They Don’t Have AC — How Will They Have AI?

The global discourse on artificial intelligence treats compute as software, assuming the hardware infrastructure is an invisible, ubiquitous commodity. It presumes that every user on earth possesses continuous electrical power, ambient thermal regulation, and high-speed fiber connectivity.

This assumption is false. Much of the world has none of these three foundational elements, and no degree of algorithmic optimization or parameter scaling supplies them.

Frontier Compute Depends on Heavy Infrastructure Nobody Mentions

Artificial intelligence is not virtual. It is an industrial process that consumes concrete, copper, electricity, and water. High-density data centers require continuous baseload power, industrial liquid cooling systems, and high-bandwidth fiber backbones. Modern frontier models demand compute clusters operating at scales that strain the electrical grids of developed nations, yet public policy discussions evaluate AI deployment as if it were a software download.

The physical footprint of this technology generates immediate local resource conflicts. Global data center water consumption is projected to reach 664 billion liters annually by 2030, up from 239 billion liters in 2024. Training a single model like GPT-3 consumed an estimated 700,000 liters of freshwater for direct evaporative cooling alone.

When placed in developing economies, these physical requirements provoke severe backlash. During Uruguay’s worst drought in 74 years—when public reservoirs dried up and municipal utilities pumped brackish water into urban taps—Google faced widespread protests over a proposed data center in Canelones designed to consume millions of liters of freshwater daily. The conflict yielded charges of “data colonialism” and forced the company to abandon its evaporative cooling plan for an air-cooled model (The Borgen Project).

A similar confrontation occurred in Santiago, Chile. In the Cerrillos district, community organization MOSACAT challenged a $200 million Google facility slated to extract 7.6 million liters of water daily from an aquifer facing severe depletion. In February 2024, an environmental court revoked the facility’s permit, forcing Google to halt construction and redesign the site with a closed-loop cooling system by September 2024 (Disconnect Blog). Compute cannot exist without physical resources, and physical resources are contested.

Seven Hundred Million People Lack Basic Electricity and Cooling

The debate around global AI access ignores basic energy deficits. Approximately 730 million people worldwide lived without electricity access in 2024. Over 80% of those without power—around 600 million people—reside in Sub-Saharan Africa, where grid extension has stalled as population growth outpaces infrastructure investment and national debt burdens restrict public capital expenditure.

Where grids exist, they frequently fail to provide stable baseload energy. In South Africa, state utility Eskom enforced multi-year rotational “load-shedding” lasting 8 to 12 hours daily during peak crises. Local IT infrastructure, mobile towers, and server hosting sites were forced to run on expensive diesel generators, causing widespread telecom degradations. Localized distribution failures and grid theft continue to compromise reliability (African Business). AI models hosted on cloud infrastructure cannot maintain availability when access networks lose power multiple times a day.

AC Household Adoption Rates (Selected Nations)
--------------------------------------------------
United States   [=========================] 90%
Japan           [=========================] 91%
Indonesia       [==] 9%
South Africa    [=] 6%
India           [=] 5%
Sub-Saharan Africa [<1] <5%
--------------------------------------------------

The thermal barrier is equally severe. Between 1.17 and 1.2 billion people in low-income urban and rural areas face extreme heat risks without access to basic cooling technology. Household air conditioning adoption stands at 5% in India, 6% in South Africa, 9% in Indonesia, and below 5% across Sub-Saharan Africa. In contrast, household AC adoption reaches 90% in the United States and 91% in Japan. Expecting populations that lack household air conditioning to operate heat-emitting personal compute hardware is a fundamental miscalculation of physical realities.

Connectivity metrics reinforce this isolation. 2.6 billion people—roughly 33% of the world’s population—remained completely offline in 2024, with 1.8 billion of the unconnected living in rural areas. Internet penetration in low-income countries is 27%, compared to 93% in high-income economies. The International Telecommunication Union estimates that achieving universal, meaningful broadband connectivity by 2030 will require $2.6 trillion to $2.8 trillion in direct infrastructure and demand-side funding. Software updates do not lay subsea cables or finance distribution transformers.

Mobile Telecom Leapfrogged Wirelines, But AI Demands Kilowatts

A common counter-argument claims that developing nations will simply “leapfrog” legacy compute infrastructure, just as Sub-Saharan Africa bypassed landline telephony in favor of mobile networks and digital wallets.

This analogy fundamentally misunderstands technological physics. Mobile telephony replaced heavy copper wire lines with lightweight cellular towers consuming 1 to 2 kilowatts of power to service low-power 2G signals. It substituted light infrastructure for heavy infrastructure.

Frontier AI requires the exact opposite: unprecedented physical density. It demands gigawatt-scale electrical generation, heavy industrial cooling plant operations, and high-throughput fiber optics. World Bank President Ajay Banga addressed this misconception directly, warning that AI requires massive physical compute and energy capacity that cannot be leapfrogged without building heavy industrial infrastructure first (Cr4fts) (VoxDev).

Leapfrogging Architecture Comparison
-----------------------------------------------------------------
Technology       Infrastructure Strategy        Energy Requirement
-----------------------------------------------------------------
Mobile (2G/3G)   Light substitution             1-2 kW per tower
Frontier AI      Heavy central aggregation      Gigawatts per cluster
-----------------------------------------------------------------

Alternative proposals to run Small Language Models (SLMs) locally on edge devices face hardware constraints. On-device inference requires high RAM capacity, fast storage controllers, and specialized Neural Processing Units. Yet over 60% of smartphones in low-income regions cost under $100, possessing 2 to 4 GB of total RAM and 32 to 64 GB of storage. Running quantized edge models rapidly drains phone batteries in regions where users rely on commercial charging kiosks or intermittent power grids. Furthermore, initial training, continuous fine-tuning, and weight distribution remain tethered to centralized cloud infrastructure.

Off-grid solar micro-grids and low-Earth-orbit satellite networks (such as Starlink) are similarly inadequate substitutes for grid-scale compute. Solar-and-battery micro-grids can charge mobile phones and power basic LED lighting, but scaling off-grid solar to power megawatt- or gigawatt-scale compute hubs is cost-prohibitive compared to utility grids (Brookings Institution). Satellite connectivity delivers wide geographic coverage, but it suffers from high latency, equipment costs that exceed average monthly household incomes, and strict bandwidth caps per geographic cell (ITU Blueprint).

Real-World AI Deployment Adapts to Low-Bandwidth Off-Grid Channels

Because frontier deployment models fail under low-resource conditions, practical AI execution in constrained environments must adopt asynchronous, low-bandwidth architectures. System architects operating in Sub-Saharan Africa and South Asia build around existing legacy channels rather than high-throughput web APIs.

In Ghana, tech enterprise Farmerline developed Darli AI to distribute agricultural advice, weather forecasts, and pest identification to smallholder farmers. The platform operates over basic voice calls and SMS in over 20 local African languages, serving over 110,000 farmers who lack smartphones or fast data connections (Agriweb). The underlying model runs centrally, but the delivery interface bypasses real-time streaming web applications entirely.

Infrastructure Constraints vs. Deployment Architectures
------------------------------------------------------------------------------------
Constraint               Standard Model                 Constrained Adapter
------------------------------------------------------------------------------------
High Latency / Offline   Continuous Cloud Streaming    Asynchronous SMS / Voice Gateway
Low Device RAM           Local SLM Processing          Edge Pre-Processing / Remote Batch
Intermittent Power       Always-On Edge Inference      Periodic Store-and-Forward Sync
------------------------------------------------------------------------------------

In Kenya, Apollo Agriculture uses satellite vegetation data, phone metadata, and machine learning models to issue micro-loans and agronomic guidance to over 200,000 smallholders. The inference occurs in remote cloud environments, while the user interaction relies on asynchronous SMS exchanges, stripping away bandwidth requirements.

Data collection platforms face similar design constraints. Karya, an engineering enterprise in India, employs rural workers to record and transcribe local-language voice datasets. To bypass rural network limitations, Karya’s mobile application operates offline on low-cost Android devices, storing audio recordings locally and syncing with central servers only when workers periodically enter areas with cellular coverage. These architectural adapters allow functional utility, but they deliver narrowed capability compared to low-latency agentic web systems.

Without Physical Infrastructure, the Global Divide Hardens into Concrete

If AI deployment requires dense physical infrastructure, its economic benefits will concentrate strictly where that infrastructure exists.

High-income economies with stable power grids, advanced liquid-cooling capacity, and universal fiber access will deploy autonomous agent networks to accelerate product development, enterprise operations, and scientific research. Economies lacking basic electricity and thermal management will remain excluded from real-time agentic systems.

This divide is structural. When an economy lacks basic grid power for 730 million citizens and household cooling for 1.2 billion people, it cannot participate in high-density compute clusters. Algorithmic efficiency gains will not bridge this gap. A model optimized to run with fewer parameters still requires electrical current to flip bits and thermal management to dissipate heat.

Without massive physical investments in power generation, transmission, cooling systems, and subsea fiber links, advanced intelligence remains trapped in the global North. The AI divide is not made of software. It is made of concrete, copper, and power grids.