....In response to this, U.S. AI companies are dramatically raising their prices and switching customer plans to per-token pricing in order to maximize their own revenues. This is only going to work against them and will lead to a mass exodus away from U.S. AI companies. In fact, that exodus has already begun (see below).
But there’s something else that’s critical to understand here.
Fighting Words Devotio...Holcomb, EllieCheck Amazon for Pricing.
You can easily download the open weights of China’s frontier AI models like DeepSeek V4, GLM 2.5, etc. But unless you have access to very expensive server hardware costing hundreds of thousands of dollars, you can’t run those models locally.
256GB AI Inference Systems Are Coming to the Desktop
In 2027, however, that picture changes dramatically. AMD has announced it will launch “Gorgon Halo” (a revamped Strix Halo) with 192GB of unified memory (160GB can be used as VRAM).