Move 4 — LOCALIZE

There is a reasonable and widely held position that in a world of open science, global supply chains and freely available models, the smart move for a developing economy is to import what already works. Why spend twenty years and a national budget re-deriving knowledge that exists, has been tested elsewhere, and can be licensed, bought or partnered into? Better to leapfrog. Take the proven thing, adapt the packaging, deploy it faster than the originators could.

That position rested on an arithmetic that held for fifty years. Building capability locally was slow and expensive, renting it was fast and cheap, and the gap between those two was wide enough that the decision usually made itself. Most organizations are still running that calculation.

The gap has narrowed considerably in the last few years, and it has narrowed in a specific place: the iteration cost of working out what is true in your own conditions. That is the entire cost centre of local re-derivation. It is also the part of the work that AI has compressed hardest. Which means the import-or-build decision that was settled in 1990 is genuinely open again, and organizations that have not revisited it are optimizing against conditions that no longer hold.

To see why that matters, it helps to look at what local re-derivation actually bought the last country that committed to it.

The savanna that was worth nothing

In the early 1970s Brazil imported food. This is easy to forget now, when Brazil feeds much of the world, but the country was a net importer of basic agricultural commodities and its rural economy was concentrated in a narrow band of good land in the south.

Meanwhile Brazil possessed roughly two million square kilometres of cerrado, a tropical savanna covering a fifth of its landmass. Agronomists considered it worthless. The soil was acidic, aluminium-toxic and almost devoid of phosphorus. Crops planted there simply failed. The received wisdom, backed by decades of temperate-zone soil science, was that this land could not be farmed.

In 1973 the Brazilian government created Embrapa, the Brazilian Agricultural Research Corporation. It was given a mandate, a budget and, critically, time. Its first significant act was not research at all. It was the decision to send hundreds of Brazilian scientists abroad for doctoral training in soil science, plant breeding and microbiology, on the understanding that they would return and work on Brazilian problems.

What those researchers came back to do could not have been imported, because nobody outside Brazil had reason to do it. They worked out the liming regimes that neutralized cerrado acidity at agricultural scale, establishing the right quantities, the right application methods and the economics of moving millions of tonnes of limestone into the interior. They cracked the phosphorus problem. They bred soybean varieties adapted to tropical daylight hours, since existing varieties were photoperiod-sensitive and would not set properly near the equator. They introduced and adapted Brachiaria grasses from Africa to turn cerrado into viable pasture. They developed nitrogen-fixing bacterial inoculants for soy that substantially reduced fertilizer requirements.

The cerrado became one of the most productive agricultural regions on the planet, and Brazil went from importing food to becoming a dominant global exporter of soy, maize, beef, cotton and coffee. That transformation rests on a body of tropical agricultural science that exists because a state institution was funded to create it and left alone long enough to do so.

What actually did the work

It is tempting to read Embrapa as a story about agricultural investment. It is not, or at least the investment is not the interesting part. Plenty of countries have spent more on agriculture with less to show for it.

The mechanism is narrower and more transferable. Knowledge derived in one set of conditions does not transfer cleanly to a materially different set. It has to be re-derived. And whoever does the re-deriving owns something that cannot be competed away, because it is specific to a place and a set of conditions that the original knowledge-holders have no commercial incentive to study.

This inverts the usual framing of technology transfer. The standard model treats local adaptation as a cost, a friction to be minimized on the way to deploying a proven solution. The Embrapa model treats local adaptation as the asset. The imported knowledge was an input, and a freely available one. The proprietary, defensible, compounding thing was what Brazilian researchers built on top of it for their own conditions.

There is one further property worth naming, because it is what connects Embrapa to the present. Re-derivation was expensive for a specific and unglamorous reason: it was gated on iteration volume. Embrapa's soybean programme required physical field trials across variety, soil and climate combinations, season after season, because there was no other way to find out what worked. Thousands of researcher-years went into narrowing a search space by brute force. The insight was never the bottleneck. The number of times you had to try something was the bottleneck.

That is precisely the constraint that has moved.

The same pattern, running now

In 2024 Lelapa AI, a research and product lab based in South Africa, released InkubaLM, a small multilingual language model for isiZulu, Yoruba, Hausa, Swahili and isiXhosa, languages spoken by well over 300 million people. It was trained from scratch on 2.4 billion tokens at 400 million parameters, small enough to run on a laptop.

The reasoning behind it is worth reading closely, because it is Embrapa's argument in a different domain. Lelapa's stated position is that while AI promises broad prosperity, the resources large models demand are out of reach for most of the world, and those models perform poorly on precisely the languages that most of the world speaks. Nobody with the capital to fix this at frontier scale has a commercial reason to prioritize it. So the capability has to be re-derived locally, under resource constraint, by people who have a reason to care.

The design choices follow from that. Smaller model, deliberately, because it has to run in environments without abundant compute. Datasets released openly, so the re-derivation compounds for other African practitioners rather than terminating in one product. When Lelapa later ran a challenge to compress the model further, it was reduced by around three quarters without losing performance, and the winning entrants were African.

The structural echo is precise. Frontier science does not serve conditions its creators have no incentive to study. The organizations that re-derive it for those conditions end up owning something specific and defensible. What differs is the timeline. Embrapa needed fifty years and a sovereign budget. Lelapa did a first useful version of its equivalent with a small team and a fraction of the compute the incumbents consider table stakes.

A shorter version of the same story: India's 1970 Patents Act recognized process but not product patents, which legally permitted Indian firms to re-derive any drug in the world provided they invented a different synthesis route. Twenty-five years of learning by copying built deep process chemistry capability. When Cipla offered triple-therapy HIV treatment at 350 dollars per patient per year in 2001, against a prevailing price above 10,000, that was not charity. It was the commercial expression of capability a generation of policy had deliberately incubated.

What specifically changed

It is worth being concrete about where the compression has actually happened, because the general claim that AI changes everything is not useful for making decisions.

The search space collapses before you spend money on it. Embrapa had to plant to find out. Crop and process models combined with satellite, soil and sensor data now let a small team simulate a large share of the combination space and reduce a physical trial programme to the trials that are genuinely informative. The same logic applies to formulation, to route-of-synthesis selection, and to any domain where the historical method was structured trial and error. Ground truth is still required. The number of times you need to reach for it has fallen substantially.

Observation stops being labour. A significant share of research and operational cost was people watching things and writing down what they saw: phenotyping plots, inspecting output, auditing compliance, monitoring dispersed sites. Most of that is now imagery and pattern recognition, which means the marginal cost of measurement falls close to zero. Cheap measurement is what makes iterative local adaptation affordable in the first place.

The knowledge assembly layer gets much cheaper. Regulatory dossiers, literature synthesis, standards documentation and contract analysis were all expensive precisely because they demanded scarce expert time on non-creative work. This is where AI is most immediately reliable, and it disproportionately benefits organizations that could never have afforded the expert headcount.

Small models make local specificity economically rational. The Lelapa case matters beyond language. When capability required frontier-scale compute, building for a small or specific market made no sense. Smaller, cheaper, task-specific models change that calculation, which means a market of 40 million speakers, or one country's regulatory regime, or one crop system, can now justify its own purpose-built capability.

What has not changed is the requirement for ground truth and for someone locally who understands the conditions well enough to know what to ask. AI compresses iteration. It does not supply judgment about which iterations matter, and organizations that mistake the first for the second will build fast in the wrong direction.

The question to ask this week

Take one capability your organization depends on and ask: if the provider walked away tomorrow, would anyone here know how to do this?

Not whether you could find a replacement provider. Whether the knowledge exists inside your own organization.

A poor answer sounds like a procurement plan. We would go to market, we have alternatives, three other firms do this. That answer describes a rented capability, which means the compounding part of whatever value it generates is accruing to someone else. It also means your negotiating position degrades over time rather than improving, because they are learning your business and you are not learning theirs.

A good answer sounds like a named person and a documented method. Two people here have done it, here is where the method is written down, here is what we learned the last three times that is not in any vendor's manual. The useful test is whether your version has diverged from the generic one. If you have adapted the imported method to your own conditions and recorded why, you have started building the asset. If you are running it exactly as delivered, you are renting.

Then ask the second question, which is the one that has changed: what did we decide not to build because it was too expensive, and when did we last check that price? Most organizations made that judgment against a cost structure from before 2022 and have not revisited it. Some of what was correctly ruled out then is now within reach of a team of five.

The move

Localize means re-deriving capability where you are, rather than renting it from where it was invented. The case for importing proven solutions is real and this is not an argument against it. It is an argument that the adaptation layer, usually treated as an irritating cost of deployment, is where the durable value sits, and that the price of building it has fallen far enough to warrant a fresh decision rather than an inherited one.

Brazil spent fifty years working out which crops its own soil would accept. The interesting question now is what an African organization can work out in five.