From a 4-Trillion-Yuan Compute Network to Agentic CPUs, the Compute Structure Is Undergoing a Fundamental Restructuring

Over the past week (October 8–15, 2026), the compute sector has seen a convergence of policy, technology, and industry developments. From the top-level design issued by the Central Committee of the Communist Party of China and the State Council, to the OCP Summit triggering new consensus on CPUs and power delivery, to large compute leasing orders being fulfilled密集? densely on the A-share market, a clear signal is emerging: compute demand is shifting comprehensively from training-driven to inference- and agent-driven, and the compute structure is undergoing a fundamental restructuring. More specifically, the hotspots of the past week are not isolated events; together they point to a new era in which compute is no longer merely the “fuel” for large-model training, but a public foundation for always-on agents, large-scale token production, and industrial intelligent transformation.

I. Policy Direction: A 4-Trillion-Yuan Compute Network Investment Kicks Off; Compute Becomes the Core Foundation of New Quality Productive Forces

On October 9, the Central Committee of the Communist Party of China and the State Council issued the Opinions on Developing New Quality Productive Forces, explicitly calling for the “full implementation of the ‘AI+’ initiative” and proposing to “build a national integrated compute network and a national blockchain network.” A responsible official from the National Development and Reform Commission further noted that, according to estimates by relevant institutions, during the “15th Five-Year Plan” period, the construction of the compute network will add RMB 4 trillion in direct investment. Considering that compute construction is mainly driven by enterprise investment, this will create huge space for private investment. The significance of this statement lies in the fact that, for the first time, the compute network has been placed on the same level as the power grid, communications network, and transportation network as a key infrastructure direction for national strategic investment.

Fiscal funding is being increased simultaneously. The Ministry of Finance has arranged RMB 550 billion in local government debt carryover quotas, of which RMB 250 billion in special bonds will focus on the construction of the compute network, power grid, communications network, and other “six networks.” Since July, the approval pace of related projects has clearly accelerated. This means that compute network construction is no longer merely an industrial policy, but an intersection of fiscal policy, investment policy, and regional policy. The involvement of special bonds will effectively alleviate the problems of large upfront investment, long payback periods, and wait-and-see sentiment among private capital in compute infrastructure, providing long-term low-cost funding for compute network construction.

At the local level, Liaoning officially launched a compute voucher subsidy policy on October 13, providing a subsidy of up to 30% of the actual amount of compute contracts, with an annual subsidy cap of RMB 2 million for the same entity. Shanghai simultaneously raised the energy-efficiency access threshold for data centers, requiring new compute facilities to have an energy utilization efficiency of no more than 1.2. One subsidy and one restriction reflect that local policies are shifting from “simply encouraging the building of compute” to “encouraging the use of compute and the good use of compute.” The essence of compute vouchers is to lower the threshold for small and medium-sized enterprises and research teams, promoting compute from “being buildable” to “being affordable and well used”; the energy-efficiency threshold means that compute construction must advance in tandem with green, low-carbon, intensive, and efficient development.

Behind the intensive rollout of policies is the continued expansion of the compute supply-demand gap. The 2026 China AI Computing Power Development Assessment Report, jointly released by IDC and Inspur, predicts that China’s intelligent compute scale will reach 2,576.5 EFLOPS in 2026, a year-on-year increase of 87.9%. Meanwhile, the global AI compute demand satisfaction rate is expected to drop from 79% in 2024 to 71% in 2027, and the supply-demand contradiction will persist for a long time. This gap is not simply “not enough cards,” but a structural contradiction: training compute remains tight, while inference compute demand grows even faster; general compute is relatively oversupplied, while intelligent compute is in short supply; demand is strong in the east, while the west has abundant green power but network latency and scheduling capabilities still need improvement.

From the perspective of investment multipliers, RMB 4 trillion in compute network investment will drive multiple links, including servers, chips, optical modules, switches, liquid cooling, power supplies, cabinets, green power, and software scheduling. The compute network is not a single data center, but a comprehensive system of “compute + network + energy + scheduling + applications.” It requires compute to be grid-connected, schedulable, measurable, and tradable like electricity. For this reason, the construction of a national integrated compute network will push the compute industry from a project-based, island-style model to a networked, service-oriented one. In the future, enterprises may not need to build large-scale intelligent computing centers themselves; instead, they can purchase compute like cloud services, or even purchase tokens like electricity.

For local economies, compute network construction will also reshape the regional industrial landscape. Western hub nodes can rely on green power and land advantages to undertake training and cold storage tasks, while eastern nodes focus on low-latency inference and real-time agent services. The combined use of policy tools such as compute vouchers, energy-efficiency thresholds, special bonds, and green power trading will push compute layout from “rushing in all at once” to “a coordinated national game.” It can be foreseen that for some time to come, compute network construction, compute scheduling platforms, green data centers, domestic chip adaptation, and industry large-model applications will become key directions for policy implementation.

II. OCP Summit: Agentic CPUs and 800V Power Delivery Reshape Data Center Architecture

From October 12 to 13, the 2026 OCP Global Summit was held in San Jose, USA, becoming the most concentrated technology bellwether event in the compute field this week. The OCP Summit has long been a bellwether for global data center hardware architecture, and the signals released at this year’s summit were particularly strong: AI data centers are moving from “GPU-centric” to “CPU-GPU collaboration,” from “12V/48V power delivery” to “800V high-voltage direct current,” and from “single-rack card stacking” to “rack-level system design.”

The most notable signal on the first day of the summit was that CPUs are returning to the core of AI data centers. The accelerated adoption of consumer-facing AI agents, represented by Meta Muse, is pushing server CPU demand into a second growth curve. AMD CEO Lisa Su said in her keynote that AI inference will account for about 60% of the market in the future, and agent-related workloads are exploding. The more critical structural change is that the ratio of data center CPUs to GPUs is gradually evolving from the traditional 1:8 to 1:4, 1:2, and even 1:1. AMD simultaneously launched Helion, a dedicated rack computing platform for agentic AI, offering both pure CPU and CPU+GPU configurations, with a pure CPU rack supporting up to 32,000 CPU cores.

The logic behind this change is not complicated. When training large models, GPUs handle large-scale matrix operations, while CPUs mainly handle data preprocessing and scheduling, so the ratio can be very low. But when agents operate, tasks are broken down into multiple steps such as planning, tool calling, retrieval, verification, execution, and reflection. CPUs need to frequently perform logical judgment, task orchestration, API calls, state management, and security isolation. GPUs excel at parallel computing, while CPUs excel at complex logic and low-latency response. As AI moves from “answering questions” to “completing tasks,” the value of CPUs is being rediscovered. Future data centers will no longer be “GPUs plugged in next to CPUs,” but CPUs and GPUs each doing their own job and working collaboratively.

Another industry consensus is the accelerated adoption of 800V high-voltage direct current power delivery. Meta, Google, and Microsoft have clearly advanced the construction of GW-level ultra-large-scale data centers, and OCP officially released technical specifications related to 800V HVDC. Institutional research reports point out that 800VDC is not a long-term concept, but will be standard in new data centers over the next 2–3 years. As single-rack power jumps from tens of kilowatts to hundreds of kilowatts or even megawatts, traditional power delivery architectures are approaching their limits in terms of copper loss, efficiency, space, and heat dissipation. 800V high-voltage direct current can reduce conversion stages, lower current, improve efficiency, and simplify in-rack power distribution. It is not only a power supply technology upgrade, but a systematic reconstruction of data center electrical architecture.

The power delivery upgrade will also drive changes across a series of industrial chains. High-voltage DC circuit breakers, solid-state transformers, silicon carbide devices, power modules, busways, liquid cooling systems, and rack-level monitoring will all benefit. At the same time, cooling solutions are evolving from air cooling to liquid cooling and immersion liquid cooling, and the integrated design of power and cooling is becoming a trend. In the future, the competitiveness of a data center will depend not only on how many GPUs it has, but also on how much effective compute each watt of electricity can produce and how many agent instances can be deployed per unit of rack space.

In terms of interconnect solutions, copper cables, NPO, and CPO are developing in parallel and being deployed in layers according to scenario. Manufacturers such as Broadcom, Nvidia, and Credo have clear positioning in different links, and the pace of technology implementation is faster than previously expected by the market. Short-distance interconnect continues to leverage the cost and power advantages of copper cables, while medium- and long-distance and ultra-high-bandwidth scenarios are accelerating toward near-packaged optics and co-packaged optics. The overall judgment conveyed by the OCP Summit is that AI data centers are entering a stage of “system-level competition.” Single-chip performance is certainly important, but power delivery, cooling, interconnect, racks, and software scheduling also determine final compute efficiency.

III. Industry Dynamics: Large Compute Leasing Orders Appear Frequently; Domestic Chips Accelerate Breakthroughs

On the industry side, compute leasing has become the most prominent market phenomenon this week. A number of listed companies on the A-share market have密集? densely landed large, long-term compute leasing agreements, with contract terms generally more detailed and performance constraints stronger. The industry’s competitive logic is shifting from competing on the scale of GPU card supply to high-quality delivery cycles of order fulfillment and performance realization. Data from the China Academy of Information and Communications Technology shows that the compute leasing market is expected to exceed RMB 260 billion in 2026, and Founder Securities estimates that the domestic compute leasing market will approach RMB 1 trillion by 2030. The explosion of compute leasing shows that inference demand and agent applications are generating real, continuous, and billable compute consumption, and compute is turning from a “capital expenditure story” into a “cash flow business.”

The business model of compute leasing is also evolving. Early compute leasing was mostly simple bare-metal rental, with customers building their own models and scheduling systems. Today, more and more compute leasing providers are beginning to offer value-added services such as model adaptation, inference optimization, token billing, elastic scheduling, and security isolation. Contract terms are shifting from short-term trials to long-term locks of more than three years; pricing mechanisms are shifting from fixed rent to “floor plus floating”; and performance clauses are shifting from vague promises to assessments based on availability, latency, and throughput. This means that the compute leasing industry is moving from a “sublandlord” model to a “compute operator” model, with core competitiveness in stable delivery, cost control, and ecosystem adaptation.

In the chip field, both domestic and international manufacturers have made important progress. Microsoft plans to deploy its new self-developed AI chip Maia 300 at gigawatt scale starting in 2027, mainly for AI inference tasks, with a target of improving inference efficiency by 40%, demonstrating a clear move to reduce dependence on Nvidia. Domestically, Biren Technology’s new-generation AI chip BR20X has been successfully lit up and entered the post-tape-out verification stage, supporting FP8 and FP4 low-precision computing and carrying the self-developed BLink 2.0 interconnect protocol. Moore Threads’ MTT S5000 completed Day-0 adaptation for StepFun’s Step 5 Preview large model on October 15. Black Sesame Technologies obtained its first RMB 300 million order for AI computing server business, equivalent to 36% of last year’s revenue, marking a substantive breakthrough in its expansion from smart automotive chips to the AI computing server field.

It is worth noting that the latest IDC report shows that in the first half of 2026, Nvidia’s share of China’s data center AI chip market has dropped to 34%, with Huawei close behind at 28%, narrowing the gap between the two to 6 percentage points. Nvidia’s AI accelerator shipments in China decreased by 44% year-on-year. Behind this data are both external control factors and reasons such as domestic chips accelerating adaptation and customers actively seeking diversified supply. The focus of competition for domestic chips is shifting from “whether it can run” to “whether it runs well, stably, and cheaply.” Day-0 adaptation, low-precision computing, self-developed interconnect, software stack compatibility, and cluster stability have become key indicators for measuring the competitiveness of domestic AI chips.

From the perspective of the industrial chain, the breakthrough of domestic compute is not a single-point breakthrough, but a systematic advance. Chips need to be collaboratively optimized with servers, operating systems, deep learning frameworks, large models, inference engines, and scheduling platforms. Manufacturers such as Biren, Moore Threads, Huawei Ascend, Cambricon, and Hygon are making efforts along different routes, while server manufacturers, cloud providers, compute leasing providers, and large-model enterprises are jointly building the ecosystem. For some time to come, the share of domestic chips in inference scenarios is expected to continue to rise, because inference has relatively lower requirements for ecosystem compatibility than training, and China has rich agent and industry application scenarios that can provide大量? a large amount of real feedback to feed back into chip and software iteration.

IV. Core Trend: Agent Competition Escalates; Inference Compute Demand Explodes

The underlying logic running through all the hotspots this week is that the large-scale operation of AI agents is reshaping the compute demand structure. Over the past two years, compute demand has mainly been driven by large-model training, manifested as centralized, phased, and ultra-large-scale cluster construction. But entering 2026, as agents move from demonstration to production, compute demand is shifting to distributed, always-on, long-chain inference. An agent is not a one-off Q&A, but a “digital employee” that continuously perceives, plans, executes, and reflects. It needs to be online for long periods, frequently call tools, access databases, execute code, and collaborate with other agents, which will greatly extend the duration of compute demand.

The enterprise-level agent layouts of overseas giants have all been unveiled: Google Cloud released Gemini Agent, supporting Gemini and Claude models, and will be compatible with various closed-source and open-source models in the future; Microsoft released a new version of Copilot as a “super app,” adding the “digital teammate” Autopilot; OpenAI released the always-on agent dots; Anthropic integrated Claude into Google Docs, Sheets, and Slides. Agents are moving from single-point tools to “general employees,” and long-chain, always-on task forms will continue to amplify inference compute consumption. In the past, a user’s single question might consume only a few hundred to a few thousand tokens; in the future, an agent completing a task may call dozens of tools, conduct multiple rounds of reflection, and generate tens of thousands or even hundreds of thousands of tokens. Compute consumption is no longer linear growth, but multiplier amplification.

This trend directly drives the migration of the center of gravity of compute consumption. AI applications are spreading from programming to four major directions: R&D, enterprise services, scientific computing, and personal assistants. Compute consumption is shifting from training to always-on inference, and the duration of demand is clearly lengthening. Tianfeng Securities points out that in the next 2–3 years, China’s compute demand will grow nearly 10 times annually, and the share of inference compute is expected to rise to 80% by 2029, becoming the main growth engine. The rise in the share of inference compute means that the compute market is moving from “a few giants training large models” to “thousands of industries running agents.” The training market has high concentration, few customers, and large single transactions; the inference market has dispersed customers, rich scenarios, and flexible billing, making it more suitable for compute leasing, cloud services, and token trading.

At the same time, the China Academy of Information and Communications Technology released the Research Report on the Development of Intelligent Computing Infrastructure for Token Services (2026), focusing on the development trends of intelligent computing infrastructure for token services. The construction of the national integrated compute network is giving rise to a new “token factory” model—Beijing No. 1 Token Factory can produce 300 billion to 500 billion tokens per day and has adapted to and connected with mainstream domestic models. This new value carrier makes it possible for compute to be measurable, divisible, priced, and tradable, marking that the compute industry is evolving from “selling compute” to “selling tokens” as a service model.

The significance of the token factory is that it transforms complex compute resources into standardized, measurable service units. In the past, customers purchasing compute needed to care about GPU models, memory, bandwidth, cluster scale, and software stacks; in the future, customers purchasing tokens will only need to care about task results, response speed, cost, and compliance. The compute network is responsible for underlying scheduling, the token factory is responsible for model inference and token production, and application developers are responsible for scenario innovation. This layered division of labor will greatly lower the threshold for AI applications and promote compute to become a public basic service like water and electricity. It can be foreseen that new links such as token billing, token trading, token quality assessment, and token security auditing will gradually mature, forming a new industrial ecosystem.

V. Impacts and Challenges: New Contradictions Under Compute Restructuring

The restructuring of compute brings huge opportunities, but also new contradictions and challenges. The first is the contradiction between power and energy consumption. GW-level data centers, 800V power delivery, and liquid cooling all point to one fact: compute competition is increasingly like energy competition. Eastern load centers face tight power, while the west has abundant green power but insufficient network latency and local demand. How to achieve “East Data West Computing,” “West-East Power Transmission,” and “computing-power coordination” through a national integrated compute network is a problem that future policy and technology must solve. Solutions such as green power trading, energy storage, virtual power plants, nuclear power supply, and source-grid-load-storage integration will accelerate.

The second is the contradiction between chip supply and ecosystem. Domestic chips are accelerating breakthroughs in inference scenarios, but the software ecosystem, cluster stability, and development toolchains still need time to mature. When customers migrate from the Nvidia ecosystem to domestic platforms, they need migration costs, adaptation cycles, and performance verification. Day-0 adaptation is an important step forward, but long-term competitiveness depends on continuous iteration and large-scale delivery. In the future, chip manufacturers, cloud providers, large-model enterprises, and industry customers need to form closer joint innovation mechanisms, using real scenarios to drive collaborative hardware-software optimization.

The third is the contradiction between compute scheduling and networking. The national integrated compute network requires compute to be grid-connected, schedulable, and tradable, but different manufacturers’ compute standards, interfaces, billing methods, and security policies are not unified. Compute scheduling platforms need to solve problems such as heterogeneous compute management, cross-domain task migration, secure data circulation, and latency-sensitive task classification. Without unified standards and an open ecosystem, the compute network can easily become “multiple islands connected,” making true global optimization difficult.

The fourth is the contradiction between business models and return on investment. Large compute leasing orders appear frequently, but the industry also worries about excessive investment, price wars, and insufficient utilization. Compute construction has a long cycle and rapid technological iteration; today’s advanced GPUs may face depreciation in two years. Compute operators need to precisely match demand and avoid “building without users.” The token billing model helps improve utilization, but it also requires operators to have stronger model adaptation and inference optimization capabilities. In the future, the compute leasing market may diverge: leading manufacturers will obtain long-term orders through scale, ecosystem, and delivery capabilities, while small and medium-sized manufacturers will turn to niche scenarios or edge compute.

The fifth is the contradiction between security and compliance. Always-on agents mean that compute systems need to handle large amounts of enterprise data, personal data, and sensitive information. Model security, data isolation, access control, audit tracing, and content compliance have become rigid needs. While providing convenience, the compute network and token factories must also establish trusted computing environments and security governance mechanisms. Otherwise, the more widespread agents become, the more security risks are amplified.

VI. Future Outlook: From Selling Compute to Selling Tokens; the Compute Network Moves Toward Public Service

In summary, the core narrative of the compute field over the past week can be summarized in three main lines: on the policy side, with the RMB 4 trillion compute network investment and the “AI+” initiative as markers, the strategic position of compute infrastructure has been further established; on the technology side, the OCP Summit confirmed two major structural trends—higher CPU ratios and 800V power delivery—and compute hardware architecture is making systematic preparations for the agent era; on the industry side, the compute leasing business model is accelerating its iteration, and domestic chips have made substantive progress in both market share and technological breakthroughs. The intersection of the three points to the same judgment: the driving engine of compute demand is shifting from training to inference, from centralized to always-on, and from compute delivery to token services.

Looking ahead, the compute industry will show several clear trends. First, compute network construction will move from policy-driven to investment implementation, with a combination of tools such as special bonds, compute vouchers, green power trading, and energy-efficiency standards, promoting compute to be grid-connected, schedulable, and tradable like electricity. Second, the share of inference compute will continue to rise, with agents, token factories, and industry large models becoming the main sources of demand, and the compute market shifting from a few giants training to thousands of industries inferring. Third, data center architecture will be restructured around agents, with CPU-GPU collaboration, 800V high-voltage direct current, liquid cooling, and optoelectronic interconnect becoming standard in new data centers. Fourth, domestic chips will accelerate penetration in inference scenarios, with ecosystem adaptation and large-scale delivery becoming key to competition. Fifth, compute services will move from “selling compute” to “selling tokens,” and billing methods, trading mechanisms, quality assessment, and security governance will gradually improve.

This transformation is not only about technology route choices, but will also profoundly affect the value distribution pattern of the compute industry chain in the coming years. Whoever can provide stable, efficient, and low-cost inference compute, whoever can convert compute into high-quality tokens, and whoever can help industry customers quickly build agent applications will be able to seize the initiative in the next stage of competition. Compute is no longer just hardware stacking, but a systematic project involving policy, energy, networking, chips, models, applications, and business models. The hotspots of the past week are precisely a concentrated preview of this systematic project.