- 目录
- 本版
- 往期
- 客户端
第A7版:经济周刊·焦点
- 数据“国家队”入场 算力购物车上线
- AI算力走进大众消费
- 高见
- 给汽车创新“上保险”
- • 第A1版:头版
- • 第A2版:要闻
- • 第A3版:跑出加速度 干出新业绩
- • 第A4版:综合
- • 第A5版:要闻
- • 第A6版:要闻
- • 第A7版:经济周刊·焦点
- • 第A8版:经济周刊·理财
- • 第A9版:湾区周刊
- • 第A10版:湾区周刊
- • 第A11版:湾区周刊·城事
- • 第A12版:湾区周刊·专题
AI算力走进大众消费
倪明、邓莉
本文字数:1898
当大模型从实验室走向市面,沿途开出大大小小的“商店”。今年“双11”,看不见的算力第一次作为商品进入消费者购物车,标志着AI服务正式从B端后台走向大众消费市面。而围绕Token,一条“生产智能—出售智能”的新产业链,已在模型厂商、聚合平台、算力工厂、运营商和大厂之间悄然成形。
文/广州日报全媒体记者倪明、邓莉
上架:算力首进“双11”购物车
往年“双11”,人们忙着囤米面、家电、服饰,购物车里全是看得见摸得着的实物。今年不一样了——除了QAIPC、AI眼镜、AI耳机等一批AI硬件新品集中亮相,看不见的算力也被放进了购物车下单,“囤Token”由此成为今年“双11”的新话题。
10月15日晚8点,天猫“双11”预售启动;天猫3C数码总经理承渊在开卖前表示:“从传统数码机型到AI硬件,再到云端算力服务,‘双11’大促正在成为科技新品最关键的孵化阵地。”这是AI商品化的标志性节点:不止硬件,模型服务也首次登上电商货架,阿里云百炼、智谱、MiniMax等厂商的Token服务,首次进入大促商品清单。
记者在天猫上搜索“Token”这一关键词发现,天猫已上线“AI空间站”充值中心,阿里云百炼、千问办公、智谱(HK2513)、MiniMax等多家大模型厂商入驻,按月卡、季卡、年卡给出不同的套餐。
算力零售化就此落地。承渊解释,把Token服务放到电商平台,就是要降低使用门槛。电商交易体系进来,是想把分散在各模型官网的注册、续订、比价整合成一个统一的消费入口,“希望把它变成像水电煤缴费一样简便的服务”。
流通:多元分销链条生长
新工艺普及,总会催生新的贸易角色。AI行业里,模型研发方、算力供给方、渠道分销方彼此分离,生产智能和售卖智能不再是同一个主体。围绕Token调用,一套复杂的流通链条已经成形。
云厂商是当下Token市面的主导者。这类企业既具备云计算基础设施,又有自研大模型和庞大的客户群体。
模型聚合平台则瞄准开发者痛点。开发者同时接入多家大模型,要处理多套账号、账单、接口,还要自行应对服务中断风险。OpenRouter把这些整合为统一服务,收取5.5%平台服务费。其联合创始人兼CEO Alex Atallah表示,“没有一个模型能最合适所有任务”,开发者需要一个能组织和管理多种模型的中立层。
最基础的是中小商家批零倒卖。一名经营小型中转站的受访者告诉记者,借助批量购买国产模型API,最低可拿到约七折的售价,再加价卖给客户,“这也会比客户直接购买API便宜”;也有中转站“囤到”显卡、搭建小机房,部署开源模型后对外出售调用服务。
Token工厂:资本热捧,盈利难题待解
比聚合平台更深一层的,是硅基流动、无问芯穹这类被称作“Token工厂”的企业。核心工作是完成大模型在异构AI芯片上的适配、推理优化与算力弹性调度,把零散算力打包成稳定可用的模型服务,也为企业给出私有化算力部署运营。
算力调度天然存在矛盾。硅基流动应对思路总监唐安波今年6月公开演讲时表示:按需要高峰准备算力,平时容易闲置;按低谷准备,高峰又接不住。推理优化和弹性调度正是为此而生。但这套模式短期很难盈利。
即便盈利承压,资本仍在加注。9月,硅基流动披露完成B+轮二期及C轮融资,2026年以来累计股权融资接近29亿元;10月12日,无问芯穹宣布完成23亿元C轮融资,成立以来累计融资43亿元。
无问芯穹投资方启明创投主管合伙人周志峰认为,推理加速和多元异构能力能帮芯片厂商、算力中心、模型公司和应用企业一起降本。但行业竞争持续压缩报价,工艺降本速度难追售价下行;客户是因售价补贴还是因不可替代性而留下,才决定“Token工厂”能走多远。
竞争升级:运营商与大厂纷纷入局
今年5月,中国电信、中国移动、中国联通三大运营商分别推出了Token(词元)套餐——有的套餐9.9元每月1000万Tokens,有的1元可得40万Tokens,甚至有办理龙虾云电脑直接送2500万Tokens。运营商还以Token作为计费基础,面向政企客户给出定制化AI应对思路。以上动作已为运营商的业绩带来实质性增量。
海外科技大厂也启动了“代销Token”模式。Google的AI开发平台Antigravity,在自家Gemini之外接入竞品Claude。它发布Antigravity时提出,希望成为智能体时代软件开发的“大本营”——模型只是工作的一部分,允许选Claude,可减少使用者为用某模型而更换工具的必要;即便任务由Claude完成,Google仍能取得订阅收入并让使用者留在自家环境。
大厂的竞争焦点,从来不是单次Token调用,而是争夺开发者和使用者的长期阵地。
- 放大
- 上一版
- 下一版
- 下载
本文来源:https://gzdaily.dayoo.com/pc/html/2026-10/23/content_872_916627.htm
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
As large language models move from laboratories to the market, various shops of all sizes have sprung up along the way. This year's Double 11 shopping festival saw invisible computing power enter consumers' shopping carts for the first time as a product, marking the official transition of AI services from B2B backends to the mainstream consumer market. Around the Token, a new supply chain — "produce intelligence, sell intelligence" — has quietly taken shape among model makers, aggregation platforms, computing power factories, telecom operators, and major tech companies. By / Ni Ming, Deng Li, Guangzhou Daily Computing Power Debuts in Double 11 Shopping Carts In previous Double 11 seasons, shoppers stocked up on rice, flour, home appliances, and clothing — only tangible goods in their carts. This year is different. Alongside a wave of new AI hardware like QAIPC, AI glasses, and AI earbuds, invisible computing power was also placed in shopping carts for purchase, making "stocking up on Tokens" the new topic of Double 11. On October 15 at 8 PM, Tmall's Double 11 presales began. Cheng Yuan, General Manager of Tmall's 3C Digital division, stated before the launch: "From traditional digital gadgets to AI hardware, and now to cloud computing services, Double 11 is becoming the most crucial incubation ground for tech newcomers." This is a landmark moment for AI commercialization: not just hardware, but model services have also appeared on e-commerce shelves for the first time. Token services from Alibaba Cloud's Bailian, Zhipu, MiniMax, and other vendors entered the promotional catalog. Searching for "Token" on Tmall, reporters found an "AI Space Station" recharge center with vendors including Alibaba Cloud Bailian, Qianwen Office, Zhipu (HK2513), and MiniMax, offering monthly, quarterly, and annual subscription plans. Retailization of computing power has arrived. Cheng Yuan explained that placing Token services on e-commerce platforms is meant to lower the barrier to entry. Integrating the e-commerce transaction system aims to consolidate fragmented registrations, renewals, and price comparisons across model vendor websites into one unified consumer gateway — "making it as simple as paying utility bills." Circulation: Diversified Distribution Chains Emerge New technology always spawns new trade roles. In the AI industry, model developers, computing power providers, and channel distributors are separated; producing intelligence and selling it are no longer the same entity. Around Token usage, a complex distribution chain has taken shape. Cloud providers are the current leaders in the Token market. These companies possess both cloud infrastructure and proprietary large models, along with massive customer bases. Model aggregation platforms target developer pain points. When developers integrate multiple large models, they must manage separate accounts, billing, and APIs, while independently handling service interruption risks. OpenRouter consolidates these into a unified service, charging a 5.5% platform fee. Co-founder and CEO Alex Atallah stated, "No single model is best for all tasks." Developers need a neutral layer to organize and manage multiple models. At the most basic level are small and medium resellers. One small relay-station operator told reporters that by purchasing domestic model APIs in bulk, they can obtain prices as low as 30% off, then markup and resell to clients — still cheaper than clients purchasing APIs directly. Some relay-station operators have even hoarded graphics cards, built small server rooms, deployed open-source models, and sold inference services. Token Factories: Hotly Sought by Capital, Profitability Remains Unsolved Deeper in the chain are companies like SiliconFlow and WuWenXinQiong, known as "Token factories." Their core work involves adapting large models to heterogeneous AI chips, optimizing inference, and managing elastic computing scheduling — bundling fragmented computing power into stable, usable model services, and providing enterprises with private computing deployment and operations. Computing scheduling inherently faces contradictions. Tang Anbo, Director of Solution Strategy at SiliconFlow, stated in a June public speech: preparing computing power for peak demand leads to idle capacity during normal periods; preparing for troughs means being unable to handle peaks. Inference optimization and elastic scheduling were born to address this. However, this model is unlikely to be profitable in the short term. Despite profitability pressures, capital continues to pour in. In September, SiliconFlow disclosed completing its B+ round Phase II and C round financing, accumulating nearly 2.9 billion yuan in equity financing since 2026. On October 12, WuWenXinQiong announced a 2.3 billion yuan C round, bringing its cumulative financing to 4.3 billion yuan since establishment. Zhou Zhifeng, Managing Partner at Qiming Venture Partners (investor in WuWenXinQiong), believes that inference acceleration and multi-heterogeneous capabilities can help chip manufacturers, computing centers, model companies, and application enterprises reduce costs together. However, industry competition continues to compress pricing, and the pace of process cost reduction struggles to keep up with falling prices. Whether customers remain due to price subsidies or irreplaceability will determine how far Token factories can go. Competition Escalates: Telecom Operators and Tech Giants Enter the Fray In May this year, China Telecom, China Mobile, and China Unicom each launched Token plans — some offering 10 million Tokens per month for 9.9 yuan, others 400,000 Tokens for 1 yuan, and some bundling 25 million Tokens free with lobster cloud computer subscriptions. Operators have also begun using Tokens as a billing basis to offer customized AI solutions for government and enterprise clients. These moves have already contributed substantial incremental revenue. Overseas tech giants have also launched "Token agency sales" models. Google's AI development platform Antigravity integrates competitor Claude alongside its own Gemini. Upon launching Antigravity, Google stated its vision of becoming the "headquarters" for software development in the agent era — models are just one part of the workflow. Allowing Claude selection reduces the need for users to switch tools for a specific model; even when tasks are completed by Claude, Google still earns subscription revenue and keeps users within its ecosystem. For tech giants, the competitive focus has never been on individual Token calls, but on securing long-term positions with developers and users.


Source: https://gzdaily.dayoo.com/pc/html/2026-10/23/content_872_916627.htm If you have any copyright concerns, please contact us at: 13951431913#139.com or submit a comment on the right side.

