weekly · 2026-09-28 — 2026-10-04
Software and systems advanced while fabrication scarcity stayed binding
Open Ascend software, a domestic 40-accelerator rack, equipment orders and manufacturing investment strengthened China's integration and scaling evidence. Huawei's own warning about scarce advanced capacity and a reported Tencent lease for foreign chips overseas showed that frontier fabrication volume and access to leading accelerators remain much less sovereign than the surrounding stack.
Published 2026-10-05 · Evidence through 2026-10-04 · Completed period
Ascend software became more inspectable as fabrication stayed constrained
DeepSeek contributors published a native Ascend 950 backend in TileLang with Ascend-specific code generation, scheduling, synchronization, vector programming and kernel examples, while Reuters reported joint work with Huawei on compute and communication libraries and a 128-chip Ascend 950 supernode solution. Public code is stronger ecosystem evidence than a closed compatibility announcement, but it does not establish broad operator coverage, production stability, adoption, power efficiency, CUDA equivalence or a matched cross-platform workload result. Huawei consumer-business chief Richard Yu separately said China's advanced-chip capacity remains very limited, that Ascend competes with smartphone processors for the same capacity, and that Chinese advanced manufacturing still relies on DUV while domestic producers work toward EUV. The statement connects Ascend supply directly to fabrication scarcity without supplying capacity, yield, node, equipment-provenance or order-fulfillment figures.
Reuters · DeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidia (source)Reuters · Huawei unveils Mate 90 phones, leans on homegrown chip design to offset US curbs (source)TileLang contributors · TileLang Ascend 950 backend (source)A reported offshore lease exposed continuing demand for foreign frontier compute
The Financial Times reported that Tencent agreed to a five-year Oracle lease covering about 100,000 advanced AI chips in Southeast Asian data centers for roughly $7 billion. If completed as reported, the arrangement would be system-level evidence that a leading Chinese AI user is still seeking foreign chips unavailable in China rather than relying only on domestic supply. Reuters could not independently verify the report, and the chip models, delivery schedule, utilization and workloads were not disclosed. It should therefore be treated as a reported procurement signal, not a confirmed deployment or performance result.
Reuters · Tencent reportedly leases 100,000 AI chips from Oracle (source)H3C added a domestic scale-up specification, not an operating comparison
H3C announced the UniPoD S81000 X1, a rack-scale system supporting 40 unnamed domestic accelerators in one scale-up domain. It claims a single-stage CLOS fabric with 448 GB/s peer-to-peer bandwidth between any two cards, 5,760 GB of unified-address memory and up to 28 PFLOP/s at FP8. The specification is concrete evidence of domestic rack and fabric integration, but H3C did not identify the accelerator or switch silicon, distinguish dense from sparse FP8, disclose power, latency, utilization, customer acceptance, shipment volume or workload results, or publish a matched comparison with NVLink, InfiniBand or a current foreign supernode. It does not establish operational parity.
H3C · H3C launches the UniPoD S81000 X1 domestic-accelerator supernode (source)Equipment demand and financing grew ahead of qualification evidence
ACM Research Shanghai reported RMB 17.073 billion of orders on hand at September 29, up 88.2% year over year. The unaudited aggregate combines future deliveries with delivered equipment awaiting revenue recognition, may not convert fully, and contains no customer, product, fab, qualification, acceptance or domestic-content split. Hwatsing completed a RMB 3.795 billion private placement and said proceeds would support a Shanghai equipment R&D and manufacturing base, Kunshan wafer-reclaim expansion and high-end equipment development. Both disclosures are meaningful demand and capital signals for domestic process equipment, but neither establishes completed capacity, qualified tools, realized output, yield, uptime or a fully domestic bill of materials.
ACM Research Shanghai · Voluntary disclosure of orders on hand (source)Hwatsing · Hwatsing completes its 2026 private placement (source)Benchmark evidence did not move
A period-end review found no matched current Chinese accelerator result against Rubin, GB300, MI355X, TPU v7 or OpenAI Jalapeño in MLPerf Training v6.0, MLPerf Inference v6.1 or InferenceX's published coverage. MLPerf provides standardized submitter-run results rather than automatically independent testing, and individual result artifacts were not exhaustively audited. TileLang code and H3C specifications improve software and systems evidence, but the week supplied no new independently controlled model-training or inference result that changes the frontier comparison. Absence of a verified public result is not evidence of incapability.
SemiAnalysis · Continuous inference benchmark dashboard (source)MLCommons · MLPerf Inference v6.1 closed submissions (source)MLCommons · MLPerf Training v6.0 submissions (source)Forecast implications
No dated forecast changes. The week strengthens the existing view that Chinese software compatibility, rack integration and equipment investment can improve before frontier single-chip efficiency or a fully sovereign production chain. Huawei's capacity statement reinforces the expectation that advanced logic volume and lithography will remain harder constraints than software or aggregate system engineering. The reported Tencent lease shows that access to foreign frontier accelerators can remain valuable even as domestic alternatives advance. H3C's rack and the equipment disclosures do not change the timelines for EUV, HBM or matched frontier training and inference evidence.
ACM Research Shanghai · Voluntary disclosure of orders on hand (source)Reuters · DeepSeek partners with Huawei to develop chip programming tools, reducing reliance on Nvidia (source)H3C · H3C launches the UniPoD S81000 X1 domestic-accelerator supernode (source)Reuters · Huawei unveils Mate 90 phones, leans on homegrown chip design to offset US curbs (source)Hwatsing · Hwatsing completes its 2026 private placement (source)SemiAnalysis · Continuous inference benchmark dashboard (source)MLCommons · MLPerf Inference v6.1 closed submissions (source)MLCommons · MLPerf Training v6.0 submissions (source)Reuters · Tencent reportedly leases 100,000 AI chips from Oracle (source)Next evidence to watch
Prioritize Ascend 950 capacity, yield and accepted customer configurations; production use and operator coverage for the new open software; confirmation, chip identities and workloads for the reported Tencent lease; named H3C deployments with fabric utilization, recovery, power and scaling results; conversion of ACM orders into accepted equipment and revenue; Hwatsing project construction, qualification and realized output; and matched training time-to-quality or inference throughput, latency and energy measurements on current Chinese and frontier foreign systems.