新智元报道
10月21日，柏林GTC大会，黄仁勋又把AI的未来讲出了新高度。
一身标志性黑皮衣的老黄站在舞台上宣布：
未来，每个人都会用AI，每家公司都会有AI赋能。
他还抛出了一个颇有意思的比喻：AI是一块「五层蛋糕」——能源、芯片、AI工厂、模型、应用，层层叠加。
在他眼里，AI已经不是软件，而是一种全新的制造业：AI工厂源源不断地生产Token，再把Token变成真正能创造价值的数字智能。
所以，企业不能只做AI的消费者，还得成为AI的生产者。
怎么生产？老黄给了一条路径：从开放模型出发，用自己的专有数据训练和优化，让AI真正掌握你的行业知识。
而且不同的行业，会由不同模态的模型来驱动。语言模型管文档，世界模型管机器人，自动驾驶、气象、物理仿真各有各的模型，没有谁能包打天下。
他还专门给智能体加了一层「知识底座」：把企业数据库里的结构化数据和文档里的非结构化数据打通，喂给智能体当上下文。
在他看来，让智能体变强的不是模型更大，而是它能读懂企业自己的知识。
现场，四只在仿真里练出来的小鸭机器人登台，坐下、起立、打盹。
这几件事放在一起，一个巨大的产业趋势已经浮现：
未来的AI，绝不会只属于几家手握超级大模型的巨头。千千万万家普通企业，也将拥有为自己业务服务的专属智能。
但问题来了。
一家工厂想训自己的质检模型，一家医院想做自己的疾病风险模型，一家能源公司想让AI替它管成千上万台设备。
这些模型拿什么训？光靠自家数据够不够？缺行业数据怎么办？训出来怎么本地部署？
如果每一家企业都要有自己的AI，谁来给它们提供源源不断的专业数据和智能资源？
就在这个拐点上，迅策科技（3317.HK）推出了TokenRouters。
它的定位是：企业AI的核心平台，以及AI实时数据及专用Token的一站式交换、匹配与服务平台，围绕优化企业模型与企业小模型落地。
说直白点，迅策想做的，是为未来成千上万个企业专属AI，搭一座能交换数据、匹配模型、流通智能资源的「超级市场」。
这背后，藏着一门更大的生意。
先讲一个迅策自己做的案子。
国内一家头部PCB厂，由于AI服务器订单暴涨，板子的层数越来越多、孔越钻越小、工艺越来越细，质检压力直接爆表。
以前是人工加传统AOI，抽检或终检。缺陷常常流到后道工序，甚至出货之后才暴露，返工、报废、客诉一个不少，异常品还一直占着设备和产能。
迅策的做法是：把这条产线从头到尾切成五六段，每一段、每一个工作岛单独做数据采集和治理，每秒生成十几万条数据。然后，给每个工作岛训一个自己的小模型，专门盯背钻、激光钻孔、电镀填孔、树脂塞孔这些关键工序。
结果：高速产线100%全检，出货良率稳定在99%以上，检测效率提升3倍以上。
而且这套方案能复制到其他产线、其他基地，成了企业接高端订单、出海的「质量通行证」。
老黄在柏林讲的物理AI、企业自己的模型，落到一条中国产线上，就是这么回事：没有什么万亿参数的大家伙，就是几个拿自家数据喂出来的小模型，盯着钻孔，比人眼准，比老设备快，还省钱。
但问题也在这儿。
这家PCB厂训小模型的「粮」，是它自己的产线数据。
它要想做得更好，缺的就不只是自己的数据了，而是整个行业的数据，是它自己没办法拥有的那部分数据。
A公司有PCB的数据，B公司有光通讯的数据，C公司有AI服务器整机的数据。
如果这些数据能合规地流通、交换，大家的小模型都能更聪明一截。
可数据这东西有个要命的特点：你给我看了，我就有了。谁敢随便交换？
答案是Token。
但这里要分清：通用Token像一个饱读诗书的大学生，啥都懂一点，能聊天、能写文章，解决的是语言理解、常识问答和内容生成问题，核心是「懂语言」；专用Token则由行业私有数据脱敏、精炼而成，每个Token都带着特定场景的业务逻辑、数据结构和合规边界，核心是「懂业务」。
TokenRouters不是通用Token的简单聚合器，而是以专用Token为核心交换资产的企业AI核心平台。
把原始数据脱敏、精炼成携带业务逻辑但不暴露原始数据的专用Token——Token可以交换，原始数据不出门。
这，就是TokenRouters要干的事。
它是给AI用的「数据集级交换」——让数据价值可交换，让智能资产触手可得。
拆开看，左边站着需要数据、AI模型、AI应用的企业，右边站着提供这些资源的合作方。
中间这个平台干三件事：让你找得到、比得清、买得对。
支撑这套逻辑的，是五大模块。
数据集，解决「用什么数据」的问题。找数据难，信息散，买回来能不能用更是没底。
TokenRouters里每份数据集都带三样东西：它是什么（名称、标签、摘要，一眼看出值不值得深挖）、它有多少（类型、规模、覆盖范围、来源、更新频率）、它怎么用（在分析、预测、开发里各管什么）。
AI模型，解决「用什么模型」的问题。模型一周出好几个，通用模型到了垂直场景总差一口气，以前只能一个个试，周期长，Token烧得快。
TokenRouters按行业、场景、类型把模型归好类，每个都写清楚能干什么、支持什么模态、适合什么场景，让「模型能干什么」和「业务要什么」对得上。
AI应用，解决「用什么工具」的问题。通用AI产品一大把，能嵌进业务流的垂直应用没几个。TokenRouters按行业、产品类型、业务环节分门别类上架，不光讲「这是什么」，更讲「怎么用、用在哪一步」。
场景库，最接地气的一个。企业最常问的是：「我们到底该在哪个环节上AI？」场景库从具体业务任务出发，拆痛点、算收益、配截图和架构图，把一句模糊的「我要做AI」翻译成一张清楚的施工图。先找准痛点，再谈选型。
智能客服，全程兜底。自然语言问答、常见问题引导、历史会话管理，页面浮窗和独立客服页双入口，搞不定的一键接人工。
五个模块串起来，是一条完整的闭环：找准场景→发现资源→评估适配→借鉴验证→全程服务。
TokenRouters是在AI资源的供需之间架了一座桥，而且是一座有路标、有交警、有服务站的桥。这座桥真正服务的，不只是「找到资源」，更是在真实场景参照中持续优化企业模型，让企业小模型从选型、适配到部署少走弯路。
TokenRouters：从场景到服务的完整闭环
回到老黄那块蛋糕。
英伟达从能源开始，一层一层往上垒。
迅策也是一层一层垒的，只不过，它垒的是数据这条线。
最底层是AI数据治理。
迅策从资产管理行业起步，十年时间把「脏活累活」干扎实，把企业内外部分散、异构、非标的原始数据，采集、清洗、标注、建模，变成能用的数据资产。
这是基座。
往上一层是TokenOS。今年5月发布，迅策把治理好的数据和具体业务场景结合，精炼成标准化、可计量、可定价的场景Token和专业模型参数。
数据从「沉睡的资源」变成「流通的资产」。
再往上是TokenCloud。9月推出，给Token生成和模型推理做算力适配：硬件选型、资源配置、模型优化与蒸馏，让模型用合适的成本落地。
然后就是TokenRouters，这一层干的是交换：把不同企业、不同行业、不同数据源之间的Token流通壁垒打通，让高质量场景Token在安全、合规、可审计的前提下跨主体交换。
最顶上一层，是企业小模型。让每一家企业都拥有自己的、可私有化部署、参数更小、推理更便宜的领域小模型。
这几层，对应着迅策的「四阶跃迁」：1.0数据治理、2.0 Token化、3.0 Token交换、4.0企业小模型。
逻辑链条极其清楚：先把数据治好，再把数据变成Token，然后让Token流通起来，最后用Token喂出企业自己的小模型。
NVIDIA研究团队2025年的论文中说道：对企业AI的专门任务来说，小模型已经够强，甚至是更优选择；跑一个70亿参数的小模型，比跑700亿到1750亿参数的大模型便宜10到30倍。
Gartner的预测更狠：到2028年，超过一半的企业级生成式AI模型将是领域专用的。
而这种专用的小模型，迅策已经在资管、能源、健康医疗等多个行业跑通了。核心壁垒都是同一件事：非公开的一手行业数据。
这也是为什么TokenRouters对迅策来说不只是一个新产品。
它意味着迅策从「AI数据基础设施提供商」变成「Token经济的基础设施运营商」。不只是修路的，而是要当收费站、当交通枢纽、当整个网络的调度中心。
商业模式也跟着改变：从过去按单客户Token调用收费，变成平台化的Token路由费、撮合服务费、合规流通费和数据资产化增值服务。
增长空间从「一条管道」变成「一张网络」。
2026年上半年，迅策营收9.67亿元，同比增长389%，成立十年来第一次在上半年盈利。
Token模式，已经开始兑现。
把老黄在柏林讲的四件事、迅策的四层蛋糕、几个行业案子放在一起看，企业AI的下一站，其实已经很清楚了。
第一，企业AI落地离不开数据治理，尤其是垂类的、高质量的、专用的数据和数据集。模型可以开源，算力可以租，唯独这个东西，只能自己攒、自己治、自己换。
第二，每一个行业、每一家公司，都会有自己的小模型。不是一个什么都懂的通才，而是一个把一件事做到极致的专才。TokenRouters的价值，也不只是把资源摆上货架，而是围绕企业真实场景持续优化企业模型，让企业AI从「能用」走向「好用、专用」。
第三，训小模型离不开数据集的交换流通。就像货币一样，每个人都可以用自己有的，去换自己需要的。但这里交换的不是通用Token，而是携带行业知识、业务规则和合规边界的专用Token。
10月20日国家数据集团在雄安挂牌，要干的也是「推动数据跨行业、跨区域流动」这件事。国家层面在搭框架，产业层面需要有人把交换真正跑起来。
第四，本地化、数据安全合规、私域训练部署，是确定的趋势。数据不出域，模型跑在本地，智能体在自己的知识库上干活。老黄在柏林说「知识和工具的所有权属于企业本身」，说的就是这个。
值得强调的是，AI网络安全与合规安全是迅策长期投入的底层能力。公司在数据治理、Token化及企业服务过程中，持续覆盖数据分级分类、访问控制、隐私保护、模型与Token流转审计、私有化部署安全等环节。TokenRouters延续这一安全合规思路，将网络安全、数据防泄漏、模型访问控制和全流程审计嵌入专用Token交换与匹配过程，使「可交换」建立在「可审计、可管控、可合规」的基础之上。此前迅策已与深圳数据交易所签署战略合作协议，并与北京国际大数据交易所订立合作框架协议，也体现了其在合规流通与数据要素安全使用方面的持续布局。
这四条线，交汇点就是TokenRouters。
在AI从「能用」走向「好用」的拐点上，一个以企业AI核心平台为定位、以优化企业模型为落点、以安全合规为底线，让数据价值可交换、让智能资产触手可得的平台，可能正是整个产业最缺的那块拼图。
编辑：所罗门
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
On 21 October, the GTC Conference in Berlin, Huang In-hoon, gave new heights to the future of AI.
On the stage, the old yellow station in the marked black leather declared:
In the future, everyone will use AI, every company will have AI.
He also threw an interesting metaphor: AI is a five-storey cake -- energy, chips, AI factories, models, applications, layers.
In his view, AI was no longer a software, but a brand-new manufacturing industry: the AI plant was constantly producing Token and turning Token into a digital intelligence that could really make value.
As a result, businesses cannot be only AI consumers, but also AI producers.
How? Wong gave us a path: from an open model, using his own proprietary data training and optimization to give AI real knowledge of your industry.
And different industries are driven by different models. Language model tube files, world model tube robots, automatic driving, weather, physical simulations have all their own models, and no one can wrap up the world.
He also added a “knowledge base” for AI agent: to feed AI agent as context by providing access to structured data and unstructured data in files in the enterprise database.
In his view, it was not the larger model that made AI agent stronger, but rather its ability to read the business itself.
At the scene, four duck robots only come on stage and sit, stand up and take a nap.
Together, a huge industrial trend has emerged:
The future of AI will never belong to just a few giants holding super-LLM. Millions of ordinary enterprises will also have the exclusive intelligence to serve their business.
But the problem came.
A factory wanted to train itself in a model of quality tests, a hospital wanted to model its own disease risk, and an energy company wanted AI to run thousands of equipment for it.
What are these models about? Is it enough for home-grown data alone? What about missing industry data? How do we deploy locally?
If each enterprise is to have its own AI, who will provide them with continuous professional data and intellectual resources?
At this point, Rapid Technology (3317.HK) launched TokenRouters.
It is located on the core platform of Enterprise AI and on the all-in-one Exchange, Matching and Service Platform of A real-time data and dedicated Token, around optimizing business models and small business models.
To put it bluntly, what is quickly envisaged is a "supermarket" for thousands of future businesses to exchange data, match models, and move smart resources.
Behind this is a bigger business.
Let's start with a quick case.
One of the country ' s head PCB plants, due to the surge in orders from the AI server, is experiencing an increase in the number of layers of plates, smaller boreholes, more sophisticated processes and direct mass-checking pressure tables.
Used to be a manual AOI, a check or a final check. Deficiencies are often found in back-track processes, even after delivery, return to work, end-of-life, numerous complaints from customers, and unusual goods continue to occupy equipment and capacity.
The prompt approach is to cut the line in five or six sections from the beginning to the end, to collect and manage data separately for each segment and for each working island, generating 100,000 data per second. Each working island is then trained in its own small model, which targets critical processes such as back drills, laser drilling, electroplating fillings, and resin plugs.
As a result: 100 per cent of high-speed production lines are fully inspected, the yield rate has stabilized at more than 99 per cent and detection efficiency has more than tripled.
Moreover, the programme can be replicated in other production lines and other bases as a high-end corporate order, quality pass to sea.
The physical AI and the company's own model in Berlin fell on a Chinese line, and that's what happened: the big guys with no trillions of parameters, just a few small models fed with their own data, staring at holes, better than people's eyes, faster than old equipment, saves money.
But that's the problem.
The PCB model "food" is its own production line.
In order for it to do better, what is missing is not just its own data, but that of the entire industry, that part of it that it cannot own.
Company A has PCB data, Company B has light communications data, and Company C has AI server integrity data.
If these data were to flow and exchange in a consistent manner, one could be smarter than the other.
But the data thing has a fatal feature: you showed it to me and I got it. Who dares to trade?
The answer is Token.
But let's be clear: General Token, like a college student who reads poetry, understands everything, talks, writes, solves language understanding, common sense questions and content generation issues, with the core being "linguistic" at its core; and Special Token, with private industry data dissension and refinement, with each Token carrying business logic, data structure and compliance boundaries with a specific scene.
TokenRouters are not a simple polymer for General Token, but a core corporate AI platform with Token as its core exchange asset.
Token-Token, which is specialized in de-sensitizing and refining raw data to carry business logic without exposing raw data, can be exchanged and raw data will not go out.
That's what Token Routers do.
It's a "data level exchange" for AI -- to make data values interchangeable and smart assets accessible.
Take a look. On the left, there are companies that need data, AI models, AI applications, and on the right, there are partners that provide these resources.
The platform in the middle does three things: you can find it, you can beat it, you can buy it right.
That logic is supported by five major modules.
Data sets, solving the problem of what kind of data. It's hard to find the data, the information is scattered, and it's even worse to buy it back.
Each data set in TokenRouters contains three items: what it is (name, label, summary, which is not worth digging at once), how it is (type, size, coverage, source, frequency of updates), how it is used (what it is in analysis, prediction, development).
AI model, solve the problem of what model. The model came out several times a week, and the generic model went to the vertical scene with a total difference of one breath, with a long cycle and Token burning fast.
TokenRouters classifies models by industry, scene, type, each with a clear picture of what it can do, what it supports, what it suits, what it can do, what it can do, and what it does.
AI application, solving the problem of what tools. There is a large number of generic AI products, with few vertical applications embedded in business streams. TokenRouters come up by industry, product type, business chain, not just "what is this" but "how and where."
The scenery library, one of the best. The most common question for companies is: "What are we supposed to do with AI?" The scenery is based on specific operational tasks, disabling points, calculating benefits, drawings and diagrams, and translating the vague phrase "I'm doing AI" into a clear construction map. Let's find some pain before we talk about options.
Smart customer service, all the way down. Natural language question-and-answer, frequently asked questions, history session management, double-entry of page floats and stand-alone guest pages, hard-to-hand.
The five modules are a complete closed loop: the search for an accurate scene to identify the suitability of the resource assessment for the validation of the entire service.
TokenRouters built a bridge between the supply of and demand for AI resources and a bridge with road signs, traffic police and service stations. The real service of the bridge is not only to "find the resources" but also to continuously optimize the business model in the context of the real scene, so that the small business model can go down the curve from selection to suitability to deployment.
TokenRouters: From scene to service full closed loop
Back to the yellow piece of cake.
Weeda begins with energy and goes up one level.
It's also a base, but it's a data line.
The bottom line is the governance of AI data.
Starting from the asset management industry, the "dirty work" has been solid for 10 years, transforming the scattered, isomeric, non-target raw data within and outside the enterprise into usable data assets.
It's the base.
Up top is TokenOS. In May of this year, publish was quick to combine well-governed data and specific business scenarios into standardized, measurable, priceable scene Token and professional model parameters.
The data has changed from "sleeping resources" to "current assets".
And up to TokenClaud. Launched in September, to calibrate Token generation and model reasoning: hardware selection, resource allocation, model optimization and distillation, allowing models to land at appropriate cost.
Then there is TokenRouters, which is a level of exchange: to break through the Token flow barriers between different businesses, industries, and different data sources, so that Token, a high-quality scene, can be exchanged across the board under conditions of security, compliance and auditability.
At the top, it's a small business model. Each enterprise will have its own small model of a sector that can be privatized and deployed, with smaller parameters and cheaper reasoning.
These layers correspond to the rapid "four steps": 1.0 data governance, 2.0 tokenization, 3.0 token exchange, 4.0 small business models.
The logical chain is perfectly clear: first fixes the data, then turns it into Token, then lets Token circulate and then feeds the company's own small model.
In a 2025 paper by the NVIDIA research team, it was stated that small models were strong enough, or even better choices, for business AI's special assignments; running a small model of 7 billion parameters was 10 to 30 times cheaper than running LLM, which ran between 70 billion and 175 billion parameters.
Gartner ' s predictions are even harsher: by 2028, more than half of the enterprise-level generation AI models will be field-specific.
This specific small model, however, is already operating in a variety of sectors, including finance, energy and health care. The core barriers are the same: non-public primary sector data.
And that's why TokenRouters is not just a new product for rapids.
It means moving quickly from an "AI data infrastructure provider" to an "infrastructure operator of the Token economy." It is not just a road repair, but a toll station, a transportation hub, a movement control centre for the entire network.
The business model has also changed: from a former single client, Token, to a platform-based Token route fee, a blended service fee, compliance traffic fees and data assetization value-added services.
The growth of space has changed from "a pipeline" to "a network".
In the first half of 2026, RMB 967 million was collected, an increase of 389 per cent over the previous year, and for the first time in the first half of the decade.
The Token model has begun to materialize.
It's clear that the next stop of the company AI is to put together four things that Wong said in Berlin, four layers of quick cake, several trade cases.
First, data governance is essential for enterprise AI to land, especially in the case of lean, high-quality, dedicated data and data sets. Models can be open-sourced, calculators can be rented, except for this thing, which can only be saved, treated and replaced.
Second, every industry and every company will have its own small model. It is not a generalist who understands everything, but a specialist who brings one thing to the extreme. The value of TokenRouters is not simply to put resources on shelf, but to continuously optimize business models around the real scene of a business, allowing business AI to move from being able to use to being “good, special”.
Third, small training models require the exchange of data sets. Like money, everyone can trade for what they need. But this exchange is not for General Token, but for Token, who carries industry knowledge, business rules and compliance boundaries.
On October 20, the National Data Group (NDG) put up a card in Yu-Anh, also to promote data flows across industries and regions. There is a framework at the national level, and there is a real need for exchange at the industrial level.
Fourth, localization, data security compliance, and private-area training deployment are identified trends. Data are not available, models run locally, and AI agent works on its knowledge base. Old Wong said in Berlin, "The ownership of knowledge and tools belongs to the enterprise itself," that's all.
It is worth emphasizing that the security of the AI network and compliance security are at the bottom of a rapid and long-term commitment. The company continuously covers data classification, access control, privacy protection, model and Token flow audits, and privatization deployment security in data governance, tokenization and business services. TokenRouters continues this approach to security compliance by embedding network security, data protection from leakage, model access control and full-process audit into a dedicated Token exchange and matching process, so that "exchangeable" is based on "auditable, manageable, and compliant". This was preceded by the signing of a strategic cooperation agreement with Shenzhen Data Exchange and a framework agreement for cooperation with the Beijing International Big Data Exchange, which also reflects its continued layout in terms of compliance flows and the safe use of data elements.
These four lines, the intersection is TokenRouters.
An AI-based platform that is located at the core of Corporate AI, where business models are optimized, where security compliance is the bottom line, where data values can be exchanged, and smart assets can be reached, is probably the missing piece of puzzle for the whole industry.
Editor: Solomon
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.