Smartphones Made the Internet Ambient

In 1999 I was running a web hosting business out of real server racks, selling dedicated machines to customers who mostly reached them from Windows desktops sitting on furniture. The endpoints were known. They lived in offices and bedrooms. They plugged into the wall. If you wanted to manage your server, you sat down at a desk and did it.

If you had told me that within fifteen years the majority of internet traffic would originate from handheld glass rectangles, in cafes and on trains and in beds at three in the morning, from people who would never see the inside of a data centre, I would have believed the direction and dismissed the pace. Handhelds existed in 1999. Palm Pilots. Early WAP phones. Java powered how many devices? None of them were on a trajectory to dissolve the distinction between being online and being awake.

Part of a series on what the DotCom era can teach us about AI. ← First: Bubbles Can Build Foundations · ← Previous: From DVD Mail to Streaming

That is exactly what happened anyway.

If streaming was the application this new internet would eventually carry, the smartphone was what made it ambient in the first place. One layer further down, and one stage earlier in the sequence.

I have spent the last decade in network operations, watching the consequences of that dissolution from the infrastructure side. The internet is no longer a place people go. It is a layer the world now runs on. That shift, which happened roughly between 2007 and 2015 and consolidated through the decade after it, is the most consequential durable residue the DotCom era left behind. It is also the single best template available for thinking about what AI becomes, if it becomes anything that lasts.

The series so far has argued three things. Post 1 set the frame: bubbles can destroy extraordinary amounts of investor capital and still finance durable infrastructure, and the DotCom crash was both real and productive. Post 2 took the retail case: Amazon did not win e-commerce by being early, but by being still standing when search, payments, logistics, trust, and generational habit change all closed at roughly the same time. Post 3 applied the same checklist reading to streaming: the disc did not lose to the internet, it lost to a checklist of conditions (broadband, compression, connected-TV penetration, delivery infrastructure, licensing, subscription psychology) that finally closed around 2010 to 2013, on top of a decade of bubble-era fibre laid in 1999 and 2000.

This post makes the next claim. The biggest change the internet delivered was not any particular product. It was the transition from destination to substrate. And that transition happened through the smartphone, via a period that was considerably rockier than most retrospectives admit.

The Ladder That Was a Curve

In May 2011, when Pew Research ran its first smartphone ownership survey, 35 percent of U.S. adults owned one. By 2025, that figure was 91 percent. Globally, the GSMA reported 4.6 billion people using mobile internet on their own device by end of 2023, with another 3.1 billion living inside coverage but not using it. The last decade of the Ericsson Mobility Report records 5.6 billion new smartphone subscriptions added between 2011 and 2020, and roughly 300 times more mobile data traffic over the same period.

Those are not numbers that describe a gradual rollout. They describe an S-curve steeper than broadband and denser than the PC. What looks in hindsight like a ladder (iPhone 2007, App Store 2008, Android 2008, 4G 2010, mobile-majority by 2015) was in real time a series of contested years where no one knew yet which phones would matter, which platforms would survive, or which app categories would hold value.

The direction was visible. The pace and the shape were not.

That is the point that most most AI coverage in 2026 is getting wrong. The direction of ambient, workflow-integrated AI is visible to almost anyone paying attention. The pace and the specific shape it will take are not. The 2024 NBER paper from Bick, Blandin, and Deming found that roughly 40 percent of U.S. adults aged 18 to 64 were already using generative AI by late 2024, while the Census Bureau’s firm-level surveys showed business adoption still in single digits. That gap between individual usage and firm-level transformation is the exact gap that existed between smartphone ownership and smartphone-native business operations in 2012. It is the gap during which positioning matters most.

The Storefront Was Not the Business

A smartphone displaying a colourful grid of app icons next to a notebook sketching a network diagram of connected services, suggesting the emergence of an app ecosystem as the substrate for new kinds of businesses.

Apple launched the App Store in July 2008 with 500 applications. By 2023, the app economy was a multi-trillion-dollar distribution layer that had reshaped retail, payments, transportation, dating, healthcare booking, and the internal tooling of almost every large business that had a workforce interacting with the outside world. The App Store was not the business. It was the substrate on which businesses that could not have existed ten years earlier grew to sizes that would have been absurd to forecast in 2008.

Uber incorporated in 2009 and ran its first rides in San Francisco in 2010. Instagram launched in October 2010. WhatsApp, Snapchat, Venmo, Airbnb, DoorDash, Robinhood, Duolingo, and most of the consumer surface of modern financial services were built on the assumption of a pocket-sized computer with a camera, a location sensor, a payment credential, and a persistent connection. None of those assumptions held in 2006. By 2013 all of them did.

The durable lesson of the app era is not that smartphones created new companies. It is that they relocated where value accrued. The storefront (the App Store itself, the handset, the operating system) captured a fraction of the economic surplus. The rest accrued to whoever built businesses that presupposed smartphone ubiquity as a starting condition rather than a feature.

The AI equivalent is already visible, but it is easy to misread for the same reason. ChatGPT, Claude, Gemini, and the other chat interfaces are storefronts. They are real, valuable, and not the durable business. The durable businesses are being built by companies whose products presuppose capable AI as a starting condition: coding environments where agentic assistance is load-bearing rather than optional, medical imaging pipelines where model-assisted reading is the default, legal research platforms where the human attorney is shepherding AI-generated analysis rather than producing it from scratch. Those products will not be called “AI products.” They will be called products, the way we stopped calling Uber a mobile-app company and started calling it a transportation company.

Trying to pick the winning foundation model is the 2026 equivalent of trying to pick the winning handset in 2008. There are people who bet correctly on Apple versus everyone else, and they did extremely well. Most people who tried to make that call got it wrong. The more reliable bet was on businesses that would benefit regardless of which handset won.

From Pocket to Layer

The cleanest analytical move available for this post is also the one that is hardest to articulate without sounding like a TED talk. The smartphone’s durable consequence was not portability. Portability was the surface feature. The durable consequence was ambient presence.

In 2005, using the internet required intentional engagement. You sat down. You opened a browser. You went somewhere. The internet was a place with edges, and between trips to that place, you were offline. Offline was a default condition most of the time.

By 2015, offline had become a state you actively had to produce. Your pocket was online by default. Your wrist might also be online. Your car increasingly was. Notifications arrived whether you solicited them or not. Location was continuously legible. Payments were a gesture. Asking a question, summoning a car, paying a bill, or calling a stranger had collapsed into sub-ten-second interactions that happened while you were walking to your next actual activity.

The internet had not simply grown. It had become ambient; it is always with us, everywhere we go. It was no longer something you used. It was something you inhabited. That is the substrate shift, and most of the people who built the best businesses of the 2010s understood it implicitly even when they did not articulate it.

Current AI, in 2026, is mostly still a place you go. You open a tab. You switch out of the tool you were working in. You paste context back and forth across the boundary. That is the 2008 version of the transition. It is real, it is useful, and it is not the final form. The final form, the thing that will actually change behaviour and markets, is AI as a capability present in the environment where work already happens, available without the ceremony of switching, context-aware without being told, and invisible enough that people stop noticing it is there.

Software development is already through that threshold in parts. Most other domains are not. Watching which ones close the gap, in what order, with what conditions aligning, is the single most useful exercise an operator can spend time on right now.

The Incumbents Did Not Miss Smartphones. Their Moats Became Their Weight.

A drawer holding old feature phones, a QWERTY keyboard handset, and a PDA with a stylus, with a modern smartphone resting on the desk above them, representing the generational turnover in mobile devices as the substrate shifted.

The smartphone transition was not a smooth march toward ambient. It was a period of sustained capital destruction, corporate collapse, and operator error that is considerably uglier in the record than the hindsight version suggests.

Nokia was the dominant handset manufacturer on earth in 2007. It held roughly 40 percent of the global mobile phone market. By 2013 it had sold its handset business to Microsoft at a price that reflected how thoroughly the franchise had been destroyed. BlackBerry peaked at nearly 50 percent U.S. smartphone share in 2009 and fell below 1 percent within six years. Palm, once synonymous with the PDA category, was sold to HP in 2010 and shut down by 2011. Microsoft’s own mobile effort, after several iterations, was effectively abandoned.

None of those companies missed the smartphone transition. They saw it clearly. Nokia had a better manufacturing operation than Apple. BlackBerry had a better keyboard, better enterprise security, and better battery life. Microsoft had every possible software advantage. They were not blind. They were comfortable. Each of them was running a business whose cost structure, organisational assumptions, and internal incentives had been built around defending a position that was about to be irrelevant. When the market shifted underneath them, none of those internal arrangements could move fast enough. The moat became the weight.

The broader app economy around them was also much rockier than the retrospective highlight reel suggests. Between 2011 and 2014 there was a wave of mobile-first consumer startups that raised venture capital on the assumption of attention metrics that later proved uneconomic. Zynga went public in December 2011 at $10 a share, reached a market cap above $10 billion in March 2012, and lost three-quarters of its value within fifteen months. Groupon’s trajectory was similar and uglier. Foursquare, Path, Viddy, Color, and a long tail of less remembered names burned through capital pursuing network-effect stories that did not compound. Venture returns for the 2011 and 2012 vintages were dragged down substantially by the mobile-consumer portion of the portfolio.

The underlying substrate was real. Much of the capital deployed to exploit it was lost. Both of those things were true at the same time, and the people who did best through the period were not the ones who guessed the right consumer app. They were the ones who positioned to operate on the new substrate whenever it settled, regardless of which specific player won the surface layer.

The current AI equivalent is already visible, and it is legacy enterprise SaaS. A long tail of vendors with entrenched positions, bloated feature sets, and unsustainable prices is being undercut by AI-fronted competitors on price and speed at once. Some of that displacement is deserved. A SaaS product that charges enterprise prices for a mediocre workflow that a subsidised AI wrapper can replicate at a lower price is a product the market is correctly repricing. Some of it is subsidy-driven and temporary, because the AI wrappers doing the damage are often operating on inference economics that do not yet reflect the real cost. But the subsidy being temporary does not save the incumbents whose cost structure depended on the old substrate continuing. That is the Nokia pattern. They are not blind. They are built for a market that is shifting, and the structure cannot adapt fast enough. The really strong enterprise tools, the ones solving a hard problem well, will come through this. The mediocre ones will not, and probably should not.

That is the pattern to watch for with AI. The capital going into foundation model training and inference capacity is historically large. Microsoft’s fiscal 2025 property additions were above 64 billion dollars. Alphabet’s 2025 capex was 91 billion, guided to 175 to 185 billion for 2026. Meta’s 2025 capex was 72 billion, guided to 115 to 135. Amazon guided to about 200 billion in 2026. Combined, the four hyperscalers are guiding to close to six hundred billion dollars of 2026 capex. Some meaningful fraction of that will prove to have been wrong in the specifics of where, when, and for which model generation, even if the substrate bet is directionally correct. The dark fibre comparison the series has used before applies here too. The infrastructure outlasts the investors. The investors often do not.

That is the rocky part the series keeps returning to. It is coming again. It is not a reason to sit out. It is a reason to notice what layer you are building on.

What Was Actually Lost

Any honest assessment of the ambient internet has to include what it cost, because the AI transition will involve similar tradeoffs and the rhetorical habit of pretending those tradeoffs do not exist is itself part of the problem.

The dedicated camera industry collapsed. PetaPixel’s coverage of the camera market records a 94 percent drop in camera shipments between 2010 and 2023, a collapse attributed almost entirely to smartphones. Paper maps, standalone GPS units, portable music players, pocket calculators, and many low-end gaming handhelds followed similar curves. That is category destruction, and for the people who worked in those industries it was not abstract.

Work-life boundary erosion is better documented than the advocates usually admit. A 2024 meta-analysis in PLOS Digital Health found that roughly 80 percent of empirical studies linked off-hours smartphone use for work to worse work-life conflict. Attention research has repeatedly shown that notification-driven environments raise stress markers, degrade sustained concentration, and produce measurable productivity costs even when users believe themselves to be more efficient.

Newspapers collapsed. Retail geography compressed. Local advertising markets that had supported a broad class of small independent businesses were hollowed out by two or three platforms whose scale and targeting precision outcompeted the local alternative almost completely. Political discourse restructured around formats that reward intensity over duration. Adolescent mental health indicators worsened materially after 2012 in ways that track smartphone and social media adoption closely. Whether that correlation is causal remains genuinely contested in the research, and readers should not take either side’s certainty at face value. The direction of the evidence, across the work of Jean Twenge, Jonathan Haidt, and a growing corpus of longitudinal studies, is consistent enough that treating smartphone-era adolescence as identical to earlier cohorts is itself a strong assumption. Whichever way the causal question eventually resolves, the change is real and the change arrived with the substrate.

None of that invalidates the underlying technology. All of it is part of the honest cost column. Ambient means pervasive means unavoidable means changes everyone, including people who did not choose the change.

The AI parallel is obvious and deserves to be named rather than gestured at. Dependency on systems whose internals are not auditable. Homogenisation of output when large populations draft through the same models. Erosion of the low-stakes friction that used to force humans to develop their own judgment. Displacement of entry-level work that functioned as a training ground for mid-career competence. None of these are hypothetical. All of them will be to AI what notification culture, platform centralisation, and attention capture were to smartphones: real costs that the technology produces alongside its real benefits, and that operators who care about their teams and their customers have to reckon with honestly rather than minimise.

Two costs more particular to the current AI moment also deserve naming. The first is subsidy. A lot of what looks like AI disruption in 2026 is not AI disruption. It is subsidised AI inference out-pricing unsubsidised software. Foundation-model providers are funding inference at prices that do not reflect its real cost, venture capital is underwriting the rest, and a mediocre AI-fronted product can undercut a strong but fairly priced software tool on that basis alone. That is not a stable configuration. When the subsidy ends, and it will, the clearing price is going to look quite different from what the current price suggests. Operators who built their economics on the assumption that inference stays at today’s price are making the same bet early mobile startups made when they assumed carrier-subsidised handsets and venture-subsidised customer acquisition would continue indefinitely. That assumption did not hold. This one will not either.

The second is AI tech debt. We used to say that the first ninety percent of the project took ninety percent of the time, and the last ten percent took another ninety percent. AI can build the first ninety percent of a product faster than most teams can type. If that ninety percent was vibe-coded without strong engineering oversight, the remaining ten per cent will take a hundred and ten per cent of the effort it would have taken to build the thing properly in the first place, and you will be lucky to recover to a most of intended functionality. Teams are already hitting that wall. Products ship fast, run into something the original generator did not understand, and the team discovers that the code it now owns was never understood by anyone, the test coverage is theatrical, the architecture does not compose, and every change produces two new bugs. That is not a reason to stop using AI for engineering. It is a reason to notice that engineering judgment has become more valuable, not less, precisely because the cost of producing a plausible-looking wrong answer has dropped to near zero.

AI’s Ambient Form

A developer working at a laptop with a code editor open, AI assistance woven into the editing surface rather than appearing as a separate chat window, illustrating ambient AI as an environmental condition of the workflow.

There is a version of this post in which the AI parallel is a gesture at the end. That is the wrong version. The smartphone story is analytically useful only if it sharpens thinking about what ambient AI actually looks like and when.

The chat window is to AI what the Palm Pilot was to the smartphone era: a real product solving real friction, training a generation of users on what the capability feels like, and building a base that the next form will inherit. Not the internet, but another product that would be subsumed by it. It is not the form that actually reshapes work. The form that reshapes work is the one that stops being a thing you go to.

What that form looks like, concretely, is already visible in the parts of the economy where AI has the deepest penetration. In modern coding environments, AI assistance is in the editor, not in a tab. It sees the repository. It sees the file you are editing, the tests you have run, the git history around the change. It offers completions as you type, not after you ask. Some portion of the work has already become AI-assisted in a way that would be awkward to roll back, not because the model is spectacular but because the integration is structural.

That same pattern, when it arrives elsewhere, will not announce itself. Your CRM will start proposing next best actions that make the obvious next action five seconds faster, and then three months later your reps will not know how to operate without them. Your document editor will start surfacing context and references you used to have to go look up, and then the step of going to look things up will feel archaic. Your imaging system at the hospital will present prior-scan comparisons and literature-derived differentials alongside the scan itself, and the radiologist will still decide, but the decision will be made inside a richer context than it was in 2024. None of those changes will feel like an “AI rollout.” They will feel like the tool got better.

That is what ambient looks like. It is already happening, unevenly, along the axes where the workflow integration problem is easiest to solve and the reliability is good enough. It will arrive in the rest of the economy in roughly the order that those two conditions close in each domain.

How to Position Through the Rocky Part

In the streaming post I listed six habits for operating through a substrate transition: watch the checklist, carry the legacy that is still paying, do not commit to the Blockbuster position, build the harness not the hero bet, protect optionality, and build through the rocky part. Those habits came out of the streaming case, but they describe the mobile transition just as well. This section is the mobile-era evidence for the same framework, with a few positioning calls that are particular to what ambient AI is doing right now.

The practical question the series has been building toward is what an operator, a career builder, or a company leader actually does with this pattern. The honest answer is not a formula. It is a set of habits that the people who came through the smartphone transition with their positions intact tended to share. They apply here too.

Stop betting on which specific model wins. Token pricing is under commodity pressure, open-weight models are closing the benchmark gap, and no one has a stable forecast for which of the current frontier providers is the same company in five years. Betting on a model is 2008’s betting on a handset. It is a bet worth making if you are a capital allocator with strong conviction and a long time horizon. For most operators it is the wrong question.

Build for the substrate instead. Ask what your product, your workflow, or your role looks like when ambient AI is an environmental condition rather than a feature. If the answer is “exactly the same, but with an AI chat widget bolted to the side,” you are building the equivalent of a 2009 mobile-responsive website for a business that is about to be disintermediated by an app-native competitor. Ambient AI rewards products that assume the capability, not products that present it.

Build the harness, not the token. Tooling, evaluation, retrieval, agent orchestration, workflow integration, and the domain-specific knowledge encoded into these layers are the sources of compounding advantage that the model layer alone does not provide. Several of the companies that did best out of the mobile era were the ones that built proprietary operational layers on top of commodity infrastructure. The same opportunity is open now.

Position your career toward roles where ambient AI raises your ceiling rather than threatens your floor. In every major technology shift I have watched up close, the people who suffered most were the ones whose work consisted of a narrow, procedural task that the new substrate could perform adequately without them. The people who did best were the ones whose work consisted of judgment, taste, relationship, orchestration, or synthesis, applied inside a substrate that made their reach larger. That is not a trick. It is a general pattern. AI does not change the pattern; it accelerates the selection.

Accept the rocky part. The capital cycle around AI is going to produce losers, including some losers currently being described as certainties. Some data centre builds will not pencil out. Some model providers will be acquired at fractions of their last round. Some product categories that look inevitable today will turn out to have been subsidised novelties. That was also true of mobile, and it did not prevent mobile from becoming the substrate we now take for granted. It is possible, and often necessary, to hold both thoughts at once: the aggregate direction is correct, and the specific bets are rough.

Build what needs to be built. The most frequent error during a substrate transition is to freeze, wait for the dust to settle, and re-enter when the shape is clearer. he shape is never as clear as people hope. It becomes obvious in retrospect, but in the present it is negotiated by the people still building. The people who did best out of the mobile transition were the ones who built during the messy years. The people who did best out of the DotCom transition were often the ones who kept building when the market was punishing everyone for building. The same choice is on offer now, in a slightly different costume.

None of this guarantees anything. The point of this series has never been that there is a safe place to stand during a technology shift of this magnitude. The point is that standing still is not the safe choice people think it is, and that the history of these transitions contains more useful guidance than the current narrative around either hype or doom gives credit for.

The internet became ambient because enough people kept building through a decade of expensive mistakes. AI will become ambient, or not, on a similar schedule and by a similar mechanism. Whether you are standing on the right side of the transition ten years from now depends less on predicting the winners and more on what you choose to build now, in the rocky part, while the shape is still being worked out.