No winner. Compare the evidence.
This page does not score the companies or collapse different business models into a single ranking. It compares the same questions side-by-side and separates reported facts from emerging thesis evidence. Figures are primarily Q2 2026 snapshots and company definitions are not always directly comparable.
At a glance
VNET
GDS
Side-by-side evidence
| Question | VNET | GDS |
|---|---|---|
| Current demand signal | Reported 862 MW H1 wholesale order wins, including a 345 MW Q2 order from a leading cloud service provider. Very rapid order accumulation; customer identities remain partly undisclosed. | Reported ~470 MW H1 bookings; FY2026 bookings target raised to 1 GW. Customers also requested 600 MW of reservations. Management says bookings are contractual take-or-pay commitments; reservations are future capacity held within the sales framework. |
| Installed / operating base | 1,007 MW wholesale capacity in service at Q2; 585 MW under construction. | 684,977 sqm area in service; 170,355 sqm under construction at Q2. GDS also described more than 2 GW of binding commitments across its broader pipeline at mid-year. |
| Utilisation / move-in | 744 MW utilised; 73.9% wholesale utilisation. Mature wholesale utilisation 92.5%; ramp-up 36.6%. Large recent capacity additions depress the blended utilisation rate while they ramp. | 79.2% area utilisation at Q2. H1 net move-in was 145 MW; management forecast 235 MW for FY2026 and more than double that level in 2027. The key test is conversion of the large backlog into billable move-in. |
| Committed demand | 970 MW committed against in-service wholesale capacity; 96.3% commitment rate. | 757 MW backlog at mid-year, up from 450 MW at the start of 2026; >2 GW total binding commitments plus 600 MW reservations disclosed by management. |
| Revenue growth | Q2 revenue RMB2.78bn, +14.2% YoY. Wholesale revenue +29.3%. | Q2 revenue RMB3.09bn, +6.5% YoY. Reported growth currently lags the acceleration in bookings because move-in takes time. |
| EBITDA / margin | Adjusted EBITDA RMB918m, +25.4%; margin 33.0% versus 30.1% a year earlier. Cash gross margin was 41.8%, down from 43.6%. | Adjusted EBITDA RMB1.406bn, +2.5%; margin 45.5% versus 47.3% a year earlier. Utility costs pressured gross margin; future pricing and returns on new-market capacity matter. |
| Capital intensity / funding | Q2 cash + restricted cash + short-term investments RMB7.21bn; short-term debt RMB4.18bn and long-term debt RMB19.24bn. Lower visibility Headline capex does not by itself reveal VNET's ultimate corporate funding burden. The funding mix for the accelerated build remains a material clarity gap. See funding comparison ↓ | Management raised FY2026 paid capex guidance to RMB10bn and described new-project funding around 60% debt / 40% equity at project level. Comparatively visible Project finance and asset recycling give investors a clearer disclosed pathway for funding the backlog. See funding comparison ↓ |
| Geography / national compute network | Large Greater Beijing / northern footprint plus expanding Inner Mongolia and other resource reserves; network heritage through DYXnet creates an additional connectivity angle. Evidence of direct participation in national/provincial compute scheduling remains a research target. | Confirmed participation in the launch of Hebei's integrated computing-power platform, described by GDS as the first formally operational provincial compute monitoring/scheduling platform. This is an evidenced orchestration connection, but workload routing and commercial economics are not yet disclosed. |
| Power / energy differentiation | CATL strategic cooperation explicitly targets compute-energy integration, including gigawatt-scale facilities, distributed compute-energy networks and zero-carbon infrastructure. Strategically distinctive; project-level economics and deployment remain the next evidence test. | Power-secured campus development and energy cost management are core operating requirements. No equivalent CATL-style strategic compute-energy relationship is established in the evidence reviewed for this page. |
| Customer visibility | Strong disclosed order scale, but several large wholesale customers are unnamed. Do not attribute unnamed MW to ByteDance or another hyperscaler without evidence. | Management says H1 bookings included significant new business from each of its three largest hyperscale customers and emerging AI leaders. Customer names and workload mix remain only partly disclosed. |
| What currently needs proving | Convert exceptional order intake into timely delivery, move-in, utilisation and cash returns; show what CATL integration changes commercially; establish whether network/control-plane capabilities create value beyond facility operations. | Convert backlog/reservations into move-ins at attractive pricing and returns; demonstrate economics of the new-market buildout; show whether provincial scheduling integration becomes commercially meaningful. |
Funding the build-out
The key comparison is not headline capex alone. It is how much infrastructure each company must deliver, how that build is financed, and how much of the capital burden ultimately sits with the listed company.
GDS — funding pathway comparatively visible
Current evidence: management has discussed project-level funding, including an indicative 60% debt / 40% equity structure for new projects, alongside asset recycling and other financing channels.
Why it matters: investors can more directly trace the path from committed demand to development capital, financing, move-in and eventual capital recycling.
Higher visibility This does not remove execution or leverage risk; it means the funding architecture is comparatively easier to assess from disclosed information.
VNET — funding need visible; funding architecture less clear
Current evidence: VNET has a very large delivery programme, existing balance-sheet resources and asset-recycling mechanisms, but the ultimate funding mix for the accelerated AI build-out is not yet disclosed with comparable clarity.
Important distinction: headline capex should not automatically be read as the amount VNET's listed-company balance sheet must ultimately fund. Customer funding/prepayments, project or JV capital, strategic capital and asset recycling could reduce the corporate burden where such structures are actually used.
Clarity gap We cannot currently quantify how much of the forthcoming build will be funded directly by VNET versus other sources.
| Funding question | VNET | GDS |
|---|---|---|
| Headline build requirement | Large and accelerating; order growth implies substantial delivery requirements. | Large backlog and reservation pipeline require sustained development capital. |
| Funding pathway visibility | Lower. Multiple potential burden reducers exist, but the mix and scale are not sufficiently disclosed to quantify. | Comparatively higher. Project financing and asset-recycling framework are more explicitly discussed. |
| Strategic funding optionality | CATL is potentially important, but any funding contribution beyond the disclosed strategic investment remains hypothetical. | Funding case relies more visibly on disclosed project-level finance and capital recycling rather than an equivalent CATL-style hypothesis. |
| What would close the gap? | Disclosure of project/JV funding, customer prepayments, CATL participation, project debt, cash timing and parent-equity contribution. | Continued disclosure showing financing closes and asset recycling keeping pace with the deployment programme. |
Technology & ecosystem relationships
These relationships matter because an AIDC operator can gain more than rack demand from the surrounding AI stack: networking, accelerator ecosystems, memory supply, cloud customers and power integration can affect how quickly capacity is deployed and how useful it is. The table deliberately distinguishes a documented relationship from a research target.
| Ecosystem node | VNET | GDS |
|---|---|---|
| Huawei | Direct evidence Documented technology relationship. Huawei published a 2026 VNET case study describing an AI-enabled hyper-connected computing network designed to connect enterprise endpoints and intelligent-computing centres. Earlier Huawei material also documents data-centre cooperation with VNET. This is stronger evidence than simple ecosystem adjacency. The next test is whether the relationship expands into material AI-cluster, cross-campus networking, SuperPoD/Peerium or 3D-AIDC deployments and measurable commercial contribution. | Research target No comparably material, current Huawei–GDS AI-infrastructure relationship has been established in the evidence reviewed for this comparison. That does not mean no Huawei equipment or interaction exists. It means this page will not infer a strategic relationship without company-specific evidence. |
| CXMT / domestic memory | Not established No material direct VNET–CXMT commercial or deployment relationship has been established in the evidence reviewed. CXMT is relevant upstream because its DDR5 product portfolio includes server applications, but supply-chain relevance is not evidence of an operator relationship. | Not established No material direct GDS–CXMT commercial or deployment relationship has been established in the evidence reviewed. A future named memory/AI-cluster deployment would therefore be new evidence, not confirmation of an assumption already embedded here. |
| Power / storage partner | Strategic CATL is a clear differentiator in the current evidence set. The relationship combines strategic investment with stated compute-energy cooperation. The unresolved issue is execution: project structure, deployed capacity, economics and whether storage/power integration improves AIDC delivery or returns. | No equivalent identified No equivalent CATL-style strategic compute-energy relationship is established here. GDS still has substantial power procurement, energy-efficiency and campus-infrastructure requirements; the distinction is the evidence for a named strategic partner. |
| Cloud / hyperscale ecosystem | Large wholesale orders provide strong demand evidence, while customer disclosure is incomplete. Volcano Engine has a confirmed cooperation route through VNET-owned DYXnet across cloud, AI, security and network bandwidth. This is an evidenced ecosystem path, not evidence that ByteDance accounts for VNET's unnamed wholesale MW. | Long-standing hyperscale relationships are directly documented, including a strategic Alibaba framework, while 2026 management commentary points to substantial bookings from its largest hyperscale customers and emerging AI leaders. The comparison should continue to separate named historical relationships from unnamed current bookings. |
| Government / compute orchestration | Research target VNET's geography and network assets make national-compute-network integration relevant, but direct provincial scheduling evidence remains to be established. | Direct evidence Hebei platform participation is documented. It provides an evidenced connection between GDS and a provincial compute monitoring/scheduling layer. Actual routed workloads, utilisation impact and commercial economics remain unproven. |
Where the comparison is genuinely different
VNET: the distinctive questions
- Can the 2026 order surge be delivered without returns being diluted by the capital required?
- Does CATL create a real compute-energy and financing advantage, or remain principally strategic optionality?
- Can VNET's network/DYXnet heritage become useful as AI workloads spread across sites?
- How quickly does new wholesale capacity move from committed to utilised?
GDS: the distinctive questions
- How quickly does the 757 MW backlog convert into billable move-ins?
- Can new-market capacity sustain pricing and the returns assumed in project financing?
- Does participation in provincial compute scheduling lead to actual workload flow or commercial advantage?
- Can asset recycling/project finance keep pace with a sharply larger deployment pipeline?
What would change this comparison?
This is the part the daily thesis engine should update. A comparison dimension changes only when new evidence changes the underlying facts—not because one company's share price moves.
| Evidence to watch | Why it matters | Where it changes the page |
|---|---|---|
| Named AI / hyperscale customer or workload | Improves visibility into demand quality and concentration. | Demand signal · customer visibility |
| Move-in / utilisation acceleration | Shows contracted demand becoming revenue-producing infrastructure. | Delivery · economics |
| Pricing / return disclosure | Separates impressive MW growth from economic value creation. | Margins · cash returns |
| New financing / asset recycling structure | Changes the amount of growth the parent can fund and who bears capital risk. | Funding · balance-sheet risk |
| Compute-energy deployment | Tests whether power/storage integration becomes an operating advantage. | VNET differentiation |
| Provincial/national scheduling usage | Tests whether an IDC becomes a node in an orchestrated compute network rather than only leased capacity. | GDS / VNET orchestration evidence |
| Huawei deployment / commercial disclosure | Tests whether VNET's documented networking relationship expands deeper into the AI data-centre stack, and whether comparable evidence emerges for GDS. | Technology ecosystem · networking |
| CXMT or other domestic-memory deployment | Would establish a direct link between an AIDC operator and the domestic memory layer rather than inferred supply-chain adjacency. | Technology ecosystem · compute/memory |
Primary evidence
VNET Q2 2026 results — 18 Aug 2026
GDS Q2 2026 results, presentation and transcript
GDS — Hebei integrated computing-power platform — 14 Sep 2026
Huawei — VNET AI-enabled hyper-connected computing network case study — Jun 2026
CXMT — DDR5 / server memory product portfolio
GDS — strategic MoU with Alibaba