Thread · HN Continuity

As much as I hate the tax avoidance and hate even more that it's Mr. "I invented Facebook so I get a tax break on $4.1B of stock options" Zuck doing this, there is a case to be made for using GPUs for training new models, because that whole endeavor is clearly pretty experimental. AI data centers are dual use after all: training and inference. And in fact, training is risky and expensive enough to justify some incentives.

The moment it is deployed in a product like a chatbot or agent it's no longer "experimental" but that's probably just a matter of accounting. I'm pretty sure these company are already tracking how much they're spending on this infra for training vs inference.

I wonder if OpenAI and Anthropic are looking at this and going, hmmmmmm....