token Inflation: A 1,800-Fold Surge Within 18 Months

The overall industry token consumption data is far more staggering, reflecting the explosive growth of AI computational demand.

China’s average daily token consumption has experienced skyrocketing growth: starting from 100 billion at the beginning of 2024, breaking 30 trillion by mid-2025, and climbing to the 180-trillion level in February 2026 — marking an 1,800-fold increase in just one and a half years. According to official data from the China Academy of Information and Communications Technology, the average daily token invocation volume soared to 140 trillion by the end of March 2026. Morgan Stanley’s industry forecast predicts that China’s AI inference token consumption will rise from about 10 trillion in 2025 to nearly 390 trillion by 2030, a 37-fold growth over five years.

The AI industry is undergoing structural token inflation: token consumption per unit of time and per individual user is growing exponentially. Every AI dialogue interaction, video semantic comprehension and intelligent code generation relies on 24/7 running GPU clusters to sustain continuous token output.

While global token production capacity keeps expanding rapidly, the token economic supply chain has exposed a fatal structural defect: sufficient production capacity but extremely inefficient circulation and distribution.

Core Crack in the Token Supply Chain: Easy Production, Hard Circulation

The token production end faces no shortage of computing power support. NVIDIA high-performance GPUs serve as the core infrastructure for token generation, and global intelligent computing centers operate as standardized token factories with stable high output. ByteDance has also raised its 2026 AI capital expenditure plan to over 200 billion yuan to expand computational capacity. At present, enterprises no longer purchase vague, abstract computing power; what they truly need and trade is the effective token output converted from computing resources.

The core pain point of the industry lies in the backward distribution system and lagged financial settlement mechanism.

The current token market is severely fragmented and isolated. Each large model vendor operates an independent closed-loop ecosystem, with opaque, unstandardized and highly volatile API pricing. Small and medium-sized computing power suppliers are trapped in operational dilemmas: they need to prepay high costs for computing resources, while revenue settlement adopts long monthly or quarterly payment cycles, resulting in severe capital occupation and enormous cash flow pressure.

A deeper structural contradiction comes from the dynamic mismatch between token supply and demand. API invocation volume surges sharply during daytime business peak hours with tight computing supply; during nighttime off-peak hours, massive GPU resources remain idle and wasted. However, the industry still adopts rigid fixed pricing all year round. Suppliers cannot obtain premium returns during peak demand, while users cannot access preferential discounts during off-peak idle periods. The entire market lacks an effective, fair price discovery mechanism.

It is precisely this industrial structural crack that Unexhub captures and targets, bringing a brand-new solution to reshape the global token trading ecosystem.

Unexhub: Building a Standardized Global Token Trading Marketplace

Essentially, Unexhub completely reshapes the distribution, circulation and pricing logic of global token resources.

It builds a professional computing power trading hub focused on token aggregation and on-demand distribution. Instead of independently producing tokens, Unexhub integrates scattered, fragmented token supply resources worldwide and performs precise dynamic distribution for global demanders. Its business model is similar to a public power trading center: it does not generate power itself, but breaks industry barriers to realize efficient flow, optimal allocation and maximum utilization of resources from production ends to consumption ends.

Instant settlement mechanism eliminates supplier cash flow pressure

Traditional token trading is restricted by long deferred settlement cycles, which suppress the capital turnover of small and medium-sized suppliers. Unexhub launches a full real-time instant settlement rule: suppliers can recover corresponding funds immediately based on the actual token supply and invocation volume. This mechanism completely zeros out long-term capital occupation and financial risks, enabling small-scale computing power providers to steadily participate in token supply without worrying about cash flow disruptions.

Zero-commission incentive gathers global high-quality industrial resources

Platform commission has long been the core profit source of traditional mid-tier token trading markets. Unexhub offers a permanent zero-commission policy for early settled cooperative suppliers. By sacrificing short-term platform profits, it rapidly gathers massive high-quality token suppliers and computing resources worldwide. When sufficient supply converges on the platform, Unexhub naturally becomes the de facto authoritative pricing center for the global token market.

Dynamic API bidding realizes real market price discovery

Unexhub’s most innovative design is the supply-demand linked API dynamic bidding model. API prices float intelligently with market changes: prices rise moderately during demand peaks to create higher profits for suppliers; prices drop during off-peak idle periods to reduce invocation costs for individual developers and enterprise users. This fundamental transformation turns tokens from unilaterally vendor-priced API services into standardized, tradable market assets with authentic price discovery capabilities.

Why Computing Power Trading Hubs Become the Ultimate Industrial Winner

TikTok’s minors’ mode algorithm upgrade is just a microcosm of the global token inflation crisis. The 7.7 million-scale content pool, 6,000 daily updated premium contents, and continuous high-intensity multimodal inference only represent the tiny tip of the iceberg of AI computational consumption.

With the large-scale deployment and popularization of AI video generation, intelligent code assistants, automated Agent workflows and enterprise-level large-scale inference services, global token demand will continue to expand exponentially. The more fragmented and complex the token supply chain becomes, the more prominent the core value of professional resource aggregation, matching and circulation platforms.

Unexhub does not develop AI models, build computing power clusters or launch end-user applications. It focuses on only one core business: making global token trading as efficient, transparent and standardized as stock trading. Instant settlement solves cash flow pain points for suppliers, zero commission gathers global high-quality supply resources, and dynamic API bidding realizes fair market price discovery.

In the era of rampant token inflation, whoever controls token circulation channels controls the lifeline of the global AI economy. As the core token exchange and computing power trading hub of the industry, Unexhub does not produce tokens, but dominates token circulation and pricing rights — and in any industrial ecosystem, the trading exchange always captures the largest market volume and long-term dividends.

Every token consumed by TikTok’s multimodal content moderation 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 noteworthy underlying logic behind the current AI computational inflation trend.

Currently, Unexhub has aggregated 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 small and medium-sized enterprises. 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 token-based real-time payment, with the lowest discount reaching 30% of the official market price, effectively reducing large-scale AI invocation costs. For suppliers, one-click platform access is available with real-time settlement of invocation revenue shares, realizing zero-delay profit realization and fully activating the vitality of the global token trading market.


中文译文 | Chinese Translation

2026年7月27日,TikTok集团对未成年人模式核心引擎——适龄推荐算法完成全面升级,正式落地多模态大语言模型(MLLM)技术,全面重构青少年内容审核、识别与分发体系。

本次升级技术逻辑严谨且完善:多模态大模型首先对海量视频内容开展跨模态语义解析,综合研判文本、画面、音频多维信息,精准判定内容认知难度与价值导向;经AI模型初步智能筛选后,由专业人工专家完成二轮复核,筛选合规优质内容纳入青少年专属内容池;最后通过适龄识别模型完成深度分类与年龄适配匹配,实现精细化适龄内容分发。

依托「人工筛选+大模型分类+适龄推荐+偏好匹配」四重防护机制,平台目前日均新增优质青少年内容约6000条,专属内容池总储量已达770万条,全方位提升青少年内容生态安全性与优质度。

表面来看,这是一次完善且优质的技术迭代,实现内容治理精细化升级。但所有AI从业者都清楚,这套精密机制的背后,隐藏着无法规避的算力成本难题:多模态大模型每一次全量内容识别、语义推理,都在海量消耗Token资源,形成难以填补的AI算力黑洞。

据行业技术测算,模型处理一条10秒1080P短视频,需消耗约35万Token算力。仅TikTok这770万条存量青少年合规内容,即便完成一次性全量遍历推理,Token消耗量就已达到天文数字。而这仅仅是单款APP的单一功能模块,足以窥见全行业Token通胀危机的严峻性。

Token通胀失控:18个月暴涨1800倍,行业算力需求爆发

放眼全行业,Token消耗的增长数据更为惊人,直观印证AI算力需求的爆发式增长。

国内日均Token消耗量呈现断崖式飙升: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已攀升至180万亿级别,18个月内暴涨1800倍。中国信息通信研究院官方数据显示,2026年3月末,全网大模型日均Token调用量已飙升至140万亿。摩根士丹利行业预测报告指出,2025年国内AI推理Token消耗量约10万亿,2030年将接近390万亿,五年增幅达37倍。

当下AI行业正经历结构性Token通胀:单位时间、单用户的Token消耗呈指数级攀升。每一次AI对话交互、每一轮视频语义理解、每一段智能代码生成,背后都是GPU集群7×24小时不间断运转,持续量产Token算力。

在全球Token产能持续扩张的同时,Token经济供应链的致命结构性缺陷彻底暴露:产能充足、流通瘫痪,生产端与消费端严重错配。

Token供应链裂痕:生产易、流通难,行业痛点凸显

当前Token生产端完全不存在算力短缺问题。英伟达高端GPU是Token量产的核心基建,全球各地智算中心如同标准化算力工厂,持续稳定输出海量Token。字节跳动更是将2026年AI资本开支计划提升至2000亿元以上,全力扩充算力产能。如今企业采购的核心标的,早已不是模糊的算力硬件,而是算力最终转化的有效Token使用权。

行业真正的核心瓶颈,集中在分发流通与资金结算两大环节。

目前Token市场高度碎片化、孤岛化,各大模型厂商各自闭环运营,API定价不透明、无标准且波动剧烈。中小算力供应商深陷经营困境:算力采购需要预先垫付高额成本,但平台收益结算沿用按月、按季度长周期账期,长期资金积压导致现金流承压,经营风险居高不下。

更深层的结构性矛盾,来自Token供需的动态错配。日间业务高峰期,企业、开发者API调用需求暴增,算力资源紧张稀缺;夜间低谷时段,大量GPU算力闲置空转、资源严重浪费。但行业始终沿用僵化的固定定价模式,无法适配市场波动:高峰期供应商无法赚取供需溢价,低谷期用户无法享受低价红利,整个市场缺失公平透明的价格发现机制。

Unexhub精准捕捉这一行业结构性裂痕,针对性破解行业痛点,重塑全球Token交易生态格局。

Unexhub:重构全球Token逻辑,打造标准化算力交易市场

Unexhub的核心价值,是彻底颠覆传统全球Token的分发、流通与定价规则,打通行业全链路壁垒。

平台打造专业化Token聚合分发算力交易枢纽,不自主生产Token,核心整合全球分散、碎片化的Token供给资源,根据企业与开发者的实时需求,实现智能调度、按需精准分发。其模式对标公共电力交易中心:不自主生产能源,却打通上下游壁垒,实现算力Token资源从生产端到消费端的高效流转、最优配置、最大化利用。

实时结算机制:清零供应商现金流压力

传统Token行业长期受困于长周期延迟结算,中小供应商资金周转严重受阻。Unexhub落地全程实时即时结算机制,供应商可依据实际Token供给量、模型调用量即时回款到账,彻底杜绝长期资金积压问题,零资金压力稳定参与市场供给,杜绝现金流断裂风险。

零佣金激励体系:汇聚全球优质算力资源

平台佣金是传统Token交易中游市场的核心盈利来源。Unexhub对早期入驻合作供应商实行永久零佣金扶持政策,主动放弃短期平台收益,快速汇聚全球海量优质算力与模型供给资源。当海量供需资源高度集聚,平台自然成为全球Token市场的权威定价中心,主导行业定价规则。

API动态竞价模式:实现市场化公平定价

平台最具创新性的核心设计,是供需联动的API动态竞价机制。价格随市场行情智能浮动:需求高峰期适度涨价,为供应商创造更高收益;算力闲置低谷期自动降价,大幅降低开发者与企业的AI调用成本。彻底改变厂商单方面定价的行业陋习,让Token从单一API服务,转变为具备真实市场价值发现能力的标准化交易标的。

算力交易枢纽:Token通胀时代的终极赢家

TikTok未成年人模式算法升级,只是全球Token通胀危机的缩影。770万条存量内容池、日均6000条优质内容更新、持续高强度多模态推理消耗,仅仅是AI算力消耗的冰山一角。

随着AI视频生成、智能代码助手、自动化智能体工作流、企业级大规模推理等应用全面落地普及,全球Token算力需求将持续指数级扩张。Token供应链越碎片化、越复杂,专业聚合、匹配、流通平台的产业价值就越核心。

Unexhub不研发大模型、不自建算力集群、不触碰终端应用,专注深耕一件事: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,动态竞价实现市场公平定价,三重优势重构行业生态。

在Token全面通胀的时代,掌控Token流通渠道,就是掌控全球AI经济的生命线。算力交易枢纽不生产Token,却掌控Token的流通与定价权,而任何产业的终极红利,永远归属掌握交易流量的核心枢纽平台。

TikTok多模态内容审核系统消耗的每一枚Token,无论源自字节自有算力还是第三方全球供给,最终都将通过专业算力交易平台完成撮合流通。这是当前AI算力浪潮中,最值得关注的产业暗线。

目前,Unexhub已汇聚全量级全球AI供应商资源,全面覆盖ChatGPT、Claude、Gemini、DeepSeek、GLM等海内外主流大模型,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。

对开发者与企业用户而言,平台支持按实际Token消耗量实付计费,最低可享官方市场价3折极致低价,大幅压降规模化AI落地成本;对供应商而言,支持一键入驻接入全渠道流量,调用分成实时结算、收益零滞后兑现,全方位激活全球Token交易市场活力。