China is betting 2 trillion yuan on AI without Nvidia. Will it build its own world?

China has just unveiled its most ambitious five-year plan for technological independence yet. It wants to pour 2 trillion yuan, around $295 billion, into national AI infrastructure. And there is one key point: a full move away from American hardware. While Washington argues over who may run its best models, Beijing is building an entire parallel world in which it needs neither American chips nor American models.

01ChapterWhat's new

China wants to build 80% of its AI grid on domestic chips and squeeze Nvidia out. According to Bloomberg reports, the state is preparing a plan worth around 2 trillion yuan (~$295bn) to connect fragmented data centers into a single national computing grid by about 2028. It is coordinated by the National Development and Reform Commission, the state itself, not individual companies.

The core of the plan is a hard rule: at least 80 percent of the technology, mainly AI chips, must come from domestic suppliers. Huawei and its Ascend line get the leading role. The goal is to finally route around US export sanctions on advanced semiconductors and build an isolated, fully self-sufficient AI ecosystem that Washington cannot reach into.

It is not just a plan on paper. Two Chinese open-weight models shipped alongside it, GLM 5.2 from Zhipu and Kimi K2.7 from Moonshot, and in coding they are catching up to the US frontier. Hardware independence and homegrown frontier models suddenly go hand in hand.

02ChapterWhat you'll appreciate most

That you have gained a cheap, downloadable alternative. Chinese open-weight models mean you are no longer tied to just two or three American companies. You can download GLM 5.2 and Kimi K2.7, run them yourself, or use a far cheaper API.

And it is an alternative nobody can switch off by your passport or reprice overnight. Open weights mean the model is yours. Once you download it, no government or company can revoke your access. That is exactly the certainty missing from closed American models today.

03ChapterWho it's for

It matters most to developers and companies paying for AI coding every month. For them the difference between the US frontier and a Chinese open-weight model is mainly a question of cost and control, not a dramatic gulf in capability.

04ChapterHow to use it in practice

Start by trying a Chinese model once on a real task, not for play. Take a concrete coding task you normally run through Opus or GPT and run it through GLM 5.2. You will see the difference in quality and price with your own eyes.

A real-world example

05ChapterRecommended tools

For everyday coding a Chinese open-weight model is plenty today, for the hardest tasks reach for the US frontier. This is a practical split, not politics. It is about which model is worth it for which job.

06ChapterSummary

The AI world is starting to split into two separate ecosystems. With a 2 trillion yuan plan China is clearly saying it does not want to depend on American hardware, and with its open-weight models it shows it has the frontier software too. Meanwhile the US lead in coding is measured in single-digit percentages, not orders of magnitude.

The good news for you is that you benefit from this contest. You have gained cheap, open, downloadable models nobody can switch off by your passport. Whether the world splits into an American and a Chinese AI bloc, or stays connected, remains to be seen. What you can do right now is test whether a Chinese open-weight model saves you most of your coding bill.

07ChapterSources

• Bloomberg: China preps $295 billion plan to fund nationwide AI buildout
• Tom's Hardware: China drafts $295 billion plan for a national AI grid running on 80% domestic chips
• Capacity: China plans state-directed AI buildout, moves to lock out Nvidia and AMD
• SWE-bench Verified leaderboard (as of 18 June 2026)
• Regolo: GLM 5.2 vs Kimi K2.7 Code coding comparison