Sunday Essay — The School Link Brazil Paid For Once

Brazil is bringing artificial intelligence into its public schools in a sensible order, and the gravest threat to that order is a budget line that has not been written yet. The pipe came first, the rules are arriving second, and the classroom tools are being held back on purpose. Plenty of school-technology programmes have run it backwards. What Brazil has not settled is who keeps paying for the pipe once the money that built it is spent, and every AI guideline the country publishes rests on that answer.

The money came from the mobile carriers. When Anatel auctioned 5G spectrum, the operators that won the 26 GHz band paid into a fund for school connectivity. That fund is administered by EACE, the Entidade Administradora da Conectividade de Escolas, and run through a programme called Aprender Conectado. Alongside it sits the federal Escolas Conectadas strategy, created by Decree 11.713 of 26 September 2023, which promised a connection for every public school in the country by the end of 2026. That deadline is seven weeks away. This essay is about what those seven weeks will and will not prove, and about the carriers who paid for the link, built it, and now, through one of their own foundations, tell schools to plan for the day it fails.

1. Who paid for the pipe

Start with the arithmetic of the auction money, because the rest of the argument hangs on it. In February 2024, as Capital Digital reported from Anatel's board agenda, the government held R$3.1 billion collected from the mobile operators that won the 26 GHz lots. The phases then being approved covered 5,320 schools in the North and Northeast at about R$653.9 million, an average of R$122,924 per school over 24 months. Nearly 3,000 of those schools needed their own power generation before a router could be switched on. Read the time unit in that figure: 24 months. The costing is a contract term. The school will need connectivity for decades.

The programme has grown since. In August 2026 EACE opened its sixth phase, 5,734 schools in 2,246 municipalities funded by an extra R$453.5 million, bringing the running total to 28,639 schools and more than 3.5 million students. A month later EACE told Teletime it had passed 30,000 schools, about 72% on fibre and 21.2% on satellite, with an average of 138 Mbps per school, or 1.7 Mbps per student. Those are the administrator's own figures as relayed by trade press; I have not found them in a published EACE report, so treat them as the operator's reading of its own network. The same piece carries the line I keep returning to. EACE's chief, Flávio Santos, said his main concern is continuity once the funds run out.

The federal side is larger and has more than one budget behind it. At the close of 2025 the Ministry of Communications counted 94,221 schools, 68.4% of the roughly 138,000 in scope, with 22,800 added in 2025 alone. It put planned investment at almost R$9 billion, R$6.5 billion of it inside the Novo PAC infrastructure programme, and more than R$3 billion spent since September 2023. The minister's line was plain: the goal is to finish connecting all of these schools in 2026.

So the pipe rests mainly on two kinds of money: a one-off auction windfall and a federal investment programme. Both are set up to pay for construction. Neither of them, as far as I can find in the public record, carries a named recurring line for keeping a rural school's satellite terminal, its inverter and its Wi-Fi access points working in 2031.

2. What the counters count

Brazil publishes more connectivity data about its schools than almost any country I can think of. That is to its credit. It also means the figures need reading carefully, because they measure different things and get quoted as if they were one series.

The second row is the one the government leads with, and it is a good indicator. The MEC's panel, published on 2 September 2026, rates schools from level 0 to 5 on electricity, monitored speed and Wi-Fi coverage. It found 98.9% with adequate power, 79.7% with speed considered adequate for teaching, and 80.7% with Wi-Fi inside the federal parameters. Those numbers are real progress.

The third row is the one I would put on a ministry's wall. Two years ago, when NIC.br's measurement centre tested schools against the standard of one megabit per second per student in the busiest shift, only about one school in nine met it. A contracted average of 1.7 Mbps per student and a measured peak of 1 Mbps are different claims, and AI tools live or die on the second. A class of forty opening a generative tool at 7:30 in the morning cares little about the monthly average. What matters to that class is whether the link holds for those twenty minutes.

My stake on this section is simple. The figure that decides whether AI can enter a Brazilian classroom is measured busiest-shift throughput, school by school, tracked over years. The country already measures it, through NIC.br's school measurement work. What it lacks is a habit of publishing that number as a trend, so that a school which falls from level 5 to level 3 in its third year shows up before its students do the noticing.

3. Rules before tools

The second step in the sequence is regulatory, and it moved this autumn. On 1 September 2026 the National Education Council approved Parecer CNE/CP 15/2026, its guidelines for the use of AI in Brazilian education, and a summary was published in the Diário Oficial on 6 October. As of this weekend it still awaits the education minister's homologation, so it has no binding force yet. The substance, as reported by Canaltech and not yet checked by me against the full text, is cautious in the right places. No AI grading of essays or written answers at any level. No discipline based solely on an AI detector. No unsupervised direct use by children up to the fifth year. Bans on emotion recognition, social scoring and the commercial use of student data. A twelve-month adaptation window after publication.

Israel Batista, a member of the council, put the intent in one sentence on a Teletime programme made in partnership with Vivo on 16 October: "A gente precisa que a IA entre na escola pela via pedagógica." We need AI to enter the school through the pedagogical door. Note who produced that programme. The carrier is now sponsoring the conversation about the rules for the tools its network carries.

The rules are arriving behind the behaviour, which is normal. Cetic.br's TIC Educação 2024 survey found that 43% of teachers were using generative AI to prepare teaching material, and 37% of students used it for school research, rising to 70% among secondary students. Only 19% of students said a teacher had given them guidance on it. The 2025 edition, which surveyed managers only, found 53% of school managers using generative AI in their own administrative work, 37% of schools allowing students to use it for tasks, and just 22% with any written guidance.

Put those beside the connectivity figures and a pattern appears. The first users of the school link are the adults: the coordinator filling in the federal systems, the teacher preparing tomorrow's lesson after the last bell. The students who use AI mostly arrive with it on their own phones and their own data plans, carried by the same carriers, outside any school policy. The CNE guidelines, once homologated, will govern the institutional route. The informal route is already open, and it runs over a mobile network rather than the school's fibre.

4. The carrier that tells schools not to rely on the carrier

The most interesting actor in this story is Fundação Telefônica Vivo, the social foundation of Brazil's largest mobile operator. In November 2025 it became the sole institutional supporter of a UNESCO UNITWIN chair in unplugged AI in education, hosted at the Federal University of Alagoas and coordinated by the Instituto IA.Edu, with R$400,000 a year committed through 2028. "Unplugged" here means teaching methods built on AI that do not depend on constant connectivity or advanced infrastructure. The chair and the foundation published a technical note on AI within Brazil's computing curriculum standard, the BNCC Computação, free to download.

A telecom company paying a university to design AI teaching that works without the telecom company sounds odd at first. I think it is the most coherent thing any carrier in the region has done on education. Network engineers design for failure. They specify what a system does when a link drops, and they call it graceful degradation. A curriculum that assumes the link is always there has no failure mode; it simply stops on the day a storm takes out the satellite dish in a riverside school in Amazonas. A curriculum built unplugged-first treats connectivity as an amplifier on top of something that already works. That is how a radio planner would think about a school, and it is a good model for a ministry.

The foundation's own reporting should be read as the foundation's own reporting. Its Informe Social 2025 says it trained 48,000 educators under Escolas Conectadas and reached about 2.2 million students in 2025, though the same page's metadata gives 2.4 million. Trained, in a foundation report, can mean anything from a ten-minute module to a semester. I would want completion and classroom-use data before treating that as capacity.

The carrier model is spreading. In May 2026 Claro Colombia launched Escuelas Conectadas Modo IA, starting with more than 150 schools, 1,000 teachers and 20,000 students, and citing an evaluation in which 98% of teachers said they felt left behind by AI. That evaluation is the company's own. The direction is clear across the region, though: the operators that sell the connection now want a voice in what runs over it in the classroom.

5. The objection from Paris and Lima

The strongest case against everything I have praised so far is old, documented, and comes from two places I take seriously.

UNESCO's 2023 Global Education Monitoring Report, Technology in education: a tool on whose terms?, is blunt about the evidence base: "A lot of the evidence comes from those trying to sell it." It cites the region's most famous cautionary tale directly: "In Peru, when over 1 million laptops were distributed without being incorporated into pedagogy, learning did not improve." The underlying study is Cristia, Ibarrarán, Cueto, Santiago and Severín in the American Economic Journal: Applied Economics (2017), a randomised evaluation of One Laptop per Child across 318 rural Peruvian primary schools. Computer access rose sharply. Math and language scores did not move.

A reader in that tradition would say Brazil is repeating Peru with better marketing: a large public infrastructure spend, a telecom foundation's figures standing in for evidence, and an AI layer bolted on because it is the decade's fashion. I think that critique deserves a straight answer, and it is mostly answered by the differences in sequence.

Peru shipped a device to a child and hoped pedagogy would follow. Brazil is shipping a link to a school building, writing rules that keep the youngest children away from unsupervised AI and keep it out of grading, and the first measurable use of the link is by teachers and managers. The unplugged chair is an explicit attempt to build the pedagogy before the tool. The country is doing, slowly and imperfectly, the thing the Peru evidence says was missing.

Where the critique lands, it lands hard. Nothing in the Brazilian record I have read shows that a connected school learns more than an unconnected one. The phased rollout is a gift for evaluation: schools connected in phase two and schools waiting for phase six differ mostly by queue position, which is close to the comparison an economist would design on purpose. If the MEC and EACE are not already using that stagger to measure learning, they are throwing away the cheapest natural experiment in Latin American education. And the same UNESCO report carries a figure that supports my worry more than its own. It estimates that connecting all schools in lower-middle-income countries would add 50% to their financing gap for national education targets. Connectivity is expensive to build. It is a permanent cost to run.

6. Year three

This brings me back to the budget line.

The auction money is a stock. It was paid once, by carriers who wanted spectrum, and it is being spent on construction and on service contracts with a fixed term. The Novo PAC is an investment programme, judged by works delivered. Neither is designed to pay the monthly bill for a satellite link at a school of forty children in Roraima in the seventh year after installation. The administrator's own chief has said, on the record, that continuity is his main concern. When the person running the programme names the risk, it should be taken at face value.

My position is this. The school connection should move into the recurring education budget, as a utility cost with its own line, the way electricity and water already sit there. It should be judged on a single published figure per school, measured busiest-shift throughput, with the level-4-and-5 share reported as a time series rather than a cumulative count. And the carriers that built the network should be contracted on that measured figure, with penalties tied to it, instead of on installed sites.

I would also make a prediction, with my reasoning attached. If no such line exists by the end of 2027, the first fall in the MEC panel's level-4-and-5 share will show up in satellite-served schools in the North, because those links carry the highest running cost per student and the thinnest local support. The CNE guidelines will then govern AI use in schools whose connections have quietly lapsed, and the informal route through students' own phones will become the only route. That is the outcome the whole careful sequence was meant to prevent.

None of this is an argument against what Brazil has done. The country took a spectrum windfall and turned it into tens of thousands of school connections, then had the discipline to write the pedagogical rules before letting AI into grading or into the hands of seven-year-olds. Its largest carrier's foundation is paying for research on how to teach AI when the network is down. That is a more deliberate record than many richer school systems can show. It is a record built on construction money, and construction money ends.

7. For the installer on the river route

You know the job better than any ministry does. A school on a tributary two hours by boat from the municipal seat, a satellite terminal on a pole you set in concrete yourself, a solar kit and an inverter because the grid never reached that far, and four access points the coordinator asked you to put where the signal used to drop. Your work order says the site is live when the test passes. Run it again at 7:30 on a Tuesday, when the morning shift has thirty-eight children in the room and the teacher has just told them to open the reading app. Write that number on the sheet, next to the one from the empty afternoon, and photograph both. Ask the coordinator who she calls when the link goes down in the wet season, and whether that number still answers in 2028. Put her answer on the sheet too. The contract that paid for your visit runs out. The school does not move.

Tarry Singh is the founder and CEO of Real AI, an enterprise AI advisory and deployment firm working with global enterprises on production agent systems, model risk, and AI sovereignty strategy. He also leads Earthscan for Energy AI startup, and is a founding contributor to the EU-funded HCAIM and PANORAIMA programmes for responsible AI education across European universities. He writes at tarrysingh.com.