No. In some sense, the article comes to the right conclusion haha. But it's probably >100x off on its central premise about output tokens costing more than input.
Even rerunning the math on my use cases with way higher input token cost doesn't change much though.
The component about requiring long context lengths to be compute-bound for attention is also quite misleading.