As large models evolve from "capable of chatting" to "capable of doing real work," and agents begin to help enterprises run through actual business processes, computing power has shifted from a nice-to-have to a hardcore necessity for enterprise development. However, computing power requirements vary significantly across different enterprises. For instance, a bank's core system requires a computing cluster to support a trillion-parameter model, while a small or micro enterprise might only need an AI workstation on a desk.
How can computing power requirements for all scenarios be met, ranging from large-scale industry deployments to lightweight applications for small and micro enterprises? At Huawei Connect 2026, the author found the answer: Yang Chaobin, Managing Director of Huawei and CEO of ICT BG, released a serialized all-in-one machine product matrix and an Agent acceleration platform. These serialized solutions, featuring tiered adaptation and out-of-the-box usability, cover computing power requirements across all scenarios.
01Industry Inflection Point: Scaled Deployment of Agents Elevates Computing Power to a "Core Necessity"
Large models are accelerating their penetration into thousands of industries, and agents are shifting from pilot verification to scaled deployment. AI is no longer just a demonstration in the laboratory; it is entering production systems and undertaking real business operations.
At the same time, enterprises face new challenges: model parameters range from tens of billions to trillions, representing a massive scale span; user concurrency demands vary widely; core businesses have increasingly stringent requirements for inference latency; coupled with multiple constraints such as deployment environments, power consumption costs, and O&M capabilities, enterprise AI computing power construction faces numerous practical difficulties.
Many enterprises easily fall into a misconception when building AI computing power: blindly pursuing large-scale computing clusters and continuously stacking hardware upwards. While the configuration may seem luxurious, it is actually superficial and results in severe resource waste.
Large clusters built blindly are prone to resource idleness, leading to low computing power utilization rates. The persistently high hardware and O&M costs are difficult to amortize, directly causing difficulties in business profitability and even continuous losses.
This is like buying clothes with only one size available; it can never fit everyone perfectly. The same applies to computing power; the key is not "how high it is stacked," but "whether it fits perfectly." Therefore, a new integrated computing power model featuring tiered adaptation and out-of-the-box usability is becoming a core necessity for the scaled deployment of agents.
02Serialized Computing Power Foundation: From Central Servers to Edge Components, Covering Computing Power for All Scenarios
How can precise matching and on-demand deployment be achieved for enterprises of different sizes and at different stages? During this conference, the author saw that Huawei has built a serialized, multi-tier all-in-one machine foundation to meet the diverse needs of enterprises for computing power facilities.
For the core businesses of large and medium-sized enterprises, the Atlas 650/650E delivers powerful performance, making it the optimal choice for private deployment. Equipped with Ascend 950 series processors, this product is the first to support mxFP4 low-precision inference, with single-card computing power reaching 1.56 PFLOPS, demonstrating strong fundamental hardware performance.
It should be pointed out that Huawei's UnifiedBus high-speed interconnect technology is not only used for super nodes but also applicable to all-in-one machines. Through dual-machine UnifiedBus direct connection, the Atlas 650/650E can achieve the industry's only 16-card Full-Mesh networking, allowing two machines to run as one, realizing optimal EP16 parallel partitioning, and further enhancing the overall business deployment benefits. In addition, when enterprises have larger business volumes and higher model evolution requirements, configurations with more than 16 cards can achieve smooth expansion through high-speed interconnection.
Benefiting from the dual empowerment of mxFP4 low-precision computing power and the 16-card UnifiedBus full-interconnect architecture, the inference performance of the Atlas 650/650E has achieved a leapfrog upgrade. Compared with conventional industry interconnect solutions, overall inference throughput is improved by over 70%, inference latency is as low as 10 milliseconds, and it supports ultra-long sequence processing of millions (1M), perfectly adapting to complex, high-concurrency AI businesses.
For small and medium-sized enterprises and lightweight applications, Huawei has opened up the Atlas 150 acceleration card, supporting partners in building flexibly configurable plug-in card servers. The Atlas 150 acceleration card is equipped with the Ascend 950PR processor, featuring a single-card computing power of 1.56 PFLOPS@mxFP4, on-chip memory of 84GB@1.4TB/s, and support for SIMD and SIMT hybrid programming, making development and adaptation more flexible. It also supports 2-card and 4-card UnifiedBus high-speed interconnection. Partners can build all-in-one machine products with flexible configurations ranging from single-card to multi-card based on the Atlas 150, providing flexible, lightweight, and easily deployable solutions for AI applications.
For edge computing scenarios with higher requirements for differentiation and customization, Huawei has opened up component products such as Kunpeng and Ascend modules. The Kunpeng 920 module forms a full matrix of Lite, Smart, and Max, with CPU cores ranging from 4 to 64. Its highly integrated design enables partners to rapidly complete complete system development, adapting to scenarios such as government services and public livelihood, intelligent manufacturing, and network security. Ascend offers the 310B and 310P series intelligent computing modules and reference designs, providing 8T to 280T computing power to support development boards, smart terminals, and embodied AI devices. To date, Huawei has collaborated with over 100 partners to develop more than 200 products across 10 major scenarios, implementing them in core industries such as finance, energy, the broad government sector, education, and healthcare.
03Platform Closed Loop: Agent Acceleration Platform Fortifies the Foundation for Industrial Deployment with Performance and Security
Having computing power in place is only the first step. The core pain points that truly constrain the intelligent transformation of enterprises are concentrated in the efficient, secure, and stable deployment of enterprise agents.
Many enterprises urgently need a dedicated system that features coordinated software and hardware optimization, secure and manageable control, and adaptation to multi-Agent inference, while also taking into account lightweight and highly efficient private deployment capabilities.
Huawei's Agent acceleration platform is built precisely for this purpose. According to the author's understanding, the platform integrates an Agent inference acceleration library, an agent management system, and an enterprise-grade security architecture, supporting full-domain access from clients, terminals, application IMs, and inference APIs.
Through on-site communication with technical staff, the author has sorted out the three core advantages of this platform.
Advantage 1: Ultimate Performance, Unlocking Efficient Computing Power for Multi-Agents through Software-Hardware Synergy
The Agent acceleration platform adopts a "native swarm architecture" design, built-in with industry-leading comprehensive capabilities for multi-agent generation, collaboration, scheduling, and evolution, capable of completing complex enterprise production tasks.
At the same time, relying on the deep optimization of A+K (Ascend + Kunpeng) software and hardware, the platform truly enables agents to run efficiently, with less memory consumption and higher performance.
First, in terms of resource utilization, the platform achieves proactive and refined management of the KV Cache based on the memory and computing power operational characteristics of multi-agent scenarios, minimizing memory waste. Coupled with a multi-level hierarchical KV Cache mechanism, the overall performance is ultimately improved by over 50%.
In addition, in terms of task acceleration, the platform achieves highly efficient asynchronous concurrent scheduling of inference and tool calls in multi-Agent collaboration scenarios, maximizing hardware performance. Through full-link underlying optimization, end-to-end task efficiency is improved by over 50%, supporting the deployment of high-concurrency, low-latency agent businesses for enterprises.
Advantage 2: Multi-Level Sandbox, Full-Link Protection, Solving the Multi-Tenant Security Challenge
Privacy protection in enterprise multi-user scenarios and full-process agent management and control are the biggest challenges.
To address the multi-user isolation issue, Huawei has built a three-tier protection system consisting of regular sandboxes, secure sandboxes, and encrypted virtual machines. Sandboxes can be selected on demand. The most secure encrypted virtual machines ensure that even platform administrators cannot read the plaintext of internal data, achieving true data security.
Furthermore, security protection is not just single-point defense, but a closed-loop management and control covering the entire business lifecycle.
In terms of full-process agent management and control, the platform covers skill listing, Agent operation, abnormal intent monitoring and interception of Agents, and supports rollback for abnormal operations. Finally, all operations are auditable, achieving complete end-to-end protection, eliminating enterprises' worries, and making enterprise-level Agents safe and usable.
Advantage 3: Out-of-the-Box Usability, Lightweight Deployment Lowers the Threshold for Enterprise Implementation
Built upon the high-performance and highly secure core foundation, the platform also considers the convenience of enterprise implementation, achieving "fine decoration delivery" for intelligent deployment. If hardware computing power is regarded as the "roughcast room" of intelligent construction, the Huawei Agent acceleration platform provides enterprises with a one-stop "move-in ready" experience.
The platform supports one-click private deployment of mainstream large models such as DeepSeek-V4-Pro/Flash, comes with pre-installed agents for general scenarios like office work and coding, and is paired with an agent management system, enabling the rapid launch of enterprise-level agent services within hours.
The author believes that the Huawei Agent acceleration platform, with its ultimate software-hardware synergistic performance and full-link closed-loop security as its two core pillars, supplemented by out-of-the-box delivery capabilities, builds a solid platform foundation for the scaled and industrialized deployment of AI in the industry.
04Case Verification: Across Thousands of Industries, Practical Application is the Ultimate Truth
In the exhibition area, the hands-on demonstrations of the full series of all-in-one machines and the Agent acceleration platform were surrounded by audiences. Compared to parameters, the author pays more attention to the effects of "practical application."
After communicating with on-site experts, the author learned that Huawei's all-in-one machines have been deployed in numerous industries. First, in the highly regulated finance and government sectors, a joint-stock bank installed its AI programming platform into an all-in-one machine for private deployment. The AI review automatically intercepts risks such as SQL injection and hardcoded keys, with a detection rate reaching 55% and a code completion adoption rate of 72%. The iteration cycle has been shortened from bi-weekly to weekly. For the development of over 30 applications in a provincial government service center, after adopting the all-in-one machine deployment, the time for handling complex requirements was compressed from one or two days to two or three hours.
Second, on 3C electronics production lines, quality inspection used to rely on a "sea of people" tactic, requiring 6 workers to conduct visual spot checks on a single line, with a missed detection rate as high as 12%. After the workshop adopted the integrated AI quality inspection machine, it can achieve synchronous detection of 8 video feeds, with a detection accuracy of 99.5% and inference latency below 100 milliseconds. The product missed detection rate has been reduced to within 0.3%, and the production line only requires 1 staff member for rechecking.
Finally, in terms of urban mobility, in a core urban area overseas, intersections are dense and the spatiotemporal distribution of traffic flow is complex, causing fixed-timing traffic lights to always "fail" during morning and evening rush hours. Based on a roadside AI intelligent computing all-in-one machine, the traffic management department can run a 14B large model at the edge, and aggregate multi-source data such as video streams, radar, and geomagnetic data in real time, predicting congestion evolution 5 to 15 minutes in advance. This equipment has reduced the average waiting time at intersections by about 17%, the daily delay index by 20%, and the urban congestion rate by 15% to 20%.
These cases demonstrate that Huawei's all-in-one machines making computing power accessible to thousands of industries is no longer just a slogan, but a referable path.
Author's Observation: Only When Computing Power "Fits Perfectly" Can AI Be Deployed Faster
From Yang Chaobin's speech at the conference to the product performance display in the exhibition area, and from the staff's introductions to the experts' case sharing, the author has found that all-in-one machines are bringing new changes to the computing power industry and the way enterprises use computing power.
First, the supply of computing power is shifting from an "arms race" of large intelligent computing centers to business-oriented tiered adaptation. In the past, measuring AI capabilities started with seeing who had the largest cluster. Today, computing power is used on demand like water and electricity, and only then can AI deployment truly reach thousands of industries.
Second, what Huawei delivers is no longer just computing power hardware, but a complete Agent capability foundation. Relying on the serialized all-in-one machines + Agent acceleration platform, Huawei has precipitated complex performance tuning, intrinsic security, and usability capabilities to the foundation side, leaving simplicity and convenience for partners and customers. This transforms AI deployment from a heavy and complex systems engineering project into out-of-the-box standardized delivery.
Third, the most crucial point for enterprises to consider is that the success or failure of AI deployment depends not on "how high the computing power is stacked," but on "whether the computing power fits perfectly." Tailoring the solution to specific needs is the only way to go far steadily.
Obviously, Huawei has achieved this. As Yang Chaobin said in his speech, Huawei will adhere to system-level innovation, build a serialized computing power foundation, remain open-source and open, jointly build an ecosystem with customers, partners, and developers, and provide new choices for the world.
Looking to the future, what Huawei's all-in-one machines aim to do is precisely to make computing power available on demand and out-of-the-box, just like water and electricity.