The escape hatch got smaller

Open weights arrived with conditions, the cheapest inference raised its prices behind a discount headline, and the money financing the whole build-out looked circular.

Top stories for week 34: 

  • ZAI's GLM 5.3 tops cyber, holds weights back 
  • Alibaba opens stripped Qwen, revenue share 
  • DeepSeek hides price hike as off-peak deal 
  • Vantage eyes $100B IPO; AI money circular 

Last week the open-weight model was the exit from vendor lock-in and from a government's reach. This week that exit narrowed three ways at once. The Chinese lab ZAI released GLM 5.3, first in the world on a leading cybersecurity benchmark, but staged the weights for a two-week safety review — restraint, it turns out, isn't only a Western habit. Alibaba published open weights for its flagship Qwen, but a stripped-down version missing image understanding and long memory, wrapped in a new license that takes a share of your revenue if you build on it. And DeepSeek, famous for rock-bottom pricing, announced "50% off-peak rates" — a headline that's actually a price hike in disguise, with the main model's output price nearly doubling off-peak and rising more than four times at peak. 

Underneath the models, the financial plumbing came into view and it looked circular. Vantage is exploring a $100 billion data center listing tied to Stargate, and NVIDIA moved to calm nervous lenders after Bloomberg mapped a web of the same handful of companies investing in each other and contractually buying each other's products. The honest read for business leaders is direct. Don't mistake open for unconditional — verify the checkpoint and read the license before you build a fallback on it. Don't mistake cheap for durable — treat when you run workloads as a real cost lever. And don't mistake enormous for stable — the ground under your vendors is more interconnected than it looks. The escape hatches are still there. They're just smaller than a week ago, and the time to map them is before you need them. 

Your weekly briefing on AI in business

Join the Future Bytes conversation

Latest insights from Future Bytes