CHINA AI
INFRASTRUCTURE
Home / VNET thesis
CHINA AI INFRASTRUCTURE · LIVING SYSTEMS RESEARCH

VNET: from data centres to compute-energy infrastructure

The full thesis: demand conversion, high-density delivery, Huawei networking, CATL’s wider ecosystem, financing and equity-value sensitivities.

AIDC comparison

How does this operator compare with the other listed China AIDC platform?

A fact-led side-by-side study now tracks demand, delivery, utilisation, economics, financing and emerging infrastructure integration without assigning a winner.

Research basis: 17 September 2026 thesis • Restored and cross-linked: 20 September 2026
Financial figures, forecasts and scenario assumptions retain their original dates. This restoration is not a fresh market-price or earnings update.

Core opportunity

Combine AI-ready sites, inter-site networking and energy partners into a more useful physical infrastructure platform.

Economic gate

Reservations must become customer move-ins, collected revenue and cash returns without excessive parent debt or dilution.

What CATL adds

Potential energy, project-capital and customer-ecosystem channels; none should be treated as an undisclosed funding commitment.

Evidence and scope

The July 2025 VNET prefab disclosure predates Huawei’s September announcement. Fast-delivery statements should be compared on like-for-like milestones; installation, contract-to-delivery and energisation are not the same interval. The original 862–867 MW order range below is retained as a working-paper reconciliation issue, not a new precision claim.

VNET Q2 release · VNET–CATL announcement · VNET prefab disclosure, July 2025 ·
Jump to a thesis section
  1. 1. What has changed since the August thesis
  2. 2. The architectural shift: estate → cluster → supernode → compute-energy fabric
  3. 3. Why VNET may be unusually well positioned
  4. 4. CATL: the thesis is broader than 'funding VNET'
  5. 5. Customer & workload ecosystem
  6. 6. Operating thesis: demand is not the main uncertainty
  7. 7. Funding and valuation: retain discipline
  8. 8. The 'motherboard' analogy: useful, with one correction
  9. 9. What would materially strengthen or break the thesis
  10. 10. Revised thesis statement
  11. 11. Evidence ledger and source discipline

1. What has changed since the August thesis

The August work correctly identified demand as increasingly de-risked and financing as the principal equity-value gate. That remains true. What is new is stronger evidence that the physical unit of AI infrastructure in China is changing. Huawei's 17 September roadmap brings Ascend 960DT forward from Q4 2027 to Q1 2027 and introduces a 4,096-card Ascend 960 supernode using NPO optical engines, while its 16 September 3D data-centre design vertically separates cooling, IT, power and backup-power layers and targets factory-style prefabrication. These developments imply much higher system-level density and tighter coupling of compute, interconnect, cooling and electricity.

August framingSeptember evidenceThesis consequence
Demand vs fundingDemand remains strong; domestic AI hardware roadmap acceleratesFunding is still critical, but the quality and technical suitability of capacity matter more.
Data-centre capacity measured mainly in MWSupernodes/SuperPoDs make the logical machine much larger and denserMW becomes an incomplete metric; power density, cooling and fabric topology become strategic.
CATL as potential funding/energy partnerHuawei architecture makes energy/storage integration structurally more importantCATL relationship gains industrial logic beyond capital provision.
VNET as neutral IDC operatorHuawei/VNET case study explicitly describes an AI-native wide-area compute networkVNET has evidence of moving from sites to a connected compute fabric.

2. The architectural shift: estate → cluster → supernode → compute-energy fabric

The most important conceptual update is that the competitive object is no longer simply the building or rack. Huawei defines the SuperPoD as a single logical machine assembled from many physical machines. Its 2025 roadmap envisaged an Atlas 960 SuperPoD with up to 15,488 NPUs; the product announced on 17 September 2026 is a 4,096-card Ascend 960 supernode. This is not necessarily a contradiction: the newly announced unit can be understood as a scale-up domain, while larger SuperPoD/SuperCluster systems can be assembled above it. The precise final topology still requires clarification from Huawei.

Huawei's new NPO design is especially relevant. The company says 5,500 Hi-ONE optical engines replace roughly 48,000 800G optical modules in the 4,096-card system, cutting power consumption by more than 550 kW and lifting system availability to 99.8%. This is a reminder that networking and power overhead become first-order economics at cluster scale. VNET's June Huawei case study is therefore more strategically important than it first appeared: the joint solution upgrades WAN bandwidth from 10GE toward 100/400/800GE, uses elastic slicing and lossless transport, and is explicitly designed to connect enterprise data with remote AI compute.

3. Why VNET may be unusually well positioned

CapabilityEvidenceWhy it matters now
Scale and pipeline1,007 MW wholesale in service at Q2; 585 MW planned delivery over the following 12 months; ~862–867 MW H1 orders plus 355 MW reservationsLarge logical AI systems need large, contiguous and repeatable power blocks.
Network layerHuawei + VNET/DYXnet AI-native compute WAN; 100/400/800GE evolution, elastic lossless transportAllows compute to be treated as a networked resource rather than isolated buildings.
Power/energy layerCATL strategic agreement for a three-layer compute-energy ecosystem and gigawatt-scale facilitiesEnergy storage, direct green power and power quality become part of the compute product.
Delivery modelExisting rapid/prefabricated delivery experience; Huawei 3D AIDC pushes factory-style standardizationFaster mechanical/electrical delivery shortens the capex-to-billing gap.
Asset recyclingMature wholesale assets have supported private REIT/ABS transactions around 13–14x EBITDA in prior workPotentially separates mature infrastructure ownership from high-growth development capital.

4. CATL: the thesis is broader than 'funding VNET'

The August funding paper focused on whether CATL-linked project equity, vendor finance, energy assets or asset recycling could reduce corporate debt accumulation. That remains a powerful equity mechanism. The new architecture adds a second mechanism: CATL can potentially improve the product itself. At very high compute density, electricity availability, backup power, storage, direct green-power connection, thermal management and grid interaction are no longer ancillary real-estate services; they are constraints on usable compute.

CATL rolePossible VNET benefitEvidence status
Strategic shareholder ecosystemLonger planning horizon and alignment around compute-energy buildoutAffiliated buyers are acquiring a large secondary stake; transaction itself does not fund VNET.
Energy/storage infrastructureShift part of non-IT capex or operating complexity away from VNETStrategic cooperation explicitly covers zero-carbon energy and compute-energy integration; project economics undisclosed.
Project/JV capitalReduce parent-level debt and dilutionPlausible structure, not yet announced.
Power-product innovationHigher-density, more dispatchable AI campuses; potentially faster grid connectionIndustrial logic strengthened by Huawei's density trajectory, but no VNET/CATL/Huawei joint project disclosed.
Customer roleCATL could itself become a compute/AI customer as industrial AI expandsOptionality only; no disclosed CATL compute contract.

5. Customer & workload ecosystem

VNET’s demand thesis should distinguish named evidence from customer inference. The company’s recent wholesale order book is large, but the customers behind most of those MW are not publicly identified. We therefore do not attribute any portion of the 862–867 MW working-paper order range or the 355 MW of reservations to ByteDance without direct evidence.

LayerWhat the evidence supportsThesis treatment
Volcano Engine ↔ VNET ecosystemVolcano Engine has a confirmed cooperation relationship with DYXnet, VNET’s wholly owned network subsidiary, covering cloud computing, AI, security and network bandwidth.Show a dotted ByteDance/Volcano Engine ↔ VNET edge as evidenced ecosystem/customer overlap.
Direct ByteDance wholesale capacityNo sufficiently strong public evidence currently identifies ByteDance as the customer behind VNET’s recent unnamed hyperscale MW awards or reservations.Do not assign MW, revenue or backlog to ByteDance.
Workload fitVNET is building high-density AI infrastructure, inter-site networking and compute-energy capability suited to large-model and leading-internet-company workloads.Treat ByteDance as an important demand-side ecosystem adjacency, not as a quantified customer assumption.

See the ByteDance infrastructure evidence and VNET overlap →

6. Operating thesis: demand is not the main uncertainty

The August Q2 model showed approximately 1.22 GW of firm plus customer-reserved demand and a 96.3% commitment rate, while mature wholesale utilization remained 92.5%. The harder execution question is move-in: after 585 MW of planned delivery, preserving the June 73.9% overall utilization rate would require utilized capacity to rise by roughly 432 MW from the Q2 level. The September Huawei announcement improves the medium-term domestic accelerator supply narrative, but it does not eliminate near-term HBM, packaging and deployment bottlenecks.

7. Funding and valuation: retain discipline

The new strategic evidence should not be used to mechanically increase EBITDA or reduce net debt. The August hypothetical funding model remains the correct bridge: at RMB5.5bn normalized EBITDA and 10x EV/EBITDA, modeled equity value ranged from roughly $14.9/ADS in a debt-heavy case to $19.8 in a CATL/asset-light case and $21.3 in a very strong funding case. Those figures are sensitivities, not targets. What changes today is the probability narrative around why external capital and partner-owned energy infrastructure could exist.

ScenarioOperating interpretationFunding interpretationWhat would validate it
StressMove-ins lag; utilization falls; hardware bottlenecks persistParent debt/equity funds most buildWeak cash conversion, rising corporate net debt, no project-level structures
Core585 MW delivery broadly achieved; reservations convert in batchesMix of corporate debt, project loans and recyclingNamed project facilities, further REIT/ABS, visible move-in acceleration
Systems-platformVNET becomes a preferred high-density neutral infrastructure layerCATL/partners own meaningful energy/project assets; VNET recycles mature assetsQuantified CATL JV economics, high-density deployments, repeatable prefabricated design, multi-site compute-network contracts

8. The 'motherboard' analogy: useful, with one correction

The systems analogy is directionally useful: the industry is moving from thinking about an estate of buildings and racks toward a system whose components are orchestrated as one machine. The correction is that the data centre does not literally become a motherboard. Rather, the boundaries between chip package, board, rack, optical fabric, cooling plant, electrical system and campus are being pulled into a single optimization problem. That favors infrastructure operators that can standardize the physical layer and expose it as fungible compute capacity.

For VNET, this is potentially more important than simply owning more MW. A neutral operator with dense power, liquid cooling, fast construction, inter-campus optical connectivity and an energy partner can become an enabling layer beneath multiple chip and cloud ecosystems. That is a more defensible strategic position than commodity colocation, but VNET still has to prove it commercially.

9. What would materially strengthen or break the thesis

SignalStrengthens thesis if...Weakens thesis if...
CATL economicsNamed projects, capital split, storage/power ownership and return framework disclosedRelationship remains branding/technology cooperation with no economic contribution
Huawei/VNET deploymentCase study expands into paid multi-site AI-compute network deploymentsRemains a reference architecture without material revenue
Move-insUtilized MW accelerates through Q4 and 2027Capacity delivery substantially outruns equipment installation
FundingLong-tenor project debt / REIT recycling limits parent net-debt growthNew common equity or expensive converts are required at depressed valuation
Domestic compute supply960/950 roadmap ships at scale and software ecosystem improvesHBM/packaging/yield constraints delay customer deployments
EconomicsHigh-density capacity earns adequate returns after power/cooling capexDensity raises capex faster than pricing and utilization compensate

10. Revised thesis statement

11. Evidence ledger and source discipline

VNET Q2 2026 note (27 Aug working paper): operating metrics, delivery bridge, utilization, financing framework and valuation baseline.

VNET hypothetical funding model (Aug 2026): debt-heavy vs balanced vs CATL/asset-light sensitivities and ADS valuation mechanics.

Huawei, 17 Sep 2026: Ascend 960DT moved to Q1 2027; 960PR to Q3 2027; 4,096-card Ascend 960 supernode using NPO/UnifiedBus.

Huawei, 16 Sep 2026: 3D data centre in Wuhu; vertically separated cooling/IT/power/backup layers and factory-style prefabrication.

Huawei/VNET case study, 2026: AI-native compute WAN using 100/400/800GE evolution, network slicing and lossless transport.

VNET/CATL strategic cooperation filing, 13 Aug 2026: three-layer compute-energy ecosystem including gigawatt-scale facilities.

Reuters, 10 Sep and 17 Sep 2026: HBM supply/cost constraints and Huawei roadmap/demand context.