The Compute Utility
Who signs a Compute Purchase Agreement — and who doesn't. A data centre needs fifteen-year revenue; its customers buy compute by the hour. The institution that carries that gap will define the intelligence economy.
Second essay in the Convergence Assets series. The first argued that in AI infrastructure the offtake agreement is the asset — and proposed the Compute Purchase Agreement as the instrument Africa's regulated demand is waiting for.
A data centre is financed over fifteen years. Its customers buy compute by the hour. That mismatch is not a contract problem — it is an institutional one. Power solved it a century ago with the utility. The intelligence economy has not solved it at all. Last week I argued that in AI infrastructure the offtake agreement is the asset, and proposed the instrument this market is missing: the Compute Purchase Agreement, pooling regulated inference demand into anchor-tenant-grade offtake. This sequel is about the harder half of that proposal — not the contract, but the counterparty. The whole art is who signs.
Start with the objection every credit committee will raise, because an instrument that cannot answer it does not deserve to exist. Why would any buyer lock in long-term compute spending while the unit cost of compute is collapsing and the economic value of AI is still being argued about? Electricity is a commodity with a century of demand history; intelligence is a technology whose price falls by the quarter. Committing to buy it for fifteen years looks less like prudent procurement than like catching a falling knife with the balance sheet. Serious readers pressed exactly this within days of the first essay — Matthew Cullinen of Renaissance Philanthropy sharpest among them — and they deserve a real answer.
So here it is, and I mean it without irony: for any end-buyer, the objection is entirely right. No bank in Lagos should sign a fifteen-year take-or-pay for inference. No ministry in Dakar should commit a decade of budget to a technology that will be unrecognisable in three years. No startup in Kisumu should carry a contract that outlives its product, its stack, and possibly its business model. If the Compute Purchase Agreement asked them to, it would be a bad instrument — a take-or-pay trap with a new name, and West Africa's power sector can tell you exactly how that story ends.
But notice the premise buried inside the objection: that the end-buyer signs. Nobody who kept their household lights on last night has ever signed a power purchase agreement — and the corporate giants who do sign them are about to prove the point. The demand-risk objection is unanswerable — for the wrong signatory.
Nobody signs a PPA for their lights
Think about how you bought electricity today. No contract. No term. No committed volume. You flipped a switch, a meter counted, and at the end of the month you will pay for what you used — willing buyer, willing seller, cancellable by walking away. By Matthew's logic, no power plant on earth should ever have been financeable against customers like you. And none ever was.
What financed the power plant was a twenty-five-year purchase agreement — but you never signed it. Between the long-lived asset and the short-horizon consumer stands an institution whose entire job is to hold that mismatch: the utility. It signs decades-long contracts on one side and sells month-to-month on the other. It is, in the deepest sense, a duration transformer — the way a bank stands between long loans and demand deposits. The retail customer never carried the contract. The utility carried it, and the utility's book of customers is what made it creditworthy enough to carry.
This is the missing character in the first essay. A Power Purchase Agreement has a generator on one side — and on the other side, not consumers, not even a ministry, but an aggregator with a portfolio. The instrument was never the whole invention. The counterparty was.
I do not make that claim from a library. I co-founded responsAbility Renewable Energy Holding — today Serengeti Energy — and we grew it to $121 million of equity capital, building, owning and operating hydro and solar plants across sub-Saharan Africa, every one of them financed against the same question this essay keeps asking: who signs, and can they carry it. No lender ever asked us how many households wanted electricity. They asked who stood behind the purchase agreement.
Now the exceptions — and they deserve to be met head-on. Corporate power purchase agreements are real: a Google or a Meta signs directly for gigawatts, no utility in between. And hyperscalers sign enormous long-term compute contracts — that is what reached investment grade in the CoreWeave facility the first essay described. But look at what qualifies a buyer for that tier: a balance sheet large enough to anchor the asset alone, and demand forecasting deep enough to be its own actuary. A corporate signer is not a counterexample to the utility; it is a company acting as its own utility — origination, aggregation and balance sheet under one roof — and that class numbers perhaps a dozen firms worldwide. So the compute market has three tiers, exactly as power does: the anchor tenant large enough to sign alone; the metered end-user who should never sign long; and, between them, the institution that aggregates everyone below anchor scale into one signature. A corporate CPA works where one balance sheet can anchor the asset. The compute utility is the counterparty where no single buyer can.
So, to say it plainly: the Compute Purchase Agreement is not signed by the bank, the ministry, or the builder. It is signed by the compute utility — the aggregator whose diversified book of regulated, institutional demand is itself the risk mitigation that makes the contract bankable. The end-buyers buy the way you buy power: metered, as consumed, free to grow and free to leave. The utility holds the duration so that no one else has to.
What made the utility creditworthy
It is worth being precise about why lenders came to trust utilities, because none of it was sentiment — and all of it transfers.
The first reason was portfolio arithmetic. Samuel Insull's Chicago insight, more than a century old, remains the most important sentence in infrastructure economics: a thousand customers who peak at different hours are, together, a better customer than any single one of them. Diversity of demand smooths the load, lifts the load factor, and turns a crowd of individually unbankable buyers into one creditworthy book. No single African bank can anchor a data centre. Four hundred regulated institutions across a corridor — banks, telcos, insurers, revenue authorities, health systems, digital-public-infrastructure operators — peak differently, grow differently, fail (rarely, and independently) differently. The portfolio is the credit enhancement. It always was.
The second reason is less romantic and more decisive: machinery. One commenter, writing as the Legibility Office, put it in a line I intend to keep: "bankable demand may ultimately depend on governable demand." That is exactly right, and it deserves to be developed. What made retail power demand financeable was never the electrification mandate — no lender ever advanced a dollar against a government's declared ambition to connect its people. What they lent against was the meter, the billing cycle, the collections record, the load forecast: the boring administrative machinery that converts a legal obligation into a payment history. Another reader made the cynic's version of the same point, comparing pooled purchase pledges to wedding-fundraiser promises — generous on the day, scarce at collection. It is a fair description of every demand-aggregation scheme that ever failed, and the answer to it is not better pledges. It is machinery that does not run on pledges: enforcement through the operating licence, payment through rails that settle before service, forecasting that treats demand as an asset to be managed rather than a hope to be cited.
Put the two commenters together and you have the institution's job description. Regulated demand is the fuel; the utility is the engine; the Compute Purchase Agreement is the contract the engine signs.
Kenya, re-read
The first essay read the stalled Microsoft–G42 campus in Kenya as the case of a missing instrument. Read it again through this lens and it sharpens: it was the case of a missing institution.
When a hyperscaler needed a creditworthy counterparty to stand behind a gigawatt of demand, it looked across the market and found no aggregator — no entity holding a diversified book of contracted African inference demand, no engine. So it asked the Treasury to impersonate one: guarantee the revenue floor, guarantee the power, become the synthetic anchor tenant. And treasuries make poor utilities. A finance ministry has no demand portfolio, no metering machinery, no ability to grow or manage the load — only a signature that converts someone else's market risk into the public's contingent liability. The Treasury declined, and — as I argued then and repeat now — it was right to.
The gap was never a clause that better lawyers could have drafted around. It was institutional. Kenya did not lack a contract; it lacked a counterparty. That is a far more useful diagnosis, because contracts are written in months, while institutions must be built — and building institutions is a thing this continent knows how to decide to do.
The escrow lesson
I have watched the synthetic-counterparty move from the other side of the table. During my years at the Global Energy Alliance for People and Planet, one of the world's largest development financiers, weighing an investment in a city-scale grid, asked for a demand guarantee approaching a quarter of the project's entire capital cost — parked in escrow to stand in for customers who had not yet materialised, with philanthropy proposed to fund it. I understood the instinct completely, because it is the instinct this essay opened with: project finance needs a creditworthy offtaker, retail consumers cannot be one, and so the financier tried to manufacture one out of an escrow account. A power purchase agreement for household demand, synthesised from donor money. It is what a serious institution does when the counterparty it needs does not exist: it asks someone — a treasury in Nairobi, a foundation in New York — to impersonate one.
Set aside that it could never scale; there is not enough philanthropy on earth to escrow a quarter of everything Africa needs to build. The deeper failure was what the guarantee did to incentives. An operator whose revenue floor is banked in advance still has equity at stake, and covenants can be written — but the gradient flips. The urgency drains out of sizing the system to real demand, out of phasing capital as load grows, out of the unglamorous work that actually fills a grid — the appliances, the tariffs, the productive uses that turn a connection into a paying customer. The structure did not de-risk the project. It tranquillised it — and it would have paid out, in advance, for precisely the failure it invited.
That is the lazy model, and the compute era is the moment to retire it — from both sides of the table at once. The lesson is not that guarantees are wrong; development finance exists to carry early risk, and I spent years deploying it. The lesson is that a guarantee should absorb the ramp, never replace the market. Development finance must stop buying demand and start underwriting its creation: concessional capital as declining first-loss on an occupancy ramp the operator is contractually obliged to grow — shrinking as the book proves itself — not an escrow that replaces the growing. And operators — the compute utility first among them — must accept the mirror obligation: prove the origination, show the book compounding, earn the next tranche with the last one's demand curve. Held to that bar, watch whom the structure conscripts. Everyone in it — the financier, the utility, the platforms above them — becomes invested in African builders shipping intelligence people keep paying for, because only real usage services the debt. The escrow bought a promise. A ramp guarantee buys a market. An instrument that cannot tell the difference will finance neither.
Where the book comes from
An engine needs fuel, so the fair question is where the utility's demand book originates. Two engines, one from law and one from code.
The first the previous essay mapped in detail: the regulated triad. Some forty-four African countries have data-protection laws, but the binding lever is sectoral — central banks that bar offshore processing of core financial and identity data and enforce it through the banking licence itself; governments classifying sovereign workloads; identity and health regimes that hard-default to in-country processing. This is demand created by sovereignty law: it cannot use the offshore escape valve, it is paid for by institutions rather than consumers, and it grows as digital public infrastructure grows. Law is what makes the demand exist. The utility is what makes it contractable — the difference between forty-four legal obligations scattered across a continent and one underwriteable book.
The second engine is origination — and here the energy-access movement's hardest lesson applies. Connection is not consumption; you can build the node and find it idle. The answer, then as now, is deliberate demand stimulation — carried this time on the operator's own book rather than escrowed away — and it is precisely why AfCEN's application layer exists. Every venture launched on NextGen through the channels Africans already use, every small firm that AfCEN Deploy equips to operate like a firm a hundred times its size, is also something else: new, metered, payable inference demand, born directly onto the utility's book. A power utility inherited its demand from a century of electrification. A compute utility, built now, originates its own — which is why it is being built as a platform rather than a procurement office.
And look at what the originated demand actually is, because this is where the compute story stops being a technology story. The largest class of African inference demand that does not yet exist as contracts is climate adaptation. Tens of millions of smallholder farms grow most of this continent's food through a destabilising climate; forecast-grade, hyper-local intelligence — when to plant, what the rains will do, which seed survives the new heat, where the market will be when the harvest comes — is adaptation infrastructure, delivered as inference, renewed every season, and paid for because it pays for itself in yield and income. The same is true of resilience: early warning for cities and watersheds, diagnostics a rural clinic renews every month, tutors that teach in a child's own language. This demand is not adjacent to the climate agenda; it is the climate agenda, expressed as workloads.
The book has a supply side too, and it holds the quietest inversion in this essay. The most valuable climate dataset in Africa has never sat in a data centre: it is held in cultures — generations of observed rainfall memory, planting calendars, drought indicators, the pharmacology of traditional medicine tested across centuries. Its holders have been rightly reticent to share it, because sharing has always meant expropriation: the knowledge left, and neither credit nor compensation came back. A rights-preserving institution inverts the bargain. Knowledge becomes quantifiable, licensable and metered like any other input — title retained by the community that holds it, payment flowing each time it is used. The same machinery that meters intelligence out can meter knowledge in; a farmer's grandmother becomes a compensated contributor to the intelligence her granddaughter consumes. None of that happens by default. It is a design choice — and the compute utility is the room where it gets made.
Deflation is the utility's friend
Now return to the strongest form of the objection: compute deflates. The cost of a unit of inference falls relentlessly — better chips, better models, better serving. Why would anyone hold a long contract for a deflating commodity?
For a single fixed-price signatory, the objection is fatal — and we do not need to imagine why. Kenyans are still paying, in their power bills, for take-or-pay contracts that froze one price against twenty years of falling costs. A sovereign signature on a long fixed commitment in a deflating market is exactly how the last infrastructure wave manufactured its regrets. If the Compute Purchase Agreement meant that, it would deserve to die in committee.
But watch what deflation does to the institution instead of the individual signer. A utility does not hold one contract; it holds a rolling book — many counterparties, staggered horizons, continuously repricing as each vintage renews. When the underlying cost of serving that book falls, the commitments already written become cheaper to honour, not more burdensome: the margin between what the book pays and what delivery costs moves in the institution's favour. Deflation is the utility's friend. Solar taught this exact lesson: module prices fell some ninety percent and the Power Purchase Agreement did not die of it — it multiplied, because every cheaper vintage made the next contract easier to sign and the market larger. The instrument thrived on the deflation that would have ruined any single frozen signature.
And deflation works the demand side too. The first essay made the Jevons argument — cheaper intelligence finds a thousand new uses — and Africa's inference base is so early that consumption growth will outrun price decline for a generation. A book measured in workloads compounds even as the price per workload falls. The falling knife no end-buyer should ever catch is, in the utility's hands, not a knife at all. It is the engine's improving fuel economy.
This is not a loophole in the first essay's argument. It is its conclusion. The offtake book is the asset — and an offtake book is a thing only a portfolio institution can safely hold.
No last mile
There is a rejoinder waiting, and it is a serious one: you are describing Eskom for compute. Africa has built demand-aggregating utilities before, and too many of them are insolvent. Why would this one end differently?
Because of where power utilities actually died. Not in the control room, and not in the purchase agreements: overwhelmingly, they died in the last mile. Technical losses down thousands of kilometres of copper. Commercial losses to theft and unmetered connections. Collections that arrive months late or never. Retail tariffs set politically below cost, so that every unit sold deepened the hole. The balance-sheet logic of the utility was never the failure; the physical and political economy of retail distribution was.
A compute utility keeps the logic and sheds the failure modes, because it has no last mile. Nothing is strung to anyone's rooftop; the distribution network is fibre that already exists and interconnection that is pure engineering. Nothing depends on a meter reader's honesty; every API call meters itself, to the token, in real time. Collections are not a department; on prepaid, mobile-money-native rails — the payment infrastructure this continent taught the world — settlement can precede service. And the counterparties are not millions of households priced in a politicised retail tariff, but regulated institutions transacting at commercial terms under the machinery described above. None of this makes the institution unkillable. Institutions fail in new ways when they stop failing in old ones — Insull's own empire ended in the holding-company collapse that handed America its modern utility regulation — which is why governance machinery is not a feature of the design but the design itself. It does mean the standard objection is aimed at the wrong century's failure modes.
The pattern deserves stating in full, because it is the series' pattern. The utility model was invented where industrial demand bound hardest, then travelled everywhere. Mobile money was invented where banking's last mile failed hardest, then travelled everywhere. The compute utility belongs to the same lineage: a model that most of the world needs — everywhere AI demand is real but institutionally fragmented, which is most of Asia, Latin America, the smaller Gulf and European markets, and Africa first and hardest — and that will therefore be perfected where it is essential before it is adopted where it is merely convenient.
Building the counterparty
The arc of these essays is deliberate. The first argued the instrument: in AI infrastructure, the offtake agreement is the asset. This one has argued the institution: an offtake agreement is only as good as its signatory, and the signatory this market lacks is the compute utility — portfolio-diversified, machinery-governed, origination-driven, holding the duration no end-buyer should ever hold. The next will argue the market: what happens when utility-grade demand books meet corridor-scale supply, and fifteen countries stop being fifteen islands.
That last sentence is not hypothetical. At the 2026 West Africa Integration and Investment Summit in Freetown, where I serve as chief operating officer, the working groups AfCEN operates are the rooms where a corridor's ministries, central banks, utilities and operators surface the regulated workloads and the platforms that originate new ones — the raw material of exactly the demand book this essay describes. This is what we are building at AfCEN: the intelligence layer that makes the book underwriteable, the platforms that grow it, and the institutional design that lets it sign. Invented at the frontier, priced for the frontier — and built, from the first line, to travel.
To the readers who built the last energy transition with me — the PPA negotiators, the utility engineers, the concessional-capital deployers: none of this was ever a departure. Every institution compute needs, your industry invented, priced and governed for a century. And every reason compute matters here runs through your agenda — adaptation for the farms that feed the continent, resilience for its cities and watersheds, compensation at last for the people who kept its climate knowledge. Climate and energy are the unlock; AI is the enabler — not the other way around. The largest new electricity load of our lifetimes is also the largest new client your playbook has ever had. I am not writing to you about a different industry. I am writing to you about your next one.
If you keep one sentence from this essay, keep this one: the Compute Purchase Agreement is the contract, the compute utility is the counterparty — and the frontier's advantage is that it must build the counterparty first, which is how it ends up owning the model everyone else adopts.
The first essay closed with a claim I will hold this whole series to: the twenty-first century will be decided by who owns the contracts underneath its intelligence. Contracts need signatories. Africa should build this one itself.
Joseph Nganga is the founder and CEO of AfCEN (Africa Climate and Energy Nexus), the infrastructure-intelligence platform built from Africa for the global energy–AI convergence, and chief operating officer of the 2026 West Africa Integration and Investment Summit. He co-founded responsAbility Renewable Energy Holding (now Serengeti Energy), the independent power producer, and previously served as Special Envoy for Mission 300 (World Bank / African Development Bank), CEO of the Africa Climate Summit, Vice-President for Africa at the Global Energy Alliance for People and Planet, and Executive Director, Power & Climate Africa at The Rockefeller Foundation. He writes The Nexus Brief on infrastructure finance, sovereign AI, and platform economics at josephnganga.com.