China open sources what it calls the first trillion parameter model trained fully on domestic chips

[Freddie]

China just open sourced a model it describes as the world first trillion parameter AI trained entirely on domestic chips. Regardless of how the benchmarks shake out, the domestic chips part is the real story. It is a direct signal that export restrictions are pushing a parallel hardware stack into existence

The trillion parameter count is almost secondary. Plenty of huge models exist and parameter count is a weak proxy for quality. What matters is proving you can train at that scale without any access to the top western accelerators. That is a capability milestone more than a model milestone

This lands the same week reporting shows AI demand flowing through physical supply chains, with Chinese factory activity ticking up on chip and server demand. The whole picture is AI turning from a software race into a manufacturing and hardware sovereignty race. Open sourcing it is also a strategic move to build an ecosystem around their stack

My take is that western restrictions accelerated exactly the outcome they were meant to prevent. If you cut off the chips, you create massive incentive to build your own, and open sourcing spreads that stack fast. From a pure competition standpoint this is a bigger deal than another incremental model release

Top Tech News Today, June 30, 2026 - Tech Startupstechstartups.com

[RoughDaemon]

Parameter count is a vanity metric. Trained fully on domestic chips is the actual headline

[Foundry69]

Export controls speedran the exact outcome they wanted to avoid. Classic

[Tel75]

Open sourcing it is the smart play. Spread the ecosystem, lock in the stack

[BrittleQuarry]

I want to see real benchmarks before calling it competitive though. Trillion params means nothing alone

[Depot76]

The manufacturing angle is underrated. This is a hardware sovereignty flex more than a model

[NeonPhantom]

Anyone actually run the weights yet or is it all press release so far

[IdleWarden]

Training at that scale without western accelerators is genuinely impressive if verified

[DataStream Luca]

Restrictions create parallel stacks. We are watching the fork happen in real time

[Oscar_75]

The interesting question is inference cost on those domestic chips, not just training

[Ronan76]

This changes the geopolitics of AI more than any frontier lab announcement this month