Research of the Day HUMBL Voice · Intelligence File 005
Research of the Day

August 2026

The Ascend Playbook

How China built a frontier AI with no Nvidia.

Intelligence FileChina AIHuawei AscendGLM-5Zhipu AI

In January 2025 the US cut Zhipu AI off from Nvidia entirely, the first Chinese AI lab ever put on the Entity List. Eleven months later it shipped a top-tier open model trained on 100,000 Chinese chips, with zero Nvidia and zero CUDA. This is exactly how, and why it matters.

100,000 Huawei Ascend chips
744B Parameters, no CUDA
#1 Open-weight, on coding
~1/6 The price of Claude
Community bonus

Got the code from the video?

Enter it with your email to unlock the deeper cut, keep a PDF of this file, and claim founding-member access on the HUMBL app.

The deeper cut

The real casualty here isn't a rival chatbot — it's the assumption that the Entity List actually stops anything. Sanctions bet that cutting off the tool stops the output. GLM-5 is the receipt that a determined team just rebuilds the tool instead, on a timeline measured in months, not decades.

Who actually loses

Not Nvidia's top line this quarter — its moat's shelf life. CUDA's value was that nobody had a reason to leave it. Zhipu just proved a full alternative stack ships in under a year when there's no other option. Every lab now has a credible reason to at least benchmark the exit.

What to watch next

Two tells. One: whether MindSpore/CANN adoption shows up outside China — that's the real test of whether this is a sovereign-stack story or a genuinely portable one. Two: GLM-5.2's inference speed on the next point release. If the tokens/sec gap to Nvidia closes even partway, 'close, not king' stops being the honest read.

Enter your code + email to unlock this file's bonus.

Not from Instagram? The code's hiding in today's video, not here.

Watch it on Instagram
01

The moat was never the chip. It was the software.

Everyone thinks the chip ban was about hardware. It was about software. For fifteen years, Nvidia's real lock on AI has not been silicon. It has been CUDA, the programming layer every AI engineer on earth learned on.

When you ban a company from Nvidia, you do not just take its chips. You take the entire toolchain: the compiler, the libraries, the debuggers, the frameworks, fifteen years of tooling and community knowledge built on top of CUDA. That is the wall Zhipu actually climbed. Every layer of the Western stack had to be swapped for a younger, thinner Chinese counterpart. The tuned kernels every lab takes for granted did not exist, so they wrote them by hand.

Being cut from Nvidia meant rebuilding fifteen years of tooling from a less mature base.

LayerWestern stackChina's stack
FrameworkPyTorchBecomes MindSpore — Huawei's framework, far less mature
Low-levelCUDABecomes CANN — China's answer to CUDA
Chip-to-chipNCCLBecomes HCCL — the library that moves data between chips
KernelsCUDA-tunedHand-written in Ascend C — the tuned versions do not exist

The hardware handicap, quantified. Per-chip FP16 throughput — the Ascend 910B is roughly A100-class and far behind the H100, so raw chips could never close the gap. Software and scale had to.

ChipFP16 throughput
Huawei Ascend 910B~320 TFLOPS
Nvidia A100~312 TFLOPS
Nvidia H100~989 TFLOPS
02

He chose to stay.

Tang Jie was born in 1977 in Sichuan, into an ordinary family. He studied Automation before switching to computer science, and earned his PhD at Tsinghua in 2006.

At the crossroads every graduate faced, go abroad or stay, he stayed. He was inspired by Wang Xuan, the scientist who leapfrogged forty years of foreign typesetting technology from inside China.

He built AMiner, a map of the world's scientific research, on a single desktop with roughly 20,000 yuan of prize money. He founded Zhipu in 2019 on three months of free rent.

In 2022 he shipped GLM-130B, a hundred-billion-parameter model, before ChatGPT existed, beating GPT-3 on some benchmarks. His whole career was one argument: China could build foundational technology from scratch. So the ban did not break his thesis. It was his thesis.

2006
AMiner

Built on one desktop.

2019
Zhipu AI founded

Three months free rent.

2022
GLM-130B

Before ChatGPT.

Jan 2025
US Entity List

Nvidia cut off overnight.

Feb 2026
GLM-5 ships

Trained on Ascend.

Jun 2026
GLM-5.2

Open under MIT.

“In the 1970s, when China's conditions were far behind, Wang Xuan solved Chinese laser typesetting and leapt forty years in one step. I was deeply inspired.”

— Tang Jie, on why he stayed
03

They could not out-spec Nvidia. So they out-engineered the ban.

Four moves. Every one exists because they had less: weaker chips, a slower interconnect, an immature toolchain, and no way to buy their way out.

#MoveWhat it does
01 · ToolchainRebuild the whole stackRan entirely on MindSpore and CANN, and hand-wrote custom operators in Ascend C because the CUDA-tuned versions simply do not exist there. Zero CUDA, zero Nvidia.
02 · SparsityLet sparsity beat the weak linkA 744B Mixture-of-Experts fires only about 44B parameters per token, plus sparse attention for long context — less data crossing the slow interconnect means the weak link hurts less.
03 · SlimeNever let a chip sit idleDecouples generating from training and runs 1,000+ practice runs in parallel — roughly 10x throughput. The crown jewel, and it's open-source.
04 · ScaleScale beats speedEach Ascend chip is roughly A100-class, well behind an H100, so they used 100,000 of them, made by SMIC on 7nm without EUV — about 15% more compute time, offset by cheaper chips and state support.

They could not buy more compute. So they built the machine that wastes none of it.

04

What it means, and what to watch.

Why it matters
  • The sovereign stack is real — silicon to framework to model, now adapted across more than forty domestic chip families. The ban created the thing it meant to prevent.
  • The gap collapsed — Chinese open models went from about a year behind the West to a matter of months, on hardware the West wrote off.
  • The price floor breaks — a near-frontier model that is free and open forces every paid lab to justify its cost, every month.
Read before you worship
  • GLM-5.2 ranks around #3 overall, behind the leading Western models. Close, not king.
  • Inference is genuinely slower, roughly 17 tokens/sec vs 25+ on Nvidia, and the CANN inference ecosystem is still immature.
  • HK$1 trillion+ market cap on ~$100M revenue — a price-to-sales ratio over 1,000x, on a multi-billion-yuan loss.
  • On the Entity List for alleged military ties, which Zhipu denies. Hosted data sits under Chinese jurisdiction.
The exchange · X, June 18 2026

Elon Musk: "A Chinese model this good? Probably Q1 2027." Tang Jie: "Won't take that long." Elon Musk: "Benchmarks aren't real intelligence. The real test shows up in revenue." Tang Jie: "Focus is all we need. Especially focusing on what intelligence truly is."

The Thesis

The ban was built to slow China down. It just proved they never needed America's chips.

Sources & methodology

GLM-5 technical documentation (bigmodel.cn).

MindVL / CUDA-vs-CANN mapping; wavespeed benchmark analysis.

news.qq.com — founder profile, the Musk exchange, IPO reporting.

36kr.com — Zhipu founding history.

NIST / AISI security notes.

cb.com.cn — IPO and market cap.

Artificial Analysis.

Industry analysis. Figures reflect reporting as of August 2026 and include Zhipu-reported claims not all independently verified. Not investment advice.

HUMBL Voice · HUMBL.Voice · August 2026

Follow @humbl.voice for the next file