算力产业今年有一个很明显的变化，用户已经不再停留在“算力规模有多大”“芯片性能提升了多少”，而是反复围绕一个更具体的词：Token。新爱算联合创始人杨智告诉我们，今年年初以来，他们客户提的问题完全变了，“客户开始更聚焦模型跑起来一秒能出多少Token？生成Token的成本是多少钱？”
这背后是算力产业的价值体系变了。AI算力的价值衡量体系，正在从训练时代的追求峰值算力，转变为推理时代的追求效率。
算力生意的逻辑变了
在杨智看来，今天的智算行业，就好像二十年前的房地产行业。“当年房地产是包工头定制化、单位分房，找个包工头把房子盖好就能住。现在很多智算中心也一样，大厂过来租机房，把设备搬过去弄好，就结束了。”
芯片是砖头，改卡生产和维修维保是装配式建筑和物业，投建管理就是房地产开发。“而新爱算这一类企业，并不参与IDC拿地建设，而是把算力卡资产注入到郑州、杭州萧山等地的智算中心里。”杨智如是说。
但房地产行业真正成熟，靠的不是盖楼，而是从交钥匙工程，走向精细化的资产运营，而这也是当下算力行业同样正在经历的一步。尤其是在AI for Science（AI4S）需求的爆发下，这种转变被进一步推动。QYResearch的统计显示，2025年全球AI for Science市场规模约为45.38亿美元，到2032年预计达到262.3亿美元，年复合增长率28.9%。
AI4S的需求也是推理市场火爆的一个映射，国家数据局局长刘烈宏在2026年数字中国建设峰会上还披露，2025年推理数据量已达101.34艾字节，首次超过训练数据量，未来推理算力需求与训练算力需求之比或将达到3:1甚至更高。
而中国信通院《词元经济发展研究报告（2026年）》也指出，2026年推理算力占AI算力总负载的比重将达到70.5%。Gartner的预测更直接地体现在钱的流向上：2026年全球推理支出将达到233亿美元，首次超过190亿美元的训练支出，占AI IaaS总支出的55%。训练和推理的“主配角”关系，在这一年正式翻转。
需求端的这种转折，已经直接传导到了采购和验收环节。杨智坦言，过去客户只看卡的型号、数量和峰值FLOPS，而现在问的是：针对自己的模型、上下文长度和并发需求，吞吐能做到什么水平？响应延迟是否稳定？每月能耗和运维成本怎么控制？出了问题业务多久能恢复？“FLOPS仍然是设备选型的基础指标，但已经不是采购决策和项目验收的终点了。客户采购的不是一组参数，而是一项可持续运行的业务能力。”杨智指出。
这也解释了为什么行业开始流行每瓦Token产出、单位成本推理延迟这类新指标。北京并行科技董事长陈健曾在中关村论坛上表示：行业最优与最差的Token性能、成本差距可达100倍；DeepSeek官网百万Token的成本，从年初的约130元，通过系统优化降到了3元以下，离线场景甚至低至0.3元。
即便如此，推理需求的“高潮”还尚未到来，对此，杨智指出，当前的第一波需求，来自C端智能体产品，一个智能体完成一整套任务，消耗的Token是普通聊天的几倍甚至上百倍，不少产品上线后Token用量直接涨6倍，服务紧急扩容10倍。而当金融、医疗、制造、能源等行业的大模型真正走进千行百业，“等大门打开的时候，会是现在的算力需求的至少100倍”。杨智判断道。
真正难的不是买卡，是让系统跑起来
如果说Token经济改变了算力的账本，那么工程化交付能力，就是这本账能不能算清楚的关键。曙光智算总裁何牧君告诉我们，如今各家基模厂商发布新模型，超算互联网“零栈适配”基本都能做到，因为国产芯片的算力兼容性越来越好，基模厂商也基本都在CUDA架构上做训练。但能跑和好用完全是两回事。
以模型上下文为例：基模基本都支持百万级（1M）上下文，但实际部署时，0-128K、128K-512K、512K-1M，各自适合不同的部署和优化方案，还要考虑PD分离怎么部署，以及并行度等问题。“这些工作量是非常大的，我们基本上从它发模型倒推，前置2-3个月的时间。”何牧君指出。
模型“点亮”只是起点，离业务稳定运行还差着大量系统工程。在此基础上，杨智把交付验证拆成了四层：第一层是设备健康，包括供电、散热部件识别和基础压力测试；第二层是集群与网络的兼容性，包括驱动固件、网络链路和通信拓扑；第三层是与客户共同定义的业务负载验证，包括模型、并发、精度和关键时延；第四层是连续运行与验收交接。“设备能通电不等于集群能稳定通信，集群能跑通测试也不等于客户的模型和流程可以达到预期。”杨智强调。
而对于好用的标准，在杨智看来，可以从四个维度评估，即：性能可预期，关键并发下不能大幅波动；运行足够稳定，长任务不掉链、不中断；故障可定位，客户不必在模组、底板、网络和软件之间反复猜测；服务可恢复，有明确的响应、备件、维修和复验机制。
在软件侧，优化的空间同样巨大。杨智把推理优化分成三层：模型层、引擎层、系统层，而“把群片一整群机器干活”的系统层，正是行业最卷的战场。DeepSeek联合清华解决的Agent KV缓存瓶颈，让线上吞吐接近翻倍；英伟达公布的公开数据显示，同样的硬件平台，仅靠软件优化，DeepSeek的Token成本一个月内最低降到原来的1/5；PD分离、FP4专家并行、卡板吞吐提升20倍，这些已全部是验证过的事实。
而这也让异构算力成为必然的选择。训练时代，拼的是超大规模集群、高速互联、成熟软件生态缺一不可，英伟达几乎一家独大；但推理赛道的游戏规则完全变了，推理比拼的是每一个Token的成本、能效、业务吞吐，“上层业务越来越不关心底层是什么卡。”杨智指出。
推理时代，巨头开始卡位战
推理的崛起，除了改变了原有的算里账本之外，也在改写着芯片产业的结构。以行业巨头英伟达为例，去年12月，英伟达宣布与推理芯片公司Groq达成约200亿美元的非排他性授权协议，获得其LPU架构等核心技术授权，创始人Jonathan Ross及核心团队集体加入英伟达，而Groq公司本身仍独立运营。这笔交易超过了英伟达2019年以约70亿美元收购Mellanox的纪录，成为其史上最大手笔。
值得注意的是交易的形式，并不是传统并购，而是“技术许可+人才迁移”。行业内普遍将其解读为一次防御性收购，“当推理对延迟、能效和确定性的要求超越通用GPU的舒适区时，英伟达选择把最有威胁的非GPU架构收编进来。”杨智进一步分析道，“英伟达拿下Groq，AMD拿到对标谷歌TPU的路线，大家在拼命争夺ASIC推理芯片，在大模型已经逐渐收敛的当下，ASIC的机会就是万亿级的蓝海市场。”2025年，AMD数据中心业务收入同比接近翻番，ASIC推理芯片创业团队大量涌现，海外和国内的信号都在指向同一个方向。
与此同时，生态协作的方式同样在进化。超算互联网的实践提供了另一种参照：他们通过“万卡集群邀测计划”，让数万名用户参与到十万卡应用的测试中，而实际客户群里用量最大的反而是10个节点以内的普惠需求，今天80卡的用户，可能几个月后就是800卡、几千卡。何牧君坦言，做算力普惠从财务账上并不如一次性卖给大客户划算，但复旦类脑团队从几十卡、一百卡起步，如今已到万卡级别，据悉，目前“万卡共创者激励计划”已经覆盖了科研、985高校和缺少IT经验的药厂、医院等多种参与者。而曙光正用商业化方式满足各方不同的利益诉求，“不光是芯片上要具备不同的能力，整个工作流流水线当中可能要处理不同的科学任务，这对系统架构，以及生态建设的需求又更进一步。”何牧君强调。
从算力账本到工程化能力，再到生态的构建，推理时代对算力产业的改动正在层层深入。衡量算力的标尺换成了Token，比拼的重心落在了交付与运维上，产业链各方的合作方式也随之改变。正如杨智所言，推理爆发前夜往往显得安静，恰恰是积蓄力量的时刻。
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。
This year, there was a clear change in the calculus industry, where users no longer stayed on “how big the calculus” and “how much the chip has improved”, but over and over again around a more specific term: Token. Young Ji, the founder of the New Love Convergence, told us that since the beginning of this year, their clients' questions have changed completely, "How many tokens do clients start to focus more on a single second?" How much is the cost of generating Token?"
Behind this is a change in the value system of the computing industry. The AIS value measurement system is shifting from the peak of the training age to the quest for efficiency in the reasoning age.
The logic of math has changed.
In Yang Ji-ji's view, today's smart business is like 20 years ago in real estate. “In that year, real estate was customised as a foreman, divided into units, and a foreman was able to build the house. Now the same is true of many centres of wisdom, and it's over when the big factory comes to rent the machine room and it's done with the equipment.”
The chip is a brick, the modified card production and maintenance protection is a assembly building and property, and construction management is real estate development. “Since this type of business is not involved in the construction of the IDC land, it is the injection of the credit card assets into the centres of wisdom in Zhengzhou, Hangzhou Xiaoshan, etc..” As Yang Jiu said.
However, the real maturity of the real estate industry is not the construction of buildings, but the movement from turnkey engineering to sophisticated asset operations, which is also a step that the current computing industry is experiencing. This shift has been given further impetus, especially in the wake of the outbreak of demand from AI for Science (AI4S). According to QYResearch, the global AI for Science market was estimated to be $4.538 billion in 2025 and is expected to reach $26.23 billion in 2032, a compound annual growth rate of 28.9 per cent.
The demand for AI4S is also a reflection of the explosion in the reasoning market, and the Director General of National Data, Liu Rehon, also disclosed at the Digital China Construction Summit in 2026 that in 2025 the amount of reasoning data had reached 101.34 Aids Z113Z, exceeding the amount of training data for the first time, and that the ratio of the need for reasoning to the need for training calculations in the future would be 3:1 or higher.
In addition, the China Institute of Information and Communication Technology ' s Economic Development Study (2026) indicated that the weight of reasoning in total AI load would reach 70.5 per cent in 2026. Gartner's projections are more directly reflected in the flow of money: global reasoning expenditures will reach $23.3 billion in 2026, exceeding $19 billion for training for the first time, or 55 per cent of total AIIAS expenditures. The “principal mix” of training and reasoning was officially reversed during the year.
This turn in demand has been channelled directly to the procurement and acceptance check chain. While Yang honestly said that in the past clients had looked only at the type, number and peak of cards, the question now was: What level can swallow up against their own models, context length and co-production needs? Are the response delays stable? How can monthly energy consumption and transportation costs be controlled? How long can we get back to business if something goes wrong? “... FLOPS remains the basic indicator for the equipment selection, but is no longer the end point of the procurement decision-making and project acceptance check. The client procures not a set of parameters, but a sustainable operational capability.” Yang Zhi pointed out.
This also explains why the industry is becoming popular with the Token output per wa, and the delay in unit cost reasoning. Chen Jian, President of the Beijing Parallel Science and Technology Board, stated at the China Guan Village Forum that the industry's best and worst Token performance and cost gap could be 100 times greater, and that the cost of DeepSeek network's million Token was reduced from about $130 at the beginning of the year to less than $3 through system optimization, and offline was even as low as 0.3.
Even so, the “climax” of the reasoning demand has not yet arrived, and Yang points out that the current first wave of demand, from C-side AI agent products, a AI agent has performed a full set of tasks, consuming token several times or more than a hundred times the normal chat, and that token usage has increased six times directly after many products are online and the service has been expanded by 10 times. And when LLM in finance, medicine, manufacturing, energy, etc. actually enters a business, “when the door is opened, it will be at least 100 times more than what is needed today for computing”. Yang Ji-ji judge.
The real problem is not buying a card, it's getting the system running.
If the Token economy changes the calculator book, then engineering delivery capacity is the key to whether the account is clear. The president of Sunset Quill told us that today's new model of publish, the Internet's “zero-barrel fit”, is almost possible, as national chips are becoming more compatible and the base modelers are largely trained in CUDA structures. But running and using is completely different.
In the case of model context: The base model basically supports a million-scale (1M) context, but when deployed, 0-128K, 128K-512K, 512K-1M are suitable for different deployments and optimization options, taking into account how PD separations are deployed, and parallelity. “This is a very large workload, and we basically push it back from it, with two to three months ahead.” What the Reverend pointed out.
The model “lighting” is only the starting point and is still short of extensive system engineering to stabilize operations. On that basis, Yang Zhi broke the delivery certification into four layers: the first level was equipment health, including power supply, heat-dispersion component identification and basic stress testing; the second level was cluster-network compatibility, including driver solids, network links and communication leapfrogging; the third level was business load validation, defined as a customer, including models, hair, accuracy and key latency; and the fourth level was continuous operation interface with acceptance check. “Equipping equipment does not mean that clusters can stabilize communications, and that clusters can run through tests does not mean that models and processes of clients can meet expectations.” Yang emphasized.
For a good standard, Yang Zhi can be assessed at four dimensions, namely: performance can be expected and critical and not subject to large fluctuations; operation is sufficiently stable, with long assignments remaining chained and uninterrupted; failure can be located without repeated speculation between the user and the module, panel, network and software; and services can be restored with a clear response, spare parts, maintenance and validation mechanism.
On the software side, the space for optimization is equally large. Yang Zhi has optimized the reasoning into three layers: model, engine, system, while the system layer of “working with a whole bunch of machines” is the biggest battleground in the industry. The Agent KVcache bottleneck, which was resolved by DeepSeek together in Qinghua, nearly doubled the online throughput; the public data released by In Weida show that the same hardware platform, based on software optimization alone, reduced the cost of Token in DeepSeek to a minimum of one-fifth within a month; and the separation of PDs, parallel FP4 experts, and a 20-fold increase in cardboard throughput, all of which have been verified.
And it makes hesitological algorithms the inevitable choice. In an era of training, where mega-clusters, high-speed interconnections, mature software ecology are indispensable, the rules of the game in England are almost unique; but the rules of the reasoning track have changed completely, and the reasoning is the cost, energy efficiency, business throughput of every Token, “the top business is becoming less and less concerned about what the bottom card is.” Yang Zhi pointed out.
In the age of reasoning, the giants began the battle of the card.
The rise of reasoning, in addition to changing the original books of accounts, is rewriting the structure of the chip industry. In December last year, for example, the industrial giant, Young Weeda, announced a non-exclusive authorization agreement of about $20 billion with the reasoned chip company Groq, with core technical mandates such as its LPU architecture, the founder of which, Jonathan Ross and his core team, joined together, while Groq itself remained independent. This deal exceeded the record of the acquisition of Mellanox in 2019, with approximately $7 billion, and became its largest hand in history.
It is noteworthy that transactions take the form not of traditional M&As, but of “technical licensing + brain migration”. It is generally interpreted within the industry as a defensive acquisition, “when the reasoning requires that delay, energy efficiency and certainty go beyond the comfort zone of the General GPU, Yvette chose to incorporate the most threatening non-GPU structures”. As Yang further analysed, “In Weidar takes Groq, AMD takes the route to Google TPU, and you fight hard for the ASC reasoning chip, and the opportunity for ASIC is the trillion-class Blue Sea market at a time when LLM has gradually grown.” In 2025, income from AMD data centre operations nearly doubled in comparison to the same year, the ASIC reasoning chip start-up team surged and signals were pointing in the same direction both overseas and domestically.
At the same time, the modalities of ecological collaboration are also evolving. The practice of overcaling the Internet provides another reference: they involve tens of thousands of users in testing for 100,000 card applications through the Huanca Cluster Invited Scheme, while the largest use in the actual client group is in the form of inclusive demand within 10 nodes, with 80 card users today, probably 800 and thousands of cards in a few months. According to her, it is not as easy to sell to a large customer as it is to do in the financial accounts, but the rehabilitation brain team, which began with dozens of cards and 100 cards, has now reached the Vanka level, and it is known that the Huanca co-founders incentive scheme now covers a wide range of participants, such as scientific research, 985 colleges and drug factories, hospitals, and others with little experience in IT. The dawn, on the other hand, is being commercialized to meet different interests, “not just the different capabilities on the chip, but the different scientific tasks that may have to be dealt with throughout the workflow current line, which is further to the need for system architecture, as well as ecological construction.”
From power books to engineering capabilities to ecological construction, changes in the computing industry in the age of reasoning are taking place in depth. Token was replaced by a calculator to measure the calculus, the focus of which was on delivery and business, and the way in which the parties in the industrial chain cooperated. As Yang Zhi said, the night before the outbreak of reasoning often seemed quiet, precisely at the moment of build-up.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.