Wireless Access and the Places That Skipped the Old Internet

In the mid-2000s I spent enough time in Tanzania to stop noticing the thing a visitor notices first: almost everyone had a cell phone. Not a landline at home. Not a shared village phone. A handset, usually a Nokia, often something from the European market the US had not seen yet. There was a real secondary market, and the phones moving through it were in most ways more interesting, more diverse, and more advanced than what people in the US were carrying. A US Razr was good currency in that market, and I traded mine more than once for something better. The network was spotty, the electricity was spottier, and the towers kept working anyway because the business of getting an SMS from one person to another had already won the argument there. There was no copper to retrofit. There was no dial-up, no DSL, no ISDN, no cable. The sequence the US and Europe had lived through was not a sequence there. It was a thing that had never happened.

Americans came through constantly (NGOs, contractors, researchers, church groups), and almost none of their phones worked. Bands did not line up. The US was a CDMA holdout while the rest of the world had moved to GSM. Carriers had no roaming agreement. A SIM card might or might not help, depending on the handset. When we could produce a working Wi-Fi connection for a visitor, the response was closer to gratitude than convenience. Being connected at all was the win. Being connected at broadband speed was a different sort of luxury, and one most of the people who lived there could not afford at any speed.

Part of a series on what the DotCom era can teach us about AI. ← Previous: Smartphones Made the Internet Ambient

I am writing this installment from Southeast Asia, which lived through a version of the same compressed sequence on a different timeline and with different outcomes.

The series has been a single argument in five cases. Post 1 set the frame. Bubbles can destroy extraordinary amounts of investor capital and still finance durable infrastructure. The DotCom crash was both real and productive. The fibre, the data centres, the payment rails, and the indexing and search plumbing that came afterwards were paid for by investors who largely did not see the returns themselves. Post 2 took the retail case. Amazon did not win e-commerce by being early. Amazon won because it was still standing when search, payments, logistics, trust, and generational habit change all closed at roughly the same time, and that alignment took two decades. Post 3 applied the same checklist reading to streaming. Netflix did not replace the disc when the internet arrived. Streaming replaced the disc when broadband, compression, connected-TV penetration, delivery infrastructure, licensing, and subscription psychology had all closed, which took another decade on top of the fibre buildout. The internet did not change how people watched until the stream reached the couch. Post 4 made the larger claim. The smartphone did not make the internet more useful. It dissolved the internet from a place you went into an ambient layer you stood inside. That transition is the dominant structural legacy of the DotCom era.

This post, the fifth, pushes on who got access on what timeline, and asks whether AI is likely to reach the people most commentators are not currently writing about.

The question is not whether AI will help rich knowledge workers in rich cities. That is already happening. The question is whether AI follows mobile’s trajectory out from the cities, or whether it hardens the gap between places that already have expert infrastructure and places that do not.

The Assumption That Dies First

The first thing the Tanzania experience broke was the assumption that landlines are easier. In rich-world storytelling, wired infrastructure is the default and wireless is the upgrade. In places that never had the wires, the opposite is obviously true. Wireless is not the upgrade. It is the only thing.

This matters for AI parallel work because the rich-world framing of AI adoption is similarly default-loaded. The implicit model is: professionals in well-equipped offices use AI as an overlay on the knowledge work they already do. The sequence runs from sophisticated existing infrastructure to AI-augmented infrastructure. But the places where AI might do the most per dollar (a village without a GP, a district without a family lawyer, a school without enough trained teachers) do not have that substrate to overlay onto. For them, an imperfect AI expert is not a downgrade from a real one. It is an upgrade from nothing. The adoption sequence is not the same sequence.

Whether that upgrade actually arrives and works is a different question, and most of this piece is spent there. But the sequencing assumption is worth naming first, because commentators keep repeating the mistake the telecom world eventually learned from: projecting a rich-world infrastructure stack onto places where the stack was never going to be built.

The Satellite and the Rent

A VSAT satellite dish mounted on a concrete rooftop in a tropical setting, cabling running down through a painted wall to a small office, with a chain-link fence and padlocked gate in the foreground, suggesting a regulated, gatekept access point rather than an open connection.

The connectivity we had in that period was satellite. Not an exotic satellite. The same physical infrastructure that served European markets. IP-over-MPEG2 was the dominant rig. A 256 kbps link in Europe cost one thing; the equivalent link in country cost four to five times as much, because the regulation required the service be delivered through approved domestic contractors. The satellite companies could not sell directly. They had to sell through a local licence holder, and the local licence holder set the price.

The multiplier was the visible part. The less visible part was access. A 256 kbps link in country delivered the same 256 kbps it delivered in Europe. Same satellite, same physical infrastructure, same provider. But only if you went through the licensed local contractor, who charged four to five times the European price and was hard to book. You were not paying more for a different product. You were paying for political permission to use the same product. The rent extraction was structural, not incidental.

I want to be careful about what this observation is and is not. It is not a story about one country or one government. The pattern held across most of the region I worked in, and variants of it held across much of the post-colonial world. Colonial administrations did not stumble into rent-extraction by accident. They built it on purpose. Concession economies were a deliberate design: a foreign operator brought the technical capability, a protected local intermediary collected the rent on behalf of the metropole, and the populations affected had no recourse to either. Independence changed the identities at the top of that structure without dismantling the structure itself, in large part because dismantling it required money, time, and trained administrative capacity that the departing colonial powers had spent decades extracting rather than building. Layer post-independence state monopolies on telecommunications onto that inheritance, layer liberalisation-era licensing regimes on top of those, and what you get is not primarily a story about individual venality. It is a story about what happens when the rent-collection architecture is the most expensive thing to dismantle, and the country inheriting it has to keep the lights on at the same time.

I met President Kikwete in 2007. He struck me as someone with the right ideas about what his country needed to become and a clear enough view of the obstacles. The obstacles were not mainly about finding better people at the top. They were about inheriting an administrative and economic structure that had been designed to concentrate gatekeeping and that resisted every attempt to open it. Good intentions at the top do not automatically reach the contractor setting the price on your satellite link. That is a harder problem than any one administration can solve in one term, and pretending otherwise is how outsiders write bad policy papers.

The AI parallel here is sharper than most commentators notice, and it is not limited to low- or middle-income countries. Regulated AI (in healthcare, legal advice, financial services, education) is going to route through licensed-intermediary structures in almost every jurisdiction, rich and poor alike. The model weights are the satellite. The certified local operator is the domestic contractor. The rent extraction will look similar wherever political permission is the scarce input and technical capability is abundant, whether that is a regional licensing regime in East Africa, a state medical board in the US, or a financial-services regulator in the EU. This is not a complaint. It is a prediction. In markets where AI is most needed but most regulated, expect a multiplier to appear between what the technology could deliver and what end users actually pay to access it.

Equipment That Would Not Translate

A mid-2000s laptop sitting on a simple wooden table in a spare, sunlit room with painted concrete walls, a dusty English-language manual beside it and no surrounding IT infrastructure visible, illustrating equipment dropped into a context without the support culture its specifications assumed.

I was working with an NGO during that period, which is to say I was working on a shoestring. The NGO strongly preferred US-shipped equipment, and I disagreed with that preference. I wanted to buy local where possible or from regional vendors who would still be there when something broke. I had contracts lined up for exactly that. Most of them got cancelled when Microsoft offered donations and software from Ireland, in English instead of Swahili.

The unit would work out of the box. That was not the problem. The problem was the layer of training, documentation, troubleshooting, and repair culture the equipment had assumed was present and was not. Manuals written for a US mid-market IT department do not survive contact with a context where nobody has ever been that IT department. The language was not the barrier. The barrier was that the material assumed a culture of repair that, in that environment, did not exist and was not being built.

This is not a critique of the people who lived there. It is a description of an economic reality. If goods flow in from offshore more cheaply than they can be produced locally (donated clothing, second-hand electronics, humanitarian-grade equipment), the incentive to develop repair capability disappears. Things worked or they did not. If they did not, you went without. The same logic applied to water pumps, generators, and laptops. The steady-state assumption was disposability, and disposability is not a defect of character. It is a rational response to the economics in front of you.

The relevance for AI is that the most important part of deploying a technology is rarely the technology itself. It is the scaffolding around it: the people who know how to prompt it well, the workflows that turn an answer into an action, the trust that gets built over many small uses, the feedback loops that catch the failures early. All of that is cultural and institutional. It does not ship in the box with the model weights. It did not ship in the box with the satellite terminals either. A lot of the current AI pilot failure in well-resourced organisations is a version of the same problem: the equipment works, the surrounding capability does not exist, and nobody has allocated budget or time for the scaffolding. In lower-resource settings, the gap between what the technology can do and what can be sustained locally is structurally larger. It does not mean the technology cannot help. It means the help will depend on whether local scaffolding gets built, and who builds it.

What Leapfrogging Actually Buys

I am not a leapfrogging purist. Compressed development timelines are, in general, good. Southeast Asia has lived through several of them and come out stronger on the whole than if it had been forced through the full Western sequence. The developed world is still struggling in rural areas with the consequences of legacy infrastructure that cannot be ripped out, because ripping it out would strand the existing user base, and because the people currently connected do not have much incentive to push for alternatives.

What leapfrogging buys is not just a better end state. It is the development of a particular kind of muscle memory. If your first experience of communication technology is a mobile handset rather than a rotary phone, your default mental model of “making a call” is different, your relationship with the asset is different, and your readiness for the next transition is different. You do not have to unlearn anything. You are not defending an investment you already made. You do not feel obliged to justify the thing you have by hesitating at the thing that comes next.

I think this is the part of the leapfrog argument that travels best. It is not that skipping a step produces a better state immediately. It is that the absence of legacy makes the next transition cheaper and faster. Mobile-first populations moved to smartphone-first with less friction than populations that had been heavily invested in feature phones. Smartphone-first populations will likely move to ambient and AI-augmented interaction with less friction than populations that are still defending their laptop-and-browser workflows. The muscle memory compounds.

It also compounds in a direction you cannot fully predict. The outcome is not a clone of the richer country’s current state. It is a different shape, because the intermediate steps that shaped rich-world habits did not happen. Mobile money in Kenya did not look like a digitised US chequing account, because it was not evolving out of one. The AI equivalent is likely to look similarly unfamiliar in places where it establishes itself first. Less like an office worker’s chatbot, more like something embedded in agricultural advisory, or community health, or informal-sector commerce, shaped by constraints that rich-world product managers are not currently designing around.

Coverage Is Not the Same as Use

Close-up of weathered hands holding an unbranded early-2000s feature phone, a thumb mid-press on the numeric keypad, a blurred open-air market with baskets of produce behind, illustrating SMS-based coordination and direct trade without any internet interface.

The most disciplined part of the current research is the distinction between coverage and meaningful use. The GSMA’s 2025 update put the coverage gap at roughly 4 per cent of the global population, with a 38 per cent usage gap on top of it. The ITU’s parallel figure says about 2.2 billion people remained offline in 2025 despite being within a signal. The bottleneck has migrated. It used to be that there was no tower. Now, in most places, there is a tower. The problem is that a tower is not enough.

What counts as meaningful use depends on who you ask, and I want to be careful here. Some of the most useful early mobile-enabled behaviour in places I worked had nothing to do with the internet. It was direct SMS between two people who could now coordinate a trade or a movement of goods without either party having to travel to find the other. The fisheries work in India and the grain-market work in Niger both show the same pattern. Reducing the cost of coordination, even without bringing the full internet with you, produces measurable welfare gains. It is not glamorous, and it does not make for good pitch decks, but it is real and it compounds.

By the formal definitions, meaningful connectivity requires the experience to be safe, affordable, and sufficient for productive use. By the definitions that matter to the people actually using the technology, it is often enough that the connection lets you do the thing you could not do before. Those definitions do not always match. The aggregate numbers in the research track the first. The lived experience tracks the second. Both are real. Writers who flatten one into the other misread the situation.

The parallel question for AI is what counts as meaningful AI use in a context where access is thin. Is it a polished chat interface? Unlikely. Is it a translated voice assistant helping a community health worker triage a symptom? Possibly. Is it crop disease identification from a phone photo, in a dialect the extension service cannot reach? Plausibly. The useful cases are smaller, more specific, and more boring than the rich-world framing suggests. They are also where the early development gains are most likely to be real.

Where the Optimism Breaks

I do not want to over-sell any of this. Every constraint that limited mobile’s actual reach applies to AI, often more sharply.

Device cost is the first and largest constraint. GSMA’s handset-affordability work estimates that a $30 smartphone would bring up to 1.6 billion more people into the affordability range. We are not close to that price point. The cheapest smartphones in markets like sub-Saharan Africa, mostly Transsion’s Tecno and Infinix lines, sit closer to fifty to eighty dollars with two gigabytes of RAM. Useful on-device AI needs more like four to eight gigabytes and a competent NPU, and that hardware tier starts above the hundred-dollar mark even in the most price-competitive markets. The cloud-dependent alternative pushes the question elsewhere: a gigabyte of mobile data still costs five to fifteen per cent of monthly income in the poorest deciles, and a cloud AI session burns through that allocation faster than most product designers have ever had to think about.

Electricity is the next. The IEA estimated around 750 million people lacked electricity access in 2023, the large majority of them in Africa. You cannot run an AI-assisted clinic workflow off a phone that cannot charge reliably. Solar and off-grid models help, but they cap the compute budget in ways the rich-world product decisions do not anticipate.

Language is the quieter killer. The quality of most large language models outside of English, Mandarin, and a narrow set of other global languages remains meaningfully worse. Minority-language support is the part most easily promised in a demo and most reliably missing in production. The populations who would benefit most from expert access are exactly the populations most likely to be served in a language the model treats as an afterthought.

Regulation can enable or strangle, and the history is not encouraging. Healthcare licensing is a patient-safety regime first, and that function is legitimate. The other licensing regimes that gate AI expansion (legal services, financial advice, education) tend to claim the same consumer-protection function while spending a meaningful fraction of their energy protecting incumbents, and the difference between those two functions is often visible only in retrospect. Markets with weak professional capacity and strong protective licensing are markets where AI expansion is most needed and most legally complicated. The satellite-contractor dynamic I described earlier is the cleanest prior.

Gender is the compounding factor. GSMA’s 2025 work still has women in low- and middle-income countries 14 per cent less likely to use mobile internet than men, with a widening mobile money gender gap. Every layer of friction in AI access lands harder on women in the same markets. Ignoring this does not make the aggregate numbers look better. It makes the aggregate numbers wrong.

None of this means AI cannot reach these markets. Mobile eventually reached them, and reached them in forms nobody fully predicted. It means the optimistic version of the story only holds if serious attention is paid to the failure modes, and those failure modes are not the ones currently attracting product investment.

What operators in these markets have to hold together is the same dual stance the earlier posts argued for in richer contexts. Do not commit to the Blockbuster position, which is to say do not build a cost structure that requires the old substrate to keep paying. And build what needs to be built, which is to say do not freeze waiting for the shape to become obvious. The shape is never obvious from inside the transition. That advice applies harder, not less, in places where the margin for error is smaller and the capital is thinner.

The Contractor Pipeline

The period I was in Tanzania overlapped the early Afghanistan and second Gulf conflicts, and a significant portion of the Americans I crossed paths with in transit were military contractors. Military personnel were generally on their own aircraft. The contractors were on the same commercial flights as the rest of us. Folks traveling for business seemed to be in the minority that decade. You ended up sharing tables with them at Schiphol, partly because Americans tended to cluster around the few English-speaking corners of any large airport, and partly because there was a decent New-York-style pizza place there. The conversations were instructive. The serious satellite gear I was ever in arm’s reach of came from that world first, and from consumer markets second.

The satellite piece in particular was a defence-shaped market. The platforms, the bands, the field-hardened gear, the encryption, all of it ran ahead of the commercial equivalents because defence customers paid for it to. What changed after the DotCom bubble is that the pipeline became bidirectional in the commodity layers. Consumer scale produced hardware and software that the military and defence world then adopted, and defence investment continued to shape the specification of what got built at the frontier. But the people sitting two seats over from me on a layover were the ones using the high-end satellite kit before any of it filtered down.

The AI pipeline is working out the same way right now. Most of the major labs are visibly serving US defence and intelligence customers. Anthropic is the recent exception, ejected from Pentagon work in early 2026 after refusing to lift restrictions on autonomous weapons and mass surveillance use. The other major labs and Big Tech vendors continued without similar conditions, and the same capabilities showing up in enterprise and consumer products are being tuned, evaluated, and hardened under contract for classified use. This is not a moral argument. It is an observational one. If you want to predict where AI capability will be pressed hardest, it is worth watching where the defence contracts are going, because historically that is where the real-world stress testing has happened, and that is where the tooling that eventually shows up in commercial settings has been matured.

What is genuinely different this time is speed. The mobile internet took roughly two decades to get from frontier communications research to a universal consumer substrate. AI capability is moving through the same cycle on a faster clock, with less separation between frontier research and consumer availability. That compression cuts both ways. It means the capability can reach development contexts faster than mobile did. It also means the errors, failure modes, and misuses arrive faster, before the scaffolding to handle them has formed.

What AI May Skip, and What It May Compound

I want to end on what I am genuinely uncertain about, because the honest answer is uncertainty.

I do not see the low-connectivity or offline populations benefiting meaningfully from AI yet. The current centre of gravity is urban, professional, English-first, and well-capitalised. That is true. It was also true of the internet in 1998 and of mobile in 2002. Neither of those stayed true. The question is not whether AI stays concentrated at the top, but what the expansion path looks like and how quickly the economics close.

Healthcare access is the area where I expect the next durable gains to appear, and the early signals are already there. A diagnostic support model running in a community clinic, second-opinion infrastructure for underserved districts, triage tools that work in local languages. These are the cases where AI competes against nothing rather than against an already-functioning expert, and the welfare arithmetic is almost unavoidably positive. Education is the next candidate, and law is close behind, though the regulatory friction there is a larger barrier than either of the other two.

The harder and more honest question is whether AI compounds existing inequalities while the distribution unfolds. It could. It almost certainly will. The people currently writing about AI are the people most visible to AI product managers, the people whose workflows get optimised first, and the people whose productivity gains get measured and publicised. Meanwhile, the populations who would benefit most from expert access at the margin remain outside the feedback loop that shapes what gets built. Mobile narrowed some gaps and widened others. AI will do the same. The mix will depend on choices (pricing, language, regulation, device cost, infrastructure) that are not inevitable and are currently being made. Even so, AI is a gateway to better things in nearly all markets. The compounding harms and the substrate gain run in parallel, not in opposition.

The practical position I hold, after enough time watching this pattern repeat: be optimistic about the eventual reach, be sceptical about the current distribution, and be careful not to pretend that because the technology exists the benefits will arrive. The internet was going to reach Tanzania. It eventually did, and on a timeline and in a form that almost nobody in 2002 predicted correctly. AI will reach the equivalent places. What it looks like when it gets there, and who it reaches first, is being decided now.

What the Series Was Arguing

Every post in this series has been the same argument in a different suit. Substrate transitions are slower, messier, and more expensive to individual investors than the headlines suggest. They are also more durable, more civilisationally consequential, and harder to reverse than the same headlines imply going the other way. The bubble is not the opposite of the revolution. The bubble is how the revolution is financed.

What matters in each case is the checklist. The internet did not replace retail until search, payments, logistics, trust, and habit had all closed. Streaming did not replace the disc until broadband, compression, connected television, delivery infrastructure, licensing, and subscription psychology had all closed. The smartphone did not dissolve the internet into the ambient layer it now is until hardware, operating systems, app economics, carrier pricing, and social behaviour had all closed. Access, the case this post is about, is closing on a similar and partly overlapping list, at different speeds in different places.

The part that matters most for AI is this. AI is sticking around. The internet did not get rolled back after 2001, even though most of the companies did. Mobile did not get rolled back after the carrier consolidation waves of the late 2000s. Streaming did not retreat to physical media after the licensing and subscription fatigue arguments of the mid-2020s. None of those shifts went backwards. It can happen, but very rarely do technological advances come out, make a splash, and then recede from history. Once a technology has crossed enough of its checklist that ordinary behaviour comes to depend on it, the behaviour pins it in place. AI is past that line in parts of software development already, and it is closing on the line in several other professional domains. Pandora’s box is open. The genie is out of the bottle. There is no version of the next decade where AI goes away, regardless of whether any particular frontier lab does.

That does not mean every current AI company is a real company, or every current AI product is a real product. Most of both will fail. That is what the DotCom pattern tells you to expect. It is also what the streaming pattern tells you to expect. What survives is whatever sits on the correct side of the checklist when the preconditions finish closing, and whatever the people using the technology continue to want once the novelty wears off.

If I had to compress the whole series into one line, it is this. Substrate transitions reward operators who stay in the game long enough to see the checklist close. Not the loudest, not the biggest, not the ones who bet on the single right product. The ones who keep building through the rocky part, who carry the legacy that is still paying while they build the thing that is still becoming, who watch the preconditions rather than the hype cycle, and who understand that the transition is not an event. The transition is a decade. Sometimes two. AI is early in that decade now. The work in front of us is what the work in front of the 1998 internet operators was, and what the work in front of the 2008 streaming operators was. It is to keep building, keep watching, and be the person still standing when the next version of normal arrives.