Deepseek & Huawei Made Local AI Run Without NVIDIA's CUDA (And Why it's Hard)

DeepSeek published six kernel libraries for Huawei Ascend chips on September 30th and reported hitting 99.8% of the hardware limit. Sixteen videos quoted that number. None of them asked what the limit was. It is 432 teraflops, and it sits in DeepSeek's own DeepGEMM-Ascend README, in the column next to the one everybody screenshotted, alongside 865 teraflops for FP8 GEMM on Ascend and 1,730 for FP4. An H100 SXM does 1,979 teraflops of dense FP8, which puts an Ascend 950's FP8 ceiling at 44% of a chip that shipped in 2022 and roughly a tenth of a B300. So 99.8% is a utilisation figure: it tells you how much of the chip got used, not how big the chip is. The fair head-to-head went out the same day, in the same repository, and almost nobody opened it. DeepSeek's FlashMLA release lists its own sparse attention kernels at 1,350 teraflops on an Nvidia B200 under CUDA 13.3 against 410 on an Ascend 950, and 1,024 against 360 on decode. Same team, same algorithm, same release date, and the Ascend number is already 95% of that chip's theoretical peak while the B200 number is nowhere near Nvidia's. What is left between them is silicon, not software. Two corrections while we are here, because most of the coverage got them backwards: CUDA's counterpart in the Huawei AI chip software stack is CANN, not a kernel library, and TileLang is not DeepSeek's project at all. It belongs to tile-ai, it is built on Apache TVM, it was created in October 2024, and a tilelang-ascend adapter has been open since September 2025. Then the parts the headlines skipped. DeepEP-Ascend holds 90 to 95% of the physical bandwidth limit only up to 32 ranks, its reduce-scatter and all-reduce kernels are "still being built", its numbers came off "a PoC HDK supplied to DeepSeek" that is "not a publicly distributed release", and it shipped with no licence file while the other five repos are MIT. Every repo needs an Ascend 950 and CANN 9.2.0 with torch_npu, so none of it runs on a 910B or a 910C, and the 950DT reaches customers in Q4 2026. The honest reading cuts the other way too, though: Huawei let an outside team measure and publish its own ceiling, which Nvidia has never allowed anyone to do, and a 950DT's 144 gigabytes at 4 terabytes a second beats an H100's 80 at 3.35 for anyone serving a very large mixture-of-experts model. The software argument for staying on CUDA is genuinely weaker than it was a week ago. The hardware argument is not, and a fab is not a compiler. CHAPTERS 00:00 The 99.8% everyone quoted 01:02 February 2025, and why I got it wrong 01:36 Six repos, matched one for one 02:41 Same import: lock-in running in reverse 03:46 99.8% of what? 04:29 431 of 432, and what an H100 actually does 05:21 Same kernel, same day: B200 1,350 vs Ascend 410 07:15 TileLang is not DeepSeek's, and isn't new 08:30 CUDA's real twin is CANN 09:12 The repo that shipped with no licence 10:55 Ascend 950 only, and nobody has one 12:51 The honest case for Huawei 14:46 What this changes on your desk 16:06 What to watch for WATCH NEXT DeepSeek's $1 Billion Is a Price Hike, Not Growth — https://youtu.be/dXieY9PwvJg What It Actually Costs to Run DeepSeek V4.1 Flash Locally? — https://youtu.be/04ePeOyRN0k DeepSeek Just Shrunk the KV Cache 437x — https://youtu.be/n-_e9bROGHo Xiaomi & DeepSeek Just Solved the Biggest Bottleneck in Your Local AI — https://youtu.be/RlRC_6hbBHc Why DeepSeek Ditched Its Own Flagship (And Then Didn't) — https://youtu.be/u00vNoI2iQc Don't Buy a Local AI Hardware Until You See This — https://youtu.be/L4UV6b2Xc64 WHAT THIS VIDEO COVERS - 99.8% is utilisation, and the hardware limit behind it is 432 teraflops BF16, 865 FP8 and 1,730 FP4 on an Ascend 950DT - DeepSeek's own FlashMLA release puts a B200 at 1,350 teraflops against the Ascend 950's 410 on the same kernel - DeepGEMM-Ascend installs under the same Python package name as DeepGEMM, so the import in your code does not change - TileLang belongs to tile-ai, is built on Apache TVM, and had Ascend support a year before this release - The expert-parallel comms library's numbers came off a pre-production Huawei box, and that repo has no licence file - Nothing in the release runs on a 910C, a 310P, llama.cpp's CANN backend, or any card you can buy today - Where Huawei genuinely wins: 144 GB at 4 TB/s against an H100's 80 GB at 3.35 TB/s #deepseek #huawei #ascend950 #cuda #cann #aihardware #machinelearning #gpu #cloudcodes