SemiAnalysis analyst explains Grok Bot’s cost in the context of Musk’s compute expansion plans

Discussing Elon Musk’s comment that Grok Bot is too expensive to run, interview participants suggested that personal AI agents could become economical in about a year. One expected serving costs to fall by more than 100x as AI becomes cheaper. This was the speakers’ forecast, not confirmation that the product has already reached those economics.

A SemiAnalysis representative attributed a possible limit on subsidizing Grok Bot to Musk’s need to finance a buildout of compute. He estimated that Musk had about $80 billion in cash on his balance sheet, of which around $20 billion would go to rockets and $60 billion could be spent on compute; Musk was also raising another $40 billion, the speaker said. The stated target was to build 13 gigawatts of compute. The conversation put the cost of building all of it at more than $400 billion, and later at more than $500 billion. The analyst also cited a gap between selling GPU compute for about $15 an hour and a total cost of ownership of roughly $4 an hour.

In his account, selling available capacity, including to Anthropic and Google, could help fund further capital expenditure, even if that meant temporarily reducing compute available to the xAI team. Another participant contrasted five-year contracts at $12–15 billion per gigawatt with a short-term Anthropic deal priced above $40 billion per gigawatt. These explanations and figures were estimates voiced in the interview, not independently verified financial data.