The number was correct to a reasonable degree under the assumptions stated by the author in the paper that tweet references since they obtained estimates from consumer grade hardware and the carbon intensity associated with average kilowatt produced in the United States not a hyperscale datacentre run using "ML best practices" although this distinction is left out of various lay media citations. The number also did not pertain to inference it was associated with training a particular model from pre-2019.

I'm sure you see the problem with assuming a very inefficient and poorly-utilized system but at the same time assuming Google scale, and multiplying those factors together.