Follow the evidence as the thesis evolves.

Taixing’s government confirms that the first-phase Zhongke compute node began operating on 30 September. It describes the facility as a scheduling node within the Zhongke compute system, with a planned Jiangsu management platform and an ambition to integrate with provincial coordination. The launch also included compute-service and ecosystem signings.

Evidence boundary: this establishes an operating facility and a stated coordination role. It does not disclose paid utilisation, cross-provider workload volumes, settlement economics or nationwide portability.

Reviewing the documentation behind the 30 September DeepSeek releases adds an important limit to yesterday’s interpretation. DeepEP-Ascend says its performance measurements used a proof-of-concept hardware development kit with additional manual configuration, not a publicly distributed release. Some communication functions remain work in progress.

What changes: public software availability is real progress, while reproducible performance on broadly available customer systems remains a separate test. This is a source qualification identified on 2 October, not another software launch.

DeepSeek has released Ascend-native versions of core infrastructure including DeepGEMM, DeepEP and FlashMLA components. The public repositories preserve familiar APIs while targeting Huawei Ascend 950-class hardware; DeepEP-Ascend reports strong measured communication bandwidth on Ascend 950DT, while the FlashMLA Ascend implementation reports high utilisation of theoretical peak for specific DeepSeek V4.1 attention kernels. This is direct, inspectable software evidence rather than a vendor roadmap. Qualification added 2 October: DeepEP’s reported measurements used a specially configured proof-of-concept development kit, not a publicly distributed release; they should not be treated as generally reproducible customer performance.

The Financial Times, followed by Reuters, reports that Tencent agreed a five-year Oracle lease worth about US$7bn for access to roughly 100,000 advanced AI chips across Oracle data centres in Southeast Asia. The chips were described as unavailable in China. Neither Tencent nor Oracle has publicly disclosed the reported transaction, so we treat the commercial details as high-quality reported evidence rather than a company-confirmed contract.

Micron’s 30 September results and earnings call point to exceptionally tight AI-memory demand. Management said the vast majority of its calendar-2027 HBM bit supply is already under agreement, with significant year-on-year price increases, while broader memory supply remains constrained. This is global evidence rather than a direct measure of Chinese HBM availability, but it strengthens the case that memory remains a valuable system bottleneck rather than a solved component problem.

On 29 September, the People’s Bank of China expanded pledged supplementary lending (PSL) support to the “six networks”, explicitly including the computing-power network (算力网) alongside new power grids and next-generation communications networks. The one-year PSL rate was cut from 1.75% to 1.5%. The PBOC also raised its technology-innovation and equipment-upgrade relending quota from RMB1.2tn to RMB1.4tn and lifted the support ratio from 60% to 100%.

Evidence surfaced in the 30 September sweep of Beijing’s first “亦企Token（词元）贷” programme, originally reported on 14 September. Six banks provided nearly RMB2bn of credit. The framework can use token production/consumption and compute-service contract value alongside technical and business factors. This is a local financing pilot, not a national standard, but it shows measurable AI activity beginning to enter credit assessment.

A HyperStrong / 海博思创 project in Baotou has begun construction using grid-side storage coupled with distributed intelligent-compute infrastructure. Xinhua says the design links energy supply, storage peak-shaving and compute scheduling and describes a dynamic “算随电走、电随算调” mechanism. This is construction-stage architecture evidence, not proof of operating economics.

China Mobile Beijing and China Mobile Cloud reported deployment of a Huawei Ascend 950DT 1,024-card SuperPoD at the Mobile Cloud Beijing node in Changping. Reported specifications are 1 EFLOPS FP8, 256TB unified memory and microsecond-scale latency, consistent with Huawei’s published 1,024-card Atlas 950 SuperPoD architecture. This strengthens commercialization evidence; utilisation, software portability and economics remain open.

State Grid Shanghai reports that 16 data centres took part in a coordinated two-hour exercise with a maximum load reduction of 97.8 MW. The process combined flexible-load adjustment, backup-power switching and cross-region compute transfer. A linked three-day electricity-price-aware inference test moved work among locations including Shanghai, Anhui and Xinjiang, handling more than 280,000 requests and more than 33 million tokens; the reported electricity cost per token fell by about 31%. A separate Zhejiang test moved part of a Jiaxing workload to Qinghai during peak periods.

The National Data Administration says China had 2.45 million PFLOPS (FP16) of intelligent compute at end-July, of which 1.45 million PFLOPS — about 60% — was inside the national monitoring scope. CCTV nevertheless says monitoring and cross-provincial invocation remain incomplete and many resources are still “islands”. Kaifeng fibre routes are being shortened for compute latency. A CAICT forum and its 2026 network-transport report add the next layer: the industry is still working to expand 1ms metro-compute coverage and pilot task-based scheduling and compute-network agents. National visibility should therefore not be read as nationally movable compute.

SINNET / 光环新网 (300383.SZ) approved up to RMB1.265bn of additional investment in its Horinger intelligent-compute project, taking planned investment to about RMB2.5bn and planned IT load to 90–120MW. The filing package also says the external 110kV supply plan is not yet approved and bank financing may fund up to 70%. This remains a separate SINNET project from GDS’s own Horinger programme: useful location-level corroboration, not evidence of a shared customer or project.

A 26 September China Business Journal report, drawing on the newly released China Token Factory Development White Paper 2026 and operator activity, frames the industry’s problem as moving from “having compute” to producing stable, competitively priced effective tokens. The underlying white paper proposes service-grade sellable token capacity, token-energy efficiency, cost per million “Good Tokens” and service quality as separate measures. This is an industry framework, not an official national standard and not proof of returns, but it aligns closely with our governing test: nameplate capacity only matters when it becomes usable, saleable output at an attractive total cost.

At HUAWEI CONNECT 2026 on 17–19 September, Huawei disclosed a system architecture explicitly designed around agentic workloads: UnifiedBus links CPUs, NPUs, memory and SSDs; OceanStor M900 provides PB-scale shared KV-cache storage; and the architecture scales from SuperPoDs toward very large clusters. Huawei said agent workloads increase CPU–NPU interaction and KV-cache pressure, making communication, memory and storage increasingly important to achieved compute utilisation.

Primary company disclosure: Huawei’s 17 September UnifiedBus architecture release describes agent-driven CPU/NPU interaction, KV-cache growth, tiered memory/storage and interconnect bottlenecks. Huawei separately presented OceanStor M900 and its Agentic SuperCluster at HUAWEI CONNECT. These are company technical claims, not independent measurements of commercial economics.

Huawei · UnifiedBus architecture · 17 Sep 2026 ↗ Huawei Connect · M900 / AI infrastructure ↗

The demand layer now connects more explicitly to the physical architecture. Longer-lived, stateful agent workflows can increase pressure on memory, KV cache, interconnect and scheduling. The thesis therefore tests models → agents → workload characteristics → memory/interconnect → scheduling → geography → power, rather than treating each layer as an independent build-out.

The competitive unit may be moving upward from accelerator nameplate performance toward useful system throughput: completed work relative to capital, power, latency, reliability and software complexity. This gives a causal explanation for why compute, storage, networking and orchestration are increasingly being designed together.

Extends / strengthens. It strengthens the system-level architecture branch and the importance of interconnect, memory and control-plane optimisation. It does not by itself prove that Huawei’s claimed performance is achieved in production or that agent demand will generate attractive infrastructure returns.

Incremental agent-driven infrastructure demand at commercial scale; portability across heterogeneous accelerators; achieved utilisation and latency outside vendor tests; and whether system-level gains reduce cost per useful completed task.

Look for independent production deployments reporting workload mix, KV-cache/storage demand, achieved throughput, accelerator utilisation, latency, power and cost per completed task. Evidence that agent workloads remain modest or that integration gains disappear outside tightly coupled vendor stacks would weaken this branch.