Export Controls Meant to Slow China Are Fueling Its AI Chip Breakout

The U.S. bet that cutting off China’s access to advanced GPUs would cripple its AI ambitions is starting to look like a spectacular miscalculation. The Biden administration’s 2022 export controls were designed to be a wall, but Chinese labs have treated them like a starting gun. The result isn’t a stalled ecosystem — it’s a rewired one. In the past few months, inference for highly performant open-weight models like Qwen and DeepSeek has begun running on domestic silicon from Huawei and Cambricon, with some training runs now making the jump too. This isn’t theoretical. It’s happening in production.

The knock-on effects are already reshaping software economics. Scarcity bred an algorithmic renaissance. DeepSeek’s Multi-head Latent Attention (MLA) and Group Relative Policy Optimization (GRPO) weren’t just clever papers — they were survival tactics that dramatically slashed compute costs. That efficiency now underpins a wave of open-source models that are cheaper to run and easier to optimize locally. Martin Casado of a16z pointed out that a significant chunk of U.S. startups are now building directly on these Chinese open-weight models, a detail that should make policymakers squirm. The intended bottleneck has become an export of innovation.

This creates a strange new dynamic. The tighter the U.S. squeezes on NVIDIA’s sales, the more aggressive the domestic full-stack deployments become. Look at Alibaba. The company isn’t just hedging; it’s building an entirely decoupled AI infrastructure. The synergy between chip vendors and model developers is creating a fast-evolving software ecosystem that challenges NVIDIA’s CUDA monopoly at its roots. Huawei’s Ascend, a chip that launched quietly in 2018, only hit its stride in 2024 and 2025 — a timeline that correlates almost perfectly with the escalation of U.S. sanctions. It’s hard to ignore that correlation.

What we’re seeing is a bifurcation of the global compute landscape. It’s no longer just about which country has the best model, but about whose hardware ecosystem sets the default. If Chinese silicon is now “sufficient” — and it’s proving itself to be exactly that — then the AI economy is heading toward a future where efficiency trumps raw power. For researchers, that’s liberating. For U.S. trade policy, which was built on the assumption of perpetual hardware dominance, it’s a quiet catastrophe unfolding in real time.