The $5.5M DeepSeek training cost is an assumed $2 GPU-hour rental rate. I went and read the three cost numbers everyone is quoting

The number that won't die is $5.576M, and what it quietly excludes is almost everything. The two things you can open right now, no arXiv needed, are Moonshot's own chat pricing page and Anthropic's models overview page. Here's the arithmetic. The DeepSeek V3 technical report, arXiv 2412.19437, reports 2.788M H800 GPU-hours and then writes: "Assuming the rental price of the H800 GPU is $2 per GPU hour, our total training costs amount to only $5.576M." That's not a disclosed cost. It's the paper d