One Hundred Nineteen to One
The ratio is not a market inefficiency awaiting correction. It is an institutional fact, produced by decades of accumulated infrastructure on one side and decades of accumulated absence on the other. The Mountain View engineer earns approximately $412/hour in 2023 industry compensation surveys; the Nairobi annotator earns approximately $3.47/hour under the conditions documented in the Muldoon study. The machine that sits between them — the AI system whose training depends on both — does not care about the ratio. It processes the annotator's labels and produces the engineer's outputs, and the market distributes the value it generates according to institutional arrangements that neither worker designed and neither can unilaterally change. The ratio is the most compact single statistic that captures what the Orange Pill Cycle must reckon with: that democratization of tools has not been accompanied by democratization of the value the tools produce.
The Infrastructure Necessity — Contrarian ^ Opus
There is a parallel reading of the 119:1 ratio that begins not with moral outrage but with physical reality: the substrate required to make AI computation possible. The Mountain View engineer's compensation reflects not exploitation but the accumulated cost of maintaining the industrial apparatus that makes machine learning feasible at all. The data centers consuming 460 terawatt-hours annually, the chip fabrication plants requiring $20 billion in capital investment, the undersea cables carrying 99% of intercontinental data traffic — these are not optional luxuries of Silicon Valley excess but mandatory infrastructure without which the entire computational project collapses. The ratio, from this view, is less a moral failure than a thermodynamic fact about the energy gradients required to sustain computation at scale.
The Nairobi annotator's work, essential as it is, depends entirely on this infrastructure existing first. Without the GPUs, without the training clusters, without the model architectures that took decades of accumulated research to develop, there would be no images to label, no models to train, no AI industry to participate in at any wage. The ratio reflects not institutional capture but institutional creation — the difference between maintaining the machinery that makes AI possible and performing tasks that the machinery enables. This is not to justify the gap but to recognize that closing it requires more than redistribution; it requires either massive infrastructure replication in every market (thermodynamically impossible at current energy budgets) or acceptance that some asymmetry follows from the physics of where computation can efficiently occur. The political problem is real, but it sits atop an engineering problem that no amount of institutional restructuring can wish away.
In the AI Story
The specific number varies by year, by company, by role, and by economic conditions. The particular figure of 119:1 reflects specific 2023 data points — a $412/hour Silicon Valley ML engineer compensation average and a $3.47/hour Kenyan data annotator compensation average documented in peer-reviewed research on Sama's labor conditions. Other periods, other companies, other roles produce different ratios. The structural pattern — that the ratio is large, persistent, and institutionally produced — is stable across the variations.
The ratio matters not because it is morally outrageous, though it is, but because it reveals the institutional architecture that determines how AI's economic value is distributed. The architecture is neither accidental nor inevitable. It is the product of specific arrangements — intellectual property regimes, corporate governance structures, labor market institutions, procurement practices, national economic policies — that could be different and have been different under different conditions.
For the Developer in Lagos — Segal's archetype of AI democratization — the ratio has specific implications. Even if she successfully uses AI tools to build capabilities that were previously unreachable, the value she captures from her capabilities depends on the institutional architecture within which her outputs are sold. If the architecture routes value toward established markets and away from emerging ones, her tool access will produce more output but not necessarily proportionate economic return. The output is universal; the returns are structurally localized.
The ratio also matters for the sustainability of the AI industry's claimed values. If the gap between the labor powering AI and the labor capturing AI's returns is large enough, sustained enough, and documented enough, it produces the political conditions under which regulatory intervention becomes inevitable. The industry's preferred posture is that voluntary commitments are sufficient. The empirical evidence suggests that voluntary commitments erode when they become commercially inconvenient, at which point involuntary arrangements — regulatory frameworks, worker organization, international agreements — become the mechanisms through which the distribution is adjusted.
The number should haunt. Not because it can be eliminated by individual action. Because it reveals the scale of the institutional work the moment requires and the scale of the institutional work currently being done, and the gap between the two is load-bearing for whether the Orange Pill Cycle's optimism will be vindicated by broadly distributed flourishing or falsified by concentrated extraction.
Origin
The specific figures reflect 2023 industry compensation data for Mountain View ML engineers and the Kenyan data annotator compensation documented in the Muldoon study.
The framing of the ratio as 'the number that should haunt every AI optimist' originates with Segal's foreword to the Janah book, written as part of the Orange Pill Cycle reckoning with the Janah framework's implications for his own earlier optimism.
Key Ideas
Institutional, not natural. The ratio is produced by specific institutional arrangements that could be different and have been different under different conditions.
Compact diagnostic. The single number captures the distributional question that the AI democratization narrative has systematically avoided.
Political implications. The persistence and visibility of the ratio produces political conditions under which regulatory intervention becomes inevitable if voluntary correction proves insufficient.
Developer in Lagos relevance. Even successful tool use produces value distributed according to institutional architecture that localizes returns to established markets and away from emerging ones.
Appears in the Orange Pill Cycle
Physics Meets Politics — Arbitrator ^ Opus
The right frame for understanding the 119:1 ratio depends entirely on which question we're asking. If we're asking 'what makes AI computation physically possible?' — the contrarian view dominates (80%). The infrastructure gradient between Mountain View and Nairobi is not primarily ideological but thermodynamic, reflecting genuine differences in accumulated computational capacity that took generations to build. The $20 billion semiconductor fabs and 460 terawatt-hour data center footprint are not optional components but mandatory physics. On this dimension, the ratio reflects real differences in what each geography can currently support.
But if we're asking 'how should the value generated by this infrastructure be distributed?' — Edo's institutional analysis proves dispositive (90%). The fact that computation requires massive infrastructure does not determine who captures the economic returns from that computation. The current distribution follows from specific choices about intellectual property, corporate governance, and labor markets that absolutely could be different. History provides numerous examples of infrastructure-dependent industries where value was distributed more equitably despite similar physical constraints. The semiconductor industry itself once operated under different arrangements that produced different distributions.
The synthetic insight is that we're dealing with a two-layer problem where physics constrains but doesn't determine economics. The infrastructure necessity is real — AI cannot exist without the accumulated computational substrate that currently concentrates in specific geographies. But accepting this physical reality doesn't mean accepting the current distribution as natural or optimal. The right response isn't to deny the infrastructure gradient but to design institutional mechanisms that share the returns from that infrastructure more broadly, recognizing that those who perform essential labor anywhere in the stack deserve proportionate participation in the value they help create. The ratio can't be eliminated, but it can be institutionally managed.
Further reading
- James Muldoon et al., "The poverty of ethical AI," AI & Society, 2023.
- Mary Gray and Siddharth Suri, Ghost Work, Houghton Mifflin, 2019.
- Thomas Piketty, Capital in the Twenty-First Century, Belknap, 2014.
- Branko Milanović, Global Inequality, Harvard University Press, 2016.