Brazil's 38 GW AI Data Center Queue Is Not a Build-Out

38 gigawatts of requests. 706 megawatts in service. Brazil has become the most credible AI data center market in Latin America, and most of its promised capacity still exists as a queue position rather than a building.

That gap is not a reason to dismiss the market. It is the variable an investor has to price.

The queue is a request, not a contract

The Ministry of Mines and Energy counted 205 operating data centers in June 2026. Investors had filed 38 GW of grid-access requests over roughly the same period. That request book equals 17.2% of Brazil's 220.7 GW of centralised generating capacity, according to Pheonix Research.

It is a demand signal. It is not a construction schedule.

JLL's August count puts the live market at 706 MW. The sector has grown sixfold since 2012 at an average of 15% a year. H1 2026 alone added 106 MW, with another 134 MW forecast by December, roughly $8 billion of investment across the year.

Both numbers can be true at once because they measure different things. The 38 GW is a stack of applications for a parecer de acesso, the access opinion a project needs before it can connect. The 706 MW is what is switched on and billing tenants.

The first real filter is PNAST, the National Transmission System Access Policy introduced in 2025. It replaced first-come, first-served queuing with access rounds and technical criteria. In the first 2026 round, the system operator selected 43 applications for the backbone grid. Thirty-eight were data centers, totaling 7,040 MW and an estimated R$159 billion.

That award is the only slice of the 38 GW with a wire behind it. It is 18.5% of the request book. Pheonix Research estimates that technical filtering leaves about 8 GW of firm requests, or 21% of what was filed.

Seven years of line, two years of building

Valor International described the mismatch in August. A data center can be built in 18 months to two years. A transmission project takes about seven years.

Brazil's wind and solar farms sit far from the load centres, so new capacity usually needs new lines. The queue is therefore a physical constraint, not an administrative delay that a faster approvals team can fix.

Ascenty chose to build its way around it. The company financed a 32-kilometre transmission line from a CPFL substation in Santa Bárbara d'Oeste to its Sumaré campus, with an ONS access opinion behind it. The first phase secured 90 MW, with room to double. Valor puts the total transmission spend at R$200 million, including the line scheduled to operate in 2027 and then transfer to CPFL.

The demand arrived faster than the plan. Chris Torto, Ascenty's CEO, said the company closed contracts in three months representing 40% of everything it had done in 15 years. Ascenty committed $1.2 billion for 150 MW of AI-ready capacity, or about $8 million per megawatt on Pheonix's arithmetic.

Renova took the opposite route. It located up to 85 MW beside its 432.6 MW Alto Sertão III wind complex in Igaporã, Bahia, and uses electricity that would otherwise be curtailed. The project required R$900 million.

As we wrote in November, the equipment that gates the grid has become a strategic asset in its own right. The $9 billion market for the parts that decide who connects is the same bottleneck, seen from the supply side.

Where the megawatts actually are

Campinas has become the country's main hub: 316 MW of inventory in August with 1% vacancy. São Paulo and Barueri together hold 269 MW. Rio de Janeiro holds 73 MW.

Fortaleza is the export story. The city has 19 MW live and 227 MW under construction, most of it pre-leased by ByteDance for a 200 MW campus at the Pecém port complex.

The construction pipeline is large by regional standards. JLL counts 660 MW under construction, 56% already pre-leased, and another 4,621 MW in the pipeline. The market is expected to double by 2030, attracting about $17.4 billion.

Density is changing the economics. Cloud-era halls were specified around 8 kW per rack. AI-ready halls now run from 60 kW to 1 MW per rack, Ascenty executives have said on the record. That shift raises the value of a megawatt and the cost of getting one late.

Terronova started work on a 300 MW campus outside São Paulo in August, with the substation scheduled to come online in December 2027 and potential to reach 1 GW. The initial phase exceeds $500 million, backed by Actis and its $1.5 billion regional plan.

Hyperscalers are already inside the market. Microsoft brought two data halls online in São Paulo in February as part of a $2.7 billion commitment. Equinix opened its SP6 facility in April for $114 million.

The award that matters

The first competitive access season under PNAST allocated 228 MW to three projects at four connection points in São Paulo state. Seven projects competed with a combined 451 MW of qualified demand. All 228 MW went to Pahlavan One, Pahlavan Three and Digital DC Brasil.

Pahlavan One and Three are part of CloudHQ's GRU campus. Phases one and two will deliver three 48 MW data centers, each with an on-site substation. Digital DC Brasil operates under NextStream, the Latin American platform created by Actis, and will use the capacity to expand its Tamboré campus.

Set that against the headline. A single competitive round moved 228 MW. The national request book is 38,000 MW. Brazil is not short of interest. It is short of deliverable connection points.

Alexandre Silveira, the Minister of Mines and Energy, framed PNAST as a state policy built on planning, technical criteria and competitive process. Marcio Rea, ONS's director-general, said the model gives investors a more predictable environment. Both statements are accurate and neither changes the arithmetic.

ONS will run two access seasons a year from 2027. That cadence is the mechanism to watch. It converts a one-off queue into a recurring, priceable process.

The three numbers to underwrite

Announced capacity, assigned capacity and commissioned capacity are three different assets. The market usually quotes the first and pays for the third.

Start with assigned megawatts. ODATA's BG03 project in Colombia is the clean example. The company announced up to 120 MW of IT capacity. ANLA's September 2026 licence for the Navarra–Bacatá transmission project records 50 MW of assigned connection capacity. Bogotá's entire metro inventory was 44.3 MW in CBRE's first-quarter count.

Then take the energization date. Terronova's 300 MW is dated December 2027. Ascenty's Sumaré 3 is scheduled for the third quarter of 2027. CloudHQ's three 48 MW buildings have no public completion date in the ONS award. A megawatt without a date is an option, not a cash flow.

Finally, check the tenant. Of the 660 MW under construction, 56% is pre-leased. That is the strongest single signal in the market, because a pre-lease transfers demand risk before the concrete is poured.

Pheonix offers a crude cross-check on capital intensity. At $8 million per AI-ready megawatt, a 300 MW announcement implies roughly $2.4 billion of spend. If the number in the deck is far below that, the site is probably not being built for accelerators.

What would change the thesis

Redata is now law. Brazil enacted the data-center tax regime as Law 15.504 on 15 September 2026. The compliance terms require 100% renewable or low-emission supply and water usage effectiveness of 0.05 litres per kWh or better, according to Pheonix Research. The incentive is no longer a pending bill that investors can hope for.

The second catalyst is process. Two access seasons a year from 2027, combined with PNAST's technical screening, should convert more of the request book into dated awards. The first season already produced a template.

The third is transmission delivery. Every line that enters service, whether built by ONS, CPFL or the developer, moves megawatts from the queue into the commissionable column. Ascenty's 2027 line and Terronova's 2027 substation are early tests.

Smaller markets remain options rather than underwriting. Uruguay has a $70 million Antel project in Canelones. Paraguay's MITIC sovereign AI pilot is a 10 MW first phase with an estimated $300 million to $500 million cost, still in technical feasibility as of August 2026. Both need a signed anchor contract or a financed first phase before they belong in a portfolio.

Brazil does not have a demand problem. It has a delivery problem. The investors who make money will be the ones who can tell the two apart before the energization date is printed.