I was listening to Ben Horowitz on a podcast recently, the kind of wide-ranging conversation where a billionaire venture capitalist explains why the world has never been better positioned for people to succeed. He was in good form. AI is going to solve auto deaths, cancer, climate change. Every child can have a super advanced amazing tutor. And then this: "I don't think we've ever seen a bigger opportunity equaliser than AI."
He said this minutes before explaining that his firm's mission is to ensure the next wave of great technology companies comes from America or allied nations.
I want to take the equaliser claim seriously. Not dismiss it with cynicism, but test it against the economics of most of the world.
- AI multiplier
A technology that amplifies whatever baseline already exists: infrastructure, purchasing power, bandwidth, institutional quality, capital access, and local market depth. If those foundations are strong, the upside compounds quickly. If they are weak, the upside remains narrow and much of the value flows elsewhere.
- Access to an AI tool is not the same as access to the economic system required to turn that tool into income, leverage, or company-building.
- The strongest examples used to prove AI's power are usually creation-tier examples, while the equaliser claim is usually made with consumption-tier examples.
- AI can still create real gains in lower-income contexts. But under the current structure, it multiplies existing advantage more than it equalises opportunity.
The silent premise
Horowitz's argument has a structure worth naming. It goes like this: AI is now accessible on any smartphone. Most people in the world have smartphones. Therefore, AI democratises opportunity. The argument is clean, intuitive, and wrong, because it confuses access to a tool with access to the economic system the tool operates in.
The restaurant owner Horowitz describes spending two hours with, mapping out how to use AI across his New York business, has suppliers, distribution infrastructure, a payments system, access to credit, a customer base with disposable income, and a regulatory environment that lets him implement what AI suggests. He has a power grid that stays on. He has broadband that does not cost a third of his income.
Strip away that stack, and "anybody who has a phone" becomes a much thinner claim.
AI is not an equaliser. It's a multiplier. And multipliers amplify whatever the baseline is.
The $90/month test
Here's what the baseline looks like for roughly half of sub-Saharan Africa. The World Bank's September 2025 poverty update puts 46% of the region's population below $3 a day, or approximately $90 a month. Sub-Saharan Africa now accounts for 73% of the world's extreme poor.
So run the equaliser claim through that economy.
To use AI, you need a smartphone. An entry-level device in sub-Saharan Africa costs 95% of average monthly income for the poorest 20%, according to the GSMA's 2025 Mobile Economy report. The GSMA is piloting $30 to $40 4G phones across six African countries in 2026. They have not launched at scale yet. For now, the phone alone is nearly a month's income.
Then you need data. One gigabyte in sub-Saharan Africa costs 14% of average income. For the poorest 40%, it costs 33%. The ITU's affordability target is below 2%. Sub-Saharan Africa is the only region where most countries still exceed the affordability threshold for 5GB of data.
Then you need the AI itself. ChatGPT Plus costs $20 a month globally, roughly 22% of a $90 monthly income. OpenAI launched ChatGPT Go in Nigeria in October 2025 at ₦7,000, or about $4.71. Credit where it is due, that is a meaningful discount. But $5 a month is still about 5% of income at the poverty line, and that is before the data you burn using it.
Then you need power. Nigeria's national grid collapsed 11 times in 2024. A study published in Energy Economics in 2025 found that reducing average outage exposure in sub-Saharan Africa to South Africa's level would raise total firm sales by 85%. For businesses without generators, the figure rises to 117%.
Add it up: device, data, electricity, subscription. The stack that Horowitz waves away with "the internet is here, you just do it" is precisely the stack that does not exist at a price point the world's poorest can bear.
Eighty-three percent of sub-Saharan Africa is covered by mobile broadband. Only 25% actually use mobile internet. The 58-point gap between coverage and usage is the gap between access and opportunity. It is the gap the equaliser story struggles to explain.
The creation tier versus the consumption tier
There is a subtler problem with the equaliser claim, and it is embedded in Horowitz's own examples.
When he talks about what AI can do, his examples are all creation-tier: a restaurant owner redesigning his operations, Cursor making programmers far more productive, companies going from zero to a billion in revenue. These are people already inside a functioning economic system, using AI to multiply their existing advantages.
When he pivots to the equaliser argument, the examples shift to the consumption tier: every child gets a tutor, everyone has a lawyer in their pocket, anyone can access powerful AI. The quiet move is that he proves AI's power through creation examples, then claims it equalises through consumption examples.
But the returns to consumption are modest without the creation infrastructure. An AI tutor can teach a kid in Kano perfect Python. If the local economy cannot absorb that skill, if there is no venture ecosystem, no employer base paying global rates, and the kid cannot get a visa to where those jobs are, the tutor is teaching into a void. The constraint was never knowledge. It was the conversion of knowledge into income within a given economic structure.
The UNDP's December 2025 report, The Next Great Divergence, puts numbers on this gap.
Two structural asymmetries, a capability gap and a vulnerability gap, mean that unequal starting points produce diverging outcomes, not converging ones.
Who captures the value
Even where AI creates value in lower-income contexts, the capture mechanism flows upstream. This is the part of the system Horowitz never discusses, despite running one of the institutions that benefits most from it.
Africa holds 0.6% of global data centre capacity. It is home to 0.04% of the world's top computers. Only 5% of AI talent on the continent has access to adequate compute resources.
A private-sector AI startup in a G7 country can iterate on its models every 30 minutes during training. The UNDP estimates its African counterpart may wait up to six days for the same cycle.
And here is the figure that does the most structural work: Andreessen Horowitz has invested a little over $20 million in African startups in total. For context, the firm manages roughly $42 billion in assets. Africa represents about 0.05% of that capital base.
Horowitz believes AI democratises opportunity globally. His firm's capital allocation says otherwise.
The leapfrog concession
I want to be honest about what works. M-Pesa lifted approximately 194,000 Kenyan households out of poverty. Mobile money now processes over $1.1 trillion in transactions annually across Africa. Rwanda's Zipline drones deliver medical supplies to remote clinics, cutting delivery times and reducing blood wastage by two-thirds. These are not theoretical gains. They are operational at scale.
But these leapfrog cases prove something narrower than the equaliser thesis. M-Pesa succeeded because it solved a local infrastructure gap, banking, using a local platform, Safaricom, within a local regulatory framework, Kenya's Central Bank. The technology was mobile. The ecosystem, value capture, and governance were Kenyan.
AI's structure is the opposite. The models are built by American companies. The compute runs on American infrastructure. The pricing is in dollars.
Out of nearly 500 AI policies, standards, and guidelines developed from 2011 to 2023, roughly two-thirds originated in the United States, Europe, or China. Only 7% came from Latin America and Africa combined.
M-Pesa is a case for locally governed technology deployment solving local problems. It is not a case for "anyone with a phone can compete globally." Horowitz conflates the two.
The Center for Global Development identifies three mechanisms by which AI widens global inequality: unequal infrastructure gives high-income countries a head start; automation enables reshoring, reducing demand for the low-cost labour that developing economies depend on; and high-income countries have greater institutional capacity to manage disruption. The second mechanism is the cruelest. AI does not just fail to equalise. It can actively pull opportunity away from poorer economies by making their comparative advantage in cheap labour less valuable.
The quiet part
Horowitz said something else in that conversation that I keep returning to. When asked about his ambition, he framed it in explicitly national terms: America won the industrial revolution, that is why America is America, and the country needs to win the AI revolution to stay America. The next great companies, he said, should come from America or allied nations.
That is a coherent national-interest position. But it does not sit cleanly beside the claim that AI is history's greatest opportunity equaliser. Either AI concentrates the future's value in America, or it distributes it globally. It cannot do both at the same time.
His father's advice, "life isn't fair", keeps surfacing through the interview. On its face, it is true enough. In this context, though, it becomes an intellectual off-ramp. Once you accept that life is unfair and that attempts to make it fair simply transfer power to whoever runs the system, you no longer have to reckon with the specific mechanisms that make it unfair and the specific people who benefit from that unfairness.
If AI is a multiplier and you are already rich, life being unfair works out well for you.
What equalisation would actually require
The honest answer is that AI equalisation is possible. Nothing about the current trajectory makes it likely.
It would require locally governed compute infrastructure, not just consumption access to US-hosted models. It would require local model development, which means the iteration cycle cannot be six days when your competitor's is 30 minutes. It would require purchasing-power pricing across the full stack, not just the subscription, but the data, the devices, and the API access that separates consumers from creators. And it would require infrastructure investment that addresses the physical layer before the application layer: stable power, affordable connectivity, and institutional capacity.
None of these are technology problems. They are capital allocation problems and governance problems. And the people with the capital, Horowitz among them, are allocating it overwhelmingly to the places that are already ahead.
The UNDP calls this trajectory the next great divergence. A plainer description is available: AI is a multiplier. If your baseline is Silicon Valley, the multiplication is extraordinary. If your baseline is $90 a month in an economy where the grid collapses 11 times a year and a gigabyte costs a third of your income, the multiplication is real but modest, and the surplus still flows to the platform owners, not to you.
That's not an equaliser. That's an amplifier with American characteristics.
- World Bank. September 2025 Global Poverty Update. September 2025.
- GSMA. The Mobile Economy Africa 2025. 2025.
- TechCabal. OpenAI brings ChatGPT Go to Nigeria, Kenya, and South Africa. October 2025.
- UNDP. The Next Great Divergence: Why AI May Deepen Inequality Between Countries. December 2025.
- Center for Global Development. Three Reasons Why AI May Widen Global Inequality. November 2024.
- Network Readiness Index / UNDP. Artificial Intelligence in the Global South. 2025.
- Aker, Jenny, and Isaac Mbiti. Mobile Phones and Economic Development in Africa. Journal of Economic Perspectives, 2010.
- StartupList Africa. Andreessen Horowitz portfolio activity in Africa. Accessed April 2026.