An April 2026 preprint, since revised, proposes treating AI’s waste heat as a “10th planetary boundary” and clocks the remaining thermal buffer before a 1.5°C overshoot at roughly 6.5 years. The number is arresting. It is also, by the paper’s own admission, a ceiling on the legacy human warming trajectory with zero AI contribution folded in. The countdown is the hook. The framing is the argument.
The arithmetic behind 6.5 years
The calculation in arXiv:2604.04956 is straightforward: take the remaining thermal headroom before the planet crosses 1.5°C (approximately 1.42 × 10²³ Joules, derived from 0.3°C of warming budget), divide by the current annual accumulation of Earth’s Energy Imbalance (approximately 2.19 × 10²² Joules per year at an EEI of 1.36 W/m², drawn from Hansen et al. 2023 and NASA CERES data). The result: ~6.5 years.
Two things to note. First, the EEI value carries uncertainty bands from its source measurements, and those bands are not propagated into the countdown arithmetic. A slightly higher or lower EEI shifts the deadline by months or years. Second, and more consequential: this figure is explicitly labeled the “legacy human curve.” It describes when the planet breaches 1.5°C from existing anthropogenic forcing alone. AI’s additional thermal contribution, which is the paper’s actual subject, is not included in the number that headlines the discussion.
This is not a minor footnote. The 6.5-year deadline is the thing being cited. The fact that it excludes the variable the paper was written to analyze is a gap worth keeping in view.
Waste heat as a planetary boundary
The paper’s core proposal is to add a “9+1” planetary boundary: net-new waste heat from exponential AI compute growth, measured against AI’s potential to reduce baseline anthropogenic waste heat through economic optimization. The framing treats computation thermodynamically: every operation dissipates heat, scaling operations scales dissipation, and there is a hard ceiling imposed by the planet’s radiative budget.
The planetary-boundaries framework, originally developed by Johan Rockström and colleagues, identifies nine Earth-system processes with defined safe operating spaces. Adding a tenth for AI compute is an argument, not a consensus. The framework itself has been criticized for threshold sharpness and unclear governance applicability. Boundaries look like bright lines in a diagram; in the physical system they are gradients with measurement noise.
The paper positions AI in a dual role: heat source and potential heat reducer. AI could optimize grid dispatch, building energy use, industrial processes, and supply chains enough to lower baseline anthropogenic waste heat. But this creates a Jevons Paradox problem: efficiency gains from AI lower the cost of computation, which increases demand for computation, which produces more waste heat. The paper identifies six interacting determinants that govern where the curve actually goes.
The six determinants and “no moderate middle ground”
According to the paper, the trajectory depends on:
- Human computing demand surge
- AI delegation’s recursive growth (agents spawning agents)
- Hardware efficiency asymptotes
- Global grid ceiling
- Economic optimization gains, subject to Jevons Paradox
- The absolute thermodynamic boundary
The paper’s sharpest claim is that there is “no moderate middle ground”: either AI’s optimization effects outpace its thermal footprint, or they don’t, and the system swings to one extreme. This is a modeling assertion, not a physical law. It depends on the interaction terms between the six variables, which the paper acknowledges are not independently resolved.
The paper also argues that alternative computing paradigms will not bail out the trajectory. Quantum computing requires cooling from 0.015K to 300K, with Carnot costs that dwarf the compute savings. Space-based computing has poor thermal radiation in vacuum. Current state-of-the-art logic gates still run approximately 10⁵ times above the Landauer limit, the theoretical minimum energy per bit erasure. These are physics constraints, not engineering problems with a funding solution.
The efficiency counterargument: AI+HW 2035
Published around the same time, a 29-author vision paper including Yann LeCun sets a 10-year target of 1000× improvement in AI training and inference efficiency. Its opening premise is direct: “AI’s energy footprint has reached environmentally and economically unsustainable levels,” and datacenters “increasingly rival nations in power demand.”
The efficiency roadmap identifies data movement as the dominant bottleneck and proposes memory-centric architectures, 3D integration, and low-complexity model architectures as the path forward. This implicitly contests the boundary paper’s claim that hardware innovation “cannot navigate out of the thermodynamic tight spot.” The two papers, landing in the same window, represent a genuine disagreement: one argues physics closes the door, the other argues engineering has not yet seriously tried to open it.
The proposed 1000× efficiency gain over a decade, per the AI+HW 2035 roadmap, would materially change the thermal trajectory. Whether that rate of improvement is achievable depends on whether the industry can sustain compound improvements across silicon, architecture, and model design simultaneously, something that has precedent in the classical computing era but faces diminishing returns as transistor scaling slows.
Where the thermodynamic lens clarifies and where it distorts
The waste-heat framing has genuine analytical value. It sidesteps the carbon-intensity-of-the-grid debate by going upstream: regardless of whether your electricity comes from solar or coal, the computation itself produces heat that must be radiated away. This is a constraint that decarbonization does not solve.
But the lens also obscures more actionable levers. Grid decarbonization still matters enormously for the carbon budget, which operates on a different timeline than the thermal budget. Inference efficiency improvements reduce both carbon and heat simultaneously. Model architecture choices (sparse models, early-exit networks, mixture-of-experts) can deliver useful compute with less energy per query. These are not thermodynamic escapes, but they are real pressure-release valves that the boundary framing tends to compress into a binary: either AI solves the problem or it doesn’t.
The data on current AI energy use is already sobering without the planetary-boundary escalation. GPT-4o’s 700 million daily queries, per 2025 estimates, consume electricity comparable to 35,000 U.S. homes, evaporate freshwater matching the annual drinking needs of 1.2 million people, and generate carbon emissions requiring a Chicago-sized forest to offset.
What the deadline does to the debate
The 6.5-year number is doing rhetorical work that the underlying math does not fully support. It moves the conversation from “is AI’s footprint material?” to “when do we cross the line?” That is a useful reframing if it drives action on inference efficiency, grid planning, and datacenter siting. It is a misleading reframing if it is read as a prediction that AI specifically will trigger a climate threshold in mid-2032.
The paper’s genuine contribution is not the countdown. It is the six-determinant framework and the identification of specific feedback loops (recursive agent delegation, Jevons Paradox on efficiency gains) that could push AI’s thermal trajectory toward the sharp end. These are testable claims about system behavior, and they deserve engagement on those terms rather than as an indictment or a defense of AI’s environmental record.
The concurrent publication of the AI+HW 2035 roadmap is the more useful tension. One group of researchers says the thermodynamic wall is close and compute scaling will hit it. Another group, overlapping in expertise, argues in the AI+HW 2035 roadmap that a decade of targeted engineering can deliver a 1000× efficiency improvement. Both cannot be right about the trajectory. Which assumption set proves more accurate will determine whether the “10th boundary” becomes a scientific framework or a historical curiosity.
Frequently Asked Questions
What happened the last time data center electricity use plateaued?
From 2005 through 2017, global data center electricity consumption stayed roughly flat because efficiency gains in servers, power supplies, and cooling offset demand growth. That equilibrium broke when AI workloads, particularly GPU-intensive training and inference, introduced a demand curve that outpaced the efficiency gains. U.S. data center electricity now accounts for about 4.4 percent of domestic supply and is projected to roughly double in the next few years.
What specific hardware changes would deliver a 1000× efficiency improvement?
The AI+HW 2035 authors identify data movement between memory and compute as the dominant energy sink, not the computation itself. Their roadmap calls for memory-centric designs that place compute directly inside memory arrays, 3D chip stacking to shorten data paths, and lower-complexity model architectures requiring fewer operations per inference. Achieving 1000× would demand compound annual improvements across all three, comparable to the rate seen during the classical Dennard scaling era that ended around 2006.
How much could agentic AI workflows increase per-query compute demand?
The paper flags recursive agent delegation as a structural amplifier but does not quantify it, because production multi-agent systems are still early. If agentic workflows become standard for enterprise tasks, each user request could trigger a cascade of sub-agent calls, multiplying compute per interaction by an estimated 5× to 50× depending on delegation depth. Current AI electricity demand projections generally assume per-query costs similar to today’s single-turn chat interactions, so a shift to agentic workflows would break those projections from the demand side regardless of hardware efficiency trends.
What is the Landauer limit and why do the two papers read it differently?
The Landauer limit is the theoretical minimum energy to erase one bit of information, roughly 2.8 × 10⁻²¹ joules at room temperature. Current logic gates operate about 100,000 times above this floor. The boundary paper treats this gap as evidence that the remaining thermal budget is too tight to wait for physics-level improvements. The AI+HW 2035 authors point to the same 10⁵ gap as room for engineering-driven gains that do not require reaching the theoretical floor, illustrating the core disagreement between the two concurrent papers.