Always-on agents reshape data center ratios; Chinese ICT infrastructure vendors are first to win orders

Persistent Agents move AI to the cloud, rewriting the CPU, storage, and network ratios in data centers. Domestic cross-platform Agents are blocked; infrastructure takes orders first, but the real quality still depends on the application layer.

In the early hours of October 22, Anthropic, in the midst of its IPO sprint, released a brand-new personal and work agent application, Claude Duet.

In the month or so before that, Meta's Muse, OpenAI's Dots, xAI's Grok Bot, and Manus Cue, built by a Chinese startup team, launched one after another, delivering their answers almost simultaneously: give AI an independent virtual computer in the cloud.

In a demo Anthropic made public, an interesting scenario was shown:

Every morning at 5:30, before the owner of a coffee chain gets up, Duet has already automatically pulled the latest data from the POS system, the Shopify online store, the customer service inbox, the Gusto payroll system, and bank accounts, generated a morning business briefing, and pointed out that in the East Bay area, due to a shipping fee adjustment, the cancellation rate rose from 2.6% to 3.4%.

The debut of Claude Duet quickly sparked a new round of discussion in the tech community: why hasn't China produced a cross-platform "Chinese version of Duet"?

If one shifts their gaze away from the battle for entry points at the application layer and looks at the lower-level industry chain, a different picture emerges—the first to feel the impact of the persistent Agent wave and the first to realize commercial value are not the AI companies in the spotlight, but the infrastructure layer beneath the surface.

An architect at a leading domestic cloud provider admitted that when the product form of AI changes from "a question-and-answer chat box" to "a 24-hour online cloud virtual machine," the compute, storage, and network ratios in data centers all have to be recalculated from scratch.

01. From chat box to cloud computer

2023 was the Chatbot stage: ChatGPT defined question-and-answer interaction, the model was responsible for generating text or code, and the actual execution actions were still completed by humans.

2024 to 2025 entered the task-based Agent stage: the first-generation Manus and Claude Code proved that AI could break down multi-step tasks and deliver results, but interaction was still single-trigger, and the session was cleared once the task ended.

The turning point came in early 2026.

The open-source framework OpenClaw (known in domestic developer circles as "crayfish" because of its icon and the homophone in its name) gained 240,000 GitHub stars in three months, and domestic giants such as Tencent, ByteDance, and Alibaba quickly followed up, creating AI assistants that "live" in users' computers.

Agents no longer occupy the user's local computer; instead, they have an independent runtime environment in the cloud (Meta even allocates a dedicated virtual machine to each Muse user), and generally introduce a "proactive inspection" mechanism—without waiting for the user to issue instructions, the system automatically wakes up in the background every few hours to check email, calendars, and business systems.

For users, it is an upgrade in experience from "passive tool" to "proactive steward"; for the underlying compute infrastructure, it means the load has fundamentally changed.

Traditional Chatbots are stateless: when a user asks a question, the GPU is invoked once, and the connection is disconnected when the answer ends, so nighttime traffic naturally falls back. Persistent Agents are stateful; taking Meta Muse as an example, the underlying layer runs multiple "work sub-agents" in parallel, while also equipped with an independent "Sentinel review agent" monitoring permissions and security in the background.

This change will directly push up the service cost per user.

On September 30, a team led by Brian Nowak, Morgan Stanley's chief internet analyst, released a unit economics model estimate for Muse: under the baseline scenario of using only 25% of the weekly free Token quota, a single active user consumes about 108 million Tokens per month, corresponding to $37.12 per month in direct hard service spending—of which $24.30 is GPU inference cost and $12.82 is CPU sandbox environment and storage overhead.

Among heavy developer groups, the numbers are even more astonishing.

OpenAI CEO Sam Altman publicly revealed in June that the engineer ranked first in internal usage at OpenAI consumed 100 billion Tokens in a single month. When persistent Agents can work 7×24 hours in the cloud, 100 billion is by no means the upper limit of monthly Token consumption.

Stretching this timeline out, one will notice an easily overlooked detail: every leap in product form is accompanied by a transfer of the compute bill.

In the Chatbot era, bills were settled per use; in the "crayfish" era, agents ran on users' own computers, and electricity and hardware depreciation were silently borne by users; by the cloud persistent stage, the bill was moved back in its entirety onto the balance sheets of AI vendors and cloud vendors.

The competitive variable for persistent Agents is shifting from "how smart the model is" to "how much it costs per month to support an active user." The so-called "proactive inspection" is essentially a cost knob: the more frequent the inspection, the more butler-like the experience, and the higher the per-user cost. Whoever can find a sustainable balance between experience and cost may be able to keep this business going.

And before this question is answered, the first to be certain of getting paid are the people selling shovels.

When massive numbers of agents run continuously in the cloud, the first thing to run short in data centers is not the GPU that has been so sought after for the past three years.

02. CPU returns to the center of the table

For the past three years, the protagonist of the AI compute market has always been the GPU.

In traditional large-model training and single-inference servers, the CPU-to-GPU ratio is usually between 1:4 and 1:8. GPUs handle high-density parallel matrix operations, while CPUs only take on tasks such as data movement and basic scheduling.

Resident agents represented by Meta Muse are breaking the existing ratio.

After an agent receives a task, it needs to go through multiple steps such as intent understanding, task decomposition, spinning up a secure sandbox, executing code, operating a browser, calling external APIs, verifying results, and reflecting and retrying. Apart from the short segment of "model inference" completed by the GPU, the vast majority of process control, tool invocation, and environment isolation are typical serial logic operations, undertaken by the CPU.

The spread of cloud virtual machines will further drive up demand for CPU cores.

For example, Meta Muse configures for each user in the cloud a Linux virtual machine with 2 vCPUs, 8GB of memory, and a 100GB SSD as standard.

Overseas hardware media Wccftech once made an estimate based on static peak values: if Muse expands to 100 million users and each person exclusively occupies physical cores without sharing, calculated on the basis of a single 126-core AMD EPYC processor, this one application alone would require 1.58 million server CPUs.

Of course, in actual engineering, cloud providers will not let virtual machines idle 24 hours a day.

But even after accounting for sleep mechanisms and multiple oversubscription ratios, when tens of millions of active users concurrently spin up multiple sub-agent temporary sandboxes during peak periods, data center demand for independent "Agent CPU racks" still grows significantly.

Bernstein judged in a research report that the CPU share of compute in agent workloads will rise from 14% in traditional large models to 50%, the CPU-to-GPU ratio is converging from 1:8 toward 1:1, and it raised its forecast for the global server CPU market size in 2030 to 223 billion US dollars.

Changes in demand have already been transmitted to the supply chain.

KeyBanc data shows that, affected by concentrated procurement by large cloud service providers, Intel and AMD's server CPU capacity for all of 2026 is basically sold out, both companies plan to raise server CPU prices by 10% to 15%, delivery lead times have lengthened from the usual 1-2 weeks to 8-12 weeks, and some models have risen by more than 40% since the beginning of the year.

Facing high CPU costs and high-concurrency sandbox demand, domestic cloud and server vendors have accelerated underlying architecture changes in 2026:

Alibaba Cloud launched the ACS Agent Sandbox based on micro-VM isolation, focusing on optimizing large-scale low-latency elastic spin-up and low-cost state retention.

Tencent Cloud open-sourced Cube Sandbox v0.5.0 in July, and together with the Arm engineering team completed native mainline support for the Arm architecture, enabling the Arm platform to match x86 in sandbox creation speed, state recovery time, and per-machine concurrency density.

H3C proposed a sandbox image pool solution based on a CXL memory pool, pre-residing commonly used operating systems and dependency libraries in a shared memory layer, so that when a new sandbox is spun up it does not need to read from disk and directly achieves zero-copy startup through memory mapping, compressing sandbox cold start time from seconds to milliseconds.

Against the backdrop of CPU price increases and lengthened delivery times, how many sandboxes a single machine can hold and how quickly they can be woken after sleep directly determine per-user cost. Industrial value is migrating from "selling more cores" to "letting each core run more agents."

03. After VRAM is overwhelmed

Along with CPUs, storage systems that have long been at the edge of data centers are also undergoing value reassessment.

In traditional cloud computing architectures, storage is mainly responsible for statically saving data, and millisecond-level read/write latency is enough to meet business needs.

But in long-horizon agent scenarios, storage directly enters the main compute path, and the core bottleneck becomes KV Cache.

Just as in the Claude Duet demo scenario, to help a user prepare for a Series A financing roadshow, the agent needs to continuously read 25 data room documents and background materials on 6 investment partners, and make dozens of rounds of tool calls.

As model context windows expand to the million level and multi-round slow thinking becomes the norm, KV Cache capacity demand under a single task chain has surged dozens of times.

GPU high-bandwidth memory (HBM), which has limited capacity and high prices, is quickly filled up, and the overflowing warm KV Cache and long-term memory data have to sink to the next tier: DRAM memory, CXL memory pools, and enterprise solid-state drives (NVMe SSDs).

In addition, hundreds of millions of cloud sandboxes, when performing "snapshot sleep and wake-up recovery," also need high-frequency read/write access to their own independent file spaces.

Wu Junjie, vice president of Huawei's data storage product line, said in a media interview that when agents access context 24/7 without interruption, the latency requirements for underlying storage have risen from the previous millisecond level to the microsecond level—any delay in data return will cause the GPU to wait idle; if the context is forcibly truncated due to insufficient cache space, a large amount of repeated computation must be performed.

The tightness of supply and demand has already been reflected in prices and financial reports.

Domestic DRAM manufacturer ChangXin Technology said in its performance forecast for the first three quarters of 2026 that, driven by AI compute demand, DRAM products are in short supply and prices have risen sharply, and the company's capacity utilization rate remains high; storage module leader Longsys expects to achieve revenue of 38 billion to 40 billion yuan in the first three quarters of 2026 (year-on-year growth of 127% to 139%).

To alleviate the "VRAM wall" and the storage isolation challenges of hundreds of millions of sandboxes, major domestic technology companies are also advancing storage architecture restructuring.

Huawei has launched its CMS (Context Memory Storage) and UCM (Unified Cache Management) solutions that support heterogeneous computing power, building a PB-level shared KV cache pool through dedicated DPU semantic offloading, reducing inference first-token latency by up to 90% and lowering the comprehensive cost per token by about 30%.

English text is being prepared.

If CPU demand depends on whether consumer-grade Agents can make their business models work, the logic for storage is more solid. Long context and multi-turn reasoning are not unique to always-on Agents; enterprise Agents, coding Agents, and even ordinary deep reasoning tasks are all driving up the scale of KV Cache. Even if growth slows for a certain star application, storage demand will not recede.

A deeper change lies in the fact that storage is "trading capacity for compute": if the cache can hold data, the GPU does less repeated computation; if it cannot, more expensive GPU time must make up for it. Storage has thus shifted from a cost item to an efficiency item, and its say in AI infrastructure keeps rising.

04. The inverted network funnel

Beyond compute and storage, communications and data center networks are also undergoing a reversal in traffic structure.

Over the past two decades, the internet was designed mainly around "people consuming content." When users browse web pages, watch videos, and download apps, data flows mainly from data centers down to terminals; in traditional scenarios, upstream traffic typically accounts for only about 10% of network traffic.

In Agent scenarios, the traffic structure is inverted.

Whether a cloud Agent continuously reads local workflows, or a user transmits high-definition screenshots, audio and video streams, and documents in real time to an Agent via a phone or smart glasses, the prerequisite for AI to execute tasks is that terminals upload context to the cloud at high frequency.

For example, at MWC 2026 earlier this year, Huawei disclosed measured data from an art museum in Shanghai: after introducing a multimodal AI agent as a guide, more than 3,000 visitors per day uploaded photos or videos of exhibits for interaction, and upstream traffic accounted for more than 60% for the first time.

Inside data centers and in cross-cloud collaboration scenarios, the density of machine-to-machine (M2M) communication is also rising in tandem.

A typical example: when Claude Duet generates a business briefing in the background, it needs to continuously call the APIs of more than ten external SaaS systems through the MCP protocol within a few minutes; and inside the data center, the separately deployed "Agent CPU racks," "GPU inference racks," and "CXL storage racks" need to synchronize state with high bandwidth and low latency at all times.

Facing expectations of hundreds of billions of Agent connections, the domestic communications and network industry chain has long begun adjusting its layout:

On the operator side, China Mobile is exploring adjusting the uplink/downlink traffic ratio to 3:2 in the 4.9GHz band to match Agent uplink demand, and has proposed the lossless intelligent computing optical network (HIC-OTN) architecture; China Unicom in September upgraded its traditional CDN to AICDN, enabling coordinated scheduling of AI Token compute power and edge content.

On the hardware and edge services side, data center demand for 800G and 1.6T high-speed optical modules continues to expand, and TrendForce estimates that the global AI optical transceiver module market will reach $26 billion in 2026.

It should be noted that although the direction of change on the network side is clear, its realization pace is likely to lag behind compute and storage. Adjusting spectrum ratios and building lossless optical networks take years, and investment must precede demand; moreover, current evidence of upstream traffic reversal still mainly comes from tests in individual scenarios.

What truly deserves attention is the role of operators.

For a long time, networks were designed around downlink bandwidth and charged based on content distribution; if Agents make uplink the main traffic, "uplink" itself may become a new billing dimension. Operators may not be the first to benefit, but they may hold the longest-term pricing leverage.

05. Final thoughts

Back to the question at the beginning.

In the short term, it will likely be difficult for the Chinese market to produce a cross-platform "Chinese version of Duet." The crux is not technology but ecosystem boundaries: as of late October, more than a hundred apps, including WeChat, Taobao, Alipay, and Meituan, had rejected screen automation operations by the Doubao phone assistant. Cross-app agency in China is first and foremost a commercial question.

This creates an intriguing mismatch:

The battle for entry points at the application layer is temporarily deadlocked in China, while the CPU expansion, storage upgrades, and compute network reconstruction driven by the global Agent wave have already allowed infrastructure vendors in China's ICT industry chain to secure orders first.

But the quality of the dividend ultimately still depends on the application layer: only when always-on Agents find a business model that can "afford users" will today's capacity expansion not become tomorrow's excess.

English text is being prepared.