Connect compute, power and business value.

Distributed compute creates value through business resilience, value-added services and personalized solutions, not performance alone.

Details

DCP coordinates workloads, computing resources and energy constraints, with project designs addressing recovery, workload migration and service continuity. Private model deployment, AI agents, software integration and operations are tailored to each customer. Decide where and when tasks run, then assess delivery against agreed acceptance criteria.

Assess a project DCP strategy

Not a replacement for every data center. A coordinated role for core, regional and edge resources.

AI accelerator and memory package · AI scientific illustration · Not measured data or product specifications — Concept

DCP distributed compute-energy system

With Mirror Realm boxes and private workstations as core nodes, DCP brings local AI, regional compute and energy coordination into one tailored delivery. Virtual power plants coordinate energy; compute centers supply regional capacity. DCP connects these capabilities rather than treating them as separate businesses.

DCP scheduling and metering

Coordinate tasks using node permissions, priorities, resource status and energy constraints. Meter usage, monitor faults and apply recovery according to the deployment plan.

Mirror Realm boxes and workstations

Local tasks take priority; join distributed workloads and resource coordination with owner permission.

Virtual power plants and energy equipment

Connect solar, storage and flexible loads for energy monitoring, authorized control and demand-response interfaces.

Compute centers and engineering delivery

Plan regional capacity, networking, power, cooling and operations to support cross-node services.

From system capabilities to customer value

More resilient operations: Limit fault impact with replication, task migration and recovery policies suited to workloads and connectivity.
Services tailored to owners: Customize knowledge, models, encrypted storage and business interfaces around the customer’s actual work.
Manage compute and energy together: Evaluate scheduling against power capacity, deadlines and energy status to manage expansion and operating costs.

Delivery and control boundaries

Confirm node ownership, data and task permissions, device protocols and energy limits; then test scheduling, metering, recovery and manual takeover. Storage encryption, sharding and end-to-end encrypted communication are configured and tested per project. Availability, savings and returns depend on deployment and acceptance, not absolute privacy or fixed-return promises.

Plan a compute-energy solution

More capacity. A sharper focus on productive use.

Capacity describes supply, not business readiness. Hardware compatibility, electricity, latency and service continuity all matter when matching a resource to a workload.

2,185 EFLOPS

China AI compute capacity · FP16

71.4%

Nationwide rack occupancy

Facilities with 10,000+ AI cards

MIIT · Digital industry, H1 2026 ↗
Period: H1 2026. Sources checked: 2026-09-19. Industry figures, not DCP installations or operating results.

Occupancy

Rack use does not directly measure how busy processors are.

Utilization

Define whether the denominator is equipment, compute capacity or workloads.

Commercial output

Verify accepted work, billable delivery, renewals and cash collection.

Why we do not claim 720,000 racks are unpowered

The atlas figures of over 1.95 million racks, around 63% occupancy, RMB 43.5 billion direct investment and RMB 200 billion induced investment refer to eight hubs in June 2024, not 2026. Multiplying 1.95 million by 37% approximates 721,500 unoccupied racks, not evidence of unpowered or currently saleable capacity. Different dates and populations do not support a like-for-like comparison.

Read the 2024 report ↗

Where AI compute is located

June 2026 · Share of nationwide AI compute capacity

East 55.9%
West 32.6%
Central 10.6%
Northeast 0.9%

Our interpretation: resource selection requires more than power prices. Customer location, workload mobility and service constraints matter. This is not a forecast of DCP market share.

Source: Digital China Summit, republishing People's Posts and Telecommunications News ↗

3 entries

Heterogeneous resources: validate chips, software and workload compatibility before scaling.

View atlas entities

JD.com · Sugon · National Supercomputing Internet

Entity names retain the original Chinese labels. Status, coordinates and ownership have not been independently verified for every entry. Inclusion of 30 entities is not a partnership claim. The Zhengzhou group includes JD's Hebi project. Regional assessment notes are our analysis, not project measurements.

One service chain, not just another facility.

Start with demand, match resources, coordinate energy and learn from operation. Scope, supported configurations and acceptance criteria are confirmed during assessment, not inferred from concept visuals.

01 / Demand first

Start with the workload

Define models, data volumes, latency, concurrency and compliance. Separate online services that need proximity from jobs that can be scheduled, paused or moved.

02 / Compute + power

Price is one constraint, not the whole decision

Evaluate compute, electrical capacity, cooling and tariffs together. Include transfer, migration and operating costs before judging dispatch value.

03 / Continuity

Contain local failure

Define isolation, backup, degradation and recovery by service tier. Regional recovery depends on consistency, spare capacity and recovery targets; uninterrupted service is not universal.

04 / Feedback

Make the next decision better informed

Use authorized operating data and traceable baselines. Compare task completion, energy per useful output and service quality. Validate, approve and provide rollback for policy changes.

Three facility types. Three procurement questions.

IDC · General compute

Web, databases and enterprise systems

Validate reliability, storage, networking and migration compatibility.

AIDC · AI compute

Model training and inference

Validate chip ecosystems, memory, interconnect, power density and cooling.

HPC · High-performance compute

Scientific computing and simulation

Validate parallel efficiency, precision, interconnect and specialist software.

A category does not imply a standard rack power, PUE or investment price. Use configuration-specific tests. Tightly coupled training cannot simply be split across arbitrary edge nodes.

DCP / NODE TOPOLOGY

From a two-system link to a governed node pool.

Two, three, and four-node layouts are starting points for topology design. DCP must validate networking, scheduling, fault isolation, power, and operations before delivery.

02 / DIRECT

Two connected nodes

Validate link compatibility, model parallelism, and throughput.

03 / RING

Three-node ring experiment

Measure hop latency, recovery, and workload placement.

04 / SWITCH

Four-node switched topology

Validate QSFP, RoCE, bandwidth, and switch settings.

These diagrams are not delivered clusters. Node memory does not automatically become a unified pool, and theoretical compute does not replace workload tests.

Explore the Mirror Realm V6 node ↗

Technology and delivery, visualized

See the capacities and workflows, then explore the complete solution.

Explore all visual guides

DCP: from private nodes to compute–power coordination

Boxes and workstations are nodes. DCP coordinates resources; virtual power plants manage flexible energy through validated interfaces.

Box / workstation
Authorized resources
DCP coordination
VPP coordination

Virtual power plants: measure to verify

Identify flexible resources from meter data, assess constraints, execute a strategy and verify the response. Savings and returns require measured data.

Meter and collect
Constraints and plan
Authorized dispatch
Verify response

Proposed workflow, not live monitoring or delivered results. Interfaces, permissions and acceptance criteria are agreed per project.

Power systems behind compute

Explore the electrical equipment and operating boundaries behind DCP integration.

AI equipment concept · Not a site photo or specification

Power distribution

Transformers and switchgear

Assess supply capacity, circuit loads and equipment status before connecting compute. DCP coordinates workloads without replacing electrical protection.

Explore integration

AI equipment concept · Not a site photo or specification

Storage and conversion

Battery storage and PCS

Evaluate storage and compute coordination within power, state-of-charge and safety limits. Capabilities depend on interfaces and site validation.

Explore integration

AI equipment concept · Not a site photo or specification

Metering and protection

Distribution and protection panels

Read authorized energy and operating data for traceable records. Protection functions and local control authority remain independent.

Explore integration

Prove one service loop. Then decide how to scale.

Begin with an existing site or edge node and a well-bounded use case. Each stage creates evidence for the next investment decision.

01
Scope & baseline
Review workloads, resources, energy bills and service requirements.
Output: scope, metric definitions and baseline report.
02
Fit & pilot
Test compatibility, networking, security and dispatch boundaries.
Output: pilot plan, risks and acceptance criteria.
03
Compare & accept
Compare cost, energy and service on equivalent workloads and periods.
Output: test evidence, variance analysis and scale recommendation.
04
Operate & expand
Roll out in stages with monitoring, rollback and recurring review.
Output: responsibilities, service agreement and optimization plan.

Investment value needs operating evidence.

Ask who uses the resources, how service is delivered and when cash is collected. Disclose planned, contracted and operating capacity separately. Industry growth is not a project return forecast.

DimensionEvidence to reviewDecision boundary

Demand: Use cases, contract term, acceptance and customer concentration; Separate interest, contracts, delivery and cash.
Unit economics: Billable output, power, network, operations, depreciation and financing; Lower power expense does not equal higher net profit.
Productivity: Occupancy, device use, job success and billable hours; Keep periods and denominators consistent.
Scale readiness: Grid access, permissions, compatibility, reserves and operators; Meet stage gates before adding capital.
Risk & exit: Obsolescence, demand, tariffs, obligations and residual value; Use scenarios, not promised fixed returns.

Public material describes business value and assessment methods, not core algorithms, policy parameters or security implementation. Further technical diligence can proceed under agreed confidentiality. Scope, price, timing and service levels are governed by the proposal and contract.

Go deeper on the question that matters.

Industry Insights · Chinese report

U.S. Compute Infrastructure

Data centers, cloud regions, chips and power. Chinese PDF and market data available.

Industry Insights · Chinese report

China’s Semiconductor Ecosystem

Cities, supply chains and compute-power demand. Read the Chinese report or download the PDF.

Architecture Dispatch Capabilities Node systems Developer kit Power trading Virtual power plant Economics Delivery Strategy

Before we begin.

Does DCP own the 63 atlas projects?

No. The atlas provides context, not a DCP pipeline, customer list, installed base or partnership roster. Unverified project budgets and progress are not presented here as facts.

Can you promise savings or a payback period?

No. Workloads, configurations, tariffs, migration costs and service constraints must be assessed against a comparable baseline. A pilot establishes whether expansion is justified.

Can the project enter power markets?

Local rules, eligibility, metering and operating responsibilities require project-specific review. System design does not confer trading authorization.

What should we bring to an assessment?

A use case, inventory, power and cooling conditions, network needs, service targets and budget. Agree confidentiality and redaction before sharing sensitive information.

Start with a real workload.

Tell us what must run, which resources you have and which cost or service constraint matters most.

Discuss a project

Loading page…

Hardware and everyday applications

Compute halls and cooling
Explore the arrangement of racks, cooling pipes and maintenance aisles. AI-generated scene concept, not a photograph or a delivered customer case.

Compute and energy coordination
Explore design connections between compute nodes, storage and energy scheduling. AI-generated scene concept, not a photograph or a delivered customer case.

Workstation experiment space
A workbench explores local models, creative tasks and equipment selection. AI-generated scene concept, not a photograph or a delivered customer case.

Connect compute, power and business value.

Distributed compute creates value through business resilience, value-added services and personalized solutions, not performance alone.

Details

Not a replacement for every data center. A coordinated role for core, regional and edge resources.

DCP distributed compute-energy system

From a two-system link to a governed node pool.

Two, three, and four-node layouts are starting points for topology design. DCP must validate networking, scheduling, fault isolation, power, and operations before delivery.

02 / DIRECT

Two connected nodes

Validate link compatibility, model parallelism, and throughput.

03 / RING

Three-node ring experiment

Measure hop latency, recovery, and workload placement.

04 / SWITCH

Four-node switched topology

Validate QSFP, RoCE, bandwidth, and switch settings.

These diagrams are not delivered clusters. Node memory does not automatically become a unified pool, and theoretical compute does not replace workload tests.

Explore the Mirror Realm V6 node ↗

Connect compute, power and business value.