What the Industry Projects vs. What the Physics Shows

A Side-by-Side Comparison of AI Growth Narratives

THE MARKET STORY

What AI companies and analysts are projecting:

Financial Projections

2025 Global AI Market: $757 billion
2030 Projection: $1.8-3.7 trillion
2033 Projection: $4.8 trillion
Economic Impact by 2030: $15.7-22.3 trillion added to global GDP
Growth Rate: 19-37% CAGR

Adoption Projections

Current Enterprise Adoption: 78% of companies using AI
Generative AI Deployment: 90% of organizations
Job Creation: 97 million AI-related jobs needed by 2025
Industry Transformation: Manufacturing gains $3.78T, Financial services $1.15T, Professional services $1.85T

Narrative

Exponential growth curves going up and to the right
AI as inevitable economic force
Massive productivity gains across all sectors
Job creation and transformation
Unlimited scaling potential

Key Assumption: Energy and infrastructure will scale to meet demand

THE PHYSICS STORY

What the energy data actually shows:

Energy Requirements

2024 Data Center Consumption: 448 TWh globally
2030 Projection: 980 TWh (119% increase in 6 years)
US Alone: 183 TWh (2024) → 426 TWh (2030) = 133% increase
By 2030: Data centers = 7.8% of US electricity (up from 4% in 2025)
By 2030: Equivalent to Japan's entire electricity consumption

Training Cost Trajectory

GPT-3: 1,287 MWh
GPT-4: 50,000 MWh (39× increase)
GPT-5 (projected): 1,500,000 MWh (30× increase)
Total Scaling: 1,165× increase across three generations

Reality Check: A human brain uses 8.76 MWh over an entire 80-year lifetime. GPT-4 training used 5,700× more energy.

Grid Constraints (Current Reality)

Utility Connection Delays: Up to 5 years
Ohio Example: Data center requests dropped from 30 GW to 13 GW after state imposed 85% energy-use-or-pay tariff
Bay Area: Data centers sitting empty, no power until 2027-2028
IEA Warning: 20% of planned projects may be delayed due to grid constraints
Required Investment: $720 billion in grid upgrades just to meet demand

What's Actually Happening

Gas turbine orders hit 20-year high (14 GW in 2024, 18 GW in H1 2025) driven entirely by data center demand
At least 10 GW of coal retirements delayed or converted to gas since early 2025
Nuclear plant restarts: ~3 GW coming back online by 2030
States implementing protective tariffs to shield residential customers from infrastructure costs
Utility analysts report data centers "underperforming expectations" due to cancellations and delays

Inference Reality

80-90% of AI electricity use is inference (running models), not training
Daily ChatGPT usage: 1 GWh (33,000 US households' daily consumption)
No economy of scale: Every query costs energy, scales linearly with adoption
If usage grows 10×, energy cost grows 10×

Key Reality: Energy infrastructure cannot scale at the rate AI companies are projecting

THE CONTRADICTION

The Arbitrage That's Closing

Why AI Looks Cheap (For Now):

Energy prices don't reflect true infrastructure costs

Data center rates are subsidized
Grid upgrade costs externalized to ratepayers
Environmental costs not priced in

Human labor costs include total metabolic overhead

Housing, healthcare, food, rest, family time
AI gets to externalize its "metabolic" costs (electricity, cooling, infrastructure)

Capital abundance masks operating costs

Venture funding absorbs losses during training
True operational costs hidden in cloud bills

Temporary regulatory environment

Infrastructure costs being socialized
But states are now pushing back (Ohio, Bay Area, others)

What Happens When:

Energy prices rise (already happening)
States refuse to socialize infrastructure costs (already happening)
Grid capacity constraints delay projects (already happening)
Environmental costs get priced in (coming)
Capital becomes scarce (cyclical)

That 100-watt human starts looking very efficient again.

THE THERMODYNAMIC REALITY

Natural Intelligence (Human)

Power: 100 watts total (brain: 20W)
Fuel: Food (ultimately solar-captured energy)
Lifetime Energy: 8.76 MWh (80 years)
Scaling: Sublinear (experience improves efficiency)
Stopping Rule: Internal (knows when enough is enough)
Context: Direct contact with reality

Simulated Intellect (AI)

Training Power: 50,000+ MWh (GPT-4)
Inference Power: 0.3 Wh per query, scales linearly
Infrastructure: Megawatts for data centers, cooling, transmission
Scaling: Linear or superlinear
Stopping Rule: None (external limits only)
Context: Symbolic space, one layer removed

Energy Ratio: GPT-4 training = 5,700 human brain lifetimes

THE PATTERN WE'RE REPEATING

The Economy We Don't Count

Before we even get to the formal economy, there's a larger economy we've chosen not to see.

The Care Economy (Invisible):

Global value: $10.8 trillion (when measured)
Household work alone: 20-50% of GDP
Who performs it: Primarily women, unpaid or underpaid
What it includes: Childcare, eldercare, emotional labor, food prep, community maintenance, teaching, nursing
Energy profile: 100-watt human scale, self-renewing
Can AI replace it? No—requires embodied presence, context-sensitivity, direct contact with reality

The Formal Economy (What We Measure):

Built on top of the care economy foundation
Only counts market transactions and government spending
Rewards abstract coordination over embodied care
GDP counts negatives (oil spills, tobacco, arms) as positives
Ignores household, volunteer, and natural economies

What This Means:

The largest economy that exists—the one that makes all other economic activity possible—operates at 100-watt human metabolic scale and is completely invisible in GDP accounting.

Meanwhile, we're building AI infrastructure that requires megawatts to automate the expensive superstructure—the abstract intellectual work that sits on top of the care foundation.

The automation pattern is backwards: AI replaces the energy-expensive intellectual work while the energy-efficient care work (which is larger and more essential) remains unchanged.

We Already Ran This Experiment

Education:

Chose abstract symbol manipulation over embodied learning
Result: 70% of teens cite anxiety/depression as major problem
High-anxiety students 2 years behind by end of elementary school

Economy:

Paid consultants ($101K) more than nurses ($93.6K)
Overvalued abstract coordination, undervalued embodied judgment
Result: 76-82% knowledge worker burnout

Productivity:

Invested trillions in IT
Knowledge workers now spend 20% of time searching for information
Process 121 emails + 226 messages daily
Email volume increased 80× since 1970s
Result: Productivity growth collapsed from 3% (1960s) to 0.7% (2020s)

What AI Inherits: All of it. The same confusion between intellect and intelligence. The same overvaluation of abstraction. The same linear or superlinear scaling. Just at data center scale, with the energy cost visible in gigawatt-hours instead of hidden in glucose and cortisol.

THE FEEDBACK LOOP

Two projections. One experiment.

Industry projects: $1-4 trillion by 2030-2033, exponential growth, $15-22 trillion economic impact

Physics shows: Energy doubling by 2030, 20% of projects delayed, 5-year waits, states pushing back, coal staying open

This isn't a competition between narratives. It's a superorganism discovering the boundaries of its phenomenon through direct encounter.

The delays aren't failures. They're information. The constraints aren't obstacles. They're feedback about what's sustainable at scale. The rising costs, the grid limitations, the protective tariffs—these are how living systems learn metabolic limits.

What the Feedback Is Revealing

The pattern that worked at one scale produces different results at another:

At individual scale:

Ignore care work, reward abstraction
Result: Some burnout, manageable

At institutional scale:

Same pattern, GDP measures only markets
Result: 76-82% burnout, productivity collapse

At infrastructure scale:

Same pattern, automated via AI
Result: Megawatt demand, grid constraints, delays

The superorganism is teaching itself: The accounting method that made care work invisible doesn't scale to data center infrastructure. The energy cost that could be hidden in glucose and unpaid labor becomes visible as gigawatt-hours and infrastructure delays.

The Foundation That Cannot Be Automated

While the experiment runs, the $10.8 trillion care economy continues:

Operating at 100 watts per person
Making all formal economic activity possible
Invisible in GDP accounting
Cannot be replaced by megawatt infrastructure

This isn't ideology. It's metabolism.

A superorganism running an experiment: can symbolic processing at data center scale deliver more value than embodied intelligence at human scale?

The feedback is arriving. Not as judgment. As constraint. As information about boundaries.

THE LEARNING PROCESS

We've already made one error: building an economic system that ignores the largest, most efficient, most essential economy—the $10.8 trillion care economy operating at 100-watt human scale.

Now we're running a second experiment: automating the expensive intellectual superstructure while calling it "the future of work."

The superorganism is discovering:

The superstructure depends on the invisible care foundation
The superstructure is energy-expensive and produces burnout
The foundation is energy-efficient and irreplaceable
AI can only touch the superstructure, not the foundation
The feedback arrives as delays, constraints, rising costs

This is how living systems learn boundaries—not through prediction, but through direct encounter with limits.

What happens next isn't a choice between paths. It's feedback integration:

Pattern A: Continue Until Constraint Forces Adjustment

Scale AI to automate formal economy
Energy demand doubles, infrastructure lags
Delays accumulate, costs rise, projects cancel
The 100-watt humans doing care work remain unchanged
System discovers: the foundation was load-bearing all along
Learning: hard constraint, forced recognition

Pattern B: Recognize Feedback Early

See delays and constraints as information about boundaries
Notice: $10.8 trillion foundation operates at 100W, can't be automated
Use AI to reduce unnecessary intellectual overhead
Support humans in the energy-efficient mode
Learning: soft constraint, adaptive response

Neither is a moral choice. Both are learning processes.

The difference is whether the superorganism integrates feedback before or after encountering hard metabolic limits.

THE PATTERN

We built an accounting method that treats care as zero and destruction as growth. Oil spills add to GDP. Parents caring for children contribute nothing.

We're now scaling that accounting error to megawatt infrastructure.

The feedback is arriving:

20% of projects delayed
5-year utility waits
States protecting grids
Coal staying open
$720B upgrades needed
Infrastructure can't keep pace

This isn't failure. It's information.

A superorganism discovering: the pattern that made care work invisible produces different feedback at data center scale.

The thermodynamics:

Natural intelligence: already built $10.8T economy at 100W per person
Simulated intellect: requires megawatts, scales linearly, hitting constraints
Care foundation: makes survival possible, can't be automated
Formal superstructure: depends on foundation, increasingly automated
Physical limits: arriving as delays, costs, infrastructure gaps

What the system is teaching itself:

The economy we didn't count is larger, more essential, and more efficient than the one we did.

The work that operates at 100 watts cannot be replaced by work that requires megawatts.

The foundation makes the superstructure possible, not the reverse.

Natural intelligence is cheap.

Simulated intellect is not.

The learning process is already underway.

The only question is how much constraint is required before the pattern becomes visible—before a superorganism recognizes its own boundaries not through theory, but through the direct experience of encountering them.

The feedback is information about what's sustainable. Whether we call it "the grid deciding" or "nature forcing us" or "hitting limits"—it's the same phenomenon: a living system learning the edges of its metabolic niche through direct encounter.

The data is clear. The constraints are real. The foundation remains unchanged.

What remains is integration: will the superorganism adjust its pattern before hard limits force the adjustment, or discover its boundaries the way all living systems eventually do—by reaching them.

Sources

AI market projections: Grand View Research, Statista, IDC, ABI Research, UNCTAD
Energy data: IEA, Goldman Sachs, S&P Global, EPRI, Bain & Company
Grid constraints: utility company reports, regulatory filings, industry analysis
Human brain data: neuroscience consensus literature
Burnout/productivity: BLS, Deloitte, Asana, NEA
Student data: Pew Research, APA

The numbers don't lie. The question is whether we're willing to see what they're telling us.