Analysis
Teravolt's argument is an arbitrage stated in dollars per megawatt-hour. A megawatt-hour of electricity consumed by an aluminum smelter generates $170 to $190 of revenue; the same megawatt-hour running AI inference generates $450 to $900, with roughly $300 of EBITDA, co-founder Laert Karaashev told The Register. The London company's plan is to buy or partner into sites that already have the grid connection -- smelters, retired thermal plants, refineries, bitcoin mines -- and convert them.
The demand case is the familiar one, with specific figures attached. Teravolt projects 410 GW of AI power demand by 2036 against roughly 170 GW available from the grid, leaving a 240 GW gap. Founder Denis Alkhazov says customers are already asking for 20,000-GPU deployments within twelve months, and the company is starting in Eastern and Southern European markets where industrial sites are cheap and interconnection queues are shorter than in Northern Virginia or Texas.
“Teravolt projects 410 GW of AI power demand by 2036 against roughly 170 GW available from the grid, leaving a 240 GW gap.”
The retrofit thesis is not new -- Crusoe built a business on stranded flare gas, TeraWulf and Core Scientific converted bitcoin mining sites into AI hosting, and Galaxy Digital turned its Helios mine into a CoreWeave campus -- but the specific move of cannibalizing heavy industry is more aggressive. A smelter's grid connection is the scarce asset; the building around it is mostly disposable. That is why Core Scientific's converted sites commanded the valuations they did, and why CoreWeave pursued the company. Pulse has previously covered Bitcoin's mining-to-AI conversion wave as the same power arbitrage at smaller scale.
The obvious objection is political, and it is not small. Converting a smelter to a data center replaces several hundred industrial jobs with a few dozen technicians, in regions where that industrial employment is the reason the grid connection exists. Amazon and Microsoft have both spent the past month siding with communities against utilities on ratepayer costs, precisely because the political backlash to data center power consumption is now a real permitting risk. Teravolt is proposing to make that tradeoff explicit rather than hide it.
The second objection is technical. Smelter loads are steady and tolerant of voltage variation; AI training loads swing hard and fast, and the power quality and cooling requirements differ substantially. Reusing the interconnection is the easy part. Reusing the substation, the switchgear and the water is not, and the capital required to bridge that gap is what determines whether the $450-per-MWh revenue figure survives contact with a construction budget.
