A Collision Course Between AI Hunger and Energy Physics
A single number from BloombergNEF's December 2025 report deserves attention: 106 gigawatts. That is the projected electricity draw from data centers by 2035 – nearly triple the 40 gigawatts consumed today. The mechanism driving this surge is straightforward: AI training and inference will account for nearly 40% of total data center compute within a decade, and the average new facility will draw well over 100 megawatts, with some exceeding 1 gigawatt.
The question of what will power these facilities is no longer speculative. It is a procurement problem, a regulatory puzzle, and an infrastructure race unfolding in real time.
The Natural Gas Bottleneck
For decades, natural gas has served as the default answer for baseload power – reliable, relatively inexpensive, and scalable. In the United States, 40% of natural gas consumed today goes toward electricity generation. But two constraints have emerged that complicate this assumption.
First, geopolitical fragility. Iranian drone strikes on Qatari natural gas infrastructure exposed supply chain vulnerabilities that energy planners had underestimated. Second, and perhaps more consequential for near-term planning: turbine availability. The waitlist for gas turbines has stretched so long that orders placed today may not be fulfilled until the early 2030s.
This creates an unusual window. By the time turbine shortages relent, competing technologies – small modular nuclear reactors (SMRs) and fusion power plants – plan to be connecting their first commercial units to the grid. The natural gas industry's delay may inadvertently subsidize its own displacement.
SMRs: Proven Physics, Unproven Economics
Small modular reactors represent the most mature alternative in this race. The underlying physics is well-established; these designs tweak existing fission reactor architectures rather than inventing new ones. Several companies have secured commitments from major technology firms.
Kairos Power, backed by Google, received approval for its Hermes 2 demonstration reactor in 2024, with construction well underway. Oklo, which merged with Sam Altman's blank check company in 2024, targets 2028 for first commercial operations. X-energy, with Amazon as an investor, aims for the early 2030s. TerraPower, founded by Bill Gates and partnered with Meta, plans commercial operations by 2030.
The constraint is cost. Nuclear power currently runs approximately $170 per megawatt-hour (MWh), according to Lazard – among the most expensive forms of new generating capacity. SMR business models depend on mass manufacturing to drive cost reductions, but this hypothesis remains untested at scale. The question is whether tech companies' willingness to sign gigawatt-scale power agreements will provide sufficient demand to trigger manufacturing efficiencies.
Fusion: Aggressive Timelines, Uncertain Delivery
Fusion power occupies a different position on the risk-reward spectrum. The technology promises abundant energy using little more than seawater as fuel, but commercial viability has remained perpetually "decades away" for decades.
That narrative may be shifting. Commonwealth Fusion Systems is on track to activate its demonstration reactor next year, with a 400-megawatt commercial reactor expected in Virginia in the early 2030s. Inertia Enterprises, building on the National Ignition Facility's breakthrough in demonstrating net energy gain from controlled fusion, hopes to begin construction on a grid-scale plant in 2030.
Then there is Helion. The Sam Altman-backed startup has committed to supplying Microsoft with electricity by 2028 and is reportedly in discussions with OpenAI for up to 5 gigawatts by 2030 and 50 gigawatts by 2035. The arithmetic is staggering: achieving those targets would require building 800 reactors by decade's end and another 7,200 in the following five years.
For context, the United States added 63 gigawatts of new generating capacity across all sources last year. If Helion delivers even a fraction of its stated ambitions, the energy market's structure would fundamentally change.
Initial fusion costs are projected around $150 per MWh – competitive with nuclear fission but still above natural gas's approximately $107 per MWh. The trajectory matters more than the starting point.
The Renewables-Plus-Storage Wildcard
While nuclear technologies compete for attention, a quieter transformation is underway. Solar paired with battery storage now ranges from $50 to $130 per MWh – overlapping with fusion, fission, and natural gas price bands. Last year, grids installed 58 gigawatt-hours of battery capacity.
The next generation of storage technologies could accelerate this trend. Form Energy recently signed a deal to provide Google with electricity from a 30 gigawatt-hour iron-air battery. XL Batteries repurposes old oil tanks to store inexpensive organic fluid, with capacity limited only by tank size and number. Because these designs avoid critical minerals like lithium, cobalt, or nickel, they promise to dramatically reduce long-duration storage costs.
This creates a scenario where baseload power – the traditional domain of fossil fuels and nuclear – faces competition from intermittent sources that have become dispatchable through storage.
The Software Layer
Hardware alone will not resolve the grid's challenges. Electricity rates rose 13% in the U.S. in 2025, driven partly by AI-related demand. Startups are now arguing that spare capacity already exists on the grid – software can help find it.
Gridcare aggregates data on transmission lines, fiber-optic connections, extreme weather, and community sentiment to identify overlooked sites for new connections. Yottar matches medium-size users with locations where known capacity exists. Base Power is building virtual power plants in Texas by leasing batteries to homeowners, aggregating their capacity to sell to the grid during peak demand. Terralayr does something similar in Germany, bundling distributed storage assets already installed.
Nvidia has partnered with EPRI (Electric Power Research Institute), a power industry R&D organization, to develop industry-specific AI models for efficiency and resiliency. Google is working with PJM Interconnection – the regional transmission organization serving Virginia, Pennsylvania, Ohio, and other states – to use AI for processing its backlog of connection requests.
These software interventions are cheaper and faster to deploy than new generation capacity. Whether they can clear reliability hurdles with traditionally conservative utilities remains the open question.
What This Means for European Observers
The developments described above are predominantly American, but the implications extend to European policymakers and investors. The EU's energy mix, grid architecture, and regulatory environment differ substantially, yet the underlying pressures – AI-driven demand growth, supply chain vulnerabilities, and the need for dispatchable clean power – are shared.
Several mechanisms warrant attention:
Procurement pathways: Tech companies are signing multi-gigawatt agreements with unproven technologies. This risk appetite may not translate directly to European public sector procurement, but it establishes price signals and deployment timelines that will influence global markets.
Regulatory arbitrage: If SMR or fusion permitting proves faster in the U.S., European competitiveness in AI infrastructure could suffer. Conversely, Europe's experience with grid-scale renewables integration may offer lessons for managing intermittency.
Software as infrastructure: The startups optimizing grid capacity represent a category of climate tech that requires less capital than generation assets but depends on regulatory access and utility cooperation. European grid operators face similar coordination challenges.
The 2035 grid will not be powered by a single technology. It will be powered by whichever combination of technologies can scale fastest, integrate most reliably, and price most competitively. The race is genuinely open – and the outcomes will shape AI's trajectory as much as any algorithmic breakthrough.
These questions – how energy infrastructure constrains or enables AI development, and what role European institutions should play – will be central to the conversation at Human x AI Europe in Vienna on May 19. The founders, investors, and policymakers gathering there are precisely the audience that needs to engage with these trade-offs before they become locked in.
Frequently Asked Questions
Q: How much electricity will data centers consume by 2035?
A: According to BloombergNEF, data centers will draw 106 gigawatts by 2035, nearly triple the 40 gigawatts consumed today. AI training and inference will account for nearly 40% of total data center compute.
Q: When will the first commercial small modular nuclear reactors come online?
A: Oklo targets 2028 for first commercial operations. TerraPower plans commercial operations by 2030, and X-energy aims for the early 2030s. Kairos Power's Hermes 2 demonstration reactor is already under construction.
Q: What is the current cost comparison between nuclear, fusion, and natural gas power?
A: Nuclear power costs approximately $170 per MWh, fusion is projected at around $150 per MWh initially, and natural gas runs about $107 per MWh according to Lazard. Solar paired with batteries ranges from $50 to $130 per MWh.
Q: What is Helion's timeline for delivering fusion power to Microsoft and OpenAI?
A: Helion aims to supply Microsoft with electricity by 2028 and is reportedly in talks with OpenAI for up to 5 gigawatts by 2030 and 50 gigawatts by 2035, requiring 800 reactors by decade's end and 7,200 more in the following five years.
Q: Why are gas turbine delays significant for the energy transition?
A: Current gas turbine orders may not be fulfilled until the early 2030s. By then, SMR and fusion startups plan to have commercial plants operational, potentially displacing natural gas before new turbines arrive.
Q: How are software startups addressing grid capacity constraints?
A: Companies like Gridcare and Yottar use data analytics to identify overlooked grid capacity. Base Power and Terralayr aggregate distributed batteries into virtual power plants. Google and Nvidia are partnering with grid operators to apply AI to connection backlogs and efficiency optimization.