AI post-training

The AI race has two bottlenecks: compute and people. One has a clear scaling path. The other doesn't.

While labs pour capital into chips, clusters, and energy infrastructure, they face a subtler constraint that money alone can't solve: the shortage of domain experts needed for post-training work. This is the phase where raw foundation models get refined into actual tools—radiology assistants, contract analysis systems, financial compliance engines.

That refinement requires a specific breed of researcher. Not just machine learning engineers, but PhD-level experts who understand clinical reasoning, case law, or regulatory frameworks well enough to evaluate model outputs, design safety protocols, and catch the edge cases that matter in deployment.

The United States doesn't have enough of these people. Latin America does.

U.S. universities awarded fewer than 3,000 Computer Science PhDs in 2023, according to the National Science Foundation. Physics and Mathematics programs added roughly 4,000 more. Nearly half are international students facing visa constraints post-graduation.

McKinsey research shows the number of U.S. workers in roles requiring explicit AI fluency grew sevenfold between 2023 and 2025—from 1 million to 7 million in two years. LinkedIn's 2026 Jobs on the Rise report makes the bottleneck concrete: AI Engineer, AI Consultant, and AI/ML Researcher rank among the five fastest-growing roles nationally.

Supply isn't keeping pace. The constraint is structural and worsening.

For years, AI talent discussions centered on machine learning engineers building foundation models. Capital-intensive work concentrated in frontier labs. Post-training shifted the equation entirely.

Pre-training is a GPU problem—massive clusters running overnight that you can solve with money. Post-training is a people problem requiring thousands of hours of expert evaluation, safety testing, and domain adaptation by researchers who genuinely understand deployment contexts.

The work is iterative. A domain expert identifies an edge case. An engineer adjusts the approach. The expert validates output. Repeat hundreds of times per model. You can't automate expert judgment at the edge cases. You need humans with deep domain knowledge.

The U.S.-based pool isn't adequate for industry scale.

While AI labs compete for constrained domestic talent, Latin America has quietly built remarkable research depth. Brazil graduates approximately 20,000 PhDs annually. Mexico contributes 3,000, Argentina 2,500, Chile 1,000—thousands in STEM fields directly relevant to AI development.

These aren't isolated programs. Researchers from USP, Unicamp, UNAM, and CONICET regularly publish in Nature, IEEE, and ACM alongside colleagues from MIT, Stanford, and Oxford. The research quality matches global standards.

Economics favor this talent pool. Equivalent expertise costs 40-60% less than comparable U.S. roles—not due to quality differences but geographic cost-of-living differentials.

Then there's timezone alignment, often overlooked but operationally critical. Brazil sits 1-4 hours from U.S. Eastern Time. A researcher in São Paulo starts their day as New York awakens. Buenos Aires teams overlap four hours daily with San Francisco.

Post-training depends on iterative loops. When problems resolve in hours rather than days, development velocity compounds.

This represents more than talent arbitrage. Three structural shifts are hitting simultaneously:

Demand escalation. Post-training has become the primary bottleneck in AI deployment, requiring PhD-level domain expertise the U.S. can't supply at scale.

Supply availability. Latin America produces researchers with precisely the quantitative and domain backgrounds post-training demands.

Infrastructure maturity. Remote work—normalized during the pandemic—eliminated geographic constraints that previously required physical relocation for cutting-edge AI contribution.

The market mismatch has an obvious solution.

The AI industry spent two years learning that compute alone doesn't create intelligence. The next lesson: intelligence without distributed domain expertise doesn't solve real-world deployment challenges.

Forward-thinking organizations are exploring distributed research models combining U.S. leadership with Latin American depth. Senior researchers direct strategy while PhD-level contributors execute specialized evaluation and refinement work.

The structural forces exist. What's missing is execution. For AI labs and research-driven companies, the question isn't whether to build distributed teams—it's whether to move while the talent market remains underutilized or wait for competition to discover the opportunity.

Follow the bottleneck shift. AI's constraint moved from compute to specialized human expertise. When scarcity patterns change, first movers capture disproportionate value. Watch where talent becomes scarce, not just where capital flows.

Geographic arbitrage compounds. 40-60% cost savings for equivalent research quality isn't marginal efficiency—it's structural advantage that compounds over development cycles. Global research teams built now create cost and scale moats that latecomers can't easily close.

Timezone overlap is undervalued infrastructure. Asynchronous communication works for code development. Post-training requires iterative expert-engineer loops. Latin America's overlap with U.S. business hours isn't coincidence—it's operational advantage.

First-mover windows close quickly. The Latin American AI talent pool remains underpriced precisely because it's not yet widely recognized. Companies building relationships with researchers in São Paulo, Buenos Aires, and Mexico City today secure hiring, cost, and relationship advantages. Market discovery eliminates arbitrage opportunities.

Breakthrough geography is unpredictable. Silicon Valley historically concentrated AI talent through proximity and prestige. Remote work broke that constraint. The next major AI contribution might come from a PhD in Buenos Aires designing model evaluation frameworks that nobody in San Francisco considered. Geography no longer limits intellectual contribution.

The constraint that matters most is shifting from chips to minds. The question is whether you'll adjust your talent strategy before or after the market prices in this reality.