前沿大模型的能力较量并未放缓，通向AGI的竞速依然在提速。与此同时，商业层面的较量也愈发关键，其重要性已比肩模型能力本身，成为另一大核心衡量标准。Anthropic按月公布的Token ARR增长速度，俨然化作观测整个行业的风向标。
在这场商业角逐中，Token收入化作了衡量实力的标尺。基于此考量，爱分析把国内最具代表性的9家AI企业，统一依据2026年8月的Token收入年化数据进行了排名，旨在呈现中国AI市场的竞争格局与演进态势。
本榜单致力于解答三大核心疑问：中国Token经济的发展现状究竟怎样，各家企业实力孰强孰弱及具体位次，以及登榜厂商各自具备哪些优势与不足。
该榜单报告正文包含四大板块：首部分为厂商排名，次部分对上榜企业逐一剖析，第三部分聚焦Token行业趋势洞察，末尾部分则阐述榜单的编制方法及数据释义。
01
榜单排名
综合统计信息与排名情况，此榜单清晰映射出当下中国Token市场的演进格局，以及AI企业的梯队分布形态。
位居榜首的字节与排名末位的百度，二者的收入差距高达十倍以上，这表明国内Token市场正呈现出高度向头部聚拢的态势。与此同时，入围的9家企业分属百亿、50亿、30亿三个量级阵营，各梯队之间存在着显著的鸿沟。
图1：中国AI厂商Token收入榜单
首个阵营为百亿俱乐部，字节、阿里与智谱成功跻身其中。即便在同一梯队，差距也相当明显：字节凭借约300亿元的年化ARR稳居榜首，不仅是排名第二的阿里（160亿元）的两倍，更是智谱的三倍之多。字节凭借在视频场景的定价主导权稳占鳌头；阿里则是唯一一家贯通芯片、云、模型与应用四大层级并成功转化为Token收入的企业；智谱则重仓Coding赛道，并在IPO后实现了商业化的强势反弹。
第二阵营处于50亿元量级，包含DeepSeek、Kimi、MiniMax与可灵四家企业。该梯队厂商皆凭借过硬的模型实力实现业绩攀升，但增长驱动力各有差异：DeepSeek得益于API价格的调升，Kimi借助K3单一模型实现了业绩的三级跳，而MiniMax与可灵则依托视频模型在海外市场大举扩张。
第三阵营为30亿元量级，腾讯与百度位居末位。这两家各有各的困境与失误，共同点在于Token业务推进迟缓，从整体规模审视，已明显处于落后状态。不过，腾讯或许还能以“马拉松才刚刚起跑”自我宽慰，而百度至今未见明显的回升势头，令人印象深刻的亮点仅剩昆仑芯而已。
02
厂商分析
2.1 字节：断崖式领先，赢在视频情形的定价权
字节凭借300亿元的年化ARR一骑绝尘，年内有望向千亿大关发起冲击。火山引擎总裁谭待回忆道：“三年前将MaaS确立为核心战略时，外界普遍觉得Token是个赔本赚吆喝的买卖。”然而如今审视，正是当年提前布局的投入才铸就了眼下的局面。
字节领跑的底座在于其视频模型能力在业内占据绝对优势，并借此确立了视频生成场景的定价主导权。其视频每秒单价较同类模型至少高出50%，且分销渠道的折扣力度极小，Seedance系列模型的折扣极少低于九折。
当然，字节亦存在自身的软肋——语言模型。字节CEO梁汝波在8月年年中全员会议上坦言，其大语言模型与海外头部玩家的差距正在扩大，并要求团队在心态上接纳暂时的落后局面，豆包大模型的下一代版本也因此延期发布。
2.2 阿里：举全栈AI之力，兑现Token收入
阿里凭借160亿元的年化ARR，位列榜单次席。阿里的核心优势在于全栈AI布局，同时也是榜单中唯一一家将收入根基构建在完整Token产业链上的企业。
依托在芯片、云、模型及应用四大维度的全面落子，阿里Token业务的财务表现十分亮眼，不仅收入丰厚、增速迅猛，利润端的表现同样可圈可点。从收入端看，今年5月至8月的3个月内，其Token业务ARR实现了近1倍的增幅；从利润端看，阿里集团CEO吴泳铭在近期的业绩说明会上表示“AI算力预计2到3年收回成本”，这暗示其Token业务的毛利率已突破50%。
在这份出色成绩单的背后，自研GPU平头哥迈入规模化量产阶段是一个关键里程碑，基于自研芯片所赋予的成本优势，阿里得以不断优化Token业务的财务模型。此外，为促进Token收入攀升，阿里特意成立了ATH事业群，其在组织资源投入力度上，也是9家上榜企业中最大的。
全栈AI布局赋予了阿里收入的广度与深度，但其短板在于开源领域的模型霸主地位已然动摇。在2025年，千问尚是开源阵营模型能力的标杆，但如今已被K3、GLM等后起之秀拉下神坛。
2.3 智谱：量价齐升标杆
位列榜单第三名的智谱估值达110亿元，堪称本榜中实现Token“量价齐升”的典型代表。步入2026年，该公司的API平均售价上扬了101%，与此同时，Token的调用规模相比年初也暴增了40倍。
实现量价齐升的关键驱动力，源自模型在AI Coding能力上的显著跃升。唐杰于2025年下半年果断选择All in AI Coding，放弃了交付沉重的ToB项目，就目前成效而言，这一赛道押注得极为精准。
智谱打算在今年底推出GLM-6.0，将核心筹码压在模型自我进化上，唐杰更是把GLM-6.0的核心理念定义为“AI训练AI”。在Token业务ARR方面，智谱官方预估年底将进一步提升50%，冲至160亿元规模，以目前进度研判，该目标极有希望提前兑现。
智谱的软肋主要集中在人才与组织架构上。清华系背景带来的高浓度学术人才，往好处看是智谱的坚实护城河，往坏处看则成了限制发展的天花板。顶尖学子的高频流动，固然能维系一时的技术工艺优势，却难以转化为企业内部沉淀的组织能力，而恰恰是这种组织能力才是维系长久领跑的关键。
2.4 DeepSeek：口碑与利润的双丰收
看似资金充沛的DeepSeek，以60亿元的年化ARR稳居第四名。作为榜单上反差最为强烈的厂商，其模型实力与毛利率均处于领跑地位，但在算力供应方面却被远远甩在后面。
DeepSeek的优势在于工艺打磨，始终对标SOTA模型。得益于V4系列的优异口碑与深度的工程优化，其API毛利率高达82.9%，在国内市场遥遥领先。按照DeepSeek的定价逻辑，单台GPU服务器仅需10个月便可回本，相比之下，阿里的投资回报周期则长达2至3年。
在算力维度，DeepSeek暴露出了阶段性的不足。整个2025年，DeepSeek在AI Infra上的开销仅有12亿元，甚至比不上头部企业单月的花费。由于供给难以满足需求，曾导致API出现大范围的服务中断。随后采取的大幅调价与峰谷分时定价策略，究其本质，都是在算力受限下的无奈之举，意图通过价格门槛来削减部分用户需求。
然而，这一算力瓶颈正被迅速填补。在今年前7个月，DeepSeek的AI Infra开支已达到110亿元，叠加上年内完成的两轮千亿元融资，算力短缺的缺口正以极快的速度被弥合。
2.5 Kimi：K3驱动的爆发式增长
Kimi凭借58亿元的年化ARR排在第五位。其爆发式增长主要归功于K3这款模型：6月份时，其ARR尚不足25亿元，但在K3问世仅仅两个月后，该数据便实现了翻倍攀升。
伴随K3的问世，其参数体量与模型效能双双迈入全球顶尖行列。作为当前全球规模最大的开源模型，K3凭借2.8万亿参数将开源模型的规格推升到了全新高度，同时其模型实力也博得了全球开发者的青睐，综合评价已与Claude等一线闭源模型并驾齐驱。此外，规划中的K4模型将进一步扩大参数规模，预计将达到5至6万亿参数的量级。
然而，盲目冲刺超大参数模型，使得Kimi的扩张深受高端算力掣肘。一方面，其模型训练离不开英伟达高端显卡的支持；另一方面，模型推理也难以在国产GPU上顺畅运行，这两大因素共同推高了API的售价。
在K3上线后的短短48小时内，Kimi的算力资源便濒临极限，不得不暂停了C端新用户的订阅服务。预计在短期内，其算力补充的速度难以追赶智谱等同业者。与Kimi不同，智谱等厂商将参数量约束在万亿以内，使得推理环节能够依托国产GPU运行，从而大幅缓解了算力供给压力。而对Kimi而言，这种劣势在短期内极难逆转。
2.6 MiniMax：视频模型强势崛起，语言模型表现欠佳
MiniMax以约48亿元的年化ARR成绩，位列榜单第六名。
今年5月，Minimax的ARR仅为30亿元，但随着开源视频生成模型H3的推出，8月份ARR直接跃升至48亿元（注：官方公布的8亿美元ARR，是根据8月最高单周数据年化得出，而非按月度数据年化）。经过推理加速优化的H3 Max，生成5秒视频耗时不到3秒，史上首次实现视频生成效率超越播放速度，这无疑为AI视频直播的商业化落地拓宽了想象边界。
除了视频模型实力的提升，Minimax的商业化进程也迎来改善。2025年，Minimax的商业化策略还是重C端、轻B端，但到了今年中旬，局势彻底倒转，B端营收占比由去年的30%猛增至80%，且企业客户总数突破了200万大关。
Minimax的弱项体现在毛利率表现上。因为H3模型主推高性价比，定价相对较低，致使其毛利率落后于Seedance和可灵这两款同属视频赛道的模型。
与此同时，Minimax的语言模型M3表现未达预期，不仅消耗了海量算力资源，却未能产出相匹配的效益，进而也拖累了整体业务的毛利率水平。
2.7 可灵：深耕海外视频领域，估值已超Minimax
可灵凭借40亿元的年化ARR位列第七，其差异化优势主要体现在海外视频赛道上。
过去一年间，可灵ARR实现了超过400%的增速，这一增长主要由海外企业客户驱动。其海外收入贡献了约八成份额，来源市场涵盖北美、韩国及东南亚等地区。在客户构成上，可灵的核心付费群体为影视制作公司和导演工作室等，累计企业客户数接近5万家。
今年7月，可灵正式从快手独立分拆，投后估值达到180亿美元，已超越已上市公司的Minimax市值。不过，高估值也意味着可灵在ARR方面面临更大挑战，公司预计今年年底冲刺10亿美元目标，并计划在2027年底达到20亿美元。
2.8 腾讯：起得早不如熬得久
腾讯的年化ARR只有32亿元，排在第八位，这与其整体的AI战略步调相吻合，呈现出一种从容不迫的姿态。在头部大厂阵营里，腾讯表现出的态度是对Token业务并不太上心。
反观其他厂商，火山把Token当作核心增长引擎，阿里甚至专门成立了ATH事业群来推进Token业务，腾讯却在这方面几乎没有明显布局。这种反差折射出腾讯内心焦虑与自我慰藉并存的复杂心态。
从焦虑的一面来看，马化腾的表述十分坦诚："一年前还以为自己已经上了船，后来才发现船是漏水的，现在感觉站是站上去了，却还没法坐稳。"登上漏水的船、坐立难安的窘迫，恰是腾讯当下心态的真实写照。
从自我安慰的角度看，腾讯觉得距离AI的终局还为时尚早，仍有调整空间。汤道生将AI竞争比喻为马拉松长跑，"起得早固然重要，但能熬得久才能把握住不断涌现的新机会"，他还指出目前"大概才跑完第一公里"。
马拉松的说法本身没问题，但腾讯在Token商业化方面似乎连起跑线都还没到。混元Hy3、Hy4通过免费API渠道大量开放，付费转化率偏低。一个尚未真正入局的长跑选手，讨论能否熬得久，似乎还言之过早。
2.9 百度：没落
百度年化ARR为25亿元，在榜单中排名末位，仅相当于字节的十二分之一。
百度是国内率先将大模型推向台前的企业，文心一言曾一度成为中文大模型的标杆。然而几年光阴过去，时过境迁，谁也未曾料到百度会走到这般境地。当然，百度仍拥有自研芯片昆仑芯，其出货量排在华为和阿里之后。
问题或许出在李彦宏的战略判断上。2024年，他曾公开表示开源模型属于智商税；而在最近的财报电话会上，李彦宏也不得不承认百度确实存在失误。
时至今日，百度的模型已鲜少出现在各类排行榜中，当其他厂商在8月集中发布新模型时，百度却按兵不动。很难让人相信，百度还能重新跻身第一梯队。
03
行业洞察
3.1 中美AI差距缩小，发展路线出现分化
中美之间AI能力的差距究竟有多大，这次的Token收入排名提供了一个客观的观察窗口。
中国追赶美国的步伐、节奏以及当前所处位置，若以Anthropic的ARR时间线作为参照，能够解读出诸多信息。Anthropic的ARR从2025年3月的14亿美元攀升至2026年7月的650亿美元，而中国几家主要AI厂商当前的发展阶段，基本都能对应到这条增长曲线上的某个节点。
图2：中美厂商ARR追赶曲线
本次排名前三的厂商中，字节的发展水平大致相当于Anthropic 2025年8月的阶段，阿里约对应2025年年中，智谱则处于2025年初的位置。换言之，中国头部AI厂商整体上落后美国大约一年时间。但中国的追赶势头十分迅猛，双方差距正在不断缩小。
这一加速趋势背后有两股力量驱动。其一是算力方面的持续投入以补足供给能力，阿里募资800亿元全部投入基础设施建设，同时DeepSeek也在加快自建算力的步伐，其位于乌兰察布的1GW算力中心预计于今年年底开始陆续上线运行，明年上半年实现全面交付。
另一股驱动力来自应用场景，中国市场的短剧、Coding、办公、政务等领域产生了海量的Token消耗场景，使得算力投入能够迅速实现商业变现。
与此同时，中美两国在AI商业化路径上的差异也日益显现。美国模型的优势在于顶尖能力，瞄准的是愿意为模型能力支付高溢价的用户群体。中国模型则胜在开源生态和性价比，面向的是庞大的腰部用户群体。这两个市场哪个空间更大尚难定论，但中美在AI领域的差距正在不断缩小。
3.2 模型涨价属于短期现象，降价才是长期主趋势
在过去半年中，模型API价格普遍上涨，对Token收入增长起到了重要推动作用。短期内模型供不应求，算力成为Token收入的天花板，只要拥有算力便不愁销路。
今年上半年，涨价幅度最大的是DeepSeek。由于算力供应受限，DeepSeek不得不通过涨价来保障高价值客户的服务体验，从而舍弃了付费能力较弱的低价值客户。相比之下，智谱因提前布局了算力储备，实现了量价齐升。在AI Infra领域，智谱先后投资了硅基流动、基流科技、无问芯穹等多家Infra厂商，极大程度地缓解并补充了算力供给。
但从长期来看，模型降价才是主要趋势，尤其是轻量级的Flash模型。
通常情况下，市场上模型调用量的绝大部分份额被Flash模型占据。以OpenRouter的调用榜单为例，DeepSeek V4 Flash、GPT-5.6 Luna、GLM 5.3 Flash、Gemini 3.7 Flash等长期稳居Top10之列。
而Flash模型将成为未来的价格杀手，例如GLM 5.3 Flash的售价不到GLM 5.3的10%，GPT-5.6 Luna仅为旗舰模型Sol的5%。
3.3 收入与调用量正呈现结构性分离趋势，视频模型与语言模型率先走向不同路径
这份榜单背后，一个趋势格外清晰：视频模型与语言模型走向分化，前者主要贡献收入，后者则主要贡献调用量。
视频模型单价较高，例如字节跳动凭借AI短剧和视频场景登顶榜单，可灵的ARR在过去一年实现了四倍增长。相比之下，语言模型单价较低但调用量庞大，贡献的主要是调用量而非实际收入。
这两类模型的商业化路径，对于未来其他类型模型的发展具有重要的参考价值。
例如，AI4S的科研模型很可能走高单价路线，而具身模型则更可能走低单价、高调用量的路线。对于厂商来说，这两条路径所需的资源和策略截然不同，赚取收入需要依靠场景和定价策略，而获取调用量则依赖生态建设和低价优势。
04
数据说明
本榜单的结论基于可追溯的数据与方法论。以下将从榜单制作方法、数据来源、研究方法三个维度，阐述榜单的构建过程。
本榜单的收入统计口径严格限定为Token计费收入，既包括通过模型API提供的调用服务，即Token Plan和Coding Plan，也包括通过MaaS平台提供的多模型服务。模型本地化部署项目、GPU云资源租赁与销售、AI应用的C端订阅等收入，均不纳入统计范围。
本榜单统一以2026年8月作为观察周期，将各厂商当月的Token收入数据进行年化处理。选取当月数据年化，旨在反映各厂商最新的收入状况，避免历史累积数据造成的滞后效应。货币换算统一采用1美元兑换7.1元人民币的汇率。榜单按照年化ARR数据从高到低进行排序。
4.2 数据来源
本榜单的数据来源于四类信源：官方披露的审计财报、管理层业绩会披露、行业权威信源，以及爱分析的一手调研。
数据制作坚持交叉验证原则。对于同一厂商的收入数据，力求在多个信源之间实现相互印证。交叉验证的目的在于确保榜单上的每个数字都能经得起公开核查。
4.3 研究方法论
本榜单的研究遵循三阶段方法论，即先建库、再验证、后评估。
建库阶段以桌面研究为主，通过收集公开财报、公告、行业研报及权威媒体报道，梳理市场上主要的Token收入厂商，从而确定上榜范围。
以上榜企业官方公布的信息为基准，逐一核对其收入数据，并厘清其披露的具体性质及统计口径。
在评估环节，综合采用了深度访谈与数据调研手段。深度访谈侧重于行业调研，受访群体囊括了Token产业链上下游的相关从业者，旨在校准对行业整体的判断。数据调研则主要针对存疑数字进行定向核实。通过将桌面研究、深度访谈与数据调研三种方法相互印证，构建起多维验证体系，从而确保每一个数字都具备多重依据支撑。
近期厂商征集
本文来源：https://www.163.com/dy/article/L6CCB9LK0531W0RG.html
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
0
The capability competition among frontier large models has not slowed down; the race toward AGI continues to accelerate. Meanwhile, the commercial battle has grown increasingly critical, its importance now rivaling model capabilities themselves and becoming another core benchmark. Anthropic's monthly Token ARR growth rate has effectively become a barometer for the entire industry.
In this commercial contest, Token revenue has become the measuring stick of strength. On this basis, AIFenxi ranked the nine most representative AI enterprises in China based on their August 2026 Token revenue annualized data, aiming to present the competitive landscape and evolutionary trends of China's AI market.
This ranking strives to answer three core questions: What is the current state of China's Token economy? How strong or weak are each enterprise, and what are their specific rankings? And what advantages and disadvantages do the listed vendors each possess?
The main body of the ranking report contains four sections: the first covers vendor rankings, the second provides individual analysis of listed companies, the third focuses on Token industry trend insights, and the final section explains the compilation methodology and data interpretation.
01
Ranking Results
Based on comprehensive statistics and rankings, this list clearly reflects the current evolution of China's Token market and the tier distribution of AI enterprises.
ByteDance, ranked first, and Baidu, ranked last, have an income gap of over ten times, indicating that China's Token market is showing a strong tendency toward concentration at the top. Meanwhile, the nine qualifying enterprises fall into three tiers—hundreds of billions, 50 billion, and 30 billion—each separated by a significant gap.
Figure 1: China AI Vendor Token Revenue Ranking
The first tier is the "hundreds-of-billions club," with ByteDance, Alibaba, and Zhipu all qualifying. Even within the same tier, the gaps are quite apparent: ByteDance, with an annualized ARR of approximately 30 billion yuan, firmly holds the top spot, double that of second-place Alibaba (16 billion yuan) and triple that of Zhipu. ByteDance leads by virtue of its pricing dominance in video scenarios; Alibaba is the only company that has integrated all four layers—chips, cloud, models, and applications—and successfully converted them into Token revenue; Zhipu has heavily invested in the AI Coding track and made a strong commercial rebound after its IPO.
The second tier is in the 5 billion yuan range, including DeepSeek, Kimi, MiniMax, and Kling. Vendors in this tier have all achieved growth driven by solid model capabilities, but the drivers of growth differ: DeepSeek benefited from an increase in API pricing, Kimi achieved a threefold performance leap thanks to its single K3 model, while MiniMax and Kling expanded aggressively in overseas markets relying on their video models.
The third tier is the 3 billion yuan range, with Tencent and Baidu at the bottom. Both face their own difficulties and mistakes, sharing the common trait of slow Token business advancement; in terms of overall scale, they are clearly lagging. Tencent might still comfort itself with the idea that "the marathon has just begun," while Baidu shows no obvious signs of recovery, with its only notable highlight being Kunlunxin chips.
02
Vendor Analysis
2.1 ByteDance: Cliff-like Lead, Winning Pricing Power in Video Scenarios
ByteDance runs away with an annualized ARR of 30 billion yuan, aiming to challenge the 100 billion mark this year. Tan Dai, President of Volcano Engine, recalled: "Three years ago when we established MaaS as a core strategy, everyone thought Token was a losing proposition that only bought you noise." Yet today, it is precisely the early investments made back then that have created the current situation.
ByteDance's lead rests on its video model capabilities, which hold an absolute advantage in the industry, thereby establishing pricing dominance in video generation scenarios. Its per-second video price is at least 50% higher than similar models, and discounts through distribution channels are minimal, with Seedance series models rarely discounted below 90% of the list price.
Of course, ByteDance also has its own weakness—language models. During ByteDance's CEO Lyu Bo's mid-year all-hands meeting in August, he admitted that the gap between its large language model and leading overseas players is widening, and required the team to mentally accept this temporary setback. Consequently, the next generation of Doubao's large model has been delayed.
2.2 Alibaba: Leveraging Full-Stack AI to Realize Token Revenue
Alibaba ranks second with an annualized ARR of 16 billion yuan. Its core advantage lies in its full-stack AI layout, and it is the only company on the list whose revenue foundation is built on a complete Token industry chain.
Relying on comprehensive moves across chips, cloud, models, and applications, Alibaba's Token business shows impressive financial performance—not only strong and fast-growing revenue, but also noteworthy profitability. On the revenue side, Alibaba's Token business ARR grew nearly 100% over the three months from May to August this year. On the profit side, Alibaba Group CEO Wu Yongming stated at a recent earnings call that "AI computing power investment will be recouped in 2 to 3 years," implying that the gross margin of its Token business has exceeded 50%.
Behind this excellent performance, a key milestone is that Tongyi's self-developed GPU has entered large-scale mass production. Leveraging the cost advantages conferred by its self-developed chips, Alibaba has been able to continuously optimize the financial model of its Token business. Additionally, to boost Token revenue, Alibaba specifically established the ATH Business Group, which has the largest organizational resource commitment among the nine listed companies.
The full-stack AI layout has given Alibaba breadth and depth in revenue, but its weakness lies in the fact that its dominant position in the open-source model arena has been shaken. In 2025, Qwen was still the benchmark for open-source model capabilities, but it has since been dethroned by rising stars such as K3 and GLM.
2.3 Zhipu: A Benchmark for Volume-and-Price Growth
Ranked third on the list, Zhipu is valued at 11 billion yuan, making it a typical example of achieving Token "volume-and-price growth" among all listed companies. In 2026, the company's average API selling price increased by 101%, while Token call volume surged 40 times compared to the beginning of the year.
The key driver behind this volume-and-price growth stems from a significant leap in the model's AI Coding capability. Tang Jie made a decisive move to go all-in on AI Coding in the second half of 2025, abandoning heavy To-B delivery projects. In terms of results, this strategic bet has proven highly precise.
Zhipu plans to launch GLM-6.0 by the end of this year, placing its core bets on the model's self-evolution. Tang Jie has defined the core philosophy of GLM-6.0 as "AI training AI." Regarding Token business ARR, Zhipu's official estimate is a further 50% increase by year-end, reaching 16 billion yuan; based on current progress, this target is highly likely to be met ahead of schedule.
Zhipu's weaknesses center on talent and organizational structure. The high concentration of academic talent derived from its Tsinghua background serves as a solid moat, but it can also become a ceiling limiting development. The high turnover of top graduates may maintain short-term technological advantages, but it cannot translate into organizational capabilities沉淀within the enterprise—which is precisely what is needed to sustain long-term leadership.
2.4 DeepSeek: A Double Harvest of Reputation and Profit
Seemingly well-funded, DeepSeek ranks fourth with a steady annualized ARR of 6 billion yuan. As the most contrasting vendor on the list, its model capabilities and gross margin both lead the field, yet it has been left far behind in computing supply.
DeepSeek's advantage lies in its craftsmanship and relentless pursuit of SOTA models. Thanks to the excellent reputation of its V4 series and deep engineering optimization, its API gross margin reaches 82.9%, far ahead of the domestic market. Under DeepSeek's pricing logic, a single GPU server can break even in just 10 months, compared to Alibaba's investment payback period of 2 to 3 years.
In terms of computing power, DeepSeek has exposed阶段性 shortcomings. Throughout 2025, DeepSeek spent only 1.2 billion yuan on AI infrastructure, less than what leading enterprises spend in a single month. Supply could not meet demand, leading to widespread API service interruptions. The subsequent sharp price increases and peak-valley time-of-use pricing strategies were essentially helpless measures under computing constraints, aiming to cut off some user demand through price thresholds.
However, this computing bottleneck is being rapidly filled. In the first seven months of this year, DeepSeek's AI infrastructure spending reached 11 billion yuan. Coupled with the two rounds of financing exceeding 100 billion yuan completed earlier this year, the computing supply gap is being closed at an extremely fast pace.
2.5 Kimi: Explosive Growth Driven by K3
Kimi ranks fifth with an annualized ARR of 5.8 billion yuan. Its explosive growth is mainly attributable to the K3 model: in June, its ARR was less than 2.5 billion yuan, but just two months after K3's launch, this figure doubled.
With the advent of K3, both its parameter scale and model performance have reached the global top tier. As the world's largest open-source model, K3 has elevated open-source model specifications to a new level with its 2.8 trillion parameters, and its model capabilities have won the favor of global developers, with comprehensive evaluations now on par with first-tier closed-source models like Claude. Additionally, the planned K4 model will further expand its parameter scale, expected to reach 5 to 6 trillion parameters.
However, the blind sprint toward ultra-large parameter models has made Kimi's expansion heavily constrained by high-end computing power. On one hand, its model training relies on NVIDIA's high-end graphics cards; on the other hand, model inference struggles to run smoothly on domestic GPUs. These two factors together have pushed up the API selling price.
Within just 48 hours of K3 going online, Kimi's computing resources nearly hit their limits, forcing it to suspend subscription services for new C-end users. It is expected that in the short term, its computing supply expansion will struggle to keep pace with peers like Zhipu. Unlike Kimi, Zhipu and other companies have constrained parameter counts to within one trillion, allowing inference to run on domestic GPUs and greatly relieving computing supply pressure. For Kimi, this disadvantage is extremely difficult to reverse in the short term.
2.6 MiniMax: Video Model Rising Strongly, Language Model Underperforming
MiniMax ranks sixth with an annualized ARR of approximately 4.8 billion yuan.
In May this year, MiniMax's ARR was only 3 billion yuan, but with the launch of its open-source video generation model H3, ARR jumped directly to 4.8 billion yuan in August (Note: The officially reported 800 million USD ARR is annualized based on the highest single-week data in August, not based on monthly data). The推理-optimized H3 Max generates a 5-second video in under 3 seconds, achieving for the first time ever that video generation speed exceeds playback speed, significantly broadening the commercialization vision for AI video live streaming.
Beyond improvements in video model capabilities, MiniMax's commercialization process has also seen positive changes. In 2025, MiniMax's commercialization strategy was heavily C-end oriented and light on B-end; by mid-year this time around, the situation had completely reversed, with B-end revenue share surging from 30% last year to 80%, and total enterprise customers surpassing 2 million.
MiniMax's weakness lies in its gross margin performance. Because the H3 model emphasizes high cost-effectiveness with relatively low pricing, its gross margin trails behind Seedance and Kling, two other video-track models.
Meanwhile, MiniMax's language model M3 failed to meet expectations. It consumed massive computing resources yet failed to deliver commensurate returns, dragging down the overall business's gross margin.
2.7 Kling: Deep Cultivation of Overseas Video, Valuation Already Surpassing MiniMax
Kling ranks seventh with an annualized ARR of 4 billion yuan, with its differentiated advantage mainly reflected in the overseas video track.
Over the past year, Kling's ARR achieved over 400% growth, primarily driven by overseas enterprise customers. Overseas revenue accounts for approximately 80% of the total, sourced from markets including North America, South Korea, and Southeast Asia. In terms of customer composition, Kling's core paying group consists of film and television production companies and director studios, with cumulative enterprise customers approaching 50,000.
In July this year, Kling was officially spun off from Kuaishou with a post-money valuation of 18 billion USD, surpassing the market cap of publicly listed MiniMax. However, the high valuation also means Kling faces greater ARR challenges. The company aims to reach 1 billion USD by the end of this year and plans to hit 2 billion USD by the end of 2027.
2.8 Tencent: It's Not About Starting Early but About Endurance
Tencent's annualized ARR is only 3.2 billion yuan, ranking eighth, consistent with its overall AI strategy pace and reflecting a calm and unhurried attitude. Among the leading tech giants, Tencent appears least concerned about Token business.
In contrast, Volcano has treated Token as a core growth engine, Alibaba even established the ATH Business Group specifically to advance Token business, while Tencent has made almost no notable moves in this area. This contrast reveals Tencent's complex mindset, mixing anxiety with self-comfort.
From the anxiety perspective, Ma Huateng's remarks were very candid: "A year ago I thought I had already gotten on board, only to later discover the boat was leaking. Now I feel like I've stood up on it but still can't sit steady." Boarding a leaky boat and feeling restless perfectly captures Tencent's current state of mind.
From the self-comfort angle, Tencent feels it is still too early for the final outcome of AI, and there is still room for adjustment. Tang Daosheng likened AI competition to a marathon: "Starting early is important, but only by enduring long can you seize the emerging opportunities," noting that they have "probably only run the first kilometer so far."
The marathon analogy itself is fine, but Tencent seems hasn't even reached the starting line in Token commercialization. Hunyuan Hy3 and Hy4 have been widely distributed through free API channels with low paid conversion rates. For a long-distance runner who hasn't truly entered the race, discussing whether they can endure is premature.
2.9 Baidu: Decline
Baidu's annualized ARR is 2.5 billion yuan, ranking last on the list, only one-twelfth of ByteDance's figure.
Baidu was the first domestic company to bring large models to the forefront, and Wenxin Yiyan was once the benchmark for Chinese large models. Yet after several years, circumstances have changed so much that no one would have predicted Baidu would end up in this position. Of course, Baidu still possesses its self-developed Kunlunxin chips, ranking behind Huawei and Alibaba in shipments.
The problem may lie in Li Yanhong's strategic judgment. In 2024, he publicly stated that open-source models were an "IQ tax." And at a recent earnings conference call, Li Yanhong had to admit that Baidu indeed made mistakes.
To date, Baidu's models rarely appear on various leaderboards, and when other vendors concentrated on releasing new models in August, Baidu remained still. It is hard to believe that Baidu can rejoin the first tier.
03
Industry Insights
3.1 China-U.S. AI Gap Narrowing, Development Paths Diverging
How large is the AI capability gap between China and the U.S.? This Token revenue ranking provides an objective observational window.
If we use Anthropic's ARR timeline as a reference, China's pace, rhythm, and current position in catching up with the U.S. can reveal considerable information. Anthropic's ARR rose from $1.4 billion in March 2025 to $65 billion in July 2026, and the current development stages of several major Chinese AI companies can basically be mapped to specific points along this growth curve.
Figure 2: China-U.S. Vendor ARR Catch-Up Curve
Among the top three vendors in this ranking, ByteDance's development level roughly corresponds to Anthropic's stage in August 2025, Alibaba to around mid-2025, and Zhipu to early 2025. In other words, China's leading AI vendors as a whole lag behind the U.S. by approximately one year. But China's catch-up momentum is very rapid, and the gap is narrowing continuously.
Two forces drive this acceleration trend. The first is continued investment in computing power to close the supply gap. Alibaba raised 80 billion yuan all invested in infrastructure, while DeepSeek is also accelerating its self-built computing power, with its 1GW computing center in Ulanqab expected to begin phased operations by the end of this year and achieve full delivery in the first half of next year.
The other driving force comes from application scenarios. China's short drama, Coding, office, and government sectors have generated massive Token consumption scenarios, enabling computing power investment to quickly realize commercial returns.
At the same time, differences in AI commercialization paths between China and the U.S. are increasingly apparent. U.S. models excel in top-tier capabilities, targeting users willing to pay high premiums for model performance. Chinese models win on open-source ecosystems and cost-effectiveness, serving a vast mid-tier user base. It remains difficult to determine which market holds greater potential, but the gap between China and the U.S. in AI is continuously narrowing.
3.2 Model Price Increases Are a Short-Term Phenomenon; Price Cuts Are the Long-Term Main Trend
Over the past six months, model API prices have generally risen, playing an important role in driving Token revenue growth. In the short term, models face supply shortages, and computing power has become the ceiling for Token revenue—as long as you have computing power, sales are not a concern.
In the first half of this year, DeepSeek saw the largest price increase. Due to restricted computing supply, DeepSeek had to raise prices to ensure service quality for high-value customers, thereby shedding low-value customers with weaker payment ability. In contrast, Zhipu, having built computing reserves in advance, achieved volume-and-price growth. In the AI infrastructure domain, Zhipu has successively invested in several infrastructure vendors including Silicon Flow, Jiliu Technology, and Wuwang Core, greatly alleviating and supplementing computing supply.
But in the long run, model price cuts are the main trend, especially for lightweight Flash models.
Normally, the vast majority of model call volume on the market is accounted for by Flash models. Taking OpenRouter's call rankings as an example, models such as DeepSeek V4 Flash, GPT-5.6 Luna, GLM 5.3 Flash, and Gemini 3.7 Flash have consistently held top 10 positions.
Flash models will become future price disruptors. For instance, GLM 5.3 Flash is priced at less than 10% of GLM 5.3, while GPT-5.6 Luna is only 5% of the flagship model Sol.
3.3 Revenue and Call Volume Are Showa Structural Divergence; Video Models and Language Models Take Different Paths First
Behind this ranking, one trend stands out clearly: video models and language models are diverging, with the former mainly contributing revenue and the latter mainly contributing call volume.
Video models have higher unit prices—for example, ByteDance topped the ranking with AI short dramas and video scenarios, and Kling's ARR quadrupled over the past year. In contrast, language models have lower unit prices but massive call volumes, contributing primarily call volume rather than actual revenue.
The commercialization paths of these two types of models provide important reference value for the development of other model types in the future.
For example, AI4S scientific research models are likely to follow the high-unit-price route, while embodied intelligence models are more likely to follow the low-unit-price, high-call-volume route. For vendors, these two paths require entirely different resources and strategies. Earning revenue relies on scenarios and pricing strategies, while gaining call volume depends on ecosystem building and low-price advantages.
04
Data Notes
The conclusions of this ranking are based on traceable data and methodology. The following will explain the ranking's construction process across three dimensions: ranking methodology, data sources, and research methods.
The revenue statistics for this ranking are strictly limited to Token billing revenue, including call services provided through model APIs—namely Token Plans and Coding Plans—as well as multi-model services provided through MaaS platforms. Model on-premises deployment projects, GPU cloud resource leasing and sales, and C-end subscriptions for AI applications are not included in the statistics.
This ranking uniformly takes August 2026 as the observation period, annualizing each vendor's monthly Token revenue data. Selecting current-month data for annualization aims to reflect the latest revenue status of each vendor and avoid the lag effect caused by historical cumulative data. Currency conversion uses a uniform exchange rate of 1 USD = 7.1 RMB. The ranking is sorted from highest to lowest based on annualized ARR data.
4.2 Data Sources
The data for this ranking comes from four types of sources: officially disclosed audited financial reports, management earnings call disclosures, authoritative industry sources, and primary research by AIFenxi.
Data compilation adheres to the principle of cross-verification. For the same vendor's revenue data, efforts are made to cross-verify across multiple sources. The purpose of cross-verification is to ensure that every number on the list can withstand public verification.
4.3 Research Methodology
The research for this ranking follows a three-stage methodology: first building a database, then verification, and finally assessment.
The database-building phase is primarily desktop research, collecting publicly available financial reports, announcements, industry research reports, and authoritative media coverage to identify the main Token revenue vendors in the market and determine the scope of listed companies.
Using the officially published information from listed companies as the baseline, each vendor's revenue data is verified one by one, and the specific nature and statistical scope of their disclosures are clarified.
In the assessment phase, in-depth interviews and data research were comprehensively adopted. In-depth interviews focused on industry research, with interviewees spanning the upstream and downstream of the Token industry chain, aiming to calibrate overall industry judgments. Data research mainly targeted disputed figures for targeted verification. By triangulating desktop research, in-depth interviews, and data research, a multi-dimensional verification system was constructed, ensuring that every figure is supported by multiple sources of evidence.
Upcoming Vendor Features
/ Reading Next /
Source: https://www.163.com/dy/article/L6CCB9LK0531W0RG.html For copyright issues, please contact: 13951431913#139.com or submit comments on the right.