Zhipu's mystery 'Niulai' model rockets to No.2 in global usage in four days
A mystery Chinese model that appeared seemingly out of nowhere has become one of the most-used AI models in the world within days, and the rankings it just cracked say something broader about where frontier computing is heading.
A mystery Chinese model called "Niulai" has jumped to No. 2 in global AI usage just four days after appearing, while DeepSeek's V4-Flash holds the top spot for a third consecutive week. Reports attribute the model to Beijing's Zhipu, which has reportedly trained it to run on roughly 100,000 China-made accelerator chips instead of Nvidia silicon — a detail independent analysts say puts fresh strain on Nvidia's compute moat even as its quarterly revenue balloons toward $100 billion. The weekly token-usage figures push China's lead over the rest of the world in model call volume to a 17th straight week, a run driven less by benchmark bragging and more by open-weights models that developers actually deploy. That gap matters: ranking first in real-world usage, not just on a leaderboard, is the clearest signal yet that Chinese open models have moved from curiosity to default infrastructure for a meaningful share of builders. We tracked Zhipu's flagship rollout earlier this month — Zhipu confirms GLM-5.3 launch as Goldman raises outlook 35%. The open question is whether "Niulai" is a genuinely new architecture or a rebranded iteration strained for attention — either way, its traction is real.
Amazon is shutting down Mechanical Turk on September 30, ending a 21-year run for the crowdsourcing platform founder Jeff Bezos once dubbed "artificial artificial intelligence." Launched in November 2005, the marketplace paid humans pennies to do micro-tasks — labeling images, transcribing audio — that computers couldn't handle yet, becoming the quiet backbone for an early generation of AI training and research. Its sunset reads as a clean historical marker: the "human in the loop" that once filled AI's labor gap is now redundant enough that Amazon is retiring the product outright. Anyone still depending on Turk for data labeling has a month to find an alternative, an abrupt reminder that the AI labor pipeline has flipped from humans paid by the task to models doing the same work instantly.
Google has begun deploying a "passthrough" URL that rewrites links in its search results, a defensive move aimed squarely at the AI companies and scrapers vacuuming up ranking data. The new google.com/goto redirects route clicks through Google's own servers, letting the search giant fingerprint and throttle traffic it suspects comes from automated tools rather than people. For a company that built its business on being crawled, the reversal is pointed: Google now wants more control over who reads its results and how, especially as rival labs harvest search data to train models. It's the latest front in the escalating scraping wars, and one of the more effective technical chokeholds — every tool that parses Google's links now has to go around a tracker Google runs itself.
What to watch: whether the "Niulai" model holds its No. 2 slot once the novelty wears off, and which lab—Zhipu or a rival—claims credit outright.
Do you think Chinese open-weight models are now a permanent part of the compute infrastructure, or is this surge a bubble? Tell us in the comments.
