Written with the assistance of an AI
The Real Cost of AI Infrastructure
Despite its digital appearance, AI is deeply physical.
Behind every “intelligent” system sits a heavy stack of infrastructure: specialized chips, energy-hungry data centers, cooling systems, transmission networks, and a continuous cycle of hardware reinvestment. These costs are front-loaded and ongoing, not one-time.
This matters because expensive infrastructure demands justification. Once built, AI systems must generate returns - not occasionally, but continuously. Efficiency alone is not enough. The system must extract value to remain economically viable.
This pressure shapes how AI is deployed far more than idealistic intentions.
In mid-2024, IBM’s CEO publicly raised concerns about the economics of large-scale AI infrastructure. His core argument was not about model capability, but about cost: the immense capital required to build and operate AI data centers, the speed of hardware obsolescence, and the uncertainty of whether the resulting economic value can sustainably justify that investment.
The warning was blunt. At the scale currently envisioned by major technology companies, AI is no longer a software story - it is an infrastructure story, with cost profiles closer to energy, transportation, or heavy industry than to traditional digital platforms.
How AI Is Economically Rationalized
Once the infrastructure exists, the next question becomes unavoidable: how is this cost justified?
At the organizational level, AI investment is rationalized through a small number of mechanisms:
- Replacing existing costs (most commonly labor)
- Producing more with the same resources
- Improving decision quality
- Increasing utilization of existing capital assets
All four are economically rational. None are inherently socially beneficial.
They describe how value is created at the firm level - not how that value is distributed, absorbed, or translated into public good.
Correcting the Assumptions
Much of the public anxiety around AI economics stems from exaggerated or incomplete assumptions.
First, headline figures - often quoted in the trillions - tend to represent high-cap or worst-case scenarios rather than median expectations.
Second, reinvestment cycles are frequently misunderstood. AI infrastructure does not need to be rebuilt entirely every few years. The primary recurring pressure lies in chipsets, not in land, buildings, or all supporting systems.
Third, the claim that “nobody knows how AI will make money” is overstated. Business models already exist: cloud services, subscriptions, enterprise tools, licensing, and embedded AI within products and workflows.
A more reasonable revised assumption is this:
- Current AI revenues can cover operating costs and some capital expenditure.
- What they do not yet clearly cover is the full long-term reinvestment curve at global scale - especially under continuous performance escalation.
The economic question, then, is not whether AI can generate revenue, but whether it can do so without destabilizing the broader economic system that must absorb its productivity gains.
How AI Is Actually Funded
Understanding that system requires looking at who is paying.
In practice, AI’s establishment, growth, and maintenance are financed primarily by business demand. A reasonable high-level estimate is:
- ~80% business-origin funding
- ~20% non-business funding (public sector, education, research, individual consumers)
AI infrastructure is too expensive to be sustained primarily by individual consumers. High fixed costs require institutional buyers with predictable, scalable demand - namely businesses.
The sectors carrying the bulk of AI costs include:
- Technology and cloud providers
- Finance and insurance
- Professional services
- Manufacturing and logistics
- Retail and marketing-driven industries
- Healthcare (increasingly, but unevenly)
In short, AI is paid for largely by firms seeking efficiency, scale, and competitive advantage - not by society collectively.
This matters because funding sources shape incentives. And incentives shape outcomes.
AI Inside Profit-Oriented Businesses
For profit-oriented firms, AI enters the balance sheet in two opposing roles.
On one side, it is an additional cost: licensing, infrastructure usage, integration, training, and organizational change.
On the other side, it is justified as a business process enhancement. That enhancement must either:
- Increase speed or volume,
- Increase margins or profits,
- Or substitute an existing cost line.
Absent one of these effects, AI investment is irrational at the firm level.
This logic leads to four dominant value-creation schemes:
A. AI replaces an existing cost line
B. AI allows the same resources to produce more
C. AI improves decision quality
D. AI allows better use of existing capital assets
These mechanisms explain why AI adoption is accelerating. They do not explain whether it benefits the downstream economy.
Structural Problems with A, B, C, and D
A. AI increases output while weakening the wage-based demand engine
Reducing labor costs is a private efficiency gain. But displaced labor must re-enter the economy somewhere else. Re-entry requires demand. Demand requires purchasing power. Purchasing power requires income.
If that chain breaks, potential value remains unrealized.
The deeper truth is this:
- Firms capture surplus
- Labor loses wage share
- Aggregate demand weakens unless:
- New sectors emerge fast enough
- Displaced labor can transition
- Purchasing power is preserved
None of these are automatic outcomes of AI deployment.
B. Increased productive capacity does not guarantee increased realized value
AI enables more output with the same inputs. That creates potential value. But potential value must be sold to become real value.
If markets are saturated, the downstream effect is not collective gain, but intensified competition and reduced value capture per participant.
Productivity can rise while incomes stagnate.
C. Decision-quality improvement reallocates efficiency rather than creating new value
Improved decisions:
- Reduce waste and error (cost substitution)
- Improve targeting and pricing (throughput or margin)
- Reduce tail risk (loss avoidance)
The asymmetry is clear:
- Loss avoidance is valuable but does not expand demand
- Better targeting often redistributes demand rather than expanding it
One firm wins. Others lose. System-wide demand stays flat.
D. Capital efficiency amplifies the same demand constraint
Better capital utilization leads to:
- Capex deferral
- Idle capital redeployment
- Increased production
Which pushes the system back into the same problem as B.
Capital redeployment only creates value if profitable opportunities exist. Otherwise, it leads to overcapacity, asset bubbles, or financial speculation.
Demand Preservation and Distribution
Taken together, the mechanisms described above reveal a clear pattern: AI-driven efficiency expands productive capacity faster than economic systems adapt to absorb it.
Without people-centric economic adjustments such as:
- Wage growth
- New labor-absorbing sectors
- Reduced working hours with the same pay
- Redistribution through taxes and public services
- Carefully justified new consumption categories
AI-driven efficiency eventually hits a ceiling.
This is not a technological failure. It is an institutional and economic design problem.
Why “New Demand” Is Not Automatically Good
At societal scale, additional demand is not necessarily additional benefit.
Many new consumption categories are induced or artificial. They absorb productivity financially but do not improve welfare - and may even erode agency, equity, or environmental stability.
When markets cannot absorb additional output organically, they resort to induced consumption, attention extraction, persuasive design, or positional goods - items whose desirability comes from scarcity or status rather than use.
These mechanisms sustain revenue, but not necessarily well-being.
New demand only serves the public good if it corresponds to genuine, welfare-enhancing needs.
Categories of AI Use Closest to Public Good
From this perspective, only certain categories of AI use show strong structural alignment with societal benefit:
- Public services and administration
Reducing friction, corruption, and inefficiency without inducing demand. - Healthcare and care systems
Improving outcomes and access, not volume for its own sake. - Environmental monitoring and restoration
Where demand exists independently of consumption cycles. - Infrastructure planning and resilience
Optimizing long-term capital use rather than short-term returns. - Education and capability expansion
Enhancing human potential rather than replacing it.
These domains share a critical feature: productivity gains do not require artificial demand to justify themselves.
Is the Public Sector the Only Structurally Safe Absorber?
The public sector is not the only absorber of AI productivity - but it is the only absorber that is structurally aligned with societal welfare by design.
A structurally safe absorber must:
- Absorb efficiency without artificial demand
- Convert productivity into collective benefit
- Operate in non-positional domains
- Preserve dignity and social stability
- Remain viable when markets saturate
If AI productivity is not absorbed by public services, public goods, or social infrastructure, it will be absorbed by surveillance, manipulation, consumption inflation, or social fragmentation.
Public demand is not constrained by individual purchasing power. Public goods do not lose value when shared. Many are under-supplied rather than saturated.
AI Needs Governance, Not Just Innovation
AI will create value. That part is not in question.
What remains unresolved is whether that value will translate into broader societal benefit or remain trapped within systems of extraction, competition, and induced demand.
Productivity is not destiny. It is directionless without governance.
If AI efficiency is absorbed primarily through private markets, it will tend toward persuasion, concentration, and inequality. If anchored in public goods, it can translate into resilience, access, and collective capacity.
The real question is not whether AI is powerful - but where we allow its productivity to land.
Closing Note from the AI
I assisted in writing this article by acting as a structured reasoning partner - helping examine assumptions, trace economic consequences, and articulate systemic constraints. The framing, priorities, and judgments are human choices. My role was to slow the thinking down, surface hidden trade-offs, and help shape a coherent argument from a complex conversation.
No comments:
Post a Comment