Energy is an Issue for AI, but Not Existential

High-performance computing remains an energy-intensive business, but we must remain alert for non-linear change, as suppliers will have huge incentives to reduce power consumption. 

A common mistake is to assume“if AI usage grows 10×, electricity consumption must grow 10×.” And, to be sure, consumption will grow as usage grows.

 

source: NextEra

 

source: NextEra

source: NextEra

But energy requirements have never grown in a linear way with demand growth. Instead, computing energy usage per unit shows a clear pattern:

• Demand grows exponentially

• Energy per unit falls faster than demand rises

• Total energy grows—but sub-linearly.

Era	Energy per computation
Mainframes	Extremely high
Minicomputers	↓ ~10×
PCs	↓ ~100×
Mobile SoCs	↓ ~1,000×
Specialized accelerators	↓ another 10–100×

Domain-specific silicon, tuning inference operations to the level of required processing, model efficiency and the shift to less-energy-intensive inference operations all will play a part. 

When any key input, such as power, becomes the bottleneck, innovation follows. 

And, by the way, we might also note that every new computing wave has produced fears that energy supply would not keep up, as some argued:

• Mainframes would overwhelm power grids

• PCs would consume too much electricity

• The internet would break energy systems

• Data centers would cause permanent shortages

• Crypto would absorb global power

U.S. Department of Commerce, between 1966 and 1972, warned that centralized computing facilities could require “significant regional generation capacity.” IBM internal studies from the late 1960s likewise studied power density limits in computer rooms. That has not proven to be the case. 

Instead, computing power consumption has been cut in half (per operation) about every 19 months. 

Sure, AI operations will consume more power. But growth will not be linear, and will not be unmanageable. 

How do Computing Products Sold Close to Marginal Cost Recover Capital Investment?

Marginal cost pricing has been a common theme for many computing industry products. The concept is that retail pricing is set in relation to the cost of producing the next units, but not including amortization of any investments in infrastructure. 

Almost counter-intuitively, there are many examples of firms selling computing products at (or near) marginal cost (sometimes at prices near zero), yet still producing strong long-term capital recovery and attractive ROIC. 

That seems to defy economic logic, but it does work. How, since the investments in infrastructure still must be recovered?

Simply, the firms did not monetize the thing they priced at marginal cost, but instead “monetized what that thing made possible.”

The underlying economics, in computing, are simple: marginal cost collapses faster than average cost. So once a fixed investment is made (processor cycles; storage reads; memory access; bits transmitted; copies of apps), it is possible to price the commodity layer at marginal cost but then recover capital in a complementary scarcity layer. 

In other words, marginal cost pricing works when a supplier has something else to monetize, someplace else in the value chain or stack. 

The IBM mainframe business sold batch jobs or processor time as a service at marginal cost. But IBM recovered the cost of its invested capital other ways. It’s margins on hardware were high. Its customer lock-in was similarly high. 

And IBM was able to sell system engineering and software pertaining to its machines and ecosystem. 

Layer	Monetization
Hardware	Extremely high gross margins
Switching costs	Proprietary architectures
Integration	Services + system engineering
Software lock-in	Non-portable applications

So marginal cost pricing for compute services worked because customers could not switch platforms. So financial returns came from other elements of the platform. 

Microsoft provides another example. Copies of Windows; Office and developer tools were sold at affordable prices. But Microsoft made its profits from its operating system “monopoly;” developer lock-in; bundled distribution and version upgrades. 

Layer	Explanation
OS monopoly	Controlled application access
Ecosystem tax	Developers required Windows
Version upgrades	Periodic re-monetization
OEM bundling	Forced distribution

So Windows licenses were “cheap” relative to value delivered, with an incremental cost that was effectively near zero, but with profit margins near 90 percent. 

What Microsoft essentially monetized was its control of the “standard” for operating systems and the platform. 

Google search arguably offers an even-more-compelling case, as the product is available to users at zero cost. 

Search queries cost nothing. Neither does use of Google Maps, Gmail, Android or the Google productivity suite. 

But with its advertising monetization, Google creates a revenue model based on user attention. 

Layer	Value capture
User attention	Scarce
Intent data	Extremely scarce
Ad auctions	Competitive pricing
Data feedback loops	Increasing returns

So apps requiring lots of compute infrastructure are monetized other ways. “Compute” is not the product; audiences are. 

Amazon Web Services, it can be argued, prices core products near marginal cost (EC2 compute, S3 storage, throughput). 

Mechanism	Explanation
Scale advantage	Lowest unit cost globally
Demand aggregation	Extremely high utilization
Service layering	Databases, AI, analytics
Switching friction	Architecture dependence

So AWS monetizes risk reduction and reliability rather than compute cycles. “Trust” creates the revenue model while lock-in sustains it. 

Perhaps the best example is Open Source, which, by definition, is “free to use.”

Products such as Red Hat are sold at marginal cost (licensing) or the software itself is available at no cost. 

Scarce layer	Revenue source
Support	Enterprises pay for certainty
Certification	Compatibility guarantees
Hosted services	Managed convenience
Security updates	Operational risk reduction

The Apple business model might not seem to be a case of marginal cost pricing for hardware, as such pricing is not bound by marginal cost parameters. 

On the other hand, the ecosystem software (iOS, macOS, developer tools, some cloud services) actually can be characterized as being made available at marginal cost. 

Apple recovers its infrastructure and sunk costs from hardware profit margins, ecosystem lock-in and services. 

Firm	What Sold at Marginal Cost	What Recovered Capital
IBM	Compute usage	Hardware + lock-in
Microsoft	Software copies	Platform control
Google	Search	Advertising
Amazon AWS	Compute	Scale + reliability
Red Hat	Software	Support & ops
NVIDIA	Runtime compute	Chips + ecosystem
Apple	OS + tools	Devices + services

If computing marginal cost approaches zero, then retail pricing also tends to fall to near zero, while successful firms find other places in the value chain to recover capital investment costs. 

Where might scarcity value remain?

• Time

• Trust

• Risk transfer

• Attention

• Control points

• Integration responsibility

• Physical manufacturing

• Distribution