The Next Four Billion

I started this series with a simple irritation: every AI argument I heard, optimist or doomer, was built on numbers nobody had checked. So I checked them. Twice with independent sources, then adversarial passes built to break my own findings. A few didn't survive, and I'll show you which ones. What's left is a curve almost nobody is reading correctly: AI is about to reach most of the human race, and the mechanism is not what the headlines say it is.

No advice in this piece. No verdicts. The ledger, and what it shows.

The number nobody checks

The share of the world's working-age population using generative AI hit 17.8% in Q1 2026, up 1.5 percentage points in that single quarter, per Microsoft's Global AI Diffusion Report. 26 economies are already past 30% adoption. The UAE leads at 70.1%. The US sits 21st at 31.3%.

Hold that quarterly pace flat, with zero acceleration, and you land near half the working-age planet by 2031. And the forces underneath the curve aren't flat. They're multiplicative, because adoption has always had 3 gates, and for the first time all 3 are collapsing at once.

Gate 1: The connection

2.17 billion people are still offline, concentrated in Southern Asia and Sub-Saharan Africa, with India alone estimated near 440 million. They're offline for a boring reason: towers and fiber never penciled out where they live.

Direct-to-cell satellites delete that math. Coverage stops requiring hardware on the ground at all. As of early 2026, Starlink has 650+ direct-to-cell satellites in orbit, live in 22 countries, with roughly 400 million people already inside the footprint. In December 2025, Airtel Africa signed to bring the service to its 174 million subscribers starting this year. AST SpaceMobile, though its launch slipped toward 2027 on rocket availability, holds agreements with 50+ mobile operators representing close to 3 billion subscribers. Amazon's Kuiper is building behind both.

The economic detail that matters: once the constellation exists, the marginal cost of covering one more village is zero. The unconnected 2 billion live almost entirely in places where infrastructure never paid for itself. Satellites don't care.

Gate 2: The device

This gate is already open, and almost nobody frames it this way. Smartphones are 89% of all handsets in use on Earth. 5.83 billion people, 70.4% of the population, carry a mobile phone. The sub-$50 Android closed this gate over the last decade while the industry's attention was elsewhere. The gap between "has a phone" and "uses AI" is now software and price, not hardware.

Gate 3: The price

Here's a number I had to correct mid-research, and the correction matters. The viral claim is that AI got "1,000x cheaper." The defensible version: matching what GPT-4 could do at its March 2023 launch cost $30 per million tokens then and costs under $0.50 on open-weight models now. That's roughly 60x in 3 years for a fixed capability level. Today's true frontier models still price at $2 to $5 per million tokens, so "frontier intelligence is nearly free" is wrong. What's right is that the floor collapsed, and the floor is what sets the business model.

At floor prices, free becomes the model. Meta ships AI inside WhatsApp's 3+ billion users at no charge, and India, with 550M+ WhatsApp users, is already Meta AI's largest market. OpenAI's roughly $4.50 tier helped push India to 100 million weekly ChatGPT users, its #2 market globally. And the pattern behind those numbers: the biggest AI user bases are no longer forming in the wealthy world.

The leapfrog

History already showed us what happens when gates open where no incumbent system exists to displace. Africa skipped landlines and went straight to mobile. It skipped branch banking and went straight to M-Pesa. Both adoptions ran faster than the originals did in rich countries, because there was nothing to replace.

AI is set up to skip the desktop era entirely: voice-first, local-language, inside apps people already own. Watch the language wires specifically. Microsoft attributes the biggest recent adoption jumps, South Korea, Thailand, Japan, largely to models getting good in those languages. Hindi, Swahili, Bahasa and Yoruba haven't had their jump yet. When they do, whole populations flip at once, because for hundreds of millions of people the language was the last closed gate.

Three roads to 2031

Share of the working-age world using generative AI:

• Bear, 35%. The pace halves as rich markets saturate, satellite rollouts keep slipping, and data stays unaffordable relative to income across Africa. Note what the pessimistic case still does: it doubles today's number.
• Base, 48%. The measured 1.5 points a quarter simply continues. Straight-line arithmetic, no acceleration assumed.
• Fast, 57%+. The local-language jumps land, direct-to-cell reaches the Airtel and AST subscriber bases on schedule, and free tiers keep riding billion-user distribution.

Honest counterweights, because a projection without them is a pitch: roughly 600 million people in Sub-Saharan Africa still lack reliable electricity, and a phone you can't charge is a paperweight. Satellite coverage existing and being priced for someone living on $3 a day are different problems. And several governments may restrict access outright. None of these bend the curve down. They decide whether 2031 looks like the bear case or the fast case.

The next 4 billion aren't hypothetical. They're in line. Which sets up the question part 2 answers: who owns the pipes they arrive through? The free tiers, the models, the clouds, and increasingly the open-model hub itself sit on a handful of balance sheets that fund each other and buy from each other. That wiring diagram is its own story, and it's next.

Sources: Microsoft Global AI Diffusion Report (May 2026) · DataReportal Digital 2026 · ITU/GSMA connectivity data · Starlink direct-to-cell reporting and the Airtel Africa agreement · AST SpaceMobile filings · OpenAI and Meta company announcements · UN population estimates.

The full series essay is on Stay Level.