Chinese models now carry a third of US AI traffic, and price is the whole story
New CNBC reporting shows Chinese models have held above 30 percent of the tokens US companies route through OpenRouter since February, peaking at 46 percent, up from an 11 percent average over the prior year. The driver is not ideology, it is invoices: open Chinese models run 60 to 90 percent cheaper than leading US systems. Startup Lindy moved 100 percent of its traffic from Claude to DeepSeek and says the switch saves millions, while Z.ai's GLM 5.2 posted the fastest adoption Vercel has tracked this year, roughly 27x token growth in its first full week. Set this against the same week US labs repriced upward, and the picture is workloads voting with their tokens. If you run inference at volume, benchmark an open model on your non-sensitive workloads this month and find out what your actual switching cost is, because your competitors are already doing that math.
Source: Original source
Anthropic signs a $19 billion, 20-year lease with a bitcoin miner
Anthropic executed a 20-year lease at TeraWulf's Justified campus in Hawesville, Kentucky, worth roughly $19 billion in contracted revenue, more than TeraWulf's entire $12 billion market value. The site will carry about 401 megawatts of critical IT load, with first capacity landing in the second half of 2027 and full ramp by early 2028. It stacks on top of Anthropic's dozen-plus US data center leases already exceeding a gigawatt. Two things worth reading here. First, a bitcoin miner converting to an AI landlord at these numbers tells you where the margin moved. Second, note the dates: the capacity crunch that has labs rationing their top models gets answered in 2027 and 2028, not this year. Plan on the current era of metered access and staged rollouts persisting well into next year, because the buildings that fix it are not built yet.
Source: Original source
DeepSeek is designing its own inference chip
Reuters reports DeepSeek has been working for about a year on an in-house chip aimed at inference, the stage where trained models answer queries, with the explicit goal of reducing reliance on both Nvidia and Huawei. The project is early: the company is in talks with partners on design, contract manufacturing and memory, and is hiring semiconductor engineers. But the direction matters more than the timeline. The lab that crashed the industry's cost floor in early 2025 is now going after the silicon layer, and inference is where the volume and the recurring spend live. If DeepSeek lands even a modest domestic inference part, the 60 to 90 percent price gap between Chinese and US models gets structurally wider, not narrower. Anyone modeling long-term inference costs on the assumption that Chinese pricing is a subsidized phase should revisit that assumption.
Source: Original source
AI governance gets its biggest week ever in Geneva, with CEOs at the table
The UN's first Global Dialogue on AI Governance ran July 6 and 7 in Geneva with delegates from 169 countries, rolling straight into the AI for Good Summit and, today, the first meeting of the UN's AI for Good Global Commission, the first UN-level body seating AI executives next to heads of state. Salesforce's Marc Benioff and Rwandan President Paul Kagame co-chair; Jensen Huang, Andy Jassy, Brad Smith and Anthropic's Jack Clark sit as members. Meanwhile Washington is finalizing its voluntary frontier-model standards framework with OpenAI, Anthropic and Google, days after asking OpenAI to stagger GPT-5.6's release. The pattern is consistent on both tracks: the rules are being written with the incumbents in the room. Voluntary frameworks have a habit of hardening into procurement requirements, so watch what gets defined as triggering a security review. That definition will set release cadence for every model you use.
Source: Original source