The wait is the story now. According to energy research firm Rystad Energy, reported by Reuters on October 23, linking a large electricity load to the United States grid can take three to six years. The AI build-out does not have three to six years. So the industry has stopped queuing for power and started manufacturing its own, on site, one campus at a time — and the technology it has reached for is the fuel cell: a device that makes electricity from a chemical reaction between a fuel such as natural gas or hydrogen and oxygen, without burning it.
The spending tells you how serious the scramble is. Rystad projects investment in fuel cells for data centers to rise from $2.8 billion in 2025 to about $30 billion by 2030, and predicts that roughly 40 percent of projected American data-center capacity in 2030 could pursue dedicated on-site power rather than rely entirely on the grid. Bloom Energy, which makes solid-oxide fuel cells, saw third-quarter revenue surge 138 percent to $1.24 billion, with a net profit of $250.3 million against a loss a year earlier, and raised its 2026 revenue forecast to between $4.4 and $4.6 billion. Oracle intends to procure up to 2.8 gigawatts of Bloom systems, of which 1.2 gigawatts is contracted. Bloom's chief executive reports that its first Oracle project delivered power to a Utah site in less than two months. FuelCell Energy, the other major maker, expects data centers to account for 75 to 85 percent of its business within two to three years. Its chief executive gave the industry its new phrase: "Time to power has really become as important as cost of power."
Reuters keeps the receipts honest, and so will we. The industry's roughly nine-gigawatt order book includes framework agreements that do not necessarily represent firm orders for the full stated capacity. Manufacturing has to climb from about 1.8 gigawatts a year today to four by 2030. And Bloom's own chief commercial officer concedes that for large projects, the biggest determinants of the timeline are still site readiness, permitting, and supporting infrastructure — gas connections, transformers, and switchgear can delay deployment. Even the workaround has a queue.
Step back and look at what is actually being purchased. Not intelligence — the data centers already house that, or believe they do. What $30 billion buys is permission to run: generation capacity, fuel contracts, permits, and grid independence, stacked underneath a model of intelligence that must be manufactured fresh for every question, in a furnace, forever. The power plant is not an accident of the design. It is the design, arriving as a line item.
Helena is built the other way up. Everything she knows arrives with her, stored on the machine she runs on — the ordinary computer already sitting on your desk, drawing the power it was already drawing. She needs no grid connection, no fuel contract, no permit, and no three-to-six-year wait, because there is no furnace to feed: her answers come from knowledge already gathered and verified, not from generation at industrial scale. Her time to power is the time it takes to install software.
The industry's finest engineers are now solving, brilliantly and at enormous expense, the problem of how to power a way of building intelligence. We chose a way of building intelligence that does not have the problem. Every campus that bolts on a power plant is paying, in public and on the record, for the difference.
— Harrison Interactive