Huawei Ascend 950DT vs H200: 86% Inference Leak — Claimed vs Verified A leaked October 2026 China Mobile test deck, relayed by ChainCatcher on November 18, 2026, claims Huawei's Ascend 950DT reaches 86% of Nvidia H200 inference performance on large-language-model serving, with 1,951 tokens per second per card on GLM 5.2 at a 16K-in/2K-out scenario versus an estimated ~2,200 tokens/s for the H200. The H200 figure was estimated from vLLM rather than measured on co-located hardware, the underlying PDF is not published on Huawei's or China Mobile's sites, and neither Huawei nor China Mobile has acknowledged the deck. The same deck cites an 8-card 950DT server bulk price of 2.97 million yuan and claims 30% better price-performance than renting H200 capacity, alongside throughput of 169 tokens/s for Ascend 910B4 and 483 tokens/s for Ascend 910C on the same row. A leaked October 2026 China Mobile test deck claims Huawei’s Ascend 950DT reaches 86% of Nvidia H200 inference performance on large-language-model serving, according to ChainCatcher’s English summary https://www.chaincatcher.com/en/article/2305844?ref=explainx published November 18, 2026. The underlying PDF is not on Huawei’s or China Mobile’s public sites; explainx.ai treats every throughput, price, and percentage as claimed until primary publication. For builders routing inference in China or evaluating domestic accelerators , the leak still matters because it names models GLM 5.2, DeepSeek V4 Pro , cluster size 64-card supernode , and per-card tokens per second beside older Ascend generations. Search interest around “ascend 950dt vs h200” and “ascend 950dt price” spiked the same day the deck circulated. This post answers those queries with source discipline : what the slide says, what outlets repeated, and what no one has published in an official PDF. What is claimed vs what is verified | Claim via ChainCatcher / TokenPost | Verified? | |---|---| | Oct 10, 2026 joint team tested at Huawei Lianqiu Lake | Only named in circulating deck | | GLM 5.2, 16K in / 2K out scenario | Same | | 950DT 1,951 tokens/s per card | Same | | H200 ~2,200 tokens/s in scenario | Estimated from vLLM , not co-located H200 run | | Overall 86% of H200 conclusion | Author conclusion in deck | | 7K/2K scenario ~82% | Same | | 8-card server 2.97M yuan 950DT bulk price | Procurement figures in deck | | 30% better price-performance vs H200 rent | Deck math using above assumptions | Verified independently: ChainCatcher published its article; TokenPost https://www.tokenpost.com/news/technology/60136?ref=explainx and other outlets relayed the same figures November 18, 2026—secondary coverage, not confirmation of the test. Not verified: Side-by-side H200 hardware in the same lab, reproducibility on your vLLM build, or Huawei acknowledgment of the deck. Throughput table from the relay ChainCatcher quotes the deck’s GLM 5.2 16K / 2K row: | Chip | Claimed total throughput per card | |---|---| | Ascend 910B4 | 169 tokens/s | | Ascend 910C | 483 tokens/s | | Ascend 950DT | 1,951 tokens/s | | Nvidia H200 estimate | ~2,200 tokens/s | The article says 950DT is about 11.5× 910B4 and 4× 910C on that row—useful as a generational jump narrative , not a guarantee on your model or batch size. Our Xing4 TeleGym Ascend 910C training report https://explainx.ai/blog/xing4-telechat-ascend-910c-agentic-models-telegym-paper-2026 covers a different slice 8,192 Ascend 910C pretrain . This leak is inference serving on a 950DT supernode , closer to DeepSeek–Huawei co-design rumors https://explainx.ai/blog/deepseek-huawei-chip-co-design-ascend-950dt-pre-training-claimed-vs-verified-2026 but with new numbers . Throughput gauge motif for claimed tokens-per-second cells in the leak relay Ascend 950DT vs H200: how searchers should read the headline People querying “ascend 950dt vs h200” want a simple ratio. The honest answer from this leak: 1. Numerator is a China Mobile internal test on Huawei silicon . 2. Denominator is a literature-style H200 estimate , not a metered competing box in the same rack. 3. Workload is GLM 5.2 at a specific context shape—your production mix may differ. That is weaker evidence than Nvidia’s own SC26 Rubin disclosures https://explainx.ai/blog/nvidia-rubin-hpc-50000-cuda-cores-nvlink-7-sc26-2026 or Artificial Analysis-style https://artificialanalysis.ai/?ref=explainx sweeps, but stronger than anonymous forum posts because the relay names dates, site, and scenario. Cost claims in the same deck ChainCatcher also lists TCO-style monthly costs after five-year amortization: | Platform | Claimed monthly cost with IDC, maintenance | |---|---| | 950DT 8-card server | ~63,600 yuan | | 910C | ~48,700 yuan | | 910B4 | ~16,100 yuan | It contrasts bare-metal rent assumptions—950DT 84,300 yuan/month vs H200 ~130,000 yuan/month —to justify ~30% better inference economics. Those rent figures are deck assumptions , not live quotes you can book today. They also ignore export rules , software stacks , and operator skill —themes in Huawei HBM sourcing reporting https://explainx.ai/blog/huawei-ascend-960dt-customers-source-own-hbm-smuggled-memory-the-information-claimed-vs-verified-2026 . Stacked memory blocks evoking bandwidth-heavy long-context inference scenarios in the deck Why the story spread November 18 HuggingFace-style aggregators and X accounts picked up the 86% headline during Ascend 910C agent-training news TeleGym paper coverage https://explainx.ai/blog/xing4-telechat-ascend-910c-agentic-models-telegym-paper-2026 and ongoing U.S.–China chip tension TSMC credit limits reporting https://explainx.ai/blog/tsmc-credit-limits-upfront-cash-ai-chip-orders-claimed-vs-verified-2026 . The timing does not validate the PDF. What builders should do before buying or switching stacks 1. Ask vendors for reproducible vLLM or CANN commands , tensor parallel width, and batch—not a single tokens/s line. 2. Run your model not only GLM 5.2 at your context distribution; the leak’s 7K/2K row already diverges ~82% . 3. Treat H200 comparison as internal estimate until you benchmark H200 or L40S baselines in your region. 4. Track official Huawei Ascend releases and China Mobile research publications—if the deck is real, a redacted version may surface later. Policy and supply context Inference parity claims do not remove U.S. export controls or Singapore routing probes https://explainx.ai/blog/singapore-nvidia-ai-chip-routing-megaspeed-ft-export-networks-2026 . Domestic clusters may win on availability and local model partnerships even when peak TFLOPS still trail Nvidia flagships. DeepSeek V4 Pro and GLM 5.2: why those models matter The leak names two flagship Chinese stacks tested the same day. GLM 5.2 at 16K/2K is the headline row because long prefills stress memory bandwidth—where Ascend marketing often focuses. DeepSeek V4 Pro appears in the same visit, tying the deck to Nvidia’s quantized DeepSeek V4 Pro drop https://explainx.ai/blog/nvidia-deepseek-v4-pro-nvfp4-huggingface-quantized-2026 and ongoing distillation politics Reuters document-theft reporting https://explainx.ai/blog/china-us-ai-model-distillation-reuters-300-documents-2026 . If you serve either model on Ascend, the leak’s lesson is not “buy 950DT.” It is “ask your operator whether their vLLM or CANN build reproduces 1,951 tokens/s at your concurrency.” One-card headline numbers rarely survive production batching. Generational comparison inside Huawei’s line The deck’s 910B4 → 910C → 950DT ladder gives believers a roadmap narrative : | Generation | Claimed tokens/s GLM 5.2 16K/2K | Role | |---|---|---| | 910B4 | 169 | Legacy baseline | | 910C | 483 | Current fleet workhorse | | 950DT | 1,951 | Next inference tier | Multiplying by 64 cards implies enormous aggregate throughput, but network collectives, KV cache sizing, and quantization change realized QPS. Our vLLM vs SGLang comparison https://explainx.ai/blog/vllm-vs-sglang-inference-server-comparison-2026 is a better starting point for server tuning than any single tokens/s cell. When the 86% figure is misleading Treat the 86% conclusion as a management summary , not a spec sheet: - It blends scenarios 89% and 82% rows without publishing error bars. - It uses an estimated H200 denominator. - It does not state precision FP8 vs FP16 , power cap, or batch size. Investors may trade on the headline anyway—see Jay Puri GPU diversion claims https://explainx.ai/blog/jay-puri-muddy-waters-cenly-chen-gpu-diversion-claimed-vs-verified-2026 for how secondary metrics move markets. Builders should stay on reproducible harness configs . Open questions for Huawei and China Mobile Until a primary PDF appears, watch for: 1. Official Ascend 950DT launch notes on huawei.com or CANN release logs. 2. China Mobile research publications citing the October visit to Lianqiu Lake. 3. Independent vLLM maintainers reproducing H200 estimates on identical context shapes. If only blogs keep circulating the deck, downgrade confidence—but do not ignore the directional jump over 910C; that aligns with Horizon-scale domestic training stories https://explainx.ai/blog/huawei-horinger-7-percent-colossus-2-epoch-ai-compute-gap-2026 . Rental economics vs capex The leak’s rent vs buy table targets CFOs choosing between bare-metal monthly assumptions. Startups on claimed AI VPS pricing https://explainx.ai/blog/ai-vps-pricing-hetzner-auction-claimed-vs-verified-2026 should not paste 84,300 yuan into a spreadsheet without contract quotes. Hyperscalers may still beat localized TCO once software labor and failover are included—similar caveats as Blackstone data-center cost claims https://explainx.ai/blog/blackstone-bxdc-5-billion-amazon-data-centers-50-percent-cheaper-claimed-vs-verified-2026 . Checklist before you trust the leak in a pitch deck | Question | Why it matters | |---|---| | Who ran the test? | China Mobile is named; vendor neutrality unclear | | Was H200 hardware in the room? | No—estimate only | | What batch and TP width? | Unstated in relays | | FP8 or BF16? | Affects tokens/s and power | | Software commit hashes? | vLLM and CANN versions move weekly | | Error bars across three runs? | Not published | Until those cells fill in, cite the story as “ChainCatcher summary of circulating deck” in internal docs, not as Huawei benchmarks . If you are outside China The leak still informs export-control strategy : Washington assumes Chinese clouds will substitute domestic inference cards when Hopper supply tightens Nvidia zero China data-center outlook https://explainx.ai/blog/nvidia-q3-fy27-zero-china-data-center-outlook-hopper-halt-november-2026 . An 86% claim does not prove parity, but it signals where operator marketing will push procurement conversations in 2027 planning cycles. Multinational teams should keep dual-stack test beds: one Ascend or alternative accelerator partition and one Nvidia partition running the same SGLang or vLLM https://explainx.ai/blog/vllm-vs-sglang-inference-server-comparison-2026 harness. Publish your own tokens/s table instead of forwarding leaked slides. Finally, remember inference is not training . The October deck discusses serving GLM and DeepSeek checkpoints. Pre-training claims on 950DT remain in a different evidence bucket earlier co-design coverage https://explainx.ai/blog/deepseek-huawei-chip-co-design-ascend-950dt-pre-training-claimed-vs-verified-2026 . A buyer optimizing fine-tuning clusters should not merge those narratives without separate MLPerf or internal training logs. Watch TSMC strategic chip financing https://explainx.ai/blog/tsmc-credit-limits-upfront-cash-ai-chip-orders-claimed-vs-verified-2026 and ByteDance Malaysia GPU loan pauses https://explainx.ai/blog/bytedance-malaysia-aolani-480m-gpu-loan-pause-export-licence-2026 as adjacent signals: capital flows react to leaked performance decks long before retail buyers see silicon in a box. For compliance officers, file the leak under market-moving unverified benchmarks : useful for threat modeling domestic clouds, not sufficient for export filings or customer SLAs without primary documentation. Journalists should attribute every percentage to “per ChainCatcher’s November 18 summary of a circulated China Mobile deck” until Huawei posts matching figures. That one clause prevents your architecture review from treating social screenshots as vendor specs, and it keeps legal teams comfortable when the deck never lands on an official domain. Related reading Primary relay: ChainCatcher, Nov 18, 2026 https://www.chaincatcher.com/en/article/2305844?ref=explainx . No Huawei or China Mobile primary document was available to explainx.ai at publication time.