The real figure is **only about 10 to 20 machines on order**. This means the next generation of AI performance will not be decided solely by chip design or software breakthroughs. It will be decided by who gets access to a few dozen machines first.