If memory is where a machine mind lives, electricity is what it eats. Every generated token is joules being converted somewhere: a turbine spinning, a transformer stepping down, coolant carrying heat off a rack of accelerators. In 2026 that metabolism became the second binding constraint of the AI era, after memory bandwidth itself, because the grid interconnection queue is now the longest supply chain in technology. The market's answer is a new food chain: thermal and power infrastructure with a $15 billion backlog, fuel cells signing $7.65 billion of bypass-the-queue deals in ninety days, merchant nuclear fleets writing twenty-year power contracts directly with hyperscalers, and a fission startup with regulatory approval and fourteen gigawatts of customer interest before its first reactor turns on. This briefing maps the calories of cognition, from the rack to the reactor, ending at the fusion frontier that would make the meal effectively free.
Vertiv supplies the thermal management, power distribution, and critical infrastructure that AI datacenters are built from, and liquid cooling is now the default design for new AI capacity. Its positioning across cooling and power keeps it attached to essentially every hyperscaler capex plan, which is why the order book has more than doubled.
Vertiv ended 2025 with a $15.0 billion backlog, up roughly 109 percent year over year with a fourth-quarter book-to-bill near 2.9x, covering 12 to 18 months of forward revenue. First-quarter 2026 revenue was $2.65 billion with adjusted EPS up 83 percent, and management raised full-year guidance to $13.5 to 14.0 billion, implying about 30 percent organic growth and 51 percent earnings growth.
Whoever wins the model race, the heat still has to leave the building; the thermal layer is paid on every buildout.
Bloom Energy's solid-oxide fuel cells generate power on-site from natural gas with high electrical efficiency, which converts the grid interconnection queue from a blocker into a bypass. Hyperscalers and utilities are buying speed: Oracle expanded to up to 2.8 gigawatts, American Electric Power signed a $2.65 billion agreement for up to a gigawatt, and Brookfield widened its AI-infrastructure power partnership with Bloom from $5 billion to $25 billion at the end of June.
Roughly $7.65 billion in datacenter-related contracts landed in a single ninety-day stretch of early 2026, an order velocity the company had never seen in its history. The Day-0 price sits about 6 percent down on the day after a monster run, so the projection assumes digestion, not extrapolation, of the current order book.
The constraint is not generation, it is connection; the company selling connection-free power gets paid until the queue clears.
Vistra operates roughly 44 gigawatts across nuclear, gas, solar, and storage, and it is converting that fleet into long-dated hyperscaler contracts: twenty-year power purchase agreements supplying more than 2,600 megawatts of zero-carbon nuclear output, including deals with Meta at its PJM nuclear sites and with Amazon Web Services. Management expects datacenter demand to meaningfully tighten power markets by late 2027 into 2028, which is exactly when uncontracted fleet capacity becomes precious.
First-quarter 2026 revenue reached $5.63 billion at a 26.6 percent operating margin with adjusted EBITDA up 20 percent year over year. The contracted-nuclear layer de-risks the story while the merchant layer keeps the upside to tightening markets, a barbell most regulated utilities cannot replicate.
Electrons are the last commodity the AI race cannot fabricate; owning the fleet that makes them is owning the kitchen.
Oklo's Aurora is a 75-megawatt liquid-metal-cooled fast reactor aimed squarely at powering compute campuses, and 2026 turned its story from concept to countdown: the NRC approved the Aurora's Principal Design Criteria in May, the Department of Energy signed the safety design agreement for the first unit at Idaho National Laboratory, and the customer pipeline reached roughly 14 gigawatts, anchored by a 12-gigawatt master agreement with Switch, a 500-megawatt Equinix letter of intent with a $25 million prepayment, and a 1.2-gigawatt campus agreement with Meta in Ohio.
The company remains pre-revenue with first deployment targeted at Idaho National Laboratory in 2027, so the position is a regulatory-execution option rather than an earnings story; the 14 gigawatts of non-binding interest quantifies the demand if, and only if, the reactors ship. Targets are correspondingly wide and speculative.
If small reactors ship on startup timelines, the datacenter stops being a grid customer and becomes its own power company.
The fusion frontier remains the only answer that makes the energy of cognition effectively unconstrained: magnet-confinement pioneers building toward net-energy pilot plants and helium-3 approaches that already signed power agreements with hyperscale buyers years ahead of first plasma-to-grid. The leading ventures remain private, funded by the same capital that builds the datacenters their output would feed.
No tradable Day-0 exists, so this trend carries a relative basis. The measurable signal is the buyer side: hyperscalers signing conditional offtake for power sources that do not yet exist is a price on desperation itself, and that desperation compounds with every model generation.
Every earlier energy transition ended with the scarce thing becoming cheap; the mind that eats electricity is betting its future on that rhyme.
This briefing is one run. Generate live macro intelligence across any sector, on demand, free.
🚀 Enter Just Signal