The Frontier Gap Is Widening. The Gap That Decides Whether Companies Get Built Has Closed.
cr4fts.com/n/jatn3
There is a genre of writing about African technology that begins with the population figures, moves to mobile money, and concludes that the continent is poised to lead. It is comforting, it is widely repeated, and it is a way of avoiding a fact that everyone working here already knows.
Africa is behind at the technological frontier. The distance is growing rather than shrinking. It will continue growing for at least a decade regardless of what anyone builds, and no amount of optimism about demographics changes the arithmetic.
We say this as a firm whose entire capital and attention are committed to building here. The argument that follows is not that the gap is overstated. It is that almost everyone has been measuring the wrong gap.
Compute is the closest thing this era has to a physical input for frontier work, and the numbers are stark. Africa hosts roughly 300 megawatts of installed data centre capacity, under two percent of the global total, spread across a little over two hundred facilities in thirty-eight countries. Continental capacity is on a path towards perhaps two gigawatts by 2030. Global capacity over the same period is forecast in the hundreds of gigawatts.
Note what that means. Africa's growth rate is strong, and a strong growth rate applied to a very small base against a strong growth rate applied to an enormous base produces a widening absolute gap every single year. The percentage figures look encouraging in a presentation. The arithmetic underneath them is divergence.
Research capacity follows the same pattern. Africa's share of global publication output rose from about 1.5 percent in 2005 to a little over three percent within the following decade, real growth from a very low base and still far below the continent's share of world population. Output is also heavily concentrated, with roughly ten countries producing the large majority of it.
Research funding is the constraint beneath both. The African Union's target is one percent of GDP spent on research and development. World Bank figures put most sub-Saharan countries between 0.1 and 0.4 percent. Kenya at around 0.8, South Africa near 0.6 and Senegal close to 0.6 are among the leaders. Egypt sits just past one percent. The countries setting the frontier spend three to five percent of economies that are many times larger. This is not a percentage gap. It is a multiple applied to a multiple, and it compounds annually.
Frontier capability is not a stock that can be acquired. It is a compounding process, and every component of it feeds the others.
Research capacity produces researchers who train more researchers. Compute infrastructure attracts the workloads that justify building more compute. Firms operating at the frontier generate the cash flows that fund the next round of frontier work, and they hire the people who would otherwise have built the local equivalent, frequently by relocating them. Each of these is a positive feedback loop and none of them has a natural stopping point.
Absent a deliberate intervention at a scale nobody is currently proposing, the correct prediction is divergence rather than convergence. Anyone forecasting otherwise is forecasting against the mechanism.
This is also precisely where the leapfrog narrative collapses, and the collapse is instructive. Leapfrogging occurred where the new technology required less infrastructure than the old one. A mobile network genuinely is cheaper to deploy than copper to every household, which is why that leapfrog happened and why it happened fast. Frontier machine learning is the exact inverse. It requires more infrastructure, more power, more capital and more specialist depth than what preceded it. There is no configuration in which the inputs are skipped, and every strategy premised on skipping them is a strategy premised on a category error.
Now the part that changes the conclusion, and it turns on a distinction that almost nobody makes.
Producing frontier capability and applying it are different activities with entirely different input requirements. Producing it demands exactly what Africa lacks: compute at scale, research depth, patient capital measured in billions. Applying it demands two things. The ability to call an interface, which costs precisely the same in Nairobi as in California. And knowledge of a specific problem, which is cheaper here than anywhere on earth.
Three years ago that distinction was academic, because applying frontier capability still meant assembling most of it yourself. Any serious application required a machine learning team, infrastructure and a training budget, which meant it required the very inputs that were absent. The distinction existed in theory and made no difference in practice.
That is no longer true, and the change was abrupt. Stanford's 2025 AI Index put the cost of querying a model at GPT-3.5 capability at twenty dollars per million tokens in November 2022 and seven cents by October 2024, a decline of more than 280 times. The same work found open-weight models narrowing the gap to closed models from eight percent to 1.7 percent on some benchmarks within a single year. That capability is available over an ordinary internet connection, to anyone with a payment card, with no institutional relationship and no permission required.
For the first time in the history of computing, the input that was the binding constraint costs the same for a team of four in Nairobi as for a team of four hundred in Seattle. The frontier gap did not close. A different gap closed, and it happens to be the one that determines whether companies can be built.
The argument is frequently misused the moment it is made, so the limits belong here rather than in a footnote.
No team in Nairobi is going to train a competitive foundation model this decade. That requires capital and compute which do not exist here at the necessary scale and will not appear on the relevant timeline. Plans that require it are not ambitious. They are simply wrong about the inputs, and they consume years of capable people's lives proving something the numbers already establish.
Nor does cheap inference make the infrastructure gap irrelevant. It makes it irrelevant to one specific activity. Anything requiring local processing at scale, anything where data cannot leave the jurisdiction, anything latency-critical enough to need nearby compute, still runs into the same wall as before.
The claim is narrow and it is the only one the evidence supports. The most valuable applications of these tools in this market will be built by people who understand this market, and that is now possible without first solving the infrastructure problem. The two constraints separated. Building an applied company here requires connectivity, an interface and deep knowledge of a specific problem. It no longer requires a national research base, and until recently it did.
This is why we spend our time on problems that look small from outside.
A clinic in Nakuru keeping patient records on paper is not a smaller version of a problem being solved in Boston. It is a different problem with different constraints, different failure modes and a different definition of success, and the people who can see it clearly are the people who have sat in that clinic and watched what happens when the register goes missing.
Their disadvantage in frontier research is total and completely irrelevant to the task. Their advantage in problem knowledge is total and decisive. That asymmetry is the entire opportunity, and it exists only because the frontier capability became purchasable rather than because anything about the research gap improved.
The corollary is uncomfortable for a lot of current activity. A company here whose advantage is technical sophistication is competing on the one dimension where the continent is weakest and the gap is widening. A company whose advantage is knowing exactly how a process fails is competing on the dimension where it is strongest and where no amount of capital elsewhere can catch up.
Of the ways this argument could fail, the infrastructure one deserves more than an acknowledgement, because it is the only one whose trajectory can be partly observed rather than guessed at.
Data centre growth on this continent is constrained by electricity rather than by capital or demand. A single large facility can draw as much power as a small city, and it needs that power continuously and at stable voltage, which is a considerably harder requirement than total generating capacity implies. Several African grids can meet the average load and cannot meet the reliability standard, which is why the projects that do proceed increasingly arrive with their own generation attached.
That detail is more revealing than the headline capacity figures. When a developer builds captive generation, usually solar with storage and a gas or diesel backstop, the project stops depending on grid reform and starts resembling an industrial site with a power plant bolted on. This is slower and far more capital intensive than plugging into a grid, and it is happening anyway, which tells you the demand is real and the patience is not infinite.
For an applied company the consequence is narrower than it first appears. Serving customers here does not currently require local compute. Inference happens wherever the provider's capacity sits, and latency across a submarine cable is tolerable for everything except genuine real-time interaction. The constraint binds only if data residency rules tighten, if regulation mandates local processing for health or financial records, or if cross-border bandwidth costs rise faster than volumes.
The regulatory variable is the one to watch, because it is the only one that can convert a non-issue into a hard requirement inside a single legislative session. A data residency rule with a short compliance window would turn an abundant input into a scarce one overnight, and it would do so in exactly the categories, health and finance, where the applied opportunity is largest. That is not a reason to avoid those categories. It is a reason to know, before starting, what the current rule is, who is proposing to change it, and what the architecture costs if it changes.
There is a third possibility that is neither right nor wrong but early, and it is the one that costs founders years.
It looks like this. The applied layer is real. Local problem knowledge is decisive. The companies are entirely buildable. And the customers cannot yet pay enough to sustain them. The clinic understands exactly why the software is better and still cannot find the money this quarter, and the following quarter, and the one after that. In that world every argument above holds and the businesses still fail, because a market that will exist in 2036 does not pay salaries in 2027.
The signal that separates the two is not enthusiasm, and it is not growth. It is renewal. First purchases can be driven by novelty, by a grant, by a relationship or by a founder's persistence. A second payment twelve months later, at full price, from a customer with other options, is evidence of nothing except value.
This is why we track renewal more closely than any growth metric, and why a company doing four hundred thousand a year with ninety percent renewal is more interesting to us than one doing two million with churn nobody wants to discuss. The first has found something. The second has found a marketing channel.
The honest position requires holding two statements that sound contradictory and are not.
Africa is behind at the frontier and falling further behind, measured in compute, in research output and in the funding that produces both. That will not resolve itself, and any plan requiring it to resolve first is not a plan. It is a wish with a timeline attached.
At the same time, the cost of applying frontier capability has fallen to the point where the binding constraint is no longer infrastructure but understanding, and understanding is the one input more abundant here than anywhere else. The gap that matters for producing technology widened. The gap that matters for building companies closed, and it closed in about three years.
The optimists who ignore the first half will build things that depend on infrastructure that is not coming. The pessimists who ignore the second half will miss that the tools finally arrived, and will keep explaining why nothing can be built here while it is being built around them. Both are reading half the evidence.
The companies worth building sit exactly between the two, and they are being started now by people who already know precisely which problem they are solving and have stopped waiting for permission to solve it.