Hangzhou published Huawei kernels on Sept. 30 and dropped Hopper from FlashMLA. The National Day address named no model and no compute figure. Moscow locked 2027–2030 metrics. Treasury republished SDN hashes. CNNIC opened an agent identity layer on DNS.
Bottom line
The shift is in repositories, not in the Great Hall. DeepSeek put Huawei Ascend 950 beside Nvidia for kernels it already runs, then broke Hopper and every model before V4.1. Xi kept innovation at the level of direction. Russia approved methods and a monitor, not a public target list. OFAC posted checksums. CNNIC reused DNS so agents can find one another.
The stack
FlashMLA is the break. Project documentation states: “We’ve released sparse attention prefill and decoding kernels for the Huawei Ascend 950 NPU, which achieve up to 410 TFlops (95% hardware peak) and 360 TFlops (83% of hardware peak) during prefill and decoding, respectively.” On a B200 with CUDA 13.3, the fused kernel reaches up to 1,460 TFlops in prefill and 950 in decoding. Then the cut. “In the 2026.09.30 release, we removed support for the Hopper architecture and for earlier models (including DeepSeek V3 / V3.2 / V4.0), and we changed the FP8 / FP4 KV cache format. This version is therefore not compatible with previous ones,” the README states. Older models need an older commit. Nvidia builds here target SM100 and SM103.
FlashMLA story · FlashMLA repo
DeepEP-Ascend matches Nvidia DeepEP’s buffer APIs. On a hand-configured Ascend 950DT with CANN 9.2.0, FP8 dispatch across eight ranks measured 373–375 GB/s. Combine ran 345–347 GB/s. Huawei’s Atlas 850E kit is the recommended baseline, public availability planned for about Oct. 15 — a vendor plan. DeepSelect, used in V3.2, V4 and V4.1, claims a 2 to 20 times gain over PyTorch’s stock topk. Ascend support arrived Sept. 30 and takes bfloat16 only. TileLang 0.1.15 adds an Ascend 950 backend marked dav-3510. TileKernels requires it and picks the chip at runtime. “Most kernels achieve performance close to the hardware’s compute or memory bandwidth limits,” the project says. “All of these kernels have already been used in our internal training and inference workloads.”
DeepEP-Ascend · DeepSelect · TileLang 0.1.15 · TileKernels
DeepSeek told Reuters the aim. “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.” It added: “TileLang was created precisely to meet this need.” The lab called TileLang “a simpler programming model” than CUDA. Huawei gave “unreserved and vigorous support.” The firms advanced a supernode of 128 Ascend 950 chips. deepseek-ai listed 44 public repos. Harness — “DeepSeek Harness: Everything is a Plugin” — sat near 241,000 stars. V3 last updated Aug. 28, 2025. R1, June 27, 2025. Sept. 30 commits sat on kernels, not weights.
Reuters, Sept. 30 · GitHub catalog · org page
Speech, metrics, hashes, names
Xi spoke Sept. 30 at the Great Hall, eve of the 77th National Day. Premier Li Qiang presided. About 800 guests, Xinhua reported. “In the face of a complex and intricate international and domestic environment, we must uphold the general principle of pursuing progress while ensuring stability,” he said, Reuters reported. The economy, he said, has shown “strong resilience and vitality.” Innovation-driven growth is the phrase tied to a good start of the 15th Five-Year Plan, 2026–2030. No model. No chip. No rule. On Taiwan, Reuters quoted him saying China should “promote the peaceful development of cross-Strait relations and advance the great cause of national reunification.”
Dmitry Grigorenko chaired the AI subcommittee Wednesday at 17:00. More than 170 indicators, 16 sectors, 32 departments. Methods for more than 100 were approved. The monitor is FGIS KI. The indicator list was not published. The English rendering says the test moves from the quantity of developments to the real effect of AI. Putin had ordered the plan.
Grigorenko story · government.ru/news/60036
OFAC posted SHA-256, SHA-384 and SHA-512 digests for human-readable sanctions PDFs, including the SDN list dated Sept. 30, 2026. “OFAC publishes hash values or message digests for its sanctions list files in order to provide a level of list content assurance,” the office says. Machine-readable files now come from the Sanctions List Service. For sdnlist.pdf the SHA-256 is b5c4515871270d94ba70ec066640c9d1925ee9c1cff13271b74106a620a4f593. Truncate the SHA-512 and a genuine PDF fails. The day’s actions title: “Counter Terrorism and Transnational Criminal Organizations Designations; Belarus, Counter Narcotics, and Libya Designations Removals.” A linked release is headed “Treasury Sanctions Financial Network of Foreign Terrorist Organization, Tren de Aragua, After Theft of Millions from U.S. Banks.”
OFAC hash story · hash page · recent actions
CNNIC opened ATI, the Intelligent Agent Trusted Infrastructure Platform, on Tuesday. DNS anchors identity. Dual certificates and transparency logs do the rest. Notice slogans: “有身份、找得到、信得过、管得住” and “一次注册，全球可发现、可验证.” Alibaba Cloud is the first vendor listed as docked. Tencent Cloud and the Shanghai Electronic Certification Authority are named, not described as finished. Invite-only testing opened July 9. No speaker quotations in the notice.
CNNIC ATI story · CNNIC notice
Take the FlashMLA update and Hopper, plus every DeepSeek model before V4.1, falls off. A stale SDN digest fails on today’s PDF. Beijing’s line is still finish the year. The chip path is in the repos.