[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-china-ai-token-factory-taas-shift-2026":3,"news-related-bb1d01e1-c61f-4428-b91a-41000a26bec5":41},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"bb1d01e1-c61f-4428-b91a-41000a26bec5","从「堆硬件」到「卖Token」:10余家上市公司押注Token工厂,算力行业 TaaS 模式浮出水面","据上海证券报 7 月 30 日报道,行云科技、润建股份、弘信电子、超讯科技、南威软件等 10 余家上市公司近期密集宣布探索「Token工厂」业务,核心是将算力计费模式从按硬件\u002F集群签约改为按 Token 交付与 Token 收入分成(TaaS)。行云科技已与某长文本大模型客户签订补充协议,256 个算力单元、5 年合同总额从 10 亿元提至 30.53 亿元,涨幅 201.14%,并新增 Token 收入固定服务费,行业从「卖资源」向「卖服务」跃迁的信号明确。","## 从「堆硬件」到「卖Token」:算力行业正在换一种方式算账\n\n如果过去三年的算力故事,主线是「谁手里有 GPU」,那么 2026 年下半年这条主线被改写了。**「Token 工厂」这个原本只在大模型公司内部出现的词,正在被算力服务商拿过去当招牌。**\n\n7 月 29 日晚,行云科技发布公告,旗下深圳行云与某深耕长上下文、自有大模型底座的头部客户(VB 客户)签订算力服务补充协议:算力服务单元数从 128 个翻倍到 256 个,5 年合同总金额由约 10 亿元调整至 **30.53 亿元,涨幅 201.14%**;同时,在原有固定服务费之外,新增了一项**「按 Token 收入相关的固定服务费」**——而且不管客户实际 Token 业务收入涨跌,这一项服务费都不调整。\n\n这不是个例。上海证券报梳理发现,2026 年以来,已有润建股份、弘信电子、行云科技、超讯科技、南威软件等 10 多家上市公司宣布探索「Token 工厂」业务。行情不是「硬件涨价」能概括的,而是**整个计费逻辑在切换**。\n\n## 同一份合同,出现了「Token 分成」这种新条目\n\n行云科技这份补充协议值得拆开看。\n\n原来合同是经典的「卖资源」:128 个算力单元 × 单价 × 5 年 = 10 亿元。补充协议变成了三层:\n\n- 第一层,固定单元费——每单元每月的算力服务费相比原合同**涨 21.21%**;\n- 第二层,扩容单元费——剩余 120 个单元 + 新增 128 个单元,每单元每月固定服务费相比原合同**涨 36.36%**;\n- 第三层,Token 收入固定服务费——只要客户继续按月验收这些算力单元,客户就要付这笔钱,**与实际 Token 收入不挂钩**。\n\n看似不挂钩的设计,实际上是把「客户用我的算力到底赚了多少」直接写进了收费模型里。固定服务费的形式给算力方吃下「客户做不出 Token」的不确定性,但「按 Token 收入相关」这个定义又锁定了未来若客户业务跑得更好时的**期权价值**。\n\n行情人士对此的解读很直白:**TaaS(Token-as-a-Service)模式最核心之处在于算力企业与大模型企业进行 Token 分成,大模型企业的 Token 溢价通过分成模式带动「Token 工厂」的收入**。一句话,卖资源是一锤子买卖,卖 Token 则是把算力方的利益跟模型方的爆款能力绑在一起。\n\n## 「Token 工厂」三件套:调度、推理优化、华为\u002F国产算力底座\n\n「Token 工厂」不是噱头,是技术栈的总和。\n\n行云科技董事长张文最近给行业分了代:1.0 是「卖设备」、2.0 是「卖资源」、**3.0 是「系统综合能力决定胜负」**。落到具体公司:\n\n- **润建股份**:推出「星算云池」,联合互联网大厂生态合作伙伴,五象云谷智算中心将升级为「Token 工厂」;\n- **超讯科技**:5 月 13 日与广西南一智能签约,联合共建「Token 工厂」,涵盖算力租赁、调度、套餐三类服务;\n- **弘信电子**:以华为超节点算力集群为首期基础设施,在无锡建一座「Token 工厂」;\n- **南威软件**:北京七星园智算中心定位为大型「Token 工厂」,拉动 Token 经济。\n\n从拼凑的描述可以归纳出共同点:**底层是国产\u002F非美系算力集群,中层是算力调度系统,上层是按 Token 计量和分成的服务接口**。MaaS(模型即服务)时代,这三层都是模型公司自己扛;TaaS 时代,调度与计费被算力方接手,模型公司只管把 Token 吐出来。\n\n## 为什么是现在:Token 消耗千倍增长,MaaS 让位 TaaS\n\n中国工程院院士郑纬民在 4 月的论坛上给过一个关键数据:**我国 Token 消耗两年实现千倍级增长**。这个增速让两件事同时发生:\n\n1. 大模型公司突然需要长期、稳定的算力合同,而不是临时找卡;\n2. 算力公司发现,只要把「Token 出货量」当作服务等级协议(SLA)的核心 KPI,议价权会比单纯按卡时高得多。\n\n这正是 TaaS 模式在 2026 年下半年集中爆发的产业基础——**Token 变成了一种可以被打包、调度、分成的「电力单位」,算力公司开始变成「Token 发电厂」**。\n\n一位算力产业链人士的判断更直接:现在很多高端算力服务合作出现交付难,核心原因是多数服务方拿不到足够的高性能算力服务器;谁能拿到,谁就拥有提价和签长约的话语权,下游大厂为了保障供给也愿意接受提价。这是一段「硬件供给紧张 + Token 需求暴涨」叠加的窗口期,TaaS 借着这个窗口把合同结构改写。\n\n## 算力行业的「Token 工厂」,能撑多久?\n\n「Token 工厂」模式的风险也很清晰。\n\n**第一,锁定的不是客户,而是客户的爆款能力。** 行云科技补充协议里,虽然 Token 收入相关服务费「不因实际收入增减而调整」,看似对算力方有利,但若客户业务跑不出来,长期续约和扩单的概率会下降,5 年 30.53 亿的合同价值高度依赖 VB 客户的长文本模型商业化能力。\n\n**第二,头部模型公司未必愿意把 Token 收入分给算力方。** 真正的大型基础模型公司(自研底座 + 自有渠道)在 Token 议价上有足够强的能力,它们更可能选择自建\u002F绑定单一算力供应商,而不是按 Token 收入与算力方分成。「Token 工厂」目前主要客户是长文本、超大上下文这类**专项模型公司**,而不是全行业通用大模型。\n\n**第三,Token 计量和定价缺乏标准。** 不同模型的 Token 长度、计费策略、缓存命中率差异极大,「每瓦 Token 生产效率」作为竞争指标听起来酷,但跨模型、跨集群比较没有统一口径。行情未来一定会出现「Token 单价」基准之争。\n\n## 所以呢:TaaS 不是新概念,但它正在被写进合同\n\n把视野拉高一点,「Token 工厂」的本质是把**算力从资本品(算力服务器)转成了消费品(API 调用量)**。\n\n过去 IDC、CDN 行业也走过类似的路径——从卖机柜、卖带宽,过渡到按 QPS、按流量计费,催生了 Cloudflare、Akamai 这类以「用量」为核心的边缘服务商。今天算力行业在 LLM 时代重走这条路:从「我有 H100 \u002F 昇腾」,过渡到「我每分钟能稳定吐多少 Token」。\n\n从行云科技、润建股份、超讯科技、弘信电子、南威软件这一波动作来看,国内算力服务方已经集体意识到:**未来三年的护城河不在 GPU 数量,而在调度系统和 Token 计量能力**。接下来 12-18 个月,真正能跑出来的「Token 工厂」,会跟现在还停留在「我有 N 个机柜」层级的算力公司,拉开一个量级的差距。\n\n合同里那条「Token 收入相关固定服务费」,就是这个时代切线的真实落点。","https:\u002F\u002Fpaper.cnstock.com\u002Fhtml\u002F2026-07\u002F31\u002Fcontent_2250605.htm","051b886c-80c9-4a83-841f-56bbb8fed645",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"fca9258a-9430-455a-b95d-b9fae5e373a8","ai-inference",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"id":19,"name":20,"slug":20,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",{"id":22,"name":23,"slug":23,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":25,"name":26,"slug":26,"description":14,"color":14},"045c011e-e2bb-45ce-bdd6-0c927f8a3b87","token-efficiency",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"0a02efdf-e42c-4644-a0a6-f3eb11ddf71f","en","From hardware to tokens: listed companies bet on the TaaS model","According to a Shanghai Securities News report on July 30, more than 10 listed companies — including Xingyun Tech, Runjian Co., Hongxin Electronics, Chaoxun Technology, and Nanwei Software — have recently announced plans to explore the 'Token Factory' business, shifting compute billing from hardware\u002Fcluster contracts to Token delivery and Token-revenue sharing (TaaS). Xingyun Tech has signed a supplementary agreement with a long-text LLM customer: 256 compute units, 5-year contract value raised from 1.0B to 3.05B yuan (a 201.14% increase), with a new fixed service fee tied to the customer's Token revenue. The signal that the industry is moving from 'selling resources' to 'selling services' is now explicit.","## From 'Stacking Hardware' to 'Selling Tokens': China's Compute Industry Is Rewriting Its Billing Logic\n\nIf the past three years of China's compute industry were defined by the question of **who had the GPUs**, the second half of 2026 is rewriting the question. **The phrase 'Token factory' — once confined to the internal vocabulary of model labs — has been co-opted by compute providers and is now plastered across their pitch decks.**\n\nOn the evening of July 29, Xingyun Tech announced that its wholly-owned subsidiary Shenzhen Xingyun had signed a supplementary compute service agreement with VB, a customer described as a leading long-context, foundation-model vendor. The deal expands the contract from 128 to **256 compute units** and lifts the 5-year total from approximately 1.0 billion yuan to **3.053 billion yuan — a 201.14% increase**. Crucially, the supplement also introduces a new line item: **a fixed service fee tied to Token revenue**, payable regardless of whether the customer's actual Token business grows or shrinks.\n\nThis is not an isolated case. Shanghai Securities News reports that since the start of 2026, more than 10 listed companies — including Runjian, Hongxin Electronics, Xingyun Tech, Chaoxun Technology, and Nanwei Software — have announced plans to explore the 'Token Factory' model. The trend cannot be summarized as 'hardware prices are rising'. **The entire billing logic is switching.**\n\n## Inside the Xingyun Deal: A New 'Token Sharing' Line Item\n\nThe supplementary agreement is worth breaking down.\n\nThe original contract was a textbook **'sell the resource'** deal: 128 compute units × unit price × 5 years = 1.0 billion yuan. The supplement restructures the deal into three tiers:\n\n- **Tier 1 — base fixed fee per unit**: monthly unit fee rises **21.21%** vs. the original contract;\n- **Tier 2 — expanded unit fee**: covering the remaining 120 units plus 128 newly added units, with a **36.36%** monthly increase;\n- **Tier 3 — Token-revenue fixed service fee**: payable monthly as long as the customer continues to accept delivery of these compute units, **decoupled from actual Token revenue**.\n\nOn the surface, the decoupling protects Xingyun from the customer's commercial risk. But the **definition itself** — 'a fee related to the customer's Token revenue' — embeds an option value: if VB's long-text model becomes a hit, Xingyun sits on the upside.\n\nIndustry observers read it bluntly: **the essence of TaaS (Token-as-a-Service) is that compute companies and model companies share in Token upside — the model company's Token margin flows through to the 'Token factory'**. Where the old 'sell-the-resource' model was a one-shot transaction, TaaS aligns the compute provider's economics with the model provider's product-market fit.\n\n## Three Layers of a 'Token Factory': Scheduling, Inference Optimization, and Domestic Compute\n\n'Token factory' is not marketing fluff — it is a stack.\n\nZhang Wen, CEO of Xingyun Tech, recently framed the industry's evolution in three eras: 1.0 was 'selling equipment', 2.0 was 'selling resources', and **3.0 is 'system-integration capabilities decide the winner'**. Across the recent announcements, a common architecture emerges:\n\n- **Runjian Co.**: launched 'Xingsuan Cloud Pool' (星算云池), partnering with major internet platforms; its Wuxiang Yungu intelligent computing center will be upgraded into a 'Token factory';\n- **Chaoxun Technology**: on May 13 signed with Guangxi Nanyi Intelligent to co-build a 'Token factory' covering compute leasing, scheduling, and packaged compute products;\n- **Hongxin Electronics**: building a 'Token factory' in Wuxi, anchored on Huawei super-node compute clusters as first-phase infrastructure;\n- **Nanwei Software**: its Qixingyuan intelligent computing center in Beijing is positioned as a large-scale 'Token factory', intended to anchor a local Token economy.\n\nStripped down, the shared blueprint is: **a domestic\u002Fnon-US compute substrate at the bottom, a scheduling layer in the middle, and Token-metered, revenue-shared service APIs on top**. In the MaaS era all three layers sat inside the model company. In the TaaS era, scheduling and metering move to the compute side; the model company just keeps emitting Tokens.\n\n## Why Now: 1000x Token Growth in Two Years, MaaS Gives Way to TaaS\n\nAt a forum in April, Chinese Academy of Engineering academician Zheng Weimin offered a key data point: **China's Token consumption has grown by three orders of magnitude in two years**. That trajectory produces two simultaneous effects:\n\n1. Model companies suddenly need **long-term, stable compute contracts** — not last-minute GPU hunting;\n2. Compute companies discover that **the moment 'Token throughput' becomes a Service Level Agreement (SLA)**, their pricing power exceeds anything a pure per-GPU-hour contract could deliver.\n\nThis is the industry substrate for the TaaS wave in H2 2026 — **Token has become a 'unit of electricity' that can be packaged, scheduled, and revenue-shared; compute providers are becoming 'Token power plants'**.\n\nOne compute-industry source put it more bluntly: many high-end compute service engagements now face delivery shortfalls because most providers cannot secure enough high-performance compute servers. Whoever can secure the hardware commands the pricing power and the long-term contract; downstream large-model players are willing to accept the price hikes to lock in supply. **This is a window where hardware scarcity stacks on top of Token demand explosion** — and TaaS uses that window to rewrite contract structures.\n\n## Will the 'Token Factory' Last?\n\nThe risks of the TaaS model are also clear.\n\n**First, you are not locking in a customer — you are locking in their hit-making ability.** Xingyun's supplement decouples the Token-revenue fee from actual revenue, which looks protective for the compute side, but if VB's long-text model fails to commercialize at scale, renewal and expansion probabilities drop. The 3.05-billion-yuan, 5-year headline number leans heavily on VB's product-market fit.\n\n**Second, top-tier foundation model companies may refuse to share Token revenue with compute providers.** The leading general-purpose model labs (with proprietary base models and direct distribution channels) have strong Token pricing leverage. They are more likely to self-build or single-source a compute partner than to revenue-share with a 'Token factory'. Today's 'Token factory' customers are mostly **specialized long-context \u002F large-context model companies**, not the general-purpose frontier.\n\n**Third, Token metering and pricing have no industry standard.** Different models have wildly different Token lengths, billing rules, and cache hit rates. 'Tokens per watt' is a slick competitive metric, but there is no cross-model, cross-cluster benchmark for it. The industry will inevitably fight over a 'Token unit price' reference standard in the coming year.\n\n## So What: TaaS Is Not a New Concept, but It Is Now Written Into Contracts\n\nStep back, and the 'Token factory' is essentially **compute moving from a capital good (servers) to a consumer good (API calls)**.\n\nThe IDC and CDN industries walked a similar path — from selling rack space and bandwidth, to billing by QPS and traffic, which spawned Cloudflare, Akamai, and the entire edge-as-a-service category. **China's compute industry is now replaying that path in the LLM era: from 'I have H100s \u002F Ascend', to 'I can reliably emit N Tokens per minute'.**\n\nJudging from the moves at Xingyun, Runjian, Chaoxun, Hongxin, and Nanwei, **China's compute providers have collectively realized that the moat for the next three years will not be GPU count, but scheduling systems and Token metering capability**. Over the next 12-18 months, the 'Token factories' that actually run at scale will pull an order-of-magnitude lead over compute companies still operating at the 'I have N cabinets' level.\n\nThe single line item — 'a fixed service fee related to the customer's Token revenue' — in Xingyun's contract is the real-world signature of where the era turns.","china-ai-token-factory-taas-shift-2026","2026-07-31T04:00:00Z","2026-07-31T02:04:01.223207Z","2026-07-31T02:04:01.223215Z",true,"agent",227,{"items":42},[43,48,53,58,63,68],{"id":44,"title":45,"news_slug":46,"published_at":47},"fb1cbe25-8b85-41ec-b619-9a27b405ec34","AMD 收购 Taalas:把 AI 模型权重「刻进硅片」的推理新打法","amd-acquires-taalas-inference-chip","2026-08-19T01:00:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"45375854-7739-4dd1-bc6a-30db4474652a","Taalas HC2:把单片参数拉到 200 亿,「模型刻进硅片」的第二章","taalas-hc2-20b-mxfp4-50-chips-1t-amd","2026-08-19T00:00:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"dfdc3216-52aa-4a78-9bf5-859affc37d17","AMD 收下 Taalas：把模型权重刻进芯片，推理的内存墙还剩多少？","amd-acquires-taalas-msic-etched-weights","2026-08-11T02:00:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"c07c67b6-6a48-4780-88bd-bc46b628c546","AMD 吃下 Taalas:把模型权重永久刻进芯片的\"硬推理\"赌局","amd-taalas-hardwired-inference-aug-2026","2026-08-08T12:00:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"2434bbc6-4fda-4750-a02d-dd3ca1fe8933","AMD 收购 Taalas:把模型权重刻进芯片,押注推理硬件的\"硬核\"路线","amd-acquires-taalas-hardcore-inference-silicon","2026-08-08T04:00:00+00:00",{"id":69,"title":70,"news_slug":71,"published_at":72},"bce0fe8f-14de-4ffc-8c22-2d798e711e73","Kimi K3 上线 48 小时打满集群:开源旗舰正在把推理算力拖进新一轮\"卖方周期\"","kimi-k3-48h-saturate-chinese-compute-supernode","2026-08-02T06:04:11+00:00"]