A CPU-only cluster with 10,368 cores and 763 teraflops of peak performance just tripled the computing capacity of Argentina's Universidad Nacional de Córdoba (UNC) — without a single new purchase. EPFL in Switzerland donated the Helvetios system, retired in favor of newer infrastructure but judged to retain real scientific value, and the handover anchored CARLA 2026, the Latin American High Performance Computing Conference held September 21–25 on UNC's campus in Córdoba.

The donation set the tone for the week. Speaker after speaker returned to one question: how can the region extract more from the machines it already runs, rather than waiting on bigger ones?

A Second Life for Helvetios

Helvetios carries 288 compute nodes, each with two 18-core Intel Xeon Gold 6140 processors, 54 TB of RAM, and 763 teraflops of peak performance, according to EPFL. Rather than recycle it, EPFL shipped nearly four metric tonnes of equipment by sea in eight boxes, with the UNC rectorate paying freight.

Antonio Russo, who leads computing platforms at EPFL, arranged the deal; Dante Paz, an astronomy researcher at UNC, oversaw transport. UNC Rector Jhon Boretto and Swiss Ambassador Andrea Semandeni joined Russo for the launch. Russo told the audience the project began with a conversation.

"Keep talking," he said, "because conversations are where ideas come from, and ideas become projects."

Nicolás Wolovick, who directs UNC's HPC center and served as CARLA 2026 general chair, framed the stakes bluntly.

"Without new hardware, technical staff lose interest, scientists stop running computations, and a downward spiral begins," he said. "I've seen this story play out dozens of times across Latin America."

Boretto tied the donation to what he called a very serious funding crisis in Argentina's science system, saying cooperation agreements let UNC keep growing supercomputing capacity when the investments it would like to make are not possible. Wolovick argued the machine serves a wider public.

"A computer is basically a universal machine," he said. "From an open platform at the Universidad Nacional de Córdoba, we are giving teraflops and terabytes of storage to whoever needs them in Argentina and in Latin America."

The Machine You Have

Florina Ciorba, associate professor at the University of Basel, delivered a keynote titled "The Machine You Have: What Performance Alone Cannot Tell Us." Her argument: the old answer to "can we go faster?" — add hardware — no longer suffices.

Her evidence came partly from SPH-EXA, the astrophysics simulation framework her group builds with collaborators at the University of Zurich. Domain decomposition consumes roughly 10 to 20 percent of every time step. Decomposing less often saved about 23 percent of wall-clock time, per the program abstract, but on one test case the result drifted visibly from the analytical solution.

"We're computing the wrong result faster," Ciorba said. "That is what performance alone doesn't tell us."

On sustainability, Ciorba argued allocation quotas reward users for burning hours rather than saving joules, and urged centers to demand energy-to-solution, not just time-to-solution, in procurement benchmarks. She also toured Helvetios two days before her talk.

"A machine that is no longer suitable for Switzerland is absolutely suitable for another university, and this is the case for Helvetios," she said. "I hope this is not the last."

Agriculture's Missing Data Layer

Tuesday's AgroHPC workshop — the first International Workshop on Agriculture Empowered by Supercomputing — brought a series that Nicolás Erdödy of Open Parallel has run at SC, ISC, and SCAsia to Latin America. DK Panda of Ohio State opened with the NSF ICICLE AI institute's tools for digital agriculture.

Juan Ignacio Cornet of Rosario-based DeepAgro described a retrofit spot-spraying system now running on more than 100 machines across Argentina, Brazil, Uruguay, and the U.S. It has covered more than a million hectares and delivers about 70 percent herbicide savings on average.

Fernando Scaramuzza, a precision agriculture researcher at Argentina's INTA for two decades, identified a problem specific to the Argentine production model: within a single season, planting, spraying, and harvest may each fall to different contractors using machines from different manufacturers, leaving records that don't line up. He proposed an offline-first, edge-to-HPC architecture that keeps field history intact and vendor-independent. Erdödy walked through failure modes for Listen to the Land, a nationwide sensor and analytics platform designed for New Zealand, and Mary-Francis LaPorte of the University of Arizona covered HPC in plant breeding.

One talk — by the author, an Intersect360 Research analyst — argued compute isn't agriculture's constraint. Intersect360's market model spans 16 industry verticals, and agriculture is not one of them, leaving the sector outside vendor roadmaps and funding lines even as camera-equipped sprayers effectively become edge inference clusters. The region already owns expensive layers in systems like Argentina's Clementina XXI and Brazil's Santos Dumont; what's missing is control of the data. The recommendation: put that data layer under agricultural cooperatives, running on public compute, with data portability as a condition of sale.

Capacity Isn't the Constraint

Wednesday's session on energy, cooling, and effective use of HPC and AI in Latin America and the Caribbean paired SCALAC's regional survey, "Indicadores de Uso HPC LAC 2025," presented by Esteban Hernández, with Intersect360's global data. The argument: power and cooling explain what hardware a center can buy, but not why an installed machine sits underused. Intersect360's figures put hyperscale AI infrastructure spending at roughly four times all enterprise AI including HPC-AI — so measured against hyperscale capital, every research center in the world looks small, and everyone struggles with cooling and energy requirements.

Sovereignty Above the Silicon

Friday's sovereignty panel, moderated by Hernández, stayed on the same track. Leandro Neumann Ciuffo of Brazil's RNP said rising memory prices and geopolitical pressure pushed Brazil's AI plan toward several technology routes rather than one supplier. "We should not put all our eggs in the same basket," he said. Isidoro Gitler of Mexico's ABACUS-Cinvestav laboratory replied: "I could just change the word Brazil for Mexico and it would sound almost like the same story," describing new public infrastructure paired with distributed-center upgrades and large-scale training.

Nayat Sánchez-Pi of Inria Chile warned against a "silicon obsession," arguing sovereignty spans data, models, governance, and people — and that adapting the EU AI Act "is not a copy paste exercise" in a region lacking Europe's regulators and formal labor markets. Rodrigo Durán of Chile's CENIA was blunt about scale: "We can't compete with Anthropic. But what we can do is contribute to have better AI for people." The region's edge, he argued, is data, starting with the Spanish and Portuguese corpus dozens of institutions built for a regional model. "We don't want HPC for AI," he said. "We want HPC for healthcare."

Gitler pointed to 15 years of SCALAC cooperation: "In good moments, they get strengthened and then maybe they become thin, but they don't disappear." Hernández closed by urging the region to move past "reggaeton AI" — AI as a word dropped into every phrase.

Across five days, the community's argument held: Latin America's next gains will come from getting more out of systems, data, and people already in place — a harder pitch to funders than a new flagship, but the more realistic one. CARLA 2027 heads to Bogotá, Colombia, and the question will travel with it.