For years, one assumption sat under the US-China chip war.

If China could not buy Nvidia's best chips or the tools needed to manufacture advanced alternatives, its AI labs would remain behind.

The restrictions did hurt. China still cannot easily match the efficiency of the best American systems.

But the workaround is now visible.

Z.ai recently revealed that the mystery model Ox Alpha was actually GLM-5.3-Flash. The company says Ox Alpha's preview traffic ran entirely on Chinese chips. SCMP reports that it processed 62 trillion tokens before launch using 100,000 domestically produced chips.

Those are company-reported numbers.

They also do not prove one Huawei chip beat one Nvidia chip.

That is not what China needed to prove.

It needed to show that the whole system could work without Nvidia.

and they kinda done it.

I am Alex. Welcome to ShortCu8 by Innov8.

Lets Dive Deep🐰🐰.

Do not compare this race chip against chip.

Compare stack against stack.

A slower chip can still power a useful AI service when the company compensates with more chips, better networking, faster storage, cheaper models and software designed around that hardware.

That is the Chinese workaround.

The controls are much wider than a ban on Nvidia GPUs.

The US Commerce Department has restricted advanced chips, high-bandwidth memory, chipmaking equipment and software used to design or manufacture advanced semiconductors.

Its stated goals include slowing China's advanced AI development and weakening its ability to build an independent semiconductor industry.

The policy is not a total wall. Since January 2026, exports of chips such as Nvidia's H200 can be reviewed case by case when buyers meet specified conditions. Access to the newest hardware remains controlled.

That made China's route more expensive.

It also made dependence on American technology a national problem China had to solve.

China's alternative is not one magical chip.

Huawei is connecting its Ascend accelerators into large clusters, then improving the network, storage and inference software around them.

In June, Huawei and China Mobile Hubei said they tested an inference system using Ascend A3 SuperPoD, OceanStor A800 storage and Unified Cache Manager on MiniMax M2.5 and GLM-5.1.

Huawei reported throughput improvements of up to 372% on long-context GLM workloads. That number is not a comparison with Nvidia. It shows how much performance can be recovered by improving the system around the chip.

The model can also be adapted to the hardware. AP reported that DeepSeek designed V4 to work with Huawei's Ascend chips.

This is how China is closing the gap:

It may require more chips and more electricity than an equivalent Nvidia setup.

But inefficient and available can be more useful than efficient and blocked.

The hardware story matters outside China because the models do not have to remain inside one Chinese product.

GLM-5.3-Flash's official model card publishes downloadable weights and local deployment instructions. Z.ai describes it as a 320-billion-parameter model with only 18 billion parameters active at once.

The company's own evaluations say it approaches Claude Opus 4.8 on some coding and agentic benchmarks at a much lower price.

The important part is portability.

Cloud providers can host the model. Researchers can inspect it. Companies can adapt it without sending every request to one American or Chinese API, subject to its licence and local law.

US restrictions were meant to make Chinese AI harder to build.

Chinese open models can make capable AI easier for everyone else to access.

India does not need to choose between permanent dependence on America and permanent dependence on China.

Competition gives us a third option:

take capable models from wherever they are built and run them through infrastructure available here.

That could mean cheaper APIs, more Indian-language adaptations and less damage when a foreign company changes its prices, blocks a country or removes a model.

India is already building part of this base. In July 2025, the government reported that the IndiaAI Compute Portal had provisioned 34,381 GPUs, with 40% pricing support and an average portal price of roughly ₹67 per GPU-hour. It also said foundation models funded through the IndiaAI Mission would be released as open source.

None of that makes us independent yet.

Using a Chinese model through a Chinese cloud merely changes which country holds the switch.

The advantage appears when an Indian provider can host the weights here, keep sensitive data here and continue serving the model even when foreign product access changes.

Now go and build something for india 🇮🇳

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Try it for FREE today

#AD2

The best voice models, now with full orchestration. Build real-time voice and chat agents on one low-latency stack: any LLM, your tools and knowledge, testing, Guardrails, and omnichannel deployment.

Start building free