This essay brackets climate science entirely and examines AI power demand as an electrical engineering and infrastructure problem.
For years, I dismissed anthropogenic global warming (AGW) as a political and rhetorical construct whose practical consequences would eventually be overtaken by some other, more immediate disruption. History suggested this was likely: large, abstract, long-horizon problems rarely get resolved on their own terms. They are usually superseded by nearer, harder constraints that reorder priorities overnight.
I assumed that disruption might be geological or astronomical — volcanic activity, an asteroid, or some other exogenous shock. What never occurred to me was that the disruptor would be self-inflicted: artificial intelligence and its unprecedented appetite for electrical power.
What follows is an engineering argument for why AI power demand is likely to overwhelm electrical grids regardless of one’s position on AGW, why efficiency gains will not solve the problem, and why current energy and climate policy is internally inconsistent in ways that cannot be sustained.
For a long time, I assumed that the climate debate would eventually be displaced by some other hard constraint. I was right about the displacement — and wrong about what would cause it.
1. Bracketing the AGW Debate
Before going any further, it is worth being explicit about what this piece is not.
It is not necessary to accept the AGW narrative in order to see the problem clearly.
Human perception of climate is local, episodic, and noisy. A modest shift in global average temperature does not guarantee noticeable change in any particular place, and many people — myself included — have not experienced a dramatic, lived change in seasonal patterns over a lifetime. Skepticism based on lived experience is not ignorance; it reflects the mismatch between statistical abstractions and phenomenological reality.
But even if AGW were entirely correct, entirely incorrect, or somewhere in between, the conclusions of this paper would not change. The contradiction I am concerned with exists independently of climate models.
2. The Core Contradiction
If carbon emissions represent an existential threat that justifies outlawing coal and natural gas, then the simultaneous build-out of the largest continuous electrical loads in human history is irrational.
AI data centers are not incremental loads. They are step functions.
A single hyperscale AI facility now demands tens to hundreds of megawatts of continuous power — equivalent to a steel mill, operating 24/7, with little tolerance for interruption. These facilities are being permitted, subsidized, and connected to grids that are already aging, capacity-constrained, and fragile.
You cannot simultaneously declare energy austerity and pursue unconstrained AI expansion. Yet that is exactly what current policy attempts to do.
3. Why Renewables Cannot Carry This Load
This is not an ideological claim. It is arithmetic and physics.
Solar and wind suffer from low capacity factors, intermittency, land-use intensity, and material demands that scale poorly. Grid-scale storage sufficient to cover multi-day or multi-week gaps does not exist at anything close to the required scale, nor is it being deployed fast enough to matter.
Renewables can contribute meaningfully at the margins. They cannot supply dense, continuous, rapidly growing baseload demand of the sort AI requires.
4. Nuclear: The Obvious Answer That Politics Blocks
Nuclear power is the only scalable, low-carbon baseload energy source that could plausibly support large-scale AI growth.
Modern Gen III+ reactors are already extremely safe by any rational engineering standard. Small modular reactors and Gen IV designs offer further promise, though timelines remain long.
The barriers are not technical. They are political, regulatory, and cultural — frozen in the fears of the late 20th century.
Absent a political reset, nuclear remains talked about and deferred rather than built.
5. Efficiency Will Not Save the Grid
The idea that compute efficiency gains (hardware or software) will reduce total energy consumption is historically naïve.
Efficiency improvements almost always trigger increased usage — a phenomenon known as Jevons’ Paradox. Compute is particularly elastic:
- Cheaper FLOPs lead to larger models
- Faster inference leads to more queries
- Better capability leads to more applications
- Strategic value drives further expansion
There is no natural mechanism that converts efficiency into restraint. Without explicit caps or rationing (bigger government, more regulation — my nemesis), efficiency gains will be reinvested into bigger, faster, smarter AI — not lower power consumption.
6. Grid Reality: The True Bottleneck
Regardless of climate beliefs, electrical grids face hard constraints:
- Aging transmission infrastructure
- Transformer manufacturing bottlenecks
- Long interconnection queues
- Thermal derating during heat waves
- Poor margins for reactive power and stability
AI data centers stress all of these simultaneously. They represent the worst possible load profile: continuous, dense, fast-growing, and interruption-intolerant.
Grid failure does not arrive as a single collapse. It arrives as price spikes, brownouts, priority loads, and quiet rationing.
7. Why Governments Accept the Risk
Governments are not confused. They are prioritizing.
AI is viewed — correctly or not — as a general-purpose strategic multiplier with military, intelligence, and economic implications. Losing leadership in AI is considered more dangerous than grid strain, higher prices, or localized outages.
As a result:
- Data centers receive priority access to power
- Emissions rules are bent or waived quietly
- Reliability is sacrificed before strategic capability
Climate policy remains rhetorically useful, but operationally flexible.
8. Displacement, Not Resolution
AI power demand does not refute AGW. It displaces it.
When policymakers face immediate, local, and undeniable constraints — blackouts, grid instability, national security pressure — abstract long-horizon optimization problems slide down the priority list.
This is how history works.
AGW did not need to be disproven to lose primacy. It only needed competition from a more immediate form of power hunger.
Conclusion
The electrical grid — not climate models — will be the ultimate constraint on AI growth.
Efficiency gains will press harder against that constraint, not relieve it. Renewables cannot carry the load alone. Nuclear remains politically blocked. Coal and gas bans collide with strategic reality.
The result is policy incoherence that will persist until forced to change by price shocks, reliability failures, or geopolitical necessity.
This outcome was predictable. The only real surprise is the identity of the disruptor.
AI, not climate, is now the problem that cannot be deferred.
No comments :
Post a Comment