⚠️ Editorial note: The open source ecosystem in China operates under a distinct institutional framework — characterized by state-led initiatives, intranet-like boundaries, and top-down governance. Readers should be aware that this context differs from the community-driven open source model common in other regions. The term “open source” as used in Chinese media may refer to practices that diverge from the conventional definition.

China Open Source Daily — 2026-10-02

🏗️ Domestic-Silicon Software-Ecosystem Open-Source Layer — DeepSeek and Huawei Open-Source the Ascend Programming Infrastructure: TileLang as an Open-Source Language Alternative to Nvidia’s CUDA, Framed as the Foundational Step of “Independent Self-Controlled GPU Software Ecosystems”

1. Reuters (2026-09-30, Che Pan and Farah Master, Beijing/Hong Kong) — “DeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidia”

2. Quartz (2026-09-30, Cris Tolomia) — “DeepSeek and Huawei are partnering to build open-source AI chip software to cut Nvidia reliance”

3. South China Morning Post (2026-09-30) — “China’s DeepSeek open-sources tools to help Huawei chips supplant Nvidia AI” — Six software modules for Ascend, including a TileLang variant with native code generation, automatic scheduling, and synchronization

4. Tom’s Hardware (2026-09-30) — “DeepSeek and Huawei release open-source Ascend AI programming tools”

DeepSeek announced on its official WeChat account on 2026-09-30 that it has open-sourced six software modules for Huawei’s Ascend platform, including compute libraries and communication libraries, and — most institutionally consequential — an Ascend-950-adapted version of TileLang, DeepSeek’s own high-level open-source DSL for AI-chip programming. DeepSeek and Huawei also jointly advanced a “supernode” solution based on 128 Ascend 950 chips optimizing both computation and communication.

The sharpest institutional-economics fact of the day is DeepSeek’s own framing of the release. Per Reuters, quoting DeepSeek’s official WeChat post directly: “To build a new generation of independent, self-controlled GPU software ecosystems, the first priority is establishing a high-level language that is universal, easy to program, and still capable of reaching the hardware’s full performance potential. TileLang was created precisely to meet this need, and offers a simpler programming model than Nvidia’s CUDA.” The Chinese phrase “independent, self-controlled GPU software ecosystem” (自主可控 GPU 软件生态) is a state-directed-industrial-policy object in Chinese policy discourse — the same “self-controlled” institutional object that this series’ September 17 briefing documented at the Guo-Ping-Huawei-supervisory-board-chair essay layer (“华为目标是成为NVIDIA” / “Huawei’s goal is to become NVIDIA”, with the sharpest line being Guo Ping’s diagnostic that “Jensen Huang’s real power is not making the fastest chip but making the world’s AI developers feel that not using NVIDIA is passing something by oneself”). The institutional-object pairing here is that the exact diagnostic Guo Ping placed at the Huawei-supervisory-board-chair-essay layer this series’ September 17 briefing documented is being operationalized at the open-source-language-layer by DeepSeek — a company Guo Ping did not authoritatively represent — rather than by Huawei itself.

TileLang’s institutional placement: an open-source language-layer-plus-CUDA-moat-alternative. TileLang is a programming-language-layer object. This is not a driver or a runtime — it is a high-level language whose value proposition, per DeepSeek, is that it is a simpler programming model than CUDA. The institutional implication is that the Ascend-software-ecosystem-institutional-object this series’ September 17 briefing documented at the Guo-Ping-CUDA-moat-institutional-mechanism layer is being operationalized at the open-source-language-layer, not at the driver-layer or the runtime-layer or the framework-layer. The language layer is institutionally the sharpest layer of the CUDA moat — CUDA is not a hardware property but a language-plus-compiler-plus-ecosystem-lock-in property, and TileLang is being positioned as the open-source alternative at that language-plus-compiler-plus-ecosystem-lock-in layer. From an institutional economics standpoint, this TileLang-open-source-language-layer / CUDA-moat-alternative pairing is a structural finding: for the first time in this series, a Chinese DSL is being operationalized at the open-source-language-layer-plus-nvidia-cuda-moat-alternative layer as a joint-release-with-the-domestic-silicon-incumbent institutional object.

The 128-Ascend-950 “supernode” solution: a training-cluster-size-plus-communication-library pairing. Per Reuters: “DeepSeek said Huawei provided full support in developing the programming infrastructure, and the two companies jointly advanced a ‘supernode’ solution based on 128 Ascend 950 chips, optimizing both computation and communication.” Per the SMCP report, the TileLang release brings native code generation, automatic scheduling, and synchronization to the Ascend platform. The institutional reading is that the “supernode” is not just a hardware-cluster-size object but a communication-library-plus-compute-library-plus-language-DSL-co-designed-cluster — the three software modules that DeepSeek is open-sourcing (compute, communication, language) are being jointly designed for the 128-chip supernode, meaning the release is not a driver-plus-SDK layer but a co-designed-training-cluster-institutional-object layer. This is a structurally new institutional-form observation about Chinese domestic-silicon ecosystem construction: the ecosystem is being built at the co-designed-software-stack-plus-hardware-cluster layer, not at the driver-plus-SDK layer.

The DeepSeek-side framing: from state-directed-capital-consolidation to state-directed-software-ecosystem-formation. This series’ September 1 briefing documented the National AI Industry Investment Fund’s $8.8B-in-capital consolidation of Moonshot at $35B and DeepSeek at $74B. The same DeepSeek surface is now being operationalized at the state-directed-software-ecosystem-formation layer — DeepSeek is not merely consolidating AI capability but is now jointly authoring the Ascend software ecosystem with Huawei. From an institutional economics standpoint, the two institutional objects — the state-directed-capital-plus-frontier-AI-lab-institutional-architecture this series’ September 1 briefing documented and the domestic-silicon-plus-open-source-software-ecosystem-institutional-object — are now being operationalized at the same TileLang-DL-DSL-plus-Ascend-950-plus-128-chip-supernode institutional-code-release surface. The institutional reading is that the frontier-AI-lab-plus-domestic-silicon-incumbent pairing is now a joint-software-ecosystem-authorship institutional object, not a hardware-procurement relationship.

The DeepSeek-Huawei relationship at this series’ prior-briefing layer. This series’ September 10 briefing documented DeepSeek’s 160,000-Ascend-950DT procurement for a 1-gigawatt inference cluster in Ulanqab Inner Mongolia ($2.56B at street price, inference-only, training remains on NVIDIA). This series’ September 17 briefing documented Guo Ping’s Huawei-supervisory-board-chair essay placing the “华为目标是成为NVIDIA” diagnostic at the essay layer. This series’ September 18 briefing documented Wang Tao’s HUAWEI CONNECT keynote placing the CANN-community-open-source-plus-Ascend-PyTorch-official-backend-plus-Atlas-960E-NPO-near-package-optics-plus-Ascend-960DT-Q1-2027-acceleration-plus-Ascend-960PR-Q3-2027-plus-Ascend-970-980-2028-2029-plus-Tau-Scaling-Law-plus-one-million-NPU-UnifiedBus-SuperCluster-plus-4.16-million-Kunpeng-developer-ecosystem-plus-90-third-party-open-source-Ascend-projects pairings at the keynote layer. This cycle’s briefing documents the co-designed-open-source-software-ecosystem layer — a fourth Huawei-Ascend-institutional-narrative surface this series’ prior briefings have not documented at, and a first-in-this-series DeepSeek-side open-source-software-ecosystem-partnership with the domestic-silicon incumbent.

The institutional-object pairing with this series’ October 1 briefing. This series’ October 1 briefing documented the Shandong-SASAC-Guantou-Jiehhe-5-year-exit-cycle-and-Higobase-20-year-old-mid-tier-domestic-replacement-database-vendor-exit surface (0.7236% equity at ¥19.1979M floor, ¥12.38M internal appraisal, ~55% premium, Higobase’s 2025 full-year net profit ¥2.97M reversing to 2026 first-five-months net loss −¥6.2953M on revenues below ¥40M) and the SOE-Digital-Technology-subsidiary-collapse-1200-to-500 surface (2020 SASAC directive peak, 2025 collapse to 500-600 nominally alive with most in loss-making or directionless operation, four documented institutional failure modes). Both surfaces documented the state-capital-portfolio-unwinding-side at the Xinchuang-era-tail and SOE-digital-technology-subsidiary tail layers. This cycle’s briefing documents the same state-capital-portfolio-institutional-object at the frontier-AI-consolidation-head-plus-domestic-silicon-open-source-software-ecosystem layer. The institutional reading is that the state-capital portfolio is being rotated from the Xinchuang-tail-mid-tier-exit-and-SOE-digital-technology-collapse layer toward the frontier-AI-consolidation-head-and-domestic-silicon-software-ecosystem-open-source layer, while simultaneously the frontier-AI-consolidation-head layer is opening its own software-ecosystem to the industry-wide layer through an open-source-language-plus-hardware-release pairing. This is the same two-endpoint-simultaneous pattern this series’ October 1 briefing documented at the Xinchuang-tail-exit-and-frontier-AI-consolidation-head layer, now operationalized at the open-source-language-layer-plus-hardware-release layer simultaneously — an institutional-form observation this series’ prior briefings have not documented at.

The sharpest institutional-economics object of the day: the open-source-label-plus-state-directed-industrial-policy-label pairing. DeepSeek’s own framing is that this release is the foundational step toward “a new generation of independent, self-controlled GPU software ecosystems” (自主可控 GPU 软件生态). This is a state-directed-industrial-policy phrase in Chinese policy discourse. The same release is being operationalized at the open-source-code-release layer — DeepSeek and Huawei are releasing the six modules as open source, not as a licensed product. From an institutional economics standpoint, this is the open-source-label / independent-self-controlled-label pairing — the same structural finding this series’ September 30 briefing documented on the Chinese-frontier-lab-open-source-agent surface (open-source-label-plus-FLOSS-commons-institutional-identity-question) and the Chinese-municipal-open-source-AI-foundation surface (Hangzhou Open Source AI Foundation plus LF AI & Data plus AAIF international-Linux-Foundation-family co-hosting), and this series’ October 1 briefing documented on the mid-tier-domestic-replacement-database-vendor surface (Higobase’s “country’s first domestic open-source database localization enterprise” label as market-positioning artifact in a Xinchuang procurement context). The open-source-label / state-directed-industrial-policy-label pairing is being operationalized on the domestic-silicon-software-ecosystem surface for the first time — a structurally new institutional-form observation in this series.

Institutional significance: The DeepSeek-Huawei-TileLang-Ascend-950 six-open-source-modules-plus-128-chip-supernode-plus-independent-self-controlled-GPU-software-ecosystem release is the first-documented instance in this series of the Ascend-software-ecosystem-institutional-object being operationalized at the open-source-language-layer-plus-nvidia-cuda-moat-alternative layer simultaneously with this series’ September 10 briefing’s DeepSeek×Ascend-950DT-160000-chip-$2.56B-inference-only-order layer, this series’ September 17 briefing’s Guo-Ping-CUDA-moat-institutional-mechanism-华为目标是成为NVIDIA essay layer, this series’ September 18 briefing’s Wang-Tao-HUAWEI-CONNECT-960DT-roadmap keynote layer, and this series’ September 1 briefing’s state-directed-capital-consolidation-of-DeepSeek layer — a four-surface-simultaneous-plus-open-source-language-layer institutional-form observation.

Sources:

📎 One Perspective, Not a Verdict

The DeepSeek-Huawei TileLang-Ascend release is institutionally the same object this series’ prior briefings have been documenting on two fronts — the state-directed-capital-plus-frontier-AI-lab-institutional-architecture layer (September 1 briefing’s $8.8B consolidation) and the domestic-silicon-plus-CUDA-moat-institutional-mechanism layer (September 17 briefing’s Guo Ping essay) — now operationalized at the open-source-language-layer-plus-hardware-release layer simultaneously. The sharpest institutional-economics question this cycle’s briefing documents but does not answer is whether the Chinese open-source-label-plus-independent-self-controlled-label pairing converts the Ascend software ecosystem into a FLOSS-commons institutional object or remains an open-source-label-plus-state-directed-industrial-policy institutional object. Reading all prior briefings together, the institutional-economics answer is that the two labels are not obviously coherent — and the same open-source-label-plus-state-directed-industrial-policy pairing that this series’ October 1 briefing documented on the Higobase-mid-tier-domestic-replacement-database-vendor surface is now being operationalized on the DeepSeek-Huawei-domestic-silicon-software-ecosystem surface simultaneously. One perspective, not a verdict.