Thirty-Eight Gigawatts Is a Queue, Not a Build Schedule

The Ministry of Mines and Energy has confirmed 38 gigawatts of data centre grid connection requests in Brazil. The number is official, accurate, and almost always used to describe the wrong thing.

It is a measurement of paperwork. Thirty-eight gigawatts equals 17.22% of the 220.7 GW of centralised generating capacity installed across the national interconnected system, and thirty-eight times the commissioned IT capacity actually operating in the country. The market that exists — the one with buildings, tenants and electricity bills — is 1,000 MW across 205 facilities in 2026. The market that has been requested is thirty-eight times larger.

Every Brazilian data centre forecast now begins with the 38 GW figure, and most of them end there too. That is where the analysis goes wrong, because the number is a queue, and a queue is a statement about who applied, not about what will stand.

What the gate actually lets through

Of the queued volume, 7,040 MW has cleared an access award in the first 2026 PNAST round. That is 18.53% of the requests filed, and the investment attached to it is estimated at R$159 billion. Ministry data record 52 access requests, of which 18 were approved and 34 remain under evaluation. Industry participants put the volume that survives technical filtering at roughly 8 GW — 21.05% of what was submitted.

If every single request received a favourable access opinion, accumulated maximum demand would reach about 13.2 GW by 2035. That is 34.74% of the queue. The remaining two-thirds is not pessimism about Brazil’s data centre market; it is the arithmetic of a system that cannot issue access opinions faster than it can move electrons to the point of connection.

The gap between 38 GW and 13.2 GW is the width of the gate. It is also, conveniently, the difference between a market sized on applications and a market sized on assets — and only the second one can be financed, insured or leased.

Eighteen months against seven years

The reason the gate is narrow has nothing to do with capital, land or demand. It is time.

A data centre in Brazil reaches operating capacity in 18 months to two years. A transmission project takes five to seven. That ratio of 3.5 to 4.7 is structural rather than cyclical, and it cannot be financed away, because the missing asset is a line the operator does not own. You cannot pay a counterparty to build faster than it is building.

Brazil’s generation is among the cleanest in the world, which turns out to be part of the problem. The wind and sun are in the Northeast, and increasingly in Bahia and Rio Grande do Norte. Load historically sat in the Southeast. A data centre arriving in interior São Paulo or Ceará is therefore not plugging into spare capacity — it is commissioning transmission.

The response tells you how the market has adapted. Where the line does not exist, operators build it: Ascenty is constructing a 32-kilometre line from a CPFL substation in Santa Bárbara d’Oeste to its Sumaré campus, secured through an ONS access opinion. Where the line exists but is congested, the load moves to the generation instead — as at Renova’s Alto Sertão III wind complex in Bahia. Neither is a workaround in the sense of avoiding the grid. Both are the grid, paid for privately, because waiting is more expensive than building.

The map the queue cannot redesign

Nowhere is the power constraint more legible than in vacancy rates. Nationally, vacancy stands at 4%, and 56% of new stock under construction is already pre-leased — a market that looks tight from the top down.

Underneath, submarket vacancy runs from 1% in Campinas to 40% in Porto Alegre. That dispersion is not a demand story. Demand in Porto Alegre is not absent; power access there is simply worse, or later, or both. When a market’s vacancy spread is forty points wide, the variable that matters is not how many companies want space. It is where the line already reaches.

The two capacity measures in circulation also deserve reconciling, because they get used against each other. The 205-facility count includes enterprise sites that are not for sale. The colocation and hyperscale estate alone measured 706 MW at mid-2026, with 106 MW delivered in the first half and 134 MW scheduled in the second — roughly 840 MW of leasable capacity at year-end. Both figures are correct. Only one of them is a market.

Growth from there is fast but not explosive, precisely because the base is already high. Commissioned IT capacity reaches 1,000 MW in 2026 and 5,576 MW by 2033 at 27.83% a year. Value compounds more slowly — USD 1.15 billion in 2026 to USD 5.00 billion in 2033, a 23.36% CAGR and USD 18.85 billion cumulative — because the market is compounding from an elevated floor rather than a standing start.

That gap between volume growth and value growth is itself informative. It implies roughly USD 4.12 million of capital per added megawatt across the window. Ascenty’s disclosed 150 MW for US$1.2 billion works out at USD 8.0 million per MW. The difference is the difference between brownfield fit-out in an existing campus and greenfield AI-ready construction, and it is the single best argument that the queue’s composition is not homogeneous: the gigawatts at the front of it are cheaper than the gigawatts at the back.

Redata is a conditional, not a gift

The tax regime that arrived in September is widely reported as an incentive. Read the habilitation terms and it is closer to a performance contract.

Redata, enacted as Law 15.504 on 15 September 2026, suspends PIS/Pasep, Cofins, IPI and import tax on qualifying electronic and ICT goods destined for a beneficiary’s fixed assets, converting the suspension to a zero rate once commitments are met and the asset is in service. To qualify, an operator must commit to at least 10% of effective processing, storage and data handling capacity made available to the domestic market — or, alternatively, a further 10% of equipment value into research and development. It must meet 100% of electricity demand from renewable or low-emission sources. It must achieve water usage effectiveness of 0.05 L/kWh or better, measured annually and publicly reported. And it must invest 2% of equipment value in research, development and innovation, with at least 40% of that directed to the North, Northeast and Centre-West.

Establishments in those three regions get the domestic-supply and reinvestment commitments cut by 20% — a regional-development clause wearing an incentive’s clothes. Benefits run five years, but the tax suspensions on PIS/Pasep, Cofins and IPI produce effects only until 31 December 2026. Fail the commitments and the bill arrives with interest and penalty, followed by suspension, cancellation and a two-year bar on re-admission.

The water term is the one to watch. A 0.05 L/kWh ceiling is achievable with closed-loop cooling and hard to hit with evaporative designs in a country where water permitting varies by municipality. It quietly rules out a class of cheap builds. The 100% renewable clause, meanwhile, is easier to satisfy on paper than in hourly practice — Brazil’s grid is clean on average and less clean at peak, and the compliance method has not yet been tested through a full year.

What would change the answer

Three things, and all three are observable.

The first is PNAST. The batch access seasons introduced in 2025 replaced first-come-first-served queuing with something closer to an auction; the first 2026 window ran from 1 to 15 June, and contested points went through a competitive process between 28 September and 8 October. Batching converts a queue into a schedule, which is exactly what the market has lacked. Whether the second round clears materially more than 7,040 MW is the single most useful number to watch.

The second is self-build. Every operator that constructs its own line converts queued megawatts into commissioned ones ahead of the transmission schedule, at a cost that only makes sense at scale. How many are admitted to do it determines how much of the 38 GW bypasses the gate entirely.

The third is Redata habilitation. A regime that demands water efficiency, hourly-matched clean supply and regional reinvestment does not stop the market — it selects the operators inside it. Watch who publishes a WUE figure and who goes quiet.

None of this supports the conclusion that Brazil’s data centre market is smaller than advertised. It supports a narrower and more useful one: the demand is real, and 38 GW is the best available evidence that it is. What is scarce is not demand, not capital, not land. It is the point at which the line arrives — and in the interior of São Paulo, the operators have stopped waiting for it.