Every country has a Compute Net Worth. Almost none have calculated it.

The index prices each country three ways. The honesty lives in reading them together.

Readiness is a discount rate on the future. Built and GDC are facts about the present. Readiness measures how underwritable an endowment is today: governance, stability, chip access under the August 2026 US export regime,³ grid, fiber, momentum, physical conditions, and capital access. Built counts the share of the resource ceiling already standing as generating capacity. GDC counts the compute itself. Nepal scores 56 percent on Readiness, has Built 3.9 percent of its ceiling, and runs roughly 10 megawatts of live datacenter capacity, a GDC of about $0.5 billion against a $7 trillion ceiling. Three different questions. The first prices the option. The other two count what stands.

The spread between the numbers is the country's reform agenda, priced. Nepal's ceiling is $5.3 to $7.1 trillion on a $45 billion economy. Its unlockable value today is roughly $535 billion, about $17,800 for every Nepali. The six trillion in between is a to-do list with line items: grid, governance, chip-access agreements, fiber, capital markets, execution. And the world as a whole is further behind than any single country: global GDC is roughly $4.7 trillion against a $662 trillion ceiling. The planet has tapped 0.7 percent of its compute net worth.

Democracy is shown and deliberately unweighted. Governance already enters through corruption scores, and the fastest movers of 2025 and 2026, the UAE, Saudi Arabia, Kazakhstan, would make any democracy weight read as editorial. The column collapses the EIU index into three honest buckets and lets you apply your own weight. The macro block is there because a compute campus is project finance wearing a hard hat. The current account tells you whether power exports or import bills dominate. Debt and reserves tell you whether the sovereign can co-invest or only concede. Broad money approximates what domestic banks could actually lend into a national buildout.

The ceiling is a ceiling. Analyst build costs run $35 to $42 billion per gigawatt today. The $60 to $80 billion band is NVIDIA's own all-in figure, and it gets re-reviewed weekly against the market.¹ Rerun everything at $35 billion if you like. The levers are cells, and the story survives.

Revenue and capture are different things. A country that only sells electricity captures the power bill, roughly 4 to 5 percent of AI-cloud revenue. Capture runs through equity, tax, land, and sovereign offtake. Owning the compute layer is the difference between selling crude and being Aramco.

Chip access is policy, and policy moves. As of August 2026 most emerging markets need no US license. That openness rests on a non-enforcement posture the GAO has already ruled procedurally defective.³ The score prices the openness and prices its fragility too.

Credit is destiny, until it moves. Paraguay reached investment grade in December 2025. Nepal took its first rating in 2024 and held BB− through a revolution. Ethiopia is still in default. The capital-access component carries exactly this.

Here is the sharpest objection to this entire index, and we will make it ourselves. Compute is valuable because energy and sites are scarce. If every country on this list built toward its ceiling, energy for compute would become abundant, and the $60 to $80 billion per gigawatt would collapse. Correct. GPU rental prices already ran the experiment: H100 hours fell roughly 80 percent from their 2023 peak before recovering.⁴ Ceilings are scarcity prices, and scarcity prices decay as supply arrives. Three things survive the decay.

The cost advantage survives. Electricity is only 4 to 5 percent of an AI cloud's cost stack. But in an abundant world margins compress toward cost, and the lowest-cost producers are the last ones standing. That is how every commodity endgame in history has resolved. The cheap-electron countries win both regimes: rents now, share later.

Demand has never once behaved. Every collapse in the unit cost of computing since the 1960s expanded total spending on computing. Cheap tokens will make the world buy more tokens, a lot more, in total dollars. Abundance grows the market it deflates.

What decays fastest is bargaining power. The premium power prices, the tax concessions, the equity stakes a host can demand: those are rents on scarcity, and they are largest right now, while rich-country grids hold 2,500 gigawatts hostage in interconnection queues and the world can finance roughly $1 trillion of data centers a year against $64 trillion of unlockable endowment.⁵ Under 2 percent of the queue gets funded this decade. The index is a queue, priced at today's scarcity, and the queue fills.

"GPUs depreciate in five years. You are pricing hundred-year rivers against five-year silicon." True. And the index prices the host, so the objection lands on the wrong asset. Power, shell, cooling, substations, and fiber make up 30 to 40 percent of the capex and live 15 to 30 years. Silicon refreshes on top of them the way aircraft refresh on an airport. Epoch's bottom-up model puts the annuity at roughly $8.5 billion per GW-year even assuming five-year silicon.¹ The endowment is a flow, and the flow survives every vintage.

"Chips are the binding constraint, and fabs are the real chokepoint." Fab capacity scaled to meet $200 billion-plus of annual NVIDIA revenue. Transmission has not: a fab takes three years, a rich-country grid interconnection now takes five to fifteen. Power is where the buildout actually fails today. Microsoft's Kenya campus stalled on electricity. Stargate UK paused over energy costs. 2,500 GW queues in the West.⁵ When two inputs are scarce, the rents flow to whoever holds the one that cannot be manufactured.

"This is a bubble, and your prices die with it." Maybe. Look at the structure of the bet, though. A host country's downside is the status quo: it keeps selling electrons at commodity prices. The option to convert costs the sovereign almost nothing to hold. And the layered math already discounts: $50 billion central, $35 billion at the bear case, where Nepal still carries a $2.9 to $3.1 trillion ceiling on a $45 billion economy.

"Inference needs to sit near users. Remote mountains cannot serve it." Training tolerates latency and follows cheap, cold, firm power. Inference is drifting toward batch and asynchronous work that tolerates it too. And the geography is kinder than the objection assumes: Nepal borders 1.4 billion Indians. The Himalaya sits next to the largest inference market that will ever exist.

"The resource curse. You are describing its next victim." The curse is the case for the framework. Weak institutions already discount the unlockable number directly, which is the score doing its job. The remedies are known and named: equity capture, sovereign funds on the Norwegian model, published offtakes. And unlike an oil enclave, a compute campus needs the grid, fiber, and institutions that spill over into everything else.

"Washington can revoke the whole premise in one rulemaking." Yes. That risk is priced at 14 percent of Readiness, and its fragility is documented in the notes rather than wished away.³ Policy risk argues for speed and for diversified alignment. It says nothing against the endowment. The rivers do not repeal.

Two years ago the sovereign AI factory was a concept. As of this month it is a category with live examples on four continents, a rollout playbook, and a growing list of casualties, and the casualties teach as much as the wins: the three stalled projects below all stalled on power, which is the entire thesis in three data points.⁹

One hundred and eight compute potentials, in plain words. Click any country to open its shareable card, or link to it directly: compute.world#nepal, compute.world#namibia, and so on.