OpenAI's projected compute bill jumped by about $256 billion since February, even as its cash burn forecast improved.
In February, OpenAI told investors its compute costs would run about $600 billion. A July presentation prepared for a computing deal, reported by the Financial Times, puts that number at $856 billion. Despite the higher infrastructure line, the company's projected negative free cash flow for 2026 through 2030 improved, falling to $278 billion from the roughly $305 billion it forecast in May.
These are internal planning numbers, not audited results, and they keep moving. The scale of the swing, a $256 billion jump in compute costs in under six months, shows how fast OpenAI's infrastructure needs are outrunning its own projections. That a cash burn estimate can improve while the underlying cost line balloons suggests other assumptions in the same presentation, likely revenue growth or financing terms, shifted just as much.
OpenAI has not published these figures itself. They keep surfacing through presentations shared with potential compute partners, which tells you as much about who is being pitched as it does about the actual math.