新智元报道
2026年，CPU成了AI世界里的「硬通货」。
全球服务器CPU集体涨价、交货期拉长，最夸张的时候订货要等半年。
原因很简单，AI不再只是「问一句答一句」的对话机器了。
Agent智能体开始自己规划任务、调用工具、运行代码、管理记忆，这些活儿全得靠CPU来扛。
CPU，这个曾经被GPU光芒掩盖的「老兵」，正在Agentic AI的浪潮中重返算力舞台的中央。
就在这个节骨眼上，鲲鹏在上海扔出了一颗重磅炸弹。
9月17日，第十一届华为全联接大会（HC 2026）在上海世博展览馆和世博中心开幕。
3场主题演讲、20场峰会、60多场论坛，展区面积超过22000平方米。规模拉满。
而在当天的鲲鹏计算产业峰会上，华为鲲鹏计算产品线总裁李义站在台上，带来了一个明确的信号：
面向Agentic AI时代，鲲鹏要做领先、开放、安全的智能体底座。
鲲鹏超节点全新升级，风冷、液冷双旗舰亮相，最大支持4096节点组网，构建百TB级统一内存池。
一句话概括这次升级的野心：以架构换算力，为世界构建新选择。
2019年鲲鹏计算产业战略发布的时候，不少人还在观望：
自研Arm架构处理器做服务器，能行吗？生态能建起来吗？谁会买？
七年过去了，数字给出了最直接的回答——
鲲鹏联合13家OEM伙伴推出了300多款整机，与7200多家ISV伙伴打造了28000多个行业解决方案，全球开发者规模超过416万。
openEuler和openGauss已经成为国内最具影响力的操作系统和数据库开源社区之一。
这些数字背后是一个清晰的逻辑：鲲鹏没有试图「颠覆」谁，而是通过硬件开放、软件开源的路线，把产业链上的伙伴一个个拉进来，共同做大蛋糕。
华鲲振宇就是一个缩影。
2020年成立之初就坚定选择鲲鹏路线，六年间推出五十多款鲲鹏算力产品，拿下鲲鹏金融应用的第一单，营收连续两年站稳百亿台阶。
昆仑技术同样如此，基于鲲鹏打造了全系列服务器产品线，预计今年营收将突破百亿。
从一个需要被「说服」的技术路线，到成为行业主流的算力选择，鲲鹏走了七年。
而现在，一个更大的变量出现了。
如果说过去几年AI的故事属于GPU，那么2026年开始，剧本正在被改写。
逻辑很简单：大模型不再只是「被问一个问题，回一段话」。
当AI从Chatbot进化为Agent，它需要自主规划任务、调度工具、执行代码、管理记忆、运行沙箱。
这些活儿，GPU干不了，得CPU来扛。
根据佐治亚理工学院与英特尔联合发表的论文《A CPU-Centric Perspective on Agentic AI》，传统AI推理场景中CPU的时间占比不到20%，而到了Agent推理场景，这个比例飙升到50%以上。
GPU和CPU的配比，也从过去的1:4历史性地走向了1:1。
这是一次算力架构的范式转移。
鲲鹏敏锐地抓住了这一点。
李义在演讲中指出，Agentic AI时代新增负载对算力架构带来三重挑战：人机交互走向多Agent交互，需要超高并发弹性架构；传统堆叠组网难以满足超低时延互联需求；大容量KV Cache缓存需要通算与智算深度协同。
换句话说，光有强大的GPU阵列还不够。
Agent时代的算力底座，必须是「通算+智算」协同的新架构。
这正是鲲鹏这次升级的核心逻辑。
鲲鹏这次在基础设施层面的升级，可以用两个关键词概括：处理器进化、超节点蜕变。
处理器方面，鲲鹏920新型号是目前业界唯一同时兼容DDR4和DDR5双内存规格的CPU。
这意味着用户可以用存量内存平滑过渡到新平台，省下一笔真金白银。
鲲鹏950则走高性能路线，高核宽比，场景化性能业界领先。
两款产品搭配，覆盖从性价比到高性能的全场景需求。
但真正让人眼前一亮的，是鲲鹏超节点的重磅升级。
过去的服务器是一台台独立的机器，靠以太网连在一起。
鲲鹏超节点的思路不同：超节点是一种在物理上由多个计算节点通过高效的灵衢互联协议紧密连接组成，具备跨物理节点的统一内存编址能力，逻辑上具备「一台计算机」特征的计算系统。
三大核心特性——内存统一编址、超大带宽（百GB级，是传统组网的10倍）、超低时延（百ns级，是传统以太网的1/10），定义了什么才是「真正的超节点」。
李义在演讲中特别强调了一点：以RoCE等协议互联、不支持「内存统一编址」能力的，不是真正意义上的超节点。
这句话虽然没有点名，但指向性很明确：鲲鹏在用技术规格划定赛道标准。
事实上，鲲鹏超节点的演进是一条清晰的「三级跳」路线——
- 年初先用电互联跑通8节点FullMesh免交换组网，验证了「内存统一编址」这个核心能力；
- 然后就是这次大会上亮相的4K光组网风冷旗舰，通过双层CLOS架构把规模一口气拉到4096节点，百TB级内存池正式落地。
- 而更值得期待的是下一跳：2027年Q3，全新液冷超节点将登场，单柜塞进128颗CPU，实现更高的部署密度和更高性能提升。
从电互联到光互联再到液冷，每一步都不是简单的「换个散热方式」，而是在密度、规模和能效上同时拧螺丝。
鲲鹏的节奏很清楚：先把架构跑通，再把规模拉大，最后把能效推到极致。
硬件再强，没有软件也是白搭。这一点鲲鹏看得很清楚。
围绕操作系统、数据库、集群软件、编译器四大方向，鲲鹏构建了一套完整的异构融合池化体系，纵向实现芯、硬、软协同，横向打通通算、智算、网络、存储的硬件资源融合。
openEuler，作为超节点的异构融合操作系统，通过池化内存和高速通信实现KV Cache毫秒级访问。
在Agentic AI场景下，这意味着智能体的「记忆」可以被近乎实时地读取和写入，不再受限于传统存储的带宽瓶颈。
openGauss，业界首个开源超节点多写数据库。2节点多写可实现570万tpmC，性能相较传统方案提升60%。
对于需要高并发事务处理的金融、电商等场景来说，这是一个实打实的竞争力。
openFuyao，作为多样化算力集群管理软件，打通超节点与云原生生态。依托URMA高速通信，业务容器通信时延下降40%以上。
简单是就是，你的Kubernetes集群里的容器，跑在超节点上比跑在传统服务器上快了接近一半。
BiSheng编译器，做异构算力亲和优化。它能结合Agent负载的动态变化，实时进行全链路调优，运行效率提升20%。
这不是传统的「编译完就不管了」，而是一个持续在线、感知负载、自适应调整的智能编译系统。
四大基础软件协同发力，把超节点从一个硬件概念变成了一个真正可用的软件平台。
架构讲得再漂亮，归根结底要看一个问题：扛不扛得住真实业务的压力？
先说一个大家都熟悉的场景。
你打开某购物App，刚搜了一双运动鞋，转头刷短视频就看到了相关广告。这背后就是生成式广告推荐系统在工作。
这套系统对时延极其敏感，每优化10毫秒就意味着0.5%—1.5%的收入增长，这是真金白银。
传统架构下，推荐系统的RPC通信占了端到端时延的大头，而四级缓存架构中远端的KV Cache访问又慢又不准。
鲲鹏超节点的解法是：灵衢互联把RPC通信提速，内存池化把上百TB的KV Cache全部拉到「身边」，让远端变近端。
效果十分显著：在线推荐端到端时延降了15%。这个方案已经在京东618大促中跑过了实战，是真正经历过流量洪峰冲刷的。
推荐系统解决的是「快」的问题，而大模型推理面对的挑战更复杂，不光要快，还要能「装得下」。
今年以来，大模型的上下文窗口从128K飞速拉长到百万Token级。
上下文越长，需要缓存的KV Cache就越多，本地存储很快就扛不住了。
传统方案下命中率不到80%，意味着每五次查询就有一次要去远端重新算。用户要等着，成本还在烧着。
鲲鹏超节点用灵衢互联构建起超大缓存池，把命中率直接拉到99%，同时靠硬件压缩引擎省下25%的存储空间。
到了Decode阶段，问题更尖锐：当对话上下文拉到256K甚至百万Token级别，一张卡需要的内存就飙过了300GB，可单机的内存条根本装不下这么多。
超节点把上百TB的内存池接进来，相当于给每张卡都开了一条直通「云端仓库」高速公路，生成速度直接翻了2到3倍，而且上下文越长、提速越猛。
如果说推理场景考验的是「记忆力」，那Agent沙箱考验的就是「爆发力」——
一个AI Agent在工作时，需要一个安全隔离的沙箱环境来执行代码、调用工具。
用强化学习训练Agent的时候，动辄要同时拉起10万个沙箱来做并行试错。
传统服务器方案启动这些沙箱要5分钟，对于需要快速迭代的训练流程来说，这个等待时间是不可接受的。
鲲鹏超节点用镜像预加载和RemoteFork技术把这个过程压到了10秒。
还有一个老问题：CPU的核数越堆越多，但单机内存的增长始终跟不上。就像工位越来越多、工区却没扩建，挤到最后总有人没地方坐。
超节点的百TB内存池等于开了一个巨大的共享工区，沙箱用的时候进来、暂停的时候腾出来，同样的资源能多跑25%的沙箱。
不过，当Agent开始大规模进入核心生产系统时，一个更根本的问题就冒出来了：
如果Agent「行为失控」怎么办？如果私密数据从Agent的记忆中泄露呢？
鲲鹏给出的回答是三层内生安全架构：
- 芯片层，处理器里内置安全可信根，支持国密算法；
- 系统层，机密虚拟机加上灵衢安全通道，把超节点内部的数据传输也锁死；
- 应用层，安全沙箱配合态势感知，实时拦截异常行为。
三层联防之下，Agent异常行为拦截准确率达到99%，而机密计算带来的性能损耗只有传统方案的五分之一。
安全是底线，但守底线不能以牺牲性能为代价——这一点鲲鹏想得很清楚。
回头看这场发布，鲲鹏的意图再清晰不过：
在全球CPU供应紧张、Agentic AI重塑算力架构的历史窗口期，鲲鹏要成为「另一种选择」。
不是替代谁，而是让市场多一条可走的路。
这条路的地基已经打了七年。
300多款整机、28000多个解决方案、416万开发者，生态的厚度不是一朝一夕堆起来的。
而今天，超大集群技术路线图已经公布，算力正在从单点突破走向体系化竞争。
鲲鹏超节点的4096节点组网和百TB内存池，不只是一组技术参数。
它代表的是一种架构级的思维方式，不是简单地堆叠更多的服务器，而是通过底层互联创新，让成千上万台机器协同工作如同一台超级计算机。
Agentic AI时代，千行百业正在被重新定义。谁能率先建好智能体运行的基础设施，谁就能在下一轮竞赛中占据先机。
鲲鹏的答案是：以架构换算力，让世界多一种选择。
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.
In 2026, CPU became "hard currency" in the AI world.
The global server CPU has increased the collective price, the delivery period has been longer and, at the most exaggerated, the order has to wait six months.
The reason is simple: AI is no longer just a dialogue machine.
Agent AI agent began to plan its own tasks, call tools, run codes, manage memory, all of which had to be carried out by CPU.
CPU, a "older" who was once masked by GPU light, is returning to the center of the math stage in the wave of Agenic AI.
On this occasion, a heavy bomb was dropped in Shanghai.
On 17 September, the eleventh session of the General Assembly (HC 2026) was opened at the Shanghai Expo and the World Centre.
Three keynote addresses, 20 summits and more than 60 forums with an area of over 22,000 square metres. Size full.
On the same day, at the Sun-Calcing Industry Summit, China stood on the stage for President Li Yi of the Sun-Calculated Product Line, giving a clear signal:
For the Age of Age Agentic, Zen is to lead, open, secure AI agent.
New upgrades were made to the supernodes, the wind cold and liquid cold double-flagged vessels were presented, with maximum support for the 4096 nodes network and the construction of a 100-TB class unified memory pool.
One sentence summarizes the ambition of this upgrade: to build a new choice for the world with the architecture convert.
In 2019, many people were watching:
Can you do that? Can you build an eco? Who's buying it?
Seven years later, the numbers give the most direct answer...
Together, 13 OEM partners launched more than 300 complete machines, and more than 28,000 industry solutions were developed with more than 7,200 ISV partners, with a global developer of over 4.16 million.
OpenEuler and openGauss have become one of the most influential operating systems and open-source communities in the country.
It is a clear logic behind these figures: rather than trying to "subverse" who, it is by bringing in the partners in the industrial chain, one by one, through an open hardware, open software route.
Joaquin Jin Woo is a microcosm.
Since its inception in 2020, it has firmly chosen its route and has launched more than 50 pieces in six years. The platinum calculator takes the first list of financial applications and stands at 10 billion steps for two consecutive years.
The same is true of Kunlun technology, which has created a full range of server product lines, and it is expected that this year ' s revenues will exceed 10 billion.
From a technical route that needs to be “convinced” to a calculator choice that is mainstream in the industry, seven years have gone by.
And now, there's a bigger variable.
If the story of AI in the last few years belongs to GPU, it's been rewritten since 2026.
The logic is simple: LLM is no longer just " asked a question, say a word."
When AI evolves from Chatbot to Agent, it needs to plan tasks, dispatch tools, implement codes, manage memory, run sandboxes.
These things, GPU can't do it. CPU will do it.
According to the paper by the Georgia Institute of Technology, A CPU-Centric Perspective on Agency AI, the CPU ratio of time in the traditional AI reasoning scene is less than 20 per cent, whereas in the Agent reasoning scene, this ratio jumped to more than 50 per cent.
The GPU and CPU ratio also moved from the past 1:4 to 1:1.
It's a paradigm shift in the architecture.
And he grabbed it very well.
In his presentation, Li pointed out that the new loads of Age Age Age of Age A posed a triple challenge to the algorithm: human interaction towards multiple Agent interactions, requiring ultra-high and elastic structures; the difficulty of traditional stacking networks to meet ultra-low latency interconnection needs; and the need for large capacity KV CacheZ96Z to work with intellectual depths.
In other words, strong GPU arrays are not enough.
The Agent-era algorithm has to be a new architecture for working together.
That is the core logic of this upgrade.
This upgrade at the infrastructure level can be summarized in two keywords: processor evolution and hypernodes.
For processors, the new 920 is currently the only CPU in the industry that is compatible with the dual memory specifications of DDT4 and DDT5.
This means that users can smooth the transition to a new platform with stock memory and save a piece of real money and silver.
鲲鹏950 takes the high-performance route, the high-nuclear width, and the scenery industry leads.
The two products combine to cover the full range of needs from value for money to high performance.
But what's really bright is that it's the height of the supernodes.
The server used to be an independent machine connected to the Ethernet.
The idea of hypernodes is different: they are a computational system that is physically closely connected by multiple computational nodes through efficient psychic interconnection protocols, with a unified memory site capability across physical nodes and logically with a "one computer" feature.
The three core characteristics - unified memory site, hyperband (100 GB, 10 times the conventional network), and ultra-low latency (100 ng, 1/10 of the traditional Ethernet) - define what is “real hypernodes”.
In his speech, Li highlighted the point that interconnecting with agreements such as RoCE, and not supporting the capacity of a single memory site, is not a genuine supernodes.
This sentence, although not named, is clear in its orientation: it defines the level of the track with technical specifications.
In fact, the evolution of the supernodes is a clear "third-rate jump" route.
- The core capacity of the `Universal Integrated Sites' was validated at the beginning of the year by means of electrical interconnections running through the FullMesh free network at eight nodes;
- Then, the 4K-ray-lighted, cold-winded flag ship that appeared at the Congress, pulled its size to 4096 nodes through a double-layered CLOS structure and officially landed at the 100-TB-class memory pool.
- What is more to be expected is the next jump: Q3 in 2027, when the new liquid cooling supernodes will be present and 128 CPUs will be placed in a single cabinet, achieving higher deployment density and higher performance.
From power to light to liquid cooling, each step is not a simple "replacement of heat" but screws both in density, size and energy efficiency.
The rhythm is clear: run through the structure, then scale up, and finally push energy efficiency to the extreme.
Hardware is so strong that no software is available. This is clear.
In the four main directions of operating systems, databases, clustering software and compilers, a complete isomer integration pool system has been constructed, with the vertical integration of core, hard, soft synergetic, horizontal computing, intelligence, network and stored hardware resources.
OpenEuler, an isomer integration operating system as a hypernodes, provides access to the KV Cache milliseconds through the pool memory and high-speed communications.
In the Agenic AI scenario, this means that AI agent's “memory” can be read and written in near real time and is no longer limited to the traditional storage of bandwidth bottlenecks.
OpenGauss, the industry ' s first open source hyper-node database. Overwriting of the two nodes would result in 5.7 million tpmC, an increase of 60 per cent over traditional programmes.
This is a solid competitive scenario for financial and electric operators that need to be dealt with at a high level.
OpenFuyao, used as a diversity calculator cluster management software to access hypernodes and clouds. latency, which is a high-speed URMA communication, decreased by more than 40 per cent.
The simple thing is, the containers in your Kubernetes cluster are almost half as fast as the traditional servers.
BiSheng compiler, isokinetic and optimized. It combines the dynamic changes in Agent ' s load with a real-time full-chain optimization with a 20 per cent increase in operating efficiency.
This is not a traditional "decompilation doesn't matter" but a continuous online, sensory load, self-adaptation smart translation system.
The four main basic software synergies have transformed the hypernodes from a hardware concept to a truly usable software platform.
The structure is so beautiful that ultimately it depends on the question: can you carry the pressure of real business?
Let's start with a scene everyone knows.
You opened up some shopping app, you searched a pair of sneakers, you turned your head on short video and you saw the ad. Behind this is the generation advertising referral system working.
The system is extremely sensitive to latency, which means 0.5-1.5 per cent income growth for every 10 milliseconds.
Under the traditional architecture, RPC communications for the recommended system are the bulk of the end-to-end latency, while access to KV Cache at the far end of the level 4 Z96Z structure is slow and off limits.
The solution to the hypernodes is to speed up the RPC communications through spiritual interconnection, and to quench the memory of hundreds of TB KV Caches to the "close" to the far end.
The effect was significant: the online referral end-to-end latency decreased by 15 per cent. This programme has run through battles at the height of 618 in Kyoto and has really been washed through the peaks of traffic.
The recommended system addresses the problem of "fast" and the challenge of LLM's reasoning is more complex, not just fast, but also "fulfillable".
Since this year, LLM's context window has grown from 128K to a million Token class.
The longer context is, the more KV Cache needs cache, the faster the local storage will be.
Traditional programmes have a hit rate of less than 80 per cent, meaning that every five inquiries are recalculated at the far end. Users wait, costs are burning.
The hypernodes were used to construct the super-large cache pool by psychic interconnection, with a direct hit rate of 99 per cent, while 25 per cent of storage space was saved by the hardware compression engine.
At the Decode stage, the problem is even more acute: when the dialogue context pulls to the level of 256 K or even millions of Token, the memory required for a card goes beyond 300 GB, but the memory strips of a single machine do not fit that much.
The hypernominal that brought hundreds of TB memory pools into the system is equivalent to a direct "unicable warehouse" highway for each card, with a direct increase of 2 to 3 times, and the longer and faster contexts.
If it's memory that the reasoning scene is testing, then the Agent Sandbox is testing flash power.
While working, an AI Agent needs a secure sandbox environment to implement codes and call tools.
When training Agent with intensive learning, it's very easy to pull up 100,000 sandboxes at the same time to make parallel mistakes.
The five-minute start-up of these sandboxes by the traditional server programme is an unacceptable waiting time for training processes that require rapid iterativeity.
The process was pressed to 10 seconds by pre-loading mirrors of the ultranodes and RemoteFork technology.
There is also an old problem: the number of CPUs is growing more and more, but the growth of single memory has not kept pace. It's like a growing number of jobs and no expansion of the construction area, and finally, there's always no place to sit.
The 100-TB memory pool at the hypernomous point is equivalent to a huge shared area, with sandboxes entering and being released when suspended, with the same resources running 25 per cent more sandboxes.
However, when Agent began to enter the core production system on a large scale, a more fundamental problem arose:
What if Agent's out of control? What if private data leaks out of Agent's memory?
The answer is three layers of internal security:
- Chip layer, with a secure and credible root in the processor to support the State secret algorithm;
- The system layer, the confidential virtual machine and the Visible Security Channel, which also locks out the transmission of data within the hypernodes;
- Application layers, security sandboxes coupled with situational awareness to intercept anomalies in real time.
Under three layers of protection, the Agent Abnormal Conduct Interception Accurate Rate reached 99 per cent, while the performance loss resulting from classified calculations was only one fifth of the conventional program.
Security is the bottom line, but keeping the bottom line cannot be at the expense of performance — a point well thought out.
Looking back at this publish, the intentions of the Zen are clear:
During the period of global CPU supply constraints and the historical window of the Agentic AI re-engineering of the computing architecture, 鲲鹏 is to become an "other option".
Not as a substitute, but as an extra road for the market.
The foundations of this road have been fought for seven years.
More than 300 complete machines, 28,000 solutions and 4.16 million developers, the ecological thickness is not stacked overnight.
Today, however, the super-cluster technology road map has been published, and calculations are moving from single-point breakthroughs to systemic competition.
The 4096 nodes network and the 100TB memory pool for the ultranodes are not just a set of technical parameters.
It represents a structured way of thinking, not simply stacking more servers, but creating a bottom-up interconnection that allows thousands of machines to work together as a supercomputer.
In the Age of Age Agentic AI, the 100th grade is being redefined. Anyone who can take the lead in building the infrastructure run by AI agent can take the lead in the next round.
The answer to this question is to give the world another choice with the convert force.
Editor: Solomon
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.