Zhipu AI's stealth model breaks records on Chinese chips, sending shares soaring

For a week, the global AI community was chasing a ghost. A model called Ox Alpha appeared on OpenRouter and OpenCode, climbing the usage charts with a ferocity no one had seen, yet its maker stayed silent. On Wednesday, the mask came off: Ox Alpha was Zhipu AI's GLM-5.3-Flash, and the real shock wasn't the performance — it was the hardware underneath. The entire system ran on a cluster of 100,000 domestically produced Chinese chips, a quiet but seismic answer to Washington's export controls.

Zhipu, one of China's so-called AI tiger companies and a rare Beijing-backed challenger to OpenAI's dominance, revealed that the model processed 62 trillion tokens before its formal release. On OpenRouter alone, it handled over 11 trillion tokens in its first three days — the platform's biggest launch ever. By Thursday, it ranked first among coding systems, accounting for 10.3 trillion tokens, or nearly 31 percent of weekly volume. The market responded instantly: Zhipu's Hong Kong-listed shares closed more than 12 percent higher at HK$1,160.

For outsiders, the significance may be lost in the numbers. But for anyone tracking Asian capital, this is a watershed. Zhipu is not just another startup; it is the flagship of Beijing's push to build a sovereign AI stack, from models to silicon. The company has long been seen as the most academically rigorous of China's major AI labs, spun out of Tsinghua University with deep state-linked backing. Its decision to train and run a frontier model entirely on domestic chips — likely Huawei's Ascend processors, though the company didn't specify — is a direct challenge to the assumption that China cannot compete without Nvidia's top-tier GPUs.

The token volume matters beyond bragging rights. It proves that Chinese-made chips can sustain real-world, high-traffic AI workloads, not just laboratory demonstrations. This is the kind of evidence that moves capital. Venture funds and institutional investors in Asia have been cautious about pouring money into Chinese AI hardware plays, worried that export controls would strangle scalability. Zhipu's stealth trial dismantles that fear with hard usage data. It also pressures rivals like Alibaba's Qwen and Baidu's Ernie to accelerate their own domestic-chip strategies, or risk looking like laggards in the new national priority.

What makes this especially potent is the timing. Beijing has been quietly funneling subsidies and policy support into domestic semiconductor alternatives, from chip design tools to advanced packaging. Zhipu's success gives policymakers a tangible proof point: the ecosystem works. For international investors, the message is clear — the China AI story is no longer solely about software genius compensating for hardware weakness. It is about a hardware-software loop that is closing faster than many in Silicon Valley expected.

The 12 percent share jump is unlikely to be the end of the move. Zhipu has signaled that GLM-5.3-Flash is just the beginning, and the company's open-weight strategy means developers worldwide will now stress-test its capabilities on Chinese silicon. If the model holds up in production environments, expect a wave of interest in Chinese chip supply chains, from foundries to memory makers. The era of assuming China's AI ambitions are capped by American export policy is over — Ox Alpha just rewrote that ceiling.