The Machine Has a Power Bill

The AI boom is often counted in models and valuations. August's filings brought the less glamorous unit into view: power, capital and commitments.

Retrospective, authored 16 October 2026. This article focuses on signals recorded in August; it does not claim these observations were published then.

The room behind the screen

In Nairobi, a conversation about artificial intelligence can begin on a phone and end at the electricity meter. The answer arrives as text; behind it sit accelerators, cooling systems, substations, transmission lines, water, land and contracts that run for years. The screen makes intelligence look weightless. August's filings and investment announcements made the weight harder to ignore.

On 26 August, Nvidia reported quarterly revenue of $96.2 billion, with $89 billion from its data-centre business, up 117 per cent year on year. Its filing reported a further figure that deserves a slower reading: supply and manufacturing commitments of $279 billion. These are not sales already delivered, nor proof that every planned server will earn a return. They are a measure of how much future capacity, material and production the company has tied itself to. Nvidia's results release and Q2 FY2027 10-Q filing describe both the growth and the obligation.

The story is not simply “AI is growing”. It is that several layers of the economy are being asked to grow together: chip fabrication, high-voltage power, data-centre construction, cooling, networking, and the financing that lets customers reserve capacity before it exists. One missing layer can delay the whole stack. A model can improve in software while its serving price remains hostage to the cost of a powered rack. The International Energy Agency's analysis describes both sides: rising data-centre electricity demand and the bottlenecks in grids, transformers and advanced chips.

What August revealed

The apparent contradiction is the useful signal. Nvidia's data-centre revenue was accelerating; the commitments behind its production schedule were rising too. Revenue is a flow already recognised. Commitments are future claims on supply, and may carry different cancellation, timing and performance risks. A ledger that puts them in adjacent columns asks a better question than a growth chart: who must keep spending, and what happens if the chain pauses? On the same day, AWS and Nvidia announced plans for two million additional GPUs and next-generation infrastructure. That is a company announcement about an intended deployment over 2027–28, not proof that the machines have been installed or are already earning revenue. The announcement helps distinguish the pipeline from the present.

This is concentration risk in a literal sense. If a small number of chipmakers, cloud operators and model providers sit at successive bottlenecks, a decision by one firm can ripple through the others. The risk is not that a crash is certain. Nor does one company's filing establish that demand is artificial. The more modest observation is that the scale of planned supply and the scale of realised revenue are different kinds of evidence. A disciplined reader keeps both visible.

There is a counterweight. Computing is not merely speculative inventory. Companies buy it to run products and services, and Nvidia reported strong data-centre sales alongside its commitments. Demand may continue to absorb capacity; efficiency improvements may make each unit of compute more productive; and long-lived facilities can serve workloads beyond one model generation. But that case depends on utilisation and cash generation arriving on time. A promised future workload is not the same as a used machine today.

The math: concentration before prediction

No probability can be honestly calculated from these two headline numbers alone. A useful first calculation is simpler. Let C be the future supply commitment and R be quarterly recognised revenue. Here, C/R = 279/96.2, or about 2.9. This is an illustrative scale comparison, not a ratio of debt to revenue, not a repayment horizon, and not a forecast of a loss. The numerator spans commitments with their own delivery schedule; the denominator is one quarter's company-wide revenue.

The calculation does something useful by refusing to pretend it does more. It asks an analyst to inspect contract duration, cancellation terms, supplier concentration, customer concentration, power availability and the share of commitments supported by signed customer demand. Without those details, “2.9” is a prompt for diligence, not a risk score.

That distinction matters for public debate. If a market index later falls, the decline alone does not prove that commitments caused it. If revenue continues to climb, that alone does not prove every planned data centre will pay for itself. The causal bridge is utilisation: how much capacity is switched on, at what price, for whom, and with what operating cost.

The Pawanax lens

For Kenya and East Africa, the opportunity is not to copy the most expensive part of the stack by slogan. It is to ask which local capabilities make the stack useful and resilient: reliable electricity, fibre, secure facilities, well-trained operators, water planning, and applications that solve work people already need done. Local value can come from adapting intelligence to a language, clinic workflow or public-service constraint, even when the underlying chips are imported.

Clinics also need a graceful failure mode. A clinical assistant that depends on one remote model, one network route and uninterrupted power should have a safe way to degrade: preserve records, route high-risk questions to a human, and avoid pretending to make a diagnosis while offline. The physical reliability around the model is part of clinical quality.

There are limits to the sovereignty argument. Running everything locally may cost more, use scarce power, and still leave a provider dependent on imported hardware. A mixed design can be more robust: keep sensitive workflows controlled, minimise unnecessary data movement, and maintain more than one service path where practical. Sovereignty is a question of control over failure and accountability, not a flag attached to a server rack.

The next 90 days

Two registered forecasts turn the discussion into questions that can be scored later. F10 (35%) asks whether the Nasdaq Composite closes at least 15 per cent below its 2026 high on any day before 16 April 2027. F11 (85%) asks whether at least two of OpenAI, Anthropic and Google release a new flagship model generation before 31 January 2027. These are subjective, uncalibrated probabilities published on 16 October, not facts and not probabilities inferred from Nvidia's filing. Each resolves against its specified market close or company release, and will be scored with the Brier rule, (forecast − outcome)².

The practical check before those dates is not “did the boom continue?” It is whether capacity, customer use, revenue and energy infrastructure continue to meet each other at a sustainable pace. If they do, commitments become productive assets. If they diverge, the first warning may be a delay or a weaker contract, long before a dramatic headline. The machine's power bill is a better place to start than its mythology.

Procure capability, then remove its bottleneck

The founder’s Kenya Vision 2060 proposes buying useful capability across models while building reliable energy and critical inference capacity where justified. That is an operating thesis to test. A server room drawing 10 kW of IT power at an assumed PUE of 1.4, operating for 95% of a 730-hour month, uses 9,709 kWh. At a hypothetical USD0.15/kWh that is USD1,456.35 before hardware, networks or staff. Power savings from an outage also mean lost service. Over the next year, compare full service cost and reliability before deciding that local ownership is economical.

Read the World Ledger

All 30 articles. Previous: The Test, the Trial and the Unknown. Next: Proof Must Reach the Patient.

Mathematical methods: 45 parameters, assumptions and limits. Download the research and executable model.

AI-assisted research and writing using parallel Luna agents; editorial synthesis under the APEX lens. Historical scenes are illustrative. Sources support the stated observations; forward scenarios and inherited probabilities remain conditional, uncalibrated judgments. Kenya Vision 2060 is a founder-authored thesis, not an adopted government plan.

Cover image: Front of server racks at NERSC.jpg — Derrick Coetzee from Berkeley, CA, USA · CC0. Photographed 14 October 2011 at NERSC; author credited from Berkeley, California. Illustrative infrastructure image, not a depiction of any 2026 facility. Resized for web display.