SpaceX's 10-gigawatt figure is built on a $100 billion-per-gigawatt model — and that number is the output of a chain of choices

Frontier inference economics have crossed a threshold where the revenue projections surfacing in analyst reports are outputs of proprietary model chains — not market data pulled from observed transactions. SemiAnalysis's 10GW-by-2027 SpaceX forecast illustrates the point: the headline figure is not sourced from a deal, a filing, or a stated price, but from a named chain of modeling choices the source itself flags, and the wire copy will not surface it.

The Tokenomics Model and Inference Simulator assume a GB300 cluster and a frontier-class model architecture. Token production is estimated using AgentX, an agentic coding benchmark built from real production coding traces. That token production is then blended between input, cache-read, cache-write, and output token costs at real workload ratios. Cost is set at roughly $12B/GW/year using a conservative $3/GPU-hr neocloud rental rate. The output: a number exceeding $100B/GW/year.

Two numbers, two stories. The headline's $300B ARR claim is built on $100B/GW/year applied to the inference capacity SpaceX is projected to host. The underlying capex math — $50B per GW across 6-10GW of build — runs $300-500B in 2027, on par with what SemiAnalysis expects from AWS and Google. These are different stories with different units. One is revenue, the other is capital. Aggregator copy will not separate them.

The exhibit, then, is not $100B/GW/year as ground truth. It is $100B/GW/year as the output of a chain a reader should be able to name: GB300 plus frontier architecture plus AgentX workload plus $3/GPU-hr cost basis plus Inference Simulator blend. Hold the 10GW headline against that chain before sizing either story, and before repeating the $300B figure as if it were a single number.