Kingsoft Cloud (KC): the heterogeneous-compute thesis

Executive summary

The KC thesis has changed materially. The original investment case was that Kingsoft Cloud was a subscale Chinese public-cloud operator receiving an unusually powerful AI demand tailwind from Xiaomi, WPS and external customers. Q2 2026 validated that demand and the first operating-profit inflection. The stronger thesis now is that KC is becoming an operating and abstraction layer for heterogeneous Chinese compute: procuring and integrating multiple accelerator types, scheduling them across clouds and regions, exposing them through MaaS/AIOS, and converting scarce hardware into usable tokens and enterprise AI applications.

The most important new evidence is not theoretical. In September 2026, Southern Medical University Zhujiang Hospital disclosed a production deployment jointly built with Huawei and Kingsoft Cloud. Huawei supplies the localized Kunpeng/Ascend hardware stack; Kingsoft Cloud supplies the AIOS software layer that uniformly manages and finely schedules Ascend NPUs alongside other GPUs. This is the cleanest direct Huawei-KC partnership structure we have found, but it is no longer isolated: the 2025 Kingsoft Government AI Appliance uses Shenzhou Kuntai's Kunpeng + Ascend compute beneath Kingsoft Cloud's AI platform and WPS applications, while Qingyang demonstrates KC scheduling compute across clouds and regions.

The evidence now supports a three-layer business model: (1) KC-owned AI compute, the most capital-intensive layer; (2) KC-managed heterogeneous compute, where customer or partner hardware sits beneath KC AIOS/StarFlow; and (3) distributed compute orchestration, where KC schedules capacity across clouds and regions it does not necessarily own. The further the mix migrates from layer 1 toward layers 2-3, the more AI activity KC can potentially control per RMB of balance-sheet capital.

We retain the discipline that this does not automatically solve the equity problem. Q2 infrastructure investment including leases was RMB3.3bn and FY26 investment remains roughly RMB15bn; the company is still financing an asset-intensive transition with only RMB4.7bn of cash/restricted cash at June. The central valuation therefore remains driven by whether high utilization, price increases, MaaS mix and project/lease financing turn the AI build into durable per-share free cash flow.

1. What has changed since the Q2 note

The Q2 note correctly identified demand, utilization and financing as the three decisive variables. Since then, the evidence has become more specific in two areas: (1) Huawei and domestic-compute systems are moving from 'available alternatives' toward a scaled ecosystem; and (2) KC has now demonstrated that its software layer can sit directly above Huawei Ascend hardware while remaining heterogeneous.

2. The core industrial thesis: China's AI stack is becoming heterogeneous

Huawei's 17 September disclosures strengthen the probability that China can scale AI without first matching Nvidia chip-for-chip. Huawei is accelerating Ascend 960DT to Q1 2027, says domestic demand for its AI equipment exceeds supply, and is scaling system architectures toward extremely large processor counts. CANN is now in sustained community-driven open-source development; Huawei says Ascend supports more than 90 leading third-party open-source projects and more than 40 models have been natively pre-trained on Ascend.

For KC, the key consequence is fragmentation. Chinese customers can increasingly deploy Nvidia where available, Ascend, other domestic accelerators, CPU/ARM infrastructure and geographically separated compute pools. The economic problem shifts from merely obtaining chips to achieving utilization, workload portability, memory allocation, scheduling, model compatibility and token economics across a mixed estate.

That is exactly the layer KC increasingly describes itself as operating in: AI engineering, RDMA networking, high-performance storage, MaaS, multi-cloud scheduling and AIOS.

3. Huawei-KC partnership structures: what is actually confirmed

3.1 Zhujiang Hospital: hardware below, KC orchestration above

On 10 September 2026, Southern Medical University Zhujiang Hospital disclosed a localized AI-native heterogeneous intelligent-compute platform jointly constructed by the hospital, Huawei and Kingsoft Cloud. The division of labour is unusually revealing: Huawei uses Kunpeng and Ascend to provide the localized hardware cluster; KC provides its one-stop AIOS platform to uniformly manage and finely schedule heterogeneous compute.

The hospital states that KC's AIOS supports Ascend NPU and other GPU resources with MB-level memory allocation and workload-aware scheduling. During the day, clinical workloads receive priority; at night idle resources are automatically redirected to model training and research computing. This is not merely a compatibility certification. It is a production example of KC acting as the control/orchestration layer above Huawei hardware.

A government procurement notice also shows Beijing Kingsoft Cloud Network Technology won the hospital's Children's Medical Center informatization project in April 2026, ranking first with a 96.40 score. The public sources do not establish that this procurement contract alone funded the later Huawei/KC AI platform, so the two should not be mechanically equated.

3.2 Qingyang: KC as a cross-cloud / cross-region scheduling layer

In December 2025, Qingyang municipal reporting described KC's 100,000-PFlops integrated compute scheduling platform as a core scheduling system for the Gansu-Qingyang national hub. The platform provides resource management, multi-cloud management, global scheduling, unified operations and AI applications. Critically, the multi-cloud layer had already connected to Huawei Cloud and Alibaba Cloud, while KC's east-west compute pools were managed under unified demand and resource scheduling.

This is strategically important because it shows that KC does not need to own every accelerator or every data centre to capture value. A credible alternative business model is to become the software, scheduling and service layer across third-party compute pools.

3.3 Kunpeng in KC's own public cloud

KC's current product catalogue lists its AC6 ARM compute instance using Huawei's Kunpeng 950, alongside the earlier Kunpeng 920 generation. This is direct evidence that Huawei compute is entering KC's own service catalogue, not only customer-private deployments.

3.4 Kingsoft group / WPS ecosystem relationship

The wider Kingsoft ecosystem has deeper Huawei links than KC alone. Huawei Cloud says it and Kingsoft collaboratively developed an operational-excellence framework for WPS covering heterogeneous compute, deterministic operations and FinOps. Separately, WPS AI has previously demonstrated large-model training and inference work on Ascend. This is strategically supportive because WPS is simultaneously one of KC's most important ecosystem workloads, but group-level Huawei cooperation must not be booked as KC revenue unless KC is explicitly named.

3.5 Government AI appliance: the same architecture appears outside healthcare

In July 2025, Kingsoft Cloud and Shenzhou Kuntai launched the Kingsoft Government AI Appliance. Shenzhou Kuntai states that the compute foundation uses Kunpeng + Ascend, while Kingsoft Cloud contributes intelligent-compute services, platform services and large-model services and WPS contributes the government-office model/application layer. The product is a private/local deployment architecture rather than ordinary public-cloud consumption.

This is strategically important because it predates Zhujiang Hospital and shows essentially the same stack through an ecosystem integrator: Huawei technology at the compute substrate, Kingsoft Cloud at the AI platform/control layer, and an industry-specific application above it. Public sources describe joint model optimization and use of Shenzhou Digital's national delivery/operations network, and indicate intended expansion into tax, finance and other government verticals. This supports repeatability, but does not establish a broad Huawei-KC corporate strategic agreement.

3.6 A three-layer KC business model

The expanded evidence suggests KC should no longer be modelled as one homogeneous cloud business. Layer 1 is owned/leased AI compute: KC procures infrastructure and monetizes GPU/NPU hours, producing the strongest capital requirement. Layer 2 is managed heterogeneous compute: customers, Huawei-ecosystem partners or other owners provide the hardware while KC supplies AIOS/StarFlow, scheduling, storage, networking and model services. Layer 3 is distributed orchestration: KC's Qingyang platform demonstrates unified resource management, multi-cloud connectivity and millisecond-level scheduling across eastern and western resource pools.

The valuation consequence is asymmetric. Layer 1 can generate substantial EBITDA but requires continuing hardware finance. Layers 2-3 could allow KC to capture software, orchestration and MaaS economics without financing every accelerator itself. The core upside question is therefore not simply how quickly AI revenue grows, but whether the incremental RMB of AI economic activity increasingly sits on third-party or customer-owned infrastructure while KC retains the control, software and service economics.

3.7 Huawei relationship: competitor, supplier and ecosystem substrate

No broad Huawei-Kingsoft Cloud strategic partnership agreement has been identified in the sources reviewed. The evidence instead points to a more nuanced operational relationship. Huawei Cloud competes with KC's cloud services; Kunpeng/Ascend can supply the compute substrate under KC software; Huawei Cloud can itself be one of the clouds connected to KC's scheduling layer; and Huawei-ecosystem server vendors can package KC software into private AI appliances. The repeatable pattern is therefore better described as a 'Huawei-compute / Kingsoft-software' architecture than as a formal corporate alliance.

4. KC's product architecture now maps unusually well to the problem

5. DeepSeek and open-model economics

KC supported DeepSeek R1/V3 in public-cloud and government-cloud scenarios in early 2025 and has continued to emphasize rapid onboarding of leading open models. The relevance is not simply that DeepSeek is popular. Efficient open models lower the cost of inference and broaden the customer base that can consume AI, while KC's model-neutral stance avoids the conflict faced by full-stack clouds that have incentives to promote proprietary models.

Management's Q2 commentary is consistent with this: StarFlow supported 120 models and more than 230 enterprise customers, while MaaS revenue increased more than twelve-fold quarter on quarter. KC explicitly argued that not owning a proprietary foundation model allows its sales organization to supply the models customers prefer.

The counterargument is important. Model efficiency can reduce compute required per task, and hyperscalers can vertically integrate. The bull thesis therefore requires Jevons-style demand expansion and/or KC taking a larger value share through MaaS, orchestration and enterprise AI rather than merely selling GPU-hours.

6. Xiaomi is simultaneously anchor customer, distribution channel and concentration risk

Revenue from the Xiaomi and Kingsoft ecosystem was RMB810m in Q2, 26% of total revenue, while first-half public-cloud revenue from the ecosystem grew 54% YoY. Shareholders approved higher connected-transaction caps, with the combined 2026-27 caps reaching RMB10bn. Xiaomi's expansion across phones, vehicles, home devices and AI creates a large internal demand surface for cloud training, inference, data and agent workloads.

But the correct valuation treatment is not to assume every edge-AI success becomes KC cloud revenue. More capable on-device inference can displace some cloud calls, Xiaomi can negotiate aggressively as an anchor customer, and related-party financing/working-capital structures can obscure underlying cash economics. The key metric is non-ecosystem growth: KC reported top-five non-ecosystem customer revenue +51% YoY in Q2.

7. Demand and pricing have crossed an important threshold

Q2 AI cloud gross billings rose 82% YoY to RMB1.327bn and represented 56% of public-cloud revenue. Management said sold compute capacity was effectively fully utilized and normally contracted long term. In July, KC implemented list-price increases of roughly 15-50% for AI-compute products and 30-50% for file storage. Q3 is therefore the first quarter in which the market can test whether scarcity translates into realized pricing rather than only published list prices.

This matters more than headline revenue growth. A cloud operator that can pass hardware scarcity and financing cost through to customers has very different economics from one forced to absorb those costs.

The Qingyang evidence also makes the scale of this architecture harder to dismiss. Municipal reporting says Kingsoft Cloud completed more than RMB5bn of fixed-asset investment and nearly RMB2bn of revenue in Qingyang during 2025, with cumulative local investment approaching RMB10bn; for 2026 it planned more than RMB6bn of additional investment and a single-enterprise compute footprint above 100,000P. Separately, Qingyang's 2026 government work report says the region hosts DeepSeek, Kimi, MiniMax and other model training/inference workloads. These are regional ecosystem facts, not proof that each workload is KC revenue, but they strengthen the strategic relevance of KC's scheduling position.

8. The financing problem remains the principal equity risk

KC's Q2 infrastructure investment including leased assets was RMB3.3bn; H1 was roughly RMB6.2bn and FY26 guidance was around RMB15bn. June cash and restricted cash were approximately RMB4.7bn. The company is therefore attempting to scale AI infrastructure much faster than internally generated accounting profit can fund it.

The financing toolkit is broader than ordinary corporate debt: equipment and finance leases, supplier terms, related-party balances, project-level lending, customer prepayments and potentially policy-linked compute financing. Our broader China AIDC work found banks increasingly willing to lend against contracted compute-service cash flows and equipment. This improves the plausibility of non-equity funding, but KC has not yet disclosed enough project-level terms to remove dilution and leverage from the bear case.

The critical distinction is economic net debt. Lease-funded servers and related-party financing still represent claims on future cash even when they do not appear as conventional bank borrowings. The valuation model therefore continues to capitalize these obligations rather than treating lease financing as free capital.

9. Evidence ladder

10. Valuation framework: value capture, not simply revenue growth

The previous $17.47 central value was built around FY27 revenue of RMB15.85bn, normalized EBITDA of RMB5.71bn and probability-weighted economic net debt of roughly RMB12.3bn. The expanded Huawei/AIOS evidence strengthens the strategic quality and repeatability of the software/orchestration layer, but it does not justify mechanically raising near-term earnings. The current scenario model therefore remains the primary valuation, while Thesis 3.1 adds a sum-of-the-parts roadmap for the point at which KC discloses enough segment economics to distinguish owned compute from managed/orchestration/MaaS activity.

We therefore use four cases. The bear case assumes that KC remains predominantly a capital-intensive compute reseller: growth continues, but margins are lower, financing claims remain high and dilution/value leakage persists. The base case assumes Q2 operating leverage persists and heterogeneous-compute orchestration becomes commercially meaningful without radically changing the earnings model. The bull case assumes MaaS/AIOS/agent mix increases margins and financing becomes increasingly project/lease/customer backed. The strategic-upside case assumes KC becomes an important neutral control plane for domestic heterogeneous compute and captures materially more software-like economics.

Arithmetic corrected on restoration: (EBITDA × EV/EBITDA − economic net debt) ÷ 7.2 ÷ 0.300bn ADS. This produces $9.63 / $18.10 / $26.39 / $35.93 from the stated assumptions. The original narrative’s approximate scenario values were inconsistent with those inputs. No operating assumption or new price target is introduced.

*Illustrative using RMB7.2/USD and approximately 300m diluted ADS-equivalent shares. These are historical scenario assumptions, not current market prices or a forecast update.

10.1 Sum-of-the-parts roadmap

If KC begins disclosing meaningful revenue/gross-profit or ARR for AIOS, StarFlow, MaaS and managed third-party compute, a single EV/EBITDA multiple will become less informative. The next model iteration should separate: (A) owned/leased AI compute, valued as capital-intensive infrastructure/cloud; (B) Xiaomi/Kingsoft ecosystem cloud; (C) AIOS/StarFlow orchestration and managed compute, valued on software/service economics; (D) MaaS/agent/application services; less (E) economic financing claims. Until those disclosures exist, assigning a software multiple to an inferred orchestration segment would create false precision.

11. What would change the valuation

12. Bottom line

KC is no longer best analyzed as a small public-cloud operator trying to keep up with Alibaba, Tencent and Huawei. The emerging role is narrower but potentially more valuable: a neutral AI operating layer that can own compute, manage customer/partner compute, and orchestrate capacity across clouds and regions. The government AI appliance and Zhujiang Hospital now provide two distinct examples of Huawei-technology compute sitting beneath Kingsoft software, while Qingyang demonstrates the distributed-control-plane extension.

Zhujiang Hospital remains the strongest direct evidence because it names Huawei and KC together and explicitly divides the stack: Huawei provides Kunpeng/Ascend infrastructure; KC provides heterogeneous AIOS control. The Government AI Appliance shows a similar architecture through Shenzhou Kuntai, and Qingyang extends the logic geographically. Taken together, this is evidence of a repeatable architecture, although not yet evidence of a broad Huawei-KC corporate alliance.

The equity question remains whether KC can capture those economics faster than it accumulates financing claims. The next decisive evidence is pricing and margin realisation, cash conversion and the financing structure supporting the next wave of infrastructure.

Sources reviewed

Kingsoft Cloud Q2 2026 results: https://ksyun.gcs-web.com/zh-hant/news-releases/news-release-details/kingsoft-cloud-announces-unaudited-second-quarter-2026-financial

KC interim results / SEC: https://www.sec.gov/Archives/edgar/data/1795589/000110465926098496/tm2623439d1_ex99-1.htm

Southern Medical University / Zhujiang Hospital, 10 Sep 2026: https://news.smu.edu.cn/info/1016/139056.htm

Qingyang government - KC 100,000 PFlops scheduling platform: https://zgqingyang.gov.cn/zt/zgsgzhqy/content_344343

Kingsoft Cloud product site / Kunpeng 950 and AI solutions: https://www.ksyun.com/

Huawei Connect 2026 Ascend/CANN ecosystem: https://www.huawei.com/en/news/2026/9/hc-wang-keynote

Huawei Cloud / Kingsoft heterogeneous-compute operations framework: https://www.huaweicloud.com/intl/en-us/news/20240920185728306.html

KC NASDAQ historical price lookup: https://ir.ksyun.com/zh-hant/stock-information/historical-price-lookup

HK 3896 17 Sep price history: https://cn.investing.com/equities/kingsoft-cloud-holdings-historical-data

Shenzhou Kuntai / Kingsoft Government AI Appliance, 17 Jul 2025: https://www.dcnetworks.com.cn/news/1035.html

Kingsoft Cloud Government AI Appliance, 16 May 2025: https://www.ksyun.com/cms/news/738.html

KC StarFlow heterogeneous scheduling documentation: https://docs.ksyun.com/documents/44150

KC StarFlow product page: https://www.ksyun.com/nv/product/KSP

Qingyang 2026 KC investment / scheduling update: https://zgqingyang.gov.cn/zw/bmdt/content_343994

Qingyang 2026 government work report: https://zgqingyang.gov.cn/gk/zfxxgk/zfxxgkml/qtfdxx35zwgkb/subject5235zwgkb/content_28283