Earlier research and evidence history (5)
A more demanding measure of progress
The latest research separates announced capacity, energised space, installed hardware, useful compute, paid utilisation and cash returns. New model marketplaces may improve access, but reported platform launches are not evidence of increased revenue or margins.
EARLIER RESEARCH · 8 OCTOBER 2026
China is building more AI capacity. Who will use it — and who will earn a return?
This week's findings point to a growing divide between older data centres with spare space and newer facilities built for demanding AI workloads. They also raise fresh questions about how DeepSeek might finance expansion, whether VNET can offer more than space and electricity, and how far distributed computing can work in practice. The 5 October research into VNET's ownership, ByteDance links and Qingyang plans remains part of that story.
Read the consolidated research → · Evidence archive → · Architecture implications →
2 OCTOBER 2026 · HISTORICAL EVIDENCE REVIEW
Useful compute: Mixed / strengthening
Taixing adds operating-node evidence. DeepSeek’s software release remains constructive, with proof-of-concept benchmark conditions now made explicit. Monitoring, marketplaces, regional scheduling and nationwide portability remain distinct stages; company cash returns and valuation assumptions are unchanged.
1 OCTOBER 2026 · HISTORICAL EVIDENCE REVIEW
Software substitution strengthens, but premium-hardware scarcity still changes where demand is served.
What moved: DeepSeek has published Ascend-native versions of core kernel and communication infrastructure; Tencent is reported to have leased roughly 100,000 advanced AI chips from Oracle across Southeast Asia; and Micron says most of its 2027 HBM supply is already committed.
Thesis effect: Huawei/Ascend software compatibility strengthens; advanced-hardware access and HBM remain binding constraints; and the domestic-demand model gains a clearer offshore escape valve. No VNET, GDS or Kingsoft Cloud revenue assumption changes without company-specific evidence.
See the 1 October evidence, sources and limits →22 SEPTEMBER 2026 · THESIS UPDATE
Coordination evidence improves; commercial value is still being tested
| Research change | Direction | Next test |
|---|---|---|
| National / provincial coordination | Stronger implementation evidence: local aggregation and cross-region platforms. | Recurring dispatch, settlement, interoperability and revenue. |
| Compute × energy | Fujian–Ningxia adds a reported operational pilot. | MW/MWh, service levels, total cost and repeatability. |
| VNET–CATL governance | Ownership and effective agreements strengthen alignment. | Project terms, funding, customers and retained returns. |
| Alibaba / KC | Capacity ambition and reported lease retained with their different evidence states. | Delivery, company confirmation, utilisation and margins. |
These are evidence updates, not forecast increases. The research record below preserves earlier dated assessments.
Read the current research →A connected system is becoming more tangible. Profitable output is still unproven.
The October evidence improves the picture of policy coordination and named technical deployment, but the economic chain still requires powered capacity, installed accelerators, paying workloads and cash returns. US AI credit repricing is a comparative risk, not evidence of equivalent Chinese financing exposure.
22 SEPTEMBER 2026 · RESEARCH UPDATE
Alibaba sets a larger infrastructure horizon
Company announcement: Alibaba targets more than 20 GW of global cloud data-centre capacity by 2032. Its V900 roadmap adds a domestic supply milestone. Read the announcement and its limits.
Reported relationship: the KC–Alibaba compute lease deserves separate attention. Review the contract report and call evidence. Revenue and valuation assumptions remain unchanged.
CATL-affiliated VNET acquisition fully completed
The remaining 70% tranche closed on 21 September, following the first 30% on 24 August. Acquisition-completion risk is resolved; the commercial reason for accelerating the final closing remains undisclosed.
Thesis implication: stronger evidence of strategic follow-through. This secondary share transfer adds no new cash to VNET and does not confirm project funding or CATL-ecosystem customer contracts.
Timeline, sources and what changes → · HKEX completion filing
Who will finance the build-out — and will the returns justify it?
The 30 September review adds policy-bank support for the computing-power network, Token-linked credit underwriting, storage–compute co-design in Baotou and a reported 1,024-card Huawei SuperPoD operator deployment. These developments strengthen the architecture case without establishing national fungibility or durable cash returns.
Still unproven: routine national workload portability, cross-provider settlement, net economic advantage and durable company cash returns.
Can the national system create durable economic value?
The scorecard is a synthesis, not a news feed. The next section asks whether the apparent architectural progress survives the harder commercial test.
Evidence cut-off: 30 September
China can combine sustained AI adoption, domestic technology and coordinated infrastructure into productive compute capacity and durable company cash flows. This dashboard tests each link in that hypothesis.
Mixed
Manufacturing and system architecture are advancing. Profitable utilisation and cash returns remain the decisive open tests.
Software usability strengthens
Public Ascend infrastructure reduces a migration barrier, while the benchmark conditions and Taixing’s bounded operating role make the next implementation tests more precise.
Read the evidence →Usable capacity lags nominal capacity
July field reporting highlights low utilisation, incomplete hardware installation and heterogeneous scheduling friction in some projects. These are constraint signals, not a national census.
Read the constraint case →The six-layer headline
The 30 September reconciliation adds policy-bank support for the computing-power network, Token-linked credit underwriting, a construction-stage storage–compute project in Baotou and a reported 1,024-card Huawei SuperPoD operator deployment. Momentum tracks recent direction; overall progress asks how far each layer supports productive, economically sustainable AI infrastructure.
| Value-chain layer / companies | Headline | Momentum | Overall progress |
|---|---|---|---|
| AI demand & models DeepSeek · Alibaba / Qwen · ByteDance / Doubao · Xiaomi | Token-linked operating data is entering credit underwriting; this is financialisation evidence, not demand proof. 2 Oct · Unchanged. The reported Tencent–Oracle lease changes where demand may be served; it does not establish stronger realised demand for the tracked operators. | → WatchBeijing’s Token-loan pilot broadens the economic layer without establishing stronger paid demand. | MixedRecurring paid use, service quality, customer diversity and realised unit economics remain open tests. |
| Control plane Kingsoft Cloud · Alibaba Cloud · Huawei · ByteDance / Volcano Engine | Orchestration is becoming operational in bounded cross-region use; economics remain unproven. 2 Oct · Strengthened — bounded implementation. Taixing adds an operating node and a stated scheduling role. Actual cross-provider dispatch volumes and economics remain undisclosed. | ↑ Stronger evidenceShanghai adds price-aware cross-region operating evidence. | EmergingPortability, billable utilisation and profit per task need validation. |
| Compute / memory Huawei · Cambricon · Iluvatar · CXMT · YMTC | System-scale domestic compute is moving further into operator deployment; HBM remains a separate constraint. 2 Oct · Strengthened — software availability. Public Ascend kernels reduce a software barrier. DeepEP benchmark results use specially configured proof-of-concept hardware; HBM supply remains a separate constraint. | ↑ ImprovingReported China Mobile deployment adds commercialization evidence for a 1,024-card Huawei SuperPoD. | Partly evidencedUtilisation, software portability, yields and total economics remain open; the deployment does not establish HBM relief. |
| Interconnect Huawei · China Mobile · VNET / DYXnet · ByteDance | System links are advancing from cluster fabrics to physical cross-region network adaptation. 2 Oct · Unchanged. The benchmark qualification limits interpretation of Ascend communication results. No new measured cross-region network improvement is established. | ↑ ImprovingKaifeng fibre-route upgrades add physical network evidence. | Partly evidencedIndependent reliability, efficiency and interoperability tests remain. |
| AIDC VNET · GDS · Huawei · ByteDance | New projects increasingly integrate compute with power and storage, but operating returns remain unproven. 2 Oct · Strengthened — local operation. Taixing adds facility-level operating evidence. It does not demonstrate improved national utilisation or VNET/GDS earnings. | ↑ Stronger architectureBaotou adds construction-stage evidence of storage, compute and scheduling being designed together. | Conversion pendingPower-secured delivery, customer acceptance and site returns are decisive. |
| Power & energy CATL · Zhongheng · Huawei Digital Power · Haier | Compute–energy integration now includes both bounded workload movement and construction-stage physical co-design. 2 Oct · Insufficient new evidence. This review establishes no new measured power-cost saving, grid-access improvement or company-level energy return. | ↑ StrengthenedBaotou extends the evidence from operating coordination toward storage + compute co-design. | Execution pendingGrid access, delivered power cost, project utilisation and cash returns remain decisive. |
Editorial assessments based on the linked research. Green momentum can coexist with amber progress. No layer is marked red merely because an outcome is unproven; no new measured deterioration is established in this snapshot. Company names indicate roles, not stock ratings.
Bull and bear variables across those layers
Domestic manufacturing
CXMT reports fifth-generation DRAM mass production. The disclosed LPDDR5X products serve mobile devices; AI-server benefits remain indirect.
Evidence, company implications & next test
No immediate procurement benefit established for KC/Alibaba or move-in benefit for VNET/GDS.
Next test: Good-die yield, shipment volumes, customer qualification and server-memory products.
Inspect supporting research →Accelerators & HBM
Domestic system options are broadening, but the 1 October review keeps memory firmly in the constraint column: Micron says the vast majority of its 2027 HBM bit supply is already under agreement. This is global evidence, not a direct measure of Chinese HBM availability.
Evidence, company implications & next test
KC/Alibaba bear hardware and migration economics; VNET/GDS depend on customer equipment availability.
Next test: Qualified HBM volumes, packaging yields and independent cluster throughput.
Inspect supporting research →System efficiency
Huawei reports system-level efficiency gains, and DeepSeek’s new Ascend-native DeepGEMM/DeepEP/FlashMLA work adds direct software evidence that important workloads can be optimised closer to the hardware. Independent end-to-end economics remain unverified.
Evidence, company implications & next test
More throughput per grid connection may help sites; MW demand per workload may fall.
Next test: Production benchmarks including power, utilisation, availability and full system cost.
Inspect supporting research →Power & national coordination
Cross-region scheduling and compute–electricity coordination now have bounded operating examples. National-scale portability and economics remain unresolved.
Evidence, company implications & next test
GDS participated in Shanghai’s coordinated data-centre event, but the source does not establish that GDS supplied the transferred workload or earned incremental scheduler revenue. Physical operators still need power-secured delivery and paid utilisation.
Next test: recurring paid cross-region workloads, net electricity savings, service-level performance, settlement and site-level returns.
Inspect supporting research →Paid demand & pricing
The 26 September review sharpens the supply-versus-demand question: industry frameworks increasingly distinguish raw token volume from service-grade effective output. That is a measurement improvement, not evidence that demand or margins have strengthened.
Evidence, company implications & next test
Relevant to every operator: available capacity must become collected revenue.
Next test: Recurring paid consumption, customer diversity, service-level output, move-ins, utilisation, realised pricing and cost per useful task.
Inspect supporting research →Financing & cash returns
PSL eligibility for the computing-power network and Beijing’s Token-linked credit pilot strengthen the financing architecture. They do not yet establish lower funding costs, project drawdowns or better cash returns for covered companies.
Evidence, company implications & next test
VNET/GDS: cash return per site. KC/Alibaba: gross profit and cash per useful workload.
Next test: Drawdowns, customer prepayments, capex, receivables, free cash flow and returns on invested capital.
Inspect supporting research →20 September correction: KC’s existing 17 September thesis contained named Huawei-related deployment evidence omitted from the site. The control-plane capability case is better evidenced; no new margin or cash-return conclusion is implied. Read the correction →
How the assessment changed
- 1 October: DeepSeek’s Ascend-native infrastructure lowers one software-friction barrier, while the reported Tencent–Oracle lease and Micron HBM tightness show that premium-compute and memory constraints remain economically consequential; domestic IDC demand is not identical to total Chinese-company AI demand.
- 30 September: policy finance moves closer to the compute architecture; Token-linked credit, storage–compute co-design and a 1,024-card SuperPoD deployment strengthen the system case while company-level cash-return assumptions remain unchanged.
- 26 September (covering 25–26 Sep): bounded power-aware workload movement, transport-layer build-out and Horinger delivery evidence strengthen execution while effective-token frameworks sharpen the economics test; no valuation assumptions change.
- 24 September: operating cross-provincial aggregation evidence strengthens the coordination branch, while field evidence on utilisation and portability makes the fungibility/economics test more demanding.
- 21 September: national coordination and compute–electricity standards strengthen the system-integration branch; GDS gains an evidenced provincial-scheduling edge. Commercial economics remain unproven.
- 20 September: CXMT manufacturing evidence improves; HBM and covered-company cash-return assessments remain unchanged. Macro demand risk receives closer scrutiny.
- 19 September: Huawei efficiency claims strengthened the system-efficiency case; independent economics remained unproven.
- 18 September: SuperPoD size comparison raised an architectural question; chip count alone did not establish regression.
Assessments are editorial judgements, not probabilities or investment recommendations. This is the latest reviewed snapshot; new evidence changes these views after review.
Follow the dated supporting evidence →The architecture only matters if it improves the economics of useful AI.
If the economics do not improve, system integration can still be technically interesting without becoming economically valuable. The final section therefore looks for evidence that could falsify the thesis.
The economic audit rejects the easy version of the thesis. Domestic compute is not automatically cheap, HBM scarcity can raise accelerator cost, and every architectural workaround carries a complexity tax. The more defensible proposition is that whole-system optimisation can reduce the amount of scarce resource required per task, raise utilisation of installed capacity and route workloads toward the cheapest adequate combination of hardware, memory, network, location and power.
25–26 September evidence: a new industry “token factory” framework separately tracks service-grade sellable token capacity, token energy efficiency, cost per million Good Tokens and service quality. We treat this as evidence that industry measurement is moving in the same direction as the thesis — away from raw capacity — not as an official standard or proof of attractive project returns. See the source and limits →
What would prove the business model works?
MW ≠ AI MW
Power becomes economically useful AI capacity only when it can support the required rack density, cooling, network, reliability, accelerator delivery and customer workload. This is the more precise bridge from the architecture to VNET, GDS, power electronics and the CATL ecosystem.
Installed compute ≠ useful compute
Fragmented capacity gains value when customers can discover, schedule and consume it. The control plane does not need to make every accelerator identical; it needs to hide enough difference to place workloads efficiently across internally optimised resource pools.
China is optimising the complete system around the resources it can obtain at scale.
The emerging architecture combines specialised compute pools, tiered memory and storage, high-bandwidth interconnect, cloud orchestration, AI-specific data centres, geographic workload placement and — at the frontier — coordination with the electricity system. Whether this closes the remaining cost and performance disadvantage in frontier components is not yet known. The investment opportunity lies partly in identifying which bottlenecks become more valuable while China tries.
What evidence would challenge our view?
Cloud and scheduler abstraction fail to hide enough hardware fragmentation for installed domestic compute to become a useful set of resource pools.
Communication, failure and synchronisation overhead erase the benefits of larger domestic systems.
HBM bandwidth/capacity shortages cannot be economically mitigated by hierarchy or pooling.
High-density power/cooling/network capabilities become commoditised and generic MW is easily upgraded.
Storage/grid coordination adds complexity without meaningful cost, reliability or deployment benefits.
The complete Chinese system consumes too much capital, power or engineering effort for the output produced.