AI chip efficiency has improved 10-fold over recent years, yet aggregate data center power demand continues to surge, exposing the paradox that Moore's Law gains are outpaced by workload growth.
Fusion reactors, liquid-cooled servers, and 600-kilowatt racks are now essential infrastructure powering the AI boom.
Everyone knows by now that AI's real constraint is power, not chips. But less acknowledged is the aggressively constructed tech stack now being deployed across every dimension to overcome power shortages—and the predictable pattern underlying it all: as hardware grows more efficient, industry actually demands more power. This is a genuine Jevons paradox unfolding at gigawatt scale.
Here's what's happening beneath the surface:
**The Chips: 10x More Efficient, 15x More Power-Hungry**
Nvidia's Vera Rubin platform, now shipping to customers, delivers up to 10x more inference performance per watt than Blackwell. Disaggregated inference—breaking inference into prefill and decode states and running them on separate hardware components (a concept Nvidia calls Dynamo)—can boost output by as much as 35x tokens per watt compared to GPU-only designs. These are genuine efficiency wins.
Yet the power consumption of state-of-the-art server racks is unprecedented. From approximately 40 kW for a Hopper rack in 2023, we're already seeing around 210 kW for a Vera Rubin NVL72 in 2024, and the projected 600 kW "Rubin Ultra" (Kyber) rack expected in 2027, with 1 MW+ units reportedly already in development. This represents a 15x increase in power draw for hardware with only a 10x leap in efficiency over just four years. Rather than reducing power consumption, increased efficiency allows engineers to pack more compute into each rack, and hyperscalers are eagerly deploying this capability to serve larger models and more demanding workloads. These racks draw the equivalent power of between 40 and 170 American homes.
This reality has transformed cooling from optimization into critical infrastructure. Fanless GPUs are now standard in Nvidia's newest systems because traditional fans cannot dissipate the massive heat generated. All systems are transitioning to warm-water direct-to-chip liquid cooling, even for previously air-cooled 8-GPU servers.
Dell'Oro predicts the data center liquid cooling market will exceed $7 billion by 2029.
Infrastructure engineers who haven't yet factored coolant distribution units (CDUs) into capacity planning will soon need to.
**The Nuclear Bet: From Press Releases to Construction**
Big Tech's interest in nuclear power has moved beyond announcements. Microsoft's acquisition of Three Mile Island, Amazon's deal for Susquehanna, Google's acquisition of the Kairos SMR, and Meta's significant nuclear portfolio are now translating into tangible projects with regulatory applications and construction underway. The timing and delivery of power remain uncertain—some SMR designs, like Kairos's molten-salt reactor, have no precedent for commercial operation anywhere in the world—but these are no longer theoretical exercises backed by hypothetical funding.
**The Fusion Bet: Science Fiction Becoming Contracts**
Three years ago, this scenario would have seemed like science fiction. Today, multiple fusion companies have entered into binding power-purchase agreements with hyperscalers—not research grants, but enforceable contracts with delivery timelines.
Helion Energy executed the first-ever fusion power purchase agreement, committing to deliver a minimum of 50 megawatts to a Microsoft data center in central Washington State by 2028. In June 2026, Helion became the first fusion company to receive operating licenses from a state regulatory body—a Radioactive Materials License and a Radioactive Air Emissions License from Washington. The company subsequently raised $465 million in its Series G round at a $15.5 billion valuation.
Commonwealth Fusion Systems is constructing Arc, a 400-megawatt commercial fusion plant, on land leased from Dominion Energy in Virginia, near the world's highest concentration of AI data centers. Google holds a power purchase agreement for 200 MW from Arc and has invested in the company twice.
Google has also funded TAE Technologies for more than a decade, targeting commercial grid-scale fusion by the early 2030s.
Eni, the Italian energy multinational, signed an off-take agreement valued at over $1 billion with Commonwealth Fusion.
No power from these facilities will likely come online before 2028, and fusion technology has famously missed its own timelines for 70 years. Many experts remain openly skeptical that fusion can become economically viable in time to support the current AI cycle. But the shift is significant: fusion companies are now entering enforceable commercial contracts with penalties for delayed delivery, not participating in simple research partnerships. For an industry historically trapped in "20 years away," this represents genuine movement—driven entirely by the relentless demand for AI compute, not climate policy.
**The Unavoidable Conclusion**
Overlaying chip efficiency trends against nuclear and fusion deployment yields an inescapable conclusion: nobody serious is betting solely on efficiency gains. Every hyperscaler pursuing 10x faster chips is simultaneously securing gigawatt-scale power contracts, fully aware that efficiency gains will be consumed by increasing scale, not enjoyed as cost savings. This isn't an engineering failure; it's what happens when the value of output (compute) grows faster than its production cost declines.
The incentive is to use ever more compute, not less.
This manifests as ever-larger models and expanding demand for agents—not efficiency improvements that reduce energy consumption.
For infrastructure teams, the planning assumption for the next five years won't be "power management will get easier." It will be: "The rack I deployed two years ago now uses one-fifth the power of the systems I'm planning today. Plan for those increases in electrical and cooling budgets, and don't assume the next chip generation will free up resources. Instead, prepare to absorb them."