AI Becomes Heavy Industry

AI appears on a screen as text, image, voice or code. Its physical form is far less ethereal.

It consists of data centres, processors, high-bandwidth memory, circuit boards, power supplies, fibre, cooling, water, transformers and power plants. It needs factories, mines, construction sites, permits and capital. A data centre is a factory that calls itself a cloud.

This factory turns electricity into computation, computation into models, and models — perhaps — into economically valuable decisions:

Electricity → compute → model capability → application → possible value

The word “possible” remains essential. Compute is not revenue, a model is not a business model, and a generated answer is not yet a productivity gain.

The physical build-out is nevertheless real. In its base case, the IEA expects global data-centre electricity consumption to more than double from roughly 415 terawatt-hours in 2024 to about 945 terawatt-hours in 2030. AI is the largest but not the only driver. China and the United States are expected to account for almost four fifths of the growth.^[6] Forecasts in this field are uncertain because chip efficiency, model design, utilisation and demand change quickly. Their scale still makes one point clear: AI has become a question for energy planning.

An old industrial truth returns to the digital economy: location matters.

A model can be offered globally, but its training and operation occur in physical places. Those places need grid capacity, firm power, cooling, data links, political stability and often water. A project may have enough chips and capital yet fail for want of a transformer, a connection agreement or local consent.

Attention goes where energy flows.

The phrase describes more than an investment trend. Where large quantities of reliable power are available, data centres, transmission projects, gas plants, renewable generation, nuclear contracts and industrial suppliers begin to cluster. AI pulls capital back into the physical world.

It also changes the political character of electricity. Households, industry and public services have long competed over price and security of supply. They are now joined by a consumer whose demand can be enormous, concentrated and backed by deep capital. A data centre may finance new generation and accelerate grid modernisation. It may also occupy scarce connection capacity, shift costs to local users or absorb firm power needed elsewhere.

The question is therefore not simply how much electricity AI consumes. It is who pays the additional cost and who receives the benefit.

If a hyperscaler enables a power plant or storage asset through a long-term contract, the region may benefit. If the public pays for grid reinforcement while most of the value accrues to a global platform company, the distribution is different. A kilowatt-hour may be physically neutral. Its contract is not.

AI is also changing the value assigned to generating technologies. Renewables can add energy quickly, but need temporal complements. Gas provides flexible power, yet binds AI to fuel and price risk. Nuclear promises continuous low-carbon electricity, but demands long planning horizons, high upfront investment and institutional competence. The scale of demand makes an ideologically pure system unlikely. The result will probably be a pragmatic mixture.

The financial wager is larger still. Companies are committing huge sums before it is clear which applications will pay reliably. Sellers of chips, grid equipment, cooling and energy may earn cash flows early. Infrastructure owners carry the risk that later revenue will justify the capital employed.

This creates a familiar asymmetry:

Capex sellers are paid for the build. Capex carriers are paid only if use eventually pays.

That asymmetry has not disappeared. But it is acquiring a new financial architecture.

On 10 August 2026, NVIDIA announced six independent financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR. Over time, they are intended to mobilise more than $500 billion of third-party capital for AI infrastructure. The partners have initially signed memoranda of understanding; final agreements, terms and deployment schedules remain outstanding.^[29]

The scope is notable. The plan is not merely to finance a chip purchase, but the chain of energy, land, grid connection, building and compute. Jensen Huang put the cost of one gigawatt of AI infrastructure at roughly $50 billion to $60 billion. Goldman Sachs wants to create a credit market backed by NVIDIA compute. BlackRock intends to draw on private and public capital as well as pension funds. KKR describes the value chain as running “from molecule to token”.

AI is therefore acquiring more than factories. It is acquiring its own capital market.

The division of labour over time is the key. Private capital can carry the first years, when a project does not yet produce a stable yield. Once construction, commissioning and utilisation create dependable cash flows, those flows can be transferred into long-duration credit and investment products. The aim is to turn a difficult stand-alone wager into standardisable infrastructure capital.

This is a particular strength of the American system. It can create strategic patience not only through public budgets but through financial engineering: risks are structured, pooled, priced and distributed across larger reservoirs of capital.

Financing, however, is still not the same as viability. The announced $500 billion is not an already funded common pool; NVIDIA says it is not putting its own balance sheet behind these platforms, and each capital provider must underwrite projects individually. David Solomon explicitly expects winners and losers; Jim Zelter returns the argument to revenue and cash flow.

The collateral value of compute is itself an assumption. It depends on systems remaining utilised, CUDA retaining its position, older hardware remaining economic and the fall in price per unit of compute not outrunning the growth in demand.

The bottleneck therefore shifts. The risk of inadequate financing becomes a risk of large-scale financialisation. If the assumptions hold, capital markets accelerate productive infrastructure. If they fail, losses are no longer confined to hyperscalers and operators; they spread into private credit, insurers, pension funds and public debt markets.

A future shake-out would not disprove AI. Railways, fibre networks and other infrastructures transformed society even though many early owners lost money. Technological usefulness and investment return are different questions.

The human question goes deeper. AI can make knowledge accessible, small businesses more productive and research faster. It can also erode capability when people delegate judgement, memory and responsibility to a system whose eloquence they mistake for certainty.

A sound human–AI relationship therefore needs a clear division of labour. AI can search, compare, simulate, translate and reveal contradictions. Humans must define goals, limits, acceptable risk and responsibility. The greater the consequences, the less human supervision may be reduced to a ceremonial click.

AI is neither a neutral hammer nor a new sovereign. It is an amplifier. It amplifies knowledge and error, productivity and waste, freedom and control.

Its largest energy question is therefore not how many terawatt-hours it consumes.

It is which human capabilities we want those terawatt-hours to amplify.

Sources and notes

1. International Energy Agency, Energy and AI — Energy Demand from AI: https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai
2. NVIDIA, NVIDIA Partners With Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to Establish AI Compute Infrastructure Financing Platforms to Mobilize Over $500 Billion of Third-Party Capital, 10 August 2026: https://nvidianews.nvidia.com/news/nvidia-partners-with-apollo-blackrock-blackstone-brookfield-goldman-sachs-and-kkr-to-establish-ai-compute-infrastructure-financing-platforms-to-mobilize-over-500-billion-of-third-party-capital; CNBC, Transcript: Becky Quick Speaks with NVIDIA’s Jensen Huang & Wall Street Leaders on $500B AI Infrastructure Push, 10 August 2026: https://pressroom.versantmedia.com/cnbc/press-releases/cnbc-exclusive-transcript-cnbcs-becky-quick-speaks-nvidias-jensen-huang-wall