--- title: "Policy and competition squeeze: will China LLMs pivot to Ascend?" type: "Topics" locale: "en" url: "https://longbridge.com/en/dolphin/post/44435654.md" description: "Last Fri., DeepSeek pushed a 4.1 update; while only a minor point release, the performance uplift is clear. Coupled with same-day headlines that DS plans to use Ascend for its next-gen pretraining and a ban issued by the U.S. Dept. of Homeland Security, the recent DS newsflow has meaningful implications for the sector and the supply chain.$Z.AI(02513.HK) $MINIMAX-W(00100.HK) core takeaways compiled. Core takeaways:1. How is the model? V4.1-Flash-1017 is not a 'China beats Opus' narrative..." datetime: "2026-10-20T11:53:29.000Z" locales: - [en](https://longbridge.com/en/dolphin/post/44435654.md) - [zh-CN](https://longbridge.com/zh-CN/dolphin/post/44435654.md) - [zh-HK](https://longbridge.com/zh-HK/dolphin/post/44435654.md) author: "[Dolphin Research](https://longbridge.com/en/dolphin.md)" generator: "portal-rs" --- # Policy and competition squeeze: will China LLMs pivot to Ascend? DeepSeek pushed an update to its V4.1 line last Fri. It was a minor in-series refresh, but the performance lift was tangible. Coupled with the same-day headlines on training its next-gen model on Ascend and the U.S. Dept. of Homeland Security's ban, the company's recent moves carry meaningful implications for the sector and the supply chain. $Z.AI(02513.HK) $MINIMAX-W(00100.HK) Key takeaways: **1\. How is the model itself?** V4.1-Flash-1017 is not a 'domestic model surpasses Opus' narrative. It is a reinforced play on cost leadership and coding agent specialization. Its most direct competitive pressure lands on domestic peers in the Flash tier and on lightweight overseas models (Haiku 5.5, GPT-6 Luna). **2\. Domestic compute substitution?** The Information cited sources that DeepSeek aims to pretrain frontier models on Ascend by year-end, with the challenge not on single cards but on communication and stability at ten-thousand-card scale. DeepSeek itself called this its 'biggest bet.' If it works, China’s frontier training could genuinely decouple from NVIDIA, but deliverables would not show up before 2027 at the earliest, and full rollout with end-to-end effects would be closer to 2028. **3\. Impact of U.S.-China policy squeeze?** The main impact hits the middle channel of third-party cloud hosting plus revenue sharing. DeepSeek monetizes via its own API and MIT open-source, and will need to maintain a frontier lead to defend its scarcity premium, with whether Ascend can underwrite next-gen pretraining a key swing factor. Peers that staked their overseas growth on AWS distribution and 2+ trillion parameter models face greater sentiment pressure on valuation. **Details below** **I. DeepSeek sparks a new cost-performance race in U.S.-China base models** **1) V4.1 Flash-1017: small version bump, big performance gains** **(1) V4.1-Flash architecture:** It is a 552 billion-parameter Mixture-of-Experts (MoE) model, but only a small subset of experts is activated per pass: roughly 8 billion parameters on read, and ~16 billion on write. With the KV cache compressed to ~890 bytes per token, reading an entire code repo does not blow up VRAM. This makes it natively suited for coding agent scenarios that read a lot and write a little. V4.1-Flash debuted on Sep 10 and the Oct 17 refresh is the 1017 build. Although it is an in-generation update, the capability uplift is material, while key cache pricing stays unchanged, effectively a spec bump at the same price. From benchmarks and field tests, V4.1 Flash-1017 is back in the Pareto-optimal zone (green area below), and in the US$0.2–0.3 single-task cost band for the Max tier, DS is notably competitive. Against similar-positioned peers, the Max tier shows a slightly better price-performance vs. Claude Haiku 5.5 Max. **(2) Pathway to comprehensive performance gains** Technically, the 1017 build keeps the Sep 10 architecture (causal encoder-decoder with MoE) and relies on targeted capability tuning to lift performance. This stack already trims attention compute and compresses KV cache usage, and with inference-side engineering, pushes cache hit rates close to 100%, routing heavy traffic through low-cost cache lanes. In short, the pathway is: **1) Reduce free-form generation, focus on executing user-specified tasks,** while cutting parameter/format errors and stabilizing structured outputs. **2) Improve retention of key context** to avoid drifting in long conversations, compressing the KV cache. **3) Shorter outputs, fewer steps:** less fluff, retries, and long-winded analysis, higher cache hit rates, and cheaper token lanes, bringing agent task cost down (per-task cost from ~RMB 3.3 to ~RMB 1.3). **4) Prioritize software engineering/endpoints/automation,** driving targeted gains in DeepSeek’s coding agent capability (DeepSWE from 74.2% → 77.0%). **5) Optimize for specific harnesses:** as noted previously, models and harnesses are not universally compatible; optimal setups come when both are in-house and fine-tuned on full trajectory data for alignment. This update adds tailored optimizations for DeepSeek Harness, CodeBuddy, WorkBuddy, and Claude Code. **(3) Best-fit use cases — Coding (light dev), agent automation (high-frequency repetitive ops), multimodal vision (3D design)** Practically, with gains in coding and visual understanding, the 1017 build materially improves on 3D workflows such as visual programming (2D modeling, website generation), 3D design, and game dev tasks. But on general reasoning, HLE (Human Last Exam) shows no obvious lift, indicating the gains were targeted: resources went to coding, agents, and visual tool tasks, not general reasoning. The official tagline is 'wishful programming' — ship full projects with fewer prompts, leaning on longer chains of thought for complex coding tasks while keeping decode speed high so wall-clock time remains short. Overall, V4.1-Flash-1017 is better suited for lightweight/entry-level coding, with some metrics comparable to GPT-6 Extra and Luna. It is not fit to carry high-quality official writing, deep creative prose, or heavy factual verification with strict hallucination tolerances, where flagship models remain superior. **(4) New monetization touchpoint — Search API starts charging** DeepSeek opened its in-house Search/Fetch endpoints at RMB 5 per 1,000 successful calls. Access via DeepSeek Harness, Claude Code, and Codex through server-side tool calls is temporarily free, but will turn per-call paid starting Nov 1. **2) The U.S.-China chase on cost-performance base models** In Sep–Oct, overseas vendors rolled out a wave of high P/Efficiency products, with some lightweight models undercutting many Chinese counterparts on list price. This validates Dolphin Research’s view that with ample compute reserves, top global labs are self-distilling low-cost models, squeezing Chinese cost-leadership models and compressing their survival space. On Oct 8, we noted Haiku 5.5 marked the first time a North American lightweight model beat most similar Chinese models on cost-performance. In this DS V4.1 Flash update, list prices were unchanged, but higher cache hit rates reduced real agent scenario costs. In agent workflows, the same context is called repeatedly, and most tokens are billed at cache rates. DeepSeek’s off-peak cache hit price is RMB 0.02 per million tokens, while many overseas vendors do not disclose hit rates. Thus, even if the list unit price looks close, the realized bill can differ by an order of magnitude. From K3’s mid-year leap, Chinese models then appeared to slow on meaningful iterations over the next 2–3 months. Some in the industry attribute this to chain-of-thought masking and anti-distillation measures overseas, which dragged Chinese catch-up speed and widened K3’s 3-month lead back to 6–12 months. That lead pertains to pretraining. According to GLM’s disclosures, Chinese models increasingly rely on post-training for gains, where data quality (with context and real feedback) matters more than raw scale. As AI coding and other productivity scenarios scale, high-quality user trajectory data for post-training is rising. However, as frontier overseas models lift pretraining scale to 5–10 trillion parameters, Chinese models still have room to run in pretraining. Most released Chinese models remain sub-1 trillion in total parameters, but more in-development models target 3+ trillion. **In the U.S.-China race, we view DS 1017 as a fresh counterattack from China on cost-performance (not a frontier model).** Industry checks suggest multiple major China models will push key updates in Oct–Nov. Beyond Minimax M3.1 announced yesterday, GLM 5.5, Kimi K3.1, Qwen 4.0, and HY-4 Pro are slated over the next 1–2 months. **II. Is domestic compute substitution imminent? Ten-thousand-card comms stability remains the swing factor** Yesterday, The Information reported DeepSeek is co-designing future AI chips with Huawei, planning to start pretraining on Ascend 950DT as early as end-2026. For DeepSeek, this is a big bet. This also echoes earlier buildout chatter: in Sep, Bloomberg said DeepSeek plans a GW-scale data center in Ulanqab, purchasing at least 160,000 950DTs (~US$2.56bn). At that time, 950DT was said to be inference-focused with training still on NVIDIA, and delivery at least 18 months out. **The discrepancy suggests training on Ascend by year-end is a directional plan for now;** actual scale-up, execution, and effect validation will need to be assessed next year. **1) Where 950DT stands** Per Huawei’s roadmap: FP8 1 PFLOPS, FP4 2 PFLOPS, 144GB in-house HBM (HiZQ 2.0), 4TB/s memory BW, 2TB/s chip-to-chip interconnect — roughly between H20 and B200. • Single card: FP8 compute is ~22% of B200, with half the memory BW. - **Interconnect:** 2TB/s chip-to-chip slightly tops B200’s 1.8TB/s. Hence Huawei’s approach is interconnect-for-scale — Atlas 950 supernode is confirmed at 1,024 cards (1 EFLOPS FP8, 256TB global unified memory space incl. CPU system memory, 3μs latency), with a roadmap toward 8,192-card full interconnect. **But for trillion-parameter MoE training, efficiency is governed by real comms throughput under large-scale expert parallelism.** As EP rises from 8 to 128, the merge phase throughput drops by ~20% when aggregating experts’ outputs. In other words, the more cards you connect, the less effective comms bandwidth per card. Within a supernode, dispatch/combine bandwidth is already near H100/H200’s intra-node NVLink. **2) Technical hurdles: stability of 10k-card interconnects** **(1) The communication wall: bandwidth decay under large EP.** Trillion-scale MoE pretraining routes tokens across cards and nodes at EP32 or higher. • Public claims: on Sep 30, DeepSeek open-sourced the Ascend version of DeepEP, measuring 375GB/s dispatch, 'near hardware limit.' But that was EP8, PoC firmware, and per-card; commercial would likely be lower. - **Actual decay:** from EP8 to EP128, dispatch falls ~15% and combine drops from 347GB/s to 278GB/s (~20%), attributed to local reduction overhead during combine and contention between remote memory access and compute for HBM. **In short, marketing cites best-case figures; engineering should budget for decayed throughput.** **(2) The stability wall: more cards, more failures.** Checks suggest Huawei cites 300 hours MTBF per 1,000-card cluster, '25% better than industry.' But failure incidence scales roughly linearly with card count; for a 16k-card cluster, that implies one failure every ~19 hours (300\*1000/16000). Checkpointing, fault isolation, and rapid recovery will be heavy lifts. NVIDIA’s 10k-card clusters face the same physics, but Ascend still trails CUDA’s mature fault-tolerance tooling. **(3) The efficiency wall: MFU not yet disclosed.** Huawei’s CANN community posted a reference practice for trillion-parameter MoE pretraining: using V4-Flash (~284B params) on a 950 supernode for validation, then partitioning plans for 550B and 1.6T, with discussion on extrapolating to 10T. Actual validation scale is far below a trillion, and card counts/MFU by scale point were not disclosed. **Reference datapoint:** Pangu 718B MoE achieved ~41% MFU on an Ascend A2 10k-card cluster. **(4) The supply wall: HBM and capacity.** Huawei management said 'supply as much as we can produce,' with bottlenecks in domestic process, memory, and optics. 950DT supernodes are still in testing; volume shipments are expected late this year or early next, with full delivery in ~18 months. - **Capacity allocation:** Some checks point to annual 950DT output in the hundreds of thousands; 160k units for DeepSeek would be ~40% of that. - **Cost:** With HBM prices surging, domestic GPU quotes are reportedly 'dual-track higher with reduced HBM configs,' potentially widening single-card performance gaps. **3) Implications for the U.S.-China compute landscape** **(1) For China:** this is the critical leap from 'good enough for inference' to 'viable for training' on domestic silicon. • Today: China’s top models still pretrain on overseas cards. For example, Kimi reportedly runs ~70% of frontier pretraining on rented GB200/GB300 abroad, and ~30% on domestic NVIDIA; domestic cards are used mainly for post-training. - **Significance:** if DeepSeek can train a frontier model on Ascend, reliance on NVIDIA for frontier training would shift from must-have to optional. - **Software stack:** DeepSeek ported DeepGEMM, FlashMLA, DeepEP, and TileLang from GPU to Ascend and open-sourced them, lowering migration cost from CUDA to Ascend. - **Capacity spillover:** top customers locking in supply will further crowd limited Ascend capacity. It will be harder for tier-2 model labs and clouds to secure cards for inference, pushing spillover demand to other domestic chips (e.g., Cambricon, Moore Threads, Biren). **(2) For overseas compute: modest near-term impact.** NVIDIA’s share in high-end China training has already been curtailed by controls and largely priced in. U.S. frontier labs are pushing to 5–10T parameters, with demand for top-end compute still exceeding supply — SemiAnalysis tracking shows B200 rental rates rose ~15% over the past week. **But medium to long term,** DeepSeek keeps driving per-token compute down via sparse activation and KV compression, and if 'good enough' domestic compute proves capable of training frontier models, the marginal efficacy of the U.S. strategy to buy time via compute controls should continue to fade. **III. Two policies: the squeeze falls mainly on third-party cloud hosting plus revenue sharing** **1) U.S.: DHS order** DHS Exclusion and Removal Order 2026-01 (signed Jul 24, published Oct 20) mandates: • Scope: federal civilian agencies must stop procuring and remove DeepSeek products and third-party products/services built on DeepSeek. National security systems have separate authorities and are excluded. - **Exceptions:** agencies may file written requests within 60 days. - **Watchlist:** Kimi and Qwen were flagged as 'of interest' but are not covered by the order. Direct revenue impact is near zero — DeepSeek has no U.S. government customers. The real bite is the 'third-party products' clause: U.S. SaaS vendors serving federal customers will avoid embedding DeepSeek going forward. This also nudges U.S. enterprise buyers, especially those with public-sector exposure, to avoid DeepSeek. Previously, several states including New York, Tennessee, and Virginia took similar steps. What bears watching is whether this logic extends from public procurement into the commercial domain. **2) China: draft technology export catalog** - **Consultation:** 23 AI technologies proposed for inclusion; comments due by Nov 15. - **Interim filing:** effective on publication day, covering ownership transfers and overseas licensing. - **Controls:** core LLM tech (weights, key checkpoints, source, architecture, training/inference acceleration) subject to restricted export; models trained on national core data are prohibited. - **Thresholds:** parameter count at 2T, training compute around 1e26 FLOPs, and a 'global top 10' capability threshold (final language per the formal release). The Ministry of Commerce drew clear lines: • Not tech export: API calls, app/web interactions, SaaS, on-device lightweight deployments. - **Under export control:** on-prem overseas deployments of restricted models; deployments on overseas self-operated clouds in relatively controllable settings may go through automatic licensing. - **Not covered:** tech already in the public domain via open-source or publications. **3) Business model impact: mainly limits the 'revenue-sharing via third-party cloud' route** Overlaying both policies shows they hit the same channel: deploying weights to overseas third-party clouds for revenue share. • China: models above thresholds deployed to overseas third-party clouds constitute 'overseas deployment' and need filing. - **U.S.:** the government and affiliates impose procurement barriers on 'products built on Chinese models.' This is precisely the primary overseas route for peers like Kimi. Its strategy deemphasizes self-run APIs in favor of distribution via AWS and other cloud marketplaces, with partners providing compliance cover; K3 is live on AWS and Oracle, and its updated open-source license requires revenue share from users with annual revenue >US$20mn. K3 has 2.8T parameters, above the 2T threshold, directly within scope of the catalog. **DeepSeek’s stock exposure is smaller, but future growth is capped:** - **Revenue mix:** primarily its own API. Revenue over the first 7 months was ~RMB 475mn, and management has verbally indicated an annualized run-rate of ~US$1bn. Its MIT license takes no revenue share; open-source serves as user acquisition. - **Scale:** the mainstay V4.1-Flash backbone is 552B parameters, and V4-Pro is 1.6T — both under the 2T threshold. Thus, the two policies have near-zero direct impact on DeepSeek’s current revenue. What it must confront are three medium/long-term issues: **Next-gen open-source requires filing:** if the next flagship trips the compute or 'global top 10' thresholds, open-sourcing would need 'file-and-release.' It is not a ban, but adds friction and could tighten later, affecting openness with overseas customers as frontier model weights become strategic assets. **Open-source user acquisition is diluted:** the U.S. ban targets 'products built on DeepSeek,' so download counts do not translate into U.S. government or large enterprise wins. Overseas revenue will skew to price-sensitive developers and SMEs, overlapping with the user pools chased by Haiku 5.5 and GPT-6 Luna via price cuts. **Business model shifts further 'tool-ward':** starting Nov 1, Search/Fetch becomes a per-call paid touchpoint, alongside Harness — an extra meter beyond tokens. This route avoids weight exports and sits outside both policies, making it the safest avenue for incremental growth. **Dolphin Research view:** markets tend to read 'U.S. ban + China export controls' as a double negative for Chinese open-source LLMs expanding overseas. We see inevitable multiple compression, but impacts differ by model: **(1) Most negative:** third-party cloud hosting, revenue sharing, and overseas on-prem. While current P&L contribution is small (e.g., Kimi and GLM just partnered with Amazon Bedrock), it clouds that growth narrative, where high-growth expectations hinge on overseas ARR. **(2) Less affected:** first-party API expansion and open-source purely for user acquisition. Models under top tech ecosystems also acquire users mainly via in-house products, limiting the impact. **For independent labs,** DeepSeek monetizes mainly via first-party APIs. Minimax is also less directly hit, as its overseas ARR comes from consumer apps and APIs, but it faces heavier pricing pressure in the base-model price war. **(3) DeepSeek’s key constraint today:** compute, not policy. Whether the next-gen model can be trained on Ascend will determine if it can upgrade from 'value leader in a given price band' to 'frontier contender,' and whether a RMB 500bn (~US$75bn) valuation narrative holds. Based on management’s verbal ARR of ~RMB 6.7bn (~US$1bn), that implies ~75x sales. Versus listed peers, especially the two already public, both DeepSeek and Kimi trade at notable premiums for scarcity value — the ability to sustain frontier model development — meaning valuation is largely underpinned by upside optionality from the next-gen model. **Conversely, as absolute leader Anthropic’s valuation shows signs of compression, overall LLM sentiment may cool.** With fierce competition and tighter controls, the premium for DeepSeek and Kimi may also be forced to narrow. \ Risk disclosure and disclaimer: [Dolphin Research Disclaimer and General Disclosure](https://support.longbridge.global/topics/misc/dolphin-disclaimer) ### Related Stocks - [MAX.US](https://longbridge.com/en/quote/MAX.US.md) - [DEEP.US](https://longbridge.com/en/quote/DEEP.US.md) - [OpenAI.NA](https://longbridge.com/en/quote/OpenAI.NA.md) - [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md) - [BYTED.NA](https://longbridge.com/en/quote/BYTED.NA.md) - [02513.HK](https://longbridge.com/en/quote/02513.HK.md) - [HZAD.SG](https://longbridge.com/en/quote/HZAD.SG.md) - [00100.HK](https://longbridge.com/en/quote/00100.HK.md) - [NVDD.US](https://longbridge.com/en/quote/NVDD.US.md) - [NVDL.US](https://longbridge.com/en/quote/NVDL.US.md) --- > **Disclaimer: This article is for reference only and does not constitute any investment advice.**