更新时间：2026-08-26 来源：互联网编辑：文道纯王点击数： 70336次

中国梦想秀

深耕Token工厂赛道硅基流动以技术与流量构筑AI算力服务新生态_我的网站

放牛班的春天

一 | 当前，AI行业加速从模型训练走向产业落地，Token工厂赛道热度持续攀升，但行业普遍存在概念超前、落地滞后的痛点。作为赛道先行者，硅基流动在2024年9月就提出“Token工厂”构想，并依托公有云MaaS模式实现行业早期落地，完成从概念创新到规模化商业运营的跨越式发展。如今，随着公司递交港交所上市申请，其完善的技术体系、成熟的商业模式与多元生态布局，也为AI算力服务行业标准化、规模化发展提供了优质参考范本。　　　前瞻布局赛道，完成Token工厂从构想至规模化落地据了解，硅基流动提出Token工厂概念，源于对AI行业发展的精准预判。团队研判，开源模型性能持续迭代升级，将大幅拓宽AI应用边界，市场推理需求将远超训练需求，而大模型规模化部署的成本、技术门槛，正是AI基础设施企业的核心价值赛道。

Photo: VCG
The US government is encouraging American companies to launch their own cyberattacks against criminal hackers, a move that certain US media claimed would bring the US closer to the practices of some of its chief cyber adversaries, including China and Russia. A Chinese expert slammed the claim as factually unfounded, saying that the US move would seriously disrupt the global cyberspace order. Washington has long projected its own conduct onto other countries, inventing adversaries to justify the militarization of cyberspace, the weaponization of artificial intelligence (AI) and even a privateer-style model of cyber piracy, said the expert.
Under a national security memorandum that US President Donald Trump signed late Wednesday, select companies would work with the Justice and Homeland Security Departments to strike foreign cybercriminal groups with hacks under certain conditions. The attacks would allow both surveillance of the criminal networks and specific types of hacking operations that could lead to disruption, manipulation or destruction of information systems and networks, including virtual and physical infrastructure, according to a report from the New York Times.
A CNN report called the move "a major policy shift that experts say could come with legal risk for the companies." It also noted that "companies can use cyber tools to surveil foreign criminals and disrupt their networks under the 'control and oversight' of the new federal program."
The US move is an attempt, under the guise of combating cybercrime, to accelerate its plan to issue "cyber privateering licenses" to private entities. Its core objective is to exploit its advantages in the cybersecurity and intelligence sectors and rapidly turn AI-enhanced cyberattack capabilities into tools for securing strategic interests and undermining the security of other nations, Xiao Xinguang, chief software architect from Chinese cybersecurity company Antiy, told the Global Times on Friday.
Analysts noted that some US media outlets routinely drag China into the frame and slander it whenever Washington unveils relevant policies, all to manufacture pretexts to legitimize America's own policies.
The New York Times report claimed that "to some extent, the new order would align the US more with some of its chief cyberantagonists, including China and Russia, where spy agencies have long relied on contract hackers working in the private sector to further their national security missions."
Xiao noted that the US has long projected its own practices onto other countries, manufacturing imaginary adversaries as scapegoats and falsely attributing to China the operational models of its own intelligence agencies, contractor networks and revolving-door system. By applying double standards and leveling unfounded accusations, Washington creates pretexts for militarizing cyberspace, weaponizing AI and even developing a cyber-privateering model.
In reality, certain US corporations have long supported US intelligence agencies in intelligence collection and offensive cyber operations, and this has been an established practice for years, Xiao said.
"Now, the US is bringing these previously covert practices into the open. As the trend develops, more US cybersecurity firms and intelligence contractors could operate under so-called privateering authorizations, posing as non-state actors while carrying out intrusions, data theft, sabotage and other disruptive cyberattacks against countries around the world," said Xiao.
。基于此，公司以“用得上、用得好、用得起”为核心价值导向，打造公有云MaaS平台，通过自研推理引擎优化性能、提升算力利用率、屏蔽底层技术壁垒，实现API一键调用，打通效能与成本的闭环，在此基础上，进一步沉淀为可复制的Token工厂解决方案，助力各类算力中心快速转型，目前已落地互联网、能源、金融、电信等多个头部行业。2025年初成为硅基流动Token工厂规模化发展的关键转折点。DeepSeek系列模型开源引爆市场推理需求，行业需求迎来指数级爆发。公司率先在华为昇腾上完成DeepSeek R1、V3满血版商业化部署，成为落地世界级大模型商业化推理服务的平台，高效承接海量市场需求。市场的爆发式增长，充分验证了Token工厂模式的可行性，也锤炼了平台的极限承载能力，其注册用户从2024年底12.7万人飙升至2025年底920万人，实现跨越式增长。这一成果背后，是硅基流动三大前瞻性行业判断的持续落地：开源模型能力将逐步追平闭源模型、推理需求量级远超训练需求、国产芯片将从“可用”迈向“好用”。当前行业发展态势持续印证该判断，开源模型生态日趋成熟，AI Agent规模化落地带动Token消耗暴涨，国产算力供应链持续完善，为Token工厂赛道长期发展筑牢根基。　　　突破多层技术壁垒，重构算力中心转型路径从传统算力中心“裸金属租赁”到Token标准化服务交付，行业存在一条高壁垒技术链路，涵盖芯片适配、模型调优、弹性调度、计量计费等核心环节。

二 | 硅基流动突破五大核心技术壁垒，将复杂技术链路封装为“一键转型Token工厂”解决方案，有效降低行业转型门槛。在芯片适配层面，公司完成多款不同品牌、不同架构芯片的规模化适配，破解异构芯片兼容难题；通过推理引擎深度调优，以动态量化、多级缓存、算子优化等技术，提升单位算力产出与响应效率。

三 | 同时，平台实现170+主流模型持续适配迭代，搭建异构算力弹性调度系统，平衡潮汐式推理需求，并重构细粒度Token计量计费体系，完成从硬件售卖到服务化交付的核心蜕变。针对算力行业普遍存在的产能过剩、算力闲置痛点，硅基流动打造“技术+流量”双引擎，构建高效算力消纳商业闭环。供给侧，公司依托自研技术提升算力产出效率，最大化激活存量闲置算力资源；需求侧，平台坐拥超千万注册用户和1.3万家企业客户，截至2026年4月日均词元吞吐量达5785亿。多元均衡的客户结构形成稳定持续的算力负载，避免碎片化流量弊端。通过智能算力调度与收益分成机制，平台精准匹配供需资源，如同搭建贯通全网的“算力电网”，让各地Token工厂的算力产能持续变现，解决算力消纳难题。

四 | 　　　多元生态落地，构建算力消纳全新商业闭环依托成熟的技术与商业模式，硅基流动持续深化产业生态合作，与光谷爱计算、贵州移动、宇信科技达成差异化战略合作，构建多维落地范式。与光谷爱计算的合作聚焦算力联合运营，赋能传统智算中心低成本转型Token工厂；与贵州移动携手打造区域智算标杆，构建“算力+技术+场景”全栈产业闭环；与宇信科技深耕垂直领域，联合打磨金融AI专属解决方案，推进核心业务场景落地。三类合作虽切入点不同，但均推动算力中心完成模式升级，收益从固定卡时租赁转向可持续Token分成，实现算力利用率、收益空间、产业地位的三重升级。伴随智能体AI规模化普及，单任务Token消耗量大幅跃升，推理成为算力消耗核心主力，产业链的关注点已从“拼算力、炼大模型”转向“拼推理、拼应用落地”。Token 工厂的竞争维度也在被重新定义：产能端从峰值应对转向持续满产运营，适配智能体的7×24小时平稳运行需求，并具备快速扩容和持续高利用率的双重能力；调度端需支持多轮推理、多模型协同等复杂场景的动态编排；服务端则需提供企业级SLA保障与Token用量管控，确保智能体深度嵌入核心业务后的稳定运行，实现服务从“可用”到“可信、可持续”的进阶。