Alok Band.

Meta shipped Muse, an always-on AI agent, on September 8th. Ten days later they opened connector access to developers. At Connect on the 23rd they added Mac computer use, voice mode, more connectors, and AI glasses.

Three weeks, and a good part of the industry has decided this is the end of the app. Gartner predicted back in January 2025 that mobile app usage could fall 25% by 2027 as AI assistants take over more functions, and Muse looks like that arriving on schedule.

I read it differently. Muse is the strongest evidence I've seen that apps survive. What dies is the grid of icons, and that distinction carries more weight than it sounds.

An app was never an interface. If you've operated a platform you know it's a bundle of obligations: a distribution right, a content license scoped by territory, a billing relationship, entitlement state, DRM keys, a certification status, a named accountable party, and a revenue share contract. Not one of those is a pixel, and every one still has to exist the moment an agent acts for you. When an agent offers to play the new season, something holds the rights, checks the subscription, enforces geography, decrypts the stream, and gets paid. An agent can be a front door. It cannot be the rights holder, and it cannot be the party that's liable.

Which is roughly what Muse demonstrates. Its connectors are a reviewed directory plus an unreviewed custom tier, with security and legal checks, Stripe payment rails, per connector permissions split between read and act, a submission process, and editors who can feature things. That's an app-store-like distribution layer. Meta set out to build the post-app interface and ended up rebuilding much of the machinery app distribution already required, inside the agent, in two weeks. Then announced Muse can drive any app on your Mac while you walk away.

You can delete the icon. You cannot delete the contract underneath it.

So the strategic question isn't whether apps persist. It's what happened to the moment of selection.

When a person picked an app, discovery was an interface problem. Grid, search box, recognizable logo. When an agent picks a connector, it becomes a ranking problem. The App Store optimized for human selection. The agent platform optimizes for machine selection. Different disciplines, different winners.

Meta didn't eliminate the distribution layer. It moved distribution from something the user browses into something the agent mediates. Ranking may become one of the most valuable surfaces in the agent economy, for the same reason retail shelves and search results did before it. Selection is where commercial leverage accumulates.

Every major agent platform will eventually have to confront this. Most won't announce it.

Three consequences I'd plan against.

Distribution investment has to move. A decade of store listings, onboarding funnels and install conversion was built around someone opening your thing. If an agent mediates the first touch, a chunk of that spend is depreciating and very few budgets reflect it. What replaces it isn't marketing work. It's product data, API quality, permission design, latency, and reliability under retry. Those sit with engineering and product rather than growth, and that reporting line matters.

Your capability description becomes a competitive asset. Being findable is not being selectable. An agent evaluates whether you can satisfy an intent using structured metadata and observed reliability. That's something you own, version and improve deliberately, and almost nobody has an owner assigned to it. I'd assign one.

And governance, which is the one I'd escalate. You now decide which service gets picked, and nobody sees you decide it. That gets monetized, then it gets questioned. Design for visibility now rather than explaining yourself later.

This gets particularly interesting on television, which is where I spend my time.

The TV home screen has spent two decades organizing entertainment around app icons, when the viewer's intent is almost always content. A game. A movie. A show they're three episodes into. Nobody ever wanted to open an app. They wanted the thing inside it, and we made them navigate a grid to get there.

An agent collapses that navigation layer, and it should. But it doesn't remove a single OS responsibility underneath: identity, entitlement, content rights, playback, DRM, advertising, measurement, certification, and the economics between OEMs, publishers and platforms. If anything it makes them harder, because they now have to resolve correctly without a person clicking through to confirm anything.

Which is the version of this argument I'd defend. The interface can disappear while the platform becomes more important.

None of this is a reason to slow down. Muse is good and the direction is right. I'd just stop merging two claims. The visible layer is collapsing, and that's fine. The contractual layer is getting more load bearing, because more decisions happen without anyone watching and someone still has to answer for them.

The question was never whether we keep the icons. It's who ranks what sits behind them, on what basis, and whether anyone outside the platform can tell.

Meta answered that for their surface in three weeks. The rest of us should decide deliberately rather than discover it in a quarterly review.

Tokens keep getting cheaper. Capability that ran in the tens of dollars per million tokens back in 2023 is well under a dollar now, and it keeps falling.

So why is everyone's bill going up?

Gartner put out a note last month projecting that inference cost per agentic workflow climbs more than fivefold through 2028. That's while model prices keep dropping. They call it the Inference Paradox, which is a decent name for it. Cheap tokens let you build more ambitious things, and the ambition eats the savings faster than the savings arrive.

Most of what I've read about this treats it as a finance story. Somebody in FinOps spots the bill, there's a cost sprint, everybody routes traffic to a smaller model, done. I don't think that's where this belongs. By the time it shows up on a finance dashboard, the decision that caused it happened months earlier in a product review, and nobody in that room was thinking about cost.

The reason it lands on product is kind of a big deal, and I don't see many people saying it plainly.

Software product management has spent roughly thirty years operating on an assumption nobody says out loud, which is that marginal cost is basically zero. You build it once. The ten thousandth user costs you nothing. Every reflex we have comes from there. Ship to everyone, turn it on by default, grow first and figure out the money later. That was correct for a very long time and it isn't correct anymore, at least not for anything agentic. Every invocation costs. The unit that matters stopped being cost per prompt and became cost per completed task.

Once that's true, "should this be an agent" isn't an experience question. It's a margin question, and it belongs to whoever owns the roadmap.

Where this gets genuinely uncomfortable is if you're on the ad-funded side, which is where I've spent most of my career. Let me just make up numbers to show the shape.

Say an AI discovery session on a TV costs three cents to complete. Say a household runs one four times a week. That's $6.24 a year on that device. If the device makes you $30 a year, one feature took 21 percent of your revenue, and content and infra and support and everything else still has to come out of what's left.

Those inputs are invented. Change any of them and the number moves. But the shape doesn't move, and the shape is the whole problem. Subscription businesses can carry a cost like that because the customer already paid and better discovery keeps them around. Ad-funded fleets can't do the same trick. ARPU is a ceiling and you don't get to raise it by being clever about the feature. Streaming, FAST channels, most of the open internet, all of it sits under that ceiling. Almost nothing written about inference cost is written from inside it. It's mostly coming out of SaaS, where the marginal user hands you a subscription on the way in.

It's worth being precise about why the costs run away, because it's the reason your pilot numbers are useless.

A chatbot question is one call. An agent isn't one call. It plans, pulls context, hits a tool, reads what came back, checks itself, sometimes starts over. Gartner's framing is that a chatbot reads and answers while an agent has to keep reasoning and questioning itself, and all of that self-questioning is metered. They put agentic workflows at five to thirty times the tokens of a normal chatbot turn. That's a wide range, which tells you the variance is the real risk, not the average.

Jevons paradox, basically. We made it cheap and then used an absurd amount more of it.

And you don't see it in a pilot, because a pilot doesn't run the loop ten thousand times a day. Gartner thinks north of 40 percent of agentic projects get cancelled by end of 2027. Cost is one of the reasons they list, along with unclear business value and weak risk controls, so I don't want to pin it all on economics. But cost is on the list.

A few things that seem to work.

Tier the models on purpose. Most agentic traffic doesn't need a frontier reasoning model and never did. Send the boring majority somewhere cheap. Gartner's conclusion is that competitive AI products end up maintaining several models rather than picking one, which sounds like more work because it is more work, and it's also where the savings live.

Give every workflow a cost ceiling per completed task before anyone writes code. Treat blowing through it like blowing a latency target. Same category of failure, same seriousness.

Put cost per completed task on the dashboard next to engagement and retention. This one is easy to skip and I'd argue it's the one that saves you. A feature with great engagement and upside-down economics isn't a win. It's a liability with good reviews.

And then the hard one, which is deciding what shouldn't be an agent at all. A lot of what's shipping as agents right now is a search query, a cache lookup, or a decent default with extra steps and an invoice attached. Reasoning should be the escalation path, not the opening move. Search first, retrieve second, predict where it earns its keep, and reason only when you actually have to. Every rung up costs more than the one below it.

The part I keep chewing on is what this does to the job.

Everybody's hiring for AI fluency, which usually means knowing what the models can do. Fine. The harder thing is knowing what they should be allowed to do, at what price, for what return. We've always allocated engineering time. Now there's a second budget, and it's intelligence, and since it costs money every time you spend it the architecture call and the margin call turn out to be the same call made by the same person. That's not the job most of us signed up for. It's a more interesting one, though.

Cheap tokens didn't make this problem go away. They made it possible to build the products that create it.

Perplexity recently blocked ads that a publisher had formatted specifically to influence AI agents browsing on behalf of users. Their word for the practice was "deceptive." What's interesting is what they were actually objecting to. This wasn't invalid traffic or some classic ad fraud scenario. It was paid influence being inserted directly into an agent's information supply, presented in a way that looked like publisher content. That distinction matters because agents don't consume advertising the way humans do. You can't really buy their attention, because they don't have any. What you're trying to influence is the decision they make on behalf of a user, and the platforms running those agents are already starting to set rules around how that can happen. So the industry's oldest question, "did a human see my ad?", is about to sit next to a very different one: "did the agent pick my product?" Those are not the same economic event, and they probably shouldn't be priced the same way either.

For the last twenty years, digital monetization has largely been about one thing: capture human attention and sell it through impressions. Viewability, attention metrics, frequency caps, reach, brand lift. Most of the machinery we built exists to narrow the gap between "an ad was rendered" and "a person actually noticed it." Agentic AI doesn't make that go away. People will still watch, scroll, stream and browse, and human attention will continue to have value, especially where you can prove it is actually there. But a second lane is opening next to it. Call it the action lane. Here, the impression is no longer the main economic unit. Brands are competing for what McKinsey calls "agent eligibility," which is basically the ability to be considered and selected when an agent is trying to complete a task for someone. That could mean a publisher gets paid when its content helps answer a query, not when someone opens a page. Or a platform takes a cut of a transaction instead of a share of an ad impression. Or a brand spends money making sure its products are structured, available and trustworthy enough for an agent to select.

In some ways, agent eligibility is just the next version of shelf space. In physical retail, companies paid to be at eye level. In search, they paid to be at the top of the results. In an agent-driven world, the competition is to be machine-readable, trustworthy and useful enough that the agent actually picks you. That's less about a media buy and more about product data, pricing, availability, reviews, policies and trust. And this isn't theoretical anymore. SSPs have already started launching agent-to-agent orchestration layers, where buyer-side agents can interact directly with sell-side agents. OpenAI moved from an ad pilot to a self-serve ads manager with CPC bidding and conversion measurement in a matter of months. At the same time, a per-query licensing model is starting to show up underneath all of this, where publishers can get paid when their content contributes to an answer rather than when a page is rendered. The point isn't that impressions disappear. It's that more and more of the value can now be tied to access, selection and outcomes instead of attention alone.

This is where I disagree a little with the broader "AI is going to destroy advertising" narrative, and a lot of that comes from spending years on the platform side of CTV. The living room is probably one of the strongest remaining places for human attention. Nobody's agent watches a movie for them. As agents mediate more activity across web and mobile, the ability to prove that an actual person is sitting there consuming something becomes more valuable. CTV has a pretty unique advantage here, because a big-screen video session in a living room still has a very high probability of representing real human attention. And scarcity usually creates pricing power. So I actually think premium CTV attention should become more valuable, not less.

But discovery and commerce are a different story. Think about some of the things people are likely to delegate: "What should we watch tonight?" "Where can I watch the Warriors game?" "Buy the jacket in this scene." "Which streaming bundle is cheapest for what we actually watch?" Those are all natural agent tasks, and the TV operating system is in a pretty unique position to intermediate many of them. Once an agent starts participating in discovery and transactions on the TV, the OS becomes the tollbooth between the attention lane and the action lane. It can influence which agents get access, what signals they see, and ultimately the economics around being eligible for selection. That puts the TV OS in a position that very few other parts of the ecosystem occupy. It can potentially monetize both sides of the same consumer journey. Human watching is attention economics: premium verified reach. Agent deciding is action economics: discovery, eligibility, query access and transactions. Same household, same screen, two currencies. That second surface is barely represented in most CTV ARPU models today, and I don't think that will stay true for long. The platforms that show up to that world with clean content metadata, consented identity and a real transaction layer are going to have a much stronger negotiating position than the ones that treat this as another ad-slot problem.

So what should CTV operators actually start doing now? Three things. First, instrument human attention and agent-mediated activity separately. If your reporting can't distinguish a human-consumed impression from an agent-mediated query, recommendation or transaction, you can't price either one properly, and eventually you'll end up applying human-attention economics to agent activity until the market reprices it for you. Second, start treating metadata and identity as monetization inputs, not just plumbing. Agents select based on structured and verifiable information, so clean metadata, consented identity, transparent pricing and machine-readable policies are what make something easy to discover and easy to select. Teams that treated data quality as a cost center may eventually realize it was actually part of the storefront. Third, decide which lane each surface really belongs in. In-stream video is attention and should be defended as premium human reach. Discovery, search and commerce are much closer to action, and those surfaces can be monetized through eligibility, access and transactions. The mistake would be to blend all of this together and assume one pricing model still works everywhere, because if attention and agent activity are priced the same way, someone is probably getting a very good deal.

The impression isn't dying. It's just losing its monopoly. For the first time since the banner ad, digital monetization may start running on two currencies at once: attention, priced on humans watching, and action, priced on agents doing. The common assumption is that AI agents are a threat to advertising. For television, I think the opposite may be true. Agents make genuine human attention scarcer and therefore more valuable, and at the same time they create a second monetization layer for the TV OS around discovery, eligibility and transactions. The platforms that understand they're operating a tollbooth, not just an ad server, will be in a position to capture both. Perplexity's move was an early warning. They didn't block those ads because agents don't matter. They blocked them because agents matter enough that the rules for influencing them are already being written. CTV operators should probably assume those same rules are coming to the living room next.

Advertising measurement is better today than it has ever been. For most of its history, the industry ran on a single panel-based number, a rough estimate everyone accepted because there was nothing better. Now we can measure real audiences, real outcomes, and real sales with a precision that would have looked like science fiction a decade ago. That is a genuine leap, and it deserves to be celebrated.

The reason is simple and encouraging. As attention spread across hundreds of platforms and screens, the industry did not sit still. It built. Privacy-safe ways to match data. First-party purchase signals tied directly to outcomes. Whole new categories of infrastructure that let companies collaborate on data without exposing anyone's raw records. Each of these solved a hard problem that used to be considered unsolvable.

Clean rooms are a great example of how far we have come. A few years ago, connecting a brand's data to a platform's data without passing identifiers around felt impossible under modern privacy rules. Clean rooms cracked it. Two parties match data in a secure space and each gets results without seeing the other's raw records. It worked so well that adoption took off. Roughly two-thirds of organizations now use clean rooms in some form, and among Fortune 500 advertisers, 78 percent have deployed or are piloting at least one in 2026, up from 34 percent in 2022. That is one of the fastest infrastructure shifts the industry has ever seen.

So the individual pieces are in remarkable shape. Each platform can now measure its own contribution with real rigor. Which brings us to the exciting part, the thing that is not solved yet, the frontier worth building toward.

We have gotten brilliant at measuring inside any single environment. The next breakthrough is measuring across all of them at once. Today a brand can run five platforms, and each one can prove its own results beautifully, but stitching those five views into one honest picture of the whole funnel is still an open problem. It shows up most at the bottom of the funnel, at conversion, where more than one platform can reasonably point to the same sale. The pieces are excellent. Making them add up to a coherent whole is the work ahead.

And here is why I find that genuinely exciting rather than discouraging. The industry has already shown it can build its way through these problems. The interoperability layer, the connective tissue that lets independent measurements be compared, deduplicated, and trusted side by side, is exactly the kind of thing that gets built once enough people decide it matters. The standards bodies are already moving on it. The IAB Tech Lab, the MRC, and the ANA are actively working on how measurement across environments should be defined and accredited. This is a problem the industry is walking toward, not away from.

In the platform work I do, this is the most interesting frontier there is. The hard, rewarding challenge is no longer measuring inside one clean environment. It is building the trusted middle layer that lets many environments reconcile with each other. That is a design problem, an identity problem, and a trust problem all at once, and solving it unlocks something valuable for everyone: a shared, credible picture of what actually worked.

The through line is optimistic. We spent the last decade making every individual piece of measurement dramatically better, and we succeeded. The next decade is about connecting those pieces into something greater than the sum of its parts. The companies that build the trusted layer in the middle, the one that lets independent numbers finally line up, will not just measure better. They will give the whole industry a clearer view of what it is actually doing.

Measurement did not lose its way. It grew up fast, in a lot of directions at once. Now comes the satisfying part: bringing it all together.

Retail media doesn't have a data problem. It has an inventory problem. And CTV may be how it solves it.

The first act of retail media was monetizing the digital shelf. Sponsored products, search results, banners inside a retailer's own app. It worked, spectacularly. Retail media is now among the fastest growing categories in digital advertising, on track to reach about 20 percent of all digital ad spend by 2030 (Omdia). The reason is simple. Retailers sit on first-party purchase data, the record of what people actually bought, and that data closes the loop between an ad and a sale in a way almost nothing else can.

But the shelf is finite. There are only so many ads you can put in front of a shopper before the experience breaks down, and as more brands crowd in for that space, prices climb and reach flattens. The data is abundant. The inventory is not.

That mismatch is the whole story of retail media's second act. The purchase data was always the valuable asset. The retailer's own website just happened to be the first place it got used. Once that data can be activated elsewhere, the addressable market stops being a homepage and becomes the open internet. This is already the fastest moving part of the market: off-site retail media is growing at roughly twice the rate of on-site (eMarketer).

That activation needs somewhere worth landing, and that is the case for CTV. Connected TV is the fastest growing major ad channel in the US, heading toward roughly 51 billion dollars by 2029 while linear spend keeps falling (eMarketer). It brings the one thing on-site inventory can't: scaled, high-attention, full-screen reach. Retail media brings the one thing television historically lacked: a direct line to purchase outcomes. For decades, TV built awareness brilliantly but struggled to connect exposure reliably to a sale. Put the two together and you get television that can be targeted on real purchase behavior and measured against real sales. That combination did not exist a few years ago.

Here is the honest part, and it's where most write-ups wave their hands. Retail purchase data does not simply "travel" across the internet. It moves through clean rooms, identity matching, and consent frameworks that are genuinely hard to get right.

Retailer measurement is often self-attributed, which is not the same as independent proof. CTV supply is fragmented and not uniformly premium. None of this is fatal, but all of it is real, and anyone selling you frictionless convergence is skipping the difficult implementation work.

Which is why I don't think the winners will simply be whoever has the most data. In the platform work I do, the hard problems are never the data itself. They're the plumbing: resolving a shopper across screens without third-party cookies, making measurement hold up across channels, and connecting commerce data to premium supply without eroding the consumer trust the whole thing depends on. The companies that solve those problems quietly, in the background, will capture a disproportionate share of the value.

Search made intent addressable. Social made interest addressable. Retail media and CTV together are making purchase addressable, on the biggest screen in the house.

That's the next phase. It's already starting.

Figures: Omdia (2025) and eMarketer (2025 to 2026).

Part 2 of 2

In Part 1, I wrote about what I saw on a recent trip. AI in India had gone quiet and ordinary. It was in the payment app that noticed something useful, the grocery order that arrived in ten minutes, the voice agent that understood me when I switched languages mid-sentence. Less spectacle, more plumbing.

This part is about what I think it means, especially if you build products for a living.

Because the thing that stayed with me on the flight back was not a single product. It was a way of thinking about AI that felt different from the one I hear every day at home. And I think it is the more useful one.

Distribution beats model quality

This is the one I keep returning to.

The models powering most of what I saw were not the absolute frontier. They did not need to be. They were good enough, and they were sitting on top of rails that already reached hundreds of millions of people. UPI for money. A handful of super apps for daily life. Cheap data and cheap phones underneath all of it.

A merely good model with that kind of reach beats a brilliant model with none.

We forget this easily in the places obsessed with the frontier, because we live inside it and assume the hard part is the intelligence. In most of the world, the hard part is getting the intelligence to a real person in a form they will actually use. India had already solved that second problem before this wave arrived. So the moment the models were ready, they had somewhere to go.

If you are building, the uncomfortable question is not whether your model is the best. It is whether you have earned the distribution to make a good-enough model matter.

The economics are genuinely different

The thing that does not survive the flight back is the cost structure.

When you serve hundreds of millions of price-sensitive users, every fraction of a cent per inference is a real constraint, not a rounding error. That forces a discipline that products in wealthier markets rarely feel. Smaller models where they suffice. Aggressive caching. Routing the easy queries to cheap paths and saving the expensive model for when it actually earns its cost.

This is not a compromise. It is a design philosophy, and I think it is the one that ends up winning.

The teams that learn to deliver intelligence cheaply at scale are building a muscle that the teams burning frontier compute on every query do not have. Scale plus low cost is not a weaker version of the AI story. In a lot of ways it is the harder and more durable one.

AI as infrastructure, not as a feature

The last shift is the easiest to miss, because it is the absence of something.

Almost nothing I used was branded as an AI product. There was no chatbot bolted onto a homepage, no sparkle icon begging to be clicked. The intelligence was load-bearing and invisible. It was inside the payment app, inside the delivery system, inside the phone line. It was infrastructure.

That is the maturity signal. When a technology stops being the feature and starts being the floor that everything else stands on, it has actually arrived. Electricity is not a feature of your house. It is the assumption underneath every other thing in it.

India is treating AI that way faster than the markets still busy slapping AI labels on everything.

For those of us building platforms, this is the real lesson. The win is not a feature users notice and praise. It is a capability users stop being able to imagine living without, precisely because they have stopped noticing it.

What this means if you are not building for India

It would be easy to read all of this as a story about one country and file it away. I think that would be a mistake.

The conditions that made India move this way are becoming more common everywhere. Inference costs are falling. Model quality is commoditizing at the good-enough tier. The differentiator is shifting away from who has the smartest model and toward who can place a good one closest to a real need, at a cost that works, inside something people already use.

That is a distribution game, an economics game, and an infrastructure game. India is simply further along in playing it because it had the rails and the scale to force the issue early.

The frontier still matters. Someone has to build the smartest models, and that work is real and hard. But most value does not get created at the frontier. It gets created in the long, unglamorous work of pushing capability out to where people actually live.

What I am taking back to my own work

I think about identity, CTV, and the infrastructure underneath open ecosystems. So three things stuck with me on the flight home.

Reach is a moat, and it is one you build before you need it, not after. Cost discipline at scale is a feature, not a constraint to apologize for. And the highest form of a technology is the one nobody talks about anymore, because it simply works.

India did not get there by having the best models. It got there by having the best distribution, the right economics, and the instinct to treat AI as plumbing rather than spectacle.

That lesson does not stay in India. It is the one I am bringing home.

I took a trip through Asia recently, India in particular. Two weeks, the usual rhythm, places I have been many times before. By now, I think I know what to expect.

This time felt different. Not because anything dramatic happened, but because of how ordinary AI had become. It was not in keynotes or headlines. It was in the things I touched every day without thinking about them.

A year ago, much of this was either absent or visibly clunky. This time, it just worked, quietly, in the background.

I build platforms for a living. CTV, identity, adtech, developer ecosystems. So I tend to notice the plumbing before I notice the product. And the plumbing in India right now is worth paying attention to.

Payments grew a brain

Everyone knows the UPI story by now. You can pay a street vendor by scanning a code, and India processes more real-time payments than anywhere else in the world. That part is old news.

What felt new this year was the layer sitting on top of it.

The payment apps were not just moving money anymore. They were starting to anticipate. Spending summaries actually made sense. Categorization was right more often than wrong. Prompts showed up in plain language and, increasingly, in the language you actually speak rather than just English.

One app flagged a recurring charge I had forgotten about before I noticed it myself.

The feature itself is not the interesting part. The interesting part is that it sits on top of payment rails that already reach hundreds of millions of people. The intelligence had somewhere to go the moment it was ready.

Ten-minute delivery, run by AI & Ops Efficiency

Quick commerce in Indian cities has crossed into a place that genuinely surprised me. Groceries in ten minutes is not a pitch-deck promise. It is the default expectation now.

You cannot do that with people alone.

What makes it work is an operations layer most users never see: demand prediction down to the neighborhood and the hour, inventory positioned in small dark stores before you have even decided to order, and routing built around Indian traffic, which is its own category of chaos.

The model quality matters, of course. But it matters less than the fact that the whole system is wired together end to end.

This is the part Western observers tend to miss. The impressive thing is not any single model. It is the orchestration.

The voice on the phone is a bot now, and you can barely tell

Customer service was where the change felt most visible to me.

Call almost any service line and you reach a voice agent first. That is not new on its own. What was new was that it could handle multiple languages, switch between them when I did, understand context, and actually resolve things instead of looping me back to a menu.

When the call did need to move to a person, the handoff was clean.

For a country with this many languages and this many people needing service at once, this is not a nice-to-have. It is the only way the math works.

The number underneath all of it

There is one fact that reframes everything above, and it is easy to say and hard to actually feel: this is happening across 1.4 billion people.

That number is not just a big market. It changes the physics of the problem.

When your potential user base is that large, the constraints flip. A feature that helps even a small fraction of users is still helping tens of millions of people, so it is worth building. A model that saves a fraction of a cent per query saves real money at that volume, so efficiency becomes a first-class design goal rather than an afterthought. And because the population spans dozens of languages, hundreds of dialects, and an enormous range of incomes and devices, anything that ships has to work for people on cheap phones and patchy networks from day one. There is no luxury of designing for the well-connected first and everyone else later.

Scale also creates its own gravity. Once a payment rail reaches that many people, it becomes the obvious place to add intelligence. Once a delivery network covers that many neighborhoods, the data to optimize it is already flowing. The ecosystem feeds itself. Each new layer has a ready-made audience and a ready-made stream of signal to learn from.

This is the part that is hard to replicate by copying a single product. You cannot import the population, the rails, or the habits. The scale is the moat.

What I keep coming back to

I left thinking the story we tell about AI is slightly wrong.

The valley conversation is obsessed with the frontier: which model is smartest, which lab is ahead, what the latest benchmark says.

India seems to be having a different conversation, mostly without saying it out loud. The question is not only how good the model is. It is how far it can reach, how cheaply it can operate, how many languages it can support, and whether it can plug into infrastructure that people already use every day.

That is a distribution question and an infrastructure question, not just a model question. And it is the one that ultimately determines whether any of this touches a real person's life.

The most interesting AI I saw in India did not announce itself as AI. It showed up as a payment app noticing something useful, a grocery order arriving in ten minutes, or a voice agent understanding me when I switched languages halfway through a sentence.

It felt ordinary. That may be the point.

In Part 2, I want to dig into why that framing matters and what builders anywhere should take from it.

Because I think the lesson travels.

Every company I've worked at had a launch culture. The countdown. The war room. The launch comms (Google Meet / Slack) channel that lights up the moment the feature flag flips. The all-hands slide with the ship date in bold. For most of my early career, the launch was the finish line. You shipped, you celebrated, you moved to the next thing.

It took me years, and a long list of shipped products, to notice a pattern I didn't like. The products with the biggest launches were rarely the ones that mattered most. And the ones that mattered most often shipped so quietly I can barely remember the day they went out.

Launching is an event. Landing is a process. They are not the same thing, and almost every incentive in our industry points at the first one.

A launch is about you. A landing is about them.

A launch is about the team. The ship date, the demo that finally works on the third try, the moment the thing is out in the world. A landing is about the user. Whether anyone adopted it. Whether they kept using it after the novelty wore off. Whether it moved a behavior or a number that was supposed to move.

You can launch perfectly and never land. I have done exactly that more than once.

The reason the gap persists is that launches are legible and landings are not. A launch has a date, a moment, a clean before-and-after you can put on a slide. A landing happens over months, in retention curves and adoption cohorts and the slow grind of removing friction nobody saw coming. It does not photograph well. Nobody throws a party in month four when a feature finally crosses the adoption line it was supposed to hit on day one.

I optimized for the launch because the launch got recognized

Early on, I treated the ship date as the goal and adoption as someone else's problem. The incentives pushed me there. Career progression and recognition were almost always tied to what you launched, not what you landed. People wore their launches like badges of honor. You listed the thing you shipped, not the thing that stuck.

Early in my career, I worked on a product that was perfect on paper. Years ahead of the industry, better than anything else in the category, and almost nobody on the team thought we could fail. On every internal scorecard, we didn't. What we got wrong was simpler and harder: we never really understood what the customer needed at that moment. We had built something impressive that solved a problem people did not yet feel. The launch was a triumph. The landing never came.

That reset how I think about my own work. The product had been flawless on every dimension I had been trained to care about. On the only one that counted, it never moved.

The products that landed taught me the opposite lesson

The ones that worked were rarely the ones with the cleanest launch moments. They were the ones where the team treated the launch as the start of the real work, not the end of it.

Years later, leading a product area across APAC, I worked on something that crossed a billion users faster than almost anyone expected. From the outside it looked like an overnight success. It wasn't. We knew exactly which KPIs mattered, we tracked them relentlessly, and we knew none of that growth would have been possible without the right partnerships on the ground. The launch was not the story. The landing was.

What that product had was not a better launch. It was a team that never stopped at the launch. We instrumented adoption obsessively. We sat with the friction instead of declaring victory. We shipped the unglamorous follow-ups that nobody noticed but that moved the curve. We earned the landing one quiet release at a time.

This is an incentive problem, not a talent problem

Most orgs staff heavily for launch and thin out right after. The launch gets the war room. The landing gets one PM and a backlog. Performance reviews capture what you shipped, not what stuck. We celebrate the team that hit the date and rarely circle back six months later to ask whether the thing they shipped is still being used.

So smart people keep optimizing for the launch, because the launch is what the organization sees and rewards. They are responding rationally to the wrong scoreboard. The fix is not to work harder. It is to change what we measure and when we walk away.

What changed in how I work

Over the years, the thing that stuck with me is that you cannot call a launch successful on launch day. The KPIs that matter show up at 7, 15, 30, 60, and 90 days. Bounce rate matters as much as sign-ups. Is engagement actually growing, or are you looking at a novelty spike that fades by week three? Those are the questions that tell you whether something landed.

So the question I lead with changed. Not "when do we launch," but "what does landed look like, and how will we know." If I can't answer that before we ship, we are not ready to ship, no matter how good the demo is. The launch date stopped being the goal and became a checkpoint on the way to the only thing that counts, which is whether the product took root.

It also changed how I staff. The instinct is to celebrate the launch team and reassign them the next morning. The better move is to keep the team on the thing through the landing, because the landing is where most of the value and almost all of the learning live.

The launch was never the hard part

I still feel the pull of the launch. Everyone does. The applause is real and the dopamine is real, and there is nothing wrong with celebrating the day the thing goes out. But the work I am proudest of now is not the stuff that launched loud. It is the stuff that landed, often so quietly that no one remembers the day it shipped, because by the time anyone looked up it had already become part of how people worked.

If you lead products, the most useful shift you can make is to move your attention, and your team's, past the launch. Decide what landed means before you ship. Staff for the months after, not just the weeks before. Measure yourself on what stuck, not on what you shipped.

After twenty-one years and four companies, that is the lesson I would go back and give myself on day one. The launch was never the hard part.

A few product conversations with industry peers this month. Different companies, different stages. Same question kept surfacing: are we showing up when AI assistants recommend our category? Nobody had data to answer it.

That just changed.

Two weeks ago, HubSpot shipped AEO Sensor, a free public dashboard tracking citation patterns, answer engine volatility, and AI-referred traffic across ChatGPT, Gemini, and Perplexity. HubSpot's own customer data tells the story. AI-referred traffic now converts at four to twenty-three times the rate of traditional organic search. The same customers have seen organic traffic drop 27% year over year while AI-referred traffic tripled.

That's the desktop story. The CTV version hasn't started being measured yet.

For three years, the honest answer to that question has been "we think so." That isn't a strategy. It's a guess we agreed to call a position.

The dashboard exposes what most companies were guessing at

Open Sensor and pull your category. The first reaction will be predictable: wait, that's our category, why aren't we higher? The next question is harder: who internally owns answering it? In most companies, nobody.

The free Sensor publishes industry-level signals. The paid product goes brand-specific. Either way, the question that was a black box for three years just turned into a chart.

The first pattern jumping out of the data is fragmentation. A brand that dominates Perplexity's citations may be invisible on ChatGPT. A category leader on Gemini may be missing entirely from the others. The "AI answer market" isn't one market. It's three or four parallel markets, and the brands winning each look different.

Most strategy decks treat AI distribution as a binary: visible or not. The reality is you're visible in some engines, absent in others, and most companies haven't been able to see which until now.

And that's the desktop and mobile version. CTV hasn't entered the measurement layer at all yet.

AEO is a product problem before it is a marketing problem

The most common misread I see is that AEO is the new SEO and therefore belongs in marketing. That framing is right in the medium term and wrong in the short term, and the short-term call is the one that matters.

In the short term, AEO is a product problem. The reason a model cites your brand is downstream of how your product is structured: documentation, changelog, API references, positioning copy, developer relations footprint. The model isn't picking favorites. It composes the best answer it can from the public corpus, and your presence in that corpus is determined by product decisions.

Companies treating AEO as content marketing will spend two years cranking out answer-optimized blog posts and move their citation rate a few percentage points. Companies treating it as a product problem will restructure their docs, APIs, pricing pages, and integration guides for model ingestion, and they'll move their citation rate by an order of magnitude in the same window.

Moving the needle on this means pulling the documentation refactor onto the product roadmap before marketing gets involved. The teams still waiting for content to fix it are already behind.

AEO Sensor is missing half the map

Worth flagging because most of the launch coverage glossed over it. AEO Sensor covers ChatGPT, Gemini, and Perplexity. Claude isn't in the dashboard. Microsoft Copilot isn't tracked. Meta AI isn't tracked. Those three engines together handle a meaningful share of the buyer queries AEO is trying to measure, and their absence in the tooling will create a misleading picture for anyone treating this instrumentation as comprehensive.

That gap is the story. The measurement layer for AI distribution is going to be as fragmented as the engines themselves. A brand could look strong on Sensor and be invisible on Claude, or missing from the tooling entirely and dominant inside a Copilot enterprise install base. If you're an executive reading a single dashboard as ground truth, you're going to make bad calls with incomplete data.

CTV is the next AEO battleground, and nobody is building for it

The AEO conversation today is shaped by what happens on desktop and mobile. A buyer asks ChatGPT, the answer surfaces brands, citation patterns get measured. That's the first surface, not the whole one.

CTV is going to look fundamentally different. When a viewer asks the TV's built-in assistant what to watch tonight, that's AEO for content. When a smart home assistant routes a purchase decision based on a brand integration in the show that just played, that's AEO for products. When AI agents browse across catalog surfaces on behalf of viewers, that's AEO at a different scale entirely.

The teams I'd bet on in CTV are already thinking about what AEO looks like on a voice query to the TV, or an agent operating inside the OS layer. The measurement tooling for those surfaces doesn't exist yet. The companies that build it will define AEO on every screen in the home for the next decade.

Most teams are about to discover their internal story doesn't match the data

The first reaction inside most product orgs is going to be uncomfortable. Teams confident about their AI distribution position will find the engines don't rate them where their internal narrative says they should. Teams that assumed they were invisible to AI users will discover surfaces where they're actually stronger than they realized.

I've been having a version of this conversation with product leaders all month. The second reaction is always the same. Who owns this internally? Usually nobody. The function hasn't been built.

Either way, the gap between internal narrative and measurable reality is going to be the most consequential strategic input of the next year. When distribution shifts overnight from "unknown" to "public dashboard," any executive review that ignores those numbers has a hole in it.

My estimate is twelve months

Categories that move from unmeasurable to measurable don't stay competitively interesting for long. SEO took five years to commoditize once analytics made organic traffic universally trackable. Programmatic attribution took roughly three years. AEO will commoditize faster, because the playbook is partially understood, the AI talent layer is in place, and the underlying technology is built on the same foundation models every product team is already using.

My estimate for the window to build real competitive advantage on AEO is twelve months. After that, every competent product organization will have instrumentation, a defined function, and a quarterly review process. The advantage today is being one of the few teams that has pulled the data, looked at it honestly, and started making product decisions around it.

What I would do this quarter

The decision-forcing question is short. Pull your AEO Sensor data this week. Compare it against what your internal narrative says about your AI distribution. If the two don't match, that gap is your strategic input for the next planning cycle.

Then three moves. Name an owner: a single point of accountability inside product, not vague joint ownership with marketing. Audit the public corpus the models are ingesting (documentation, changelog, copy, third-party reviews). Find what's outdated, missing, or wrong. Instrument AEO as a quarterly metric tied to roadmap decisions, not a one-time audit.

The gap between the companies that move and those that don't will be visible by year-end. The right way to treat this is a six-week sprint, not a year-long initiative. The dashboard is public. The data is comparable. The window is short.

The companies that don't look at it this quarter are likely to find, in twelve months, that their competitors did.

Nearly 2 decades in big tech and startup, and I've watched a handful of these resets up close. Mobile rewrote the consumer playbook. Cloud broke enterprise pricing. Streaming dismantled the bundle. Each one felt like the end of something at the time, but most turned out to be evolutions of a model that still recognized its old self.

This one isn't.

The playbook I learned at Google and Netflix, the one most companies still hand to new PMs, is getting hollowed out fast. Not "shifting." Not "evolving." The actual mechanics of why products win or lose are changing, and the people thriving in 2026 are the ones who noticed early.

Most leaders accept the product strategy is changing. They're not accepting that the product leader executing it is changing in parallel. You can't fix one without the other, and companies trying to are about to learn an expensive lesson.

Your moats are getting repriced

Strategy decks have leaned on the same four words forever. Distribution moat. Data moat. Network effects. Switching costs. For most of my career those words meant something, because the economics behind them were real. Buyers were hard to reach. Proprietary data was rare. Pulling a customer off your platform took quarters of engineering plus a procurement cycle.

Take distribution. I have seen companies spending enormous money on sales, SEO, partnerships, content, paid acquisition. The logic was clean: whoever fights hardest for attention wins. That was true when attention was acquired through search engines and inside sales calls. It's a lot less true when buyers start by asking ChatGPT, Claude, or Gemini. If your product isn't on the shortlist the model returns, your distribution budget is funding a channel that's getting routed around.

Data was supposed to be the most durable moat in software. Then foundation models trained on the open internet collapsed the marginal value of anything that wasn't structurally unique. The data moats that still hold up have one of three things going for them: a closed feedback loop, regulated access, or behavioral signals you can't reasonably infer from public sources. Without one of those, what you have isn't a moat. It's a CSV with extra steps.

Switching costs are softer than they look too. Migrations that used to take quarters get done over weekends now because AI-assisted engineering compresses the integration work. The lock-in that justifies your renewal motion is being quietly unlocked by the same technology your roadmap is trying to use.

The question I keep landing on isn't "how do we add AI." Everyone's asking that and it isn't interesting. The interesting question is: which of our advantages only existed because intelligence used to be expensive? That answer is what's actually getting repriced. In most strategy meetings I've sat in this year, that question gets very carefully avoided.

Timing is doing more work than most teams admit

You can get the moat question right and still build the wrong thing if you read the era wrong.

Vine launched in 2013, was beloved, viral, and Twitter shut it down four years later. TikTok rebuilt the same idea in 2018 with the AI-driven feed, mobile bandwidth, and creator infrastructure Vine never had, and is now one of the most valuable consumer platforms in the world. Quibi entered the same space in 2020 with $1.75 billion and Hollywood-grade content, and folded in six months because TikTok, YouTube Shorts, and Instagram Reels had already eaten the market. Same idea, three eras, three completely different outcomes. The variable wasn't the customer or the team. It was the moment.

Jasper is the AI-era version. Raised $125M at a $1.5B valuation as a polished writing tool sitting on top of GPT-3. ChatGPT shipped into the same use case weeks later. The moat was gone before the Series A had cleared the bank.

A great AI product in 2022 turns into a commodity feature by 2026. Entire categories now appear and saturate inside a single planning cycle. The leaders who win consistently hold three things in their head at once: what just became possible, what just became commoditized, and what hasn't moved at all. Most teams obsess over the first. The expensive mistakes come from missing the second.

The product leader is changing too

The role itself is shifting under us.

For about fifteen years, product management was essentially a translation layer. Engineering, design, business, ops, legal, sales: all had their own dialects. The PM sat in the middle, gathered context, wrote things down, aligned stakeholders, and kept the machine moving. The role assumed context was expensive to collect and synthesis was a specialized craft.

Neither assumption is true anymore. Anyone on the team can generate a competitive analysis, summarize feedback, draft a roadmap, or prototype an interface in an afternoon. If your value was the synthesis layer, you're being routed around the same way buyers are routing around old distribution channels.

What's left is judgment. When five teammates can each generate a credible-looking analysis with different conclusions, the most valuable person in the room is the one whose call you trust. That comes from reps, taste, scar tissue, and a willingness to be wrong in public. Not prompts.

The technical bar has moved too. I'm not saying every PM needs to be an engineer. I'm saying the line between "technical enough" and "not technical enough" has moved a long way. The PMs gaining influence aren't writing decks. They're building working prototypes with AI coding tools and putting a functional thing in front of stakeholders by Wednesday. A prototype communicates conviction in a way a deck never does.

The senior product leader in 2026 is doing work that used to be split across six different roles:

• Product manager, framing the problem and scoping the bet
• UX designer, mocking up the experience
• UX researcher, pressure-testing assumptions with users
• Engineer, building a working prototype
• Data analyst, pulling the signal that justifies the call
• And increasingly, some data scientist work, designing experiments and reading results

None of those crafts have disappeared. They're still real disciplines at scale. But the senior leader who can credibly do enough of each to move a decision forward by Friday, without convening a four-person meeting first, is operating at a different altitude than the one who can only coordinate. The bar moved a couple of years ago. It's just becoming impossible to ignore.

Power has shifted in lockstep. When an exec can use an AI agent over a weekend to analyze metrics, draft positioning, and put together a roadmap alternative, the Monday review stops being a presentation and becomes a negotiation between competing AI-assisted visions. Authority that used to come from sitting on top of the information flow doesn't really exist anymore. What's left is judgment when the visions diverge.

Both halves or you lose

Most companies are doing half of this work, confidently. Some are rolling out AI tools aggressively while leaving their assumptions about competitive advantage untouched. Others are rewriting strategy decks while leaving the organizational model exactly where it was in 2021. Either feels productive. Neither is enough.

The leaders I'd bet on are doing both at once. They're asking which moats actually survive a repriced market, and which parts of their own playbook are now decorative. They're not waiting for permission to reposition the product or themselves.

The product leaders who'll matter next decade won't be differentiated by knowing how to use AI tools. Everyone will. The differentiation comes from rethinking, almost from scratch, what a product leader is for in a world where intelligence, synthesis, and execution have all gotten dramatically cheaper.

Timing matters more. Judgment matters more. So does the ability to keep updating your own model of what the job even is.

The playbook changed.

The harder question is whether we're willing to change with it.

The app grid on your smart TV has not meaningfully changed since 10+ years. It is about to be obsoleted, and the industry that built it is watching the wrong signal.

Reports this past week indicate that OpenAI is partnering with Qualcomm, MediaTek, and Luxshare to ship a smartphone designed around AI agents instead of apps, with mass production reportedly targeted for early 2027. Sam Altman wrote publicly that "it feels like a good time to seriously rethink how operating systems and user interfaces are designed."

That sentence will be quoted a thousand times this year in smartphone analysis. Almost no one is applying it to connected television.

They should be.

I have spent several years close to the platform layer of CTV, and what struck me about the OpenAI announcement is how directly it maps to a structural question the industry is about to face. The smartphone world is being challenged to rethink its interface model. CTV will face the same challenge, and it will face it with higher stakes and a faster timeline than most product roadmaps account for.

The App Grid Was Built for a Simpler Era

Boot a smart TV today and you see what you would have seen a decade ago. A grid of streaming tiles. A featured row. A few curated recommendations. The user picks an app, searches inside it, watches.

This model worked when the underlying assumption held: the user knew which app had what they wanted, switching cost was acceptable, and the platform's job was to make the grid easy to navigate.

Those assumptions are collapsing. The average US household subscribes to four streaming services with access to several more through ad-supported tiers. The user no longer knows which service holds which content. Switching cost is measurable in seconds of lost engagement and minutes of session abandonment. The app grid was a sensible interface when the content library was simple. It is a hostile interface when the user has to mentally map thousands of titles across a dozen services.

The Agentic Shift Hits CTV Harder Than Mobile

The OpenAI phone is provocative because it challenges fifteen years of iPhone interaction design. But mobile already has search, voice assistants, and deep linking. The change is real but evolutionary.

CTV is structurally different in ways that make the shift hit harder.

The input is constrained. On mobile, you can type. On CTV, every interaction passes through a directional pad and a few buttons, or through voice. The friction of navigating an app grid on a television is dramatically higher. Agentic interaction on CTV is not a luxury improvement. It is a release of decades of accumulated friction.

The content fragmentation is worse. CTV apps compete for the same minute of viewing time with massively overlapping libraries and inconsistent licensing windows. An agent that knows where the content lives can route the request without making the user solve a puzzle.

The session economics are different. The average CTV session is over an hour. Every second spent navigating the grid is a second the platform is not monetizing and the user is one second closer to walking away.

The household model breaks the grid. Multiple users share the same TV with different preferences, different subscriptions, and different intent at different times of day. An agent that adapts to the household and the moment is structurally better suited to the way the screen is actually used.

When agentic AI hits CTV, it will not feel like a feature upgrade. It will feel like a reset of how the living room works.

What Actually Changes When the Agent Becomes the Interface

The agent does not eliminate apps. Streaming services still deliver content, manage rights, and own their relationship with the viewer. What changes is the interaction layer above them.

When the user says "show me something light and funny, around 30 minutes" and the agent surfaces three options from across the household's available services with one-tap playback, the discovery experience has fundamentally moved out of the app and into the platform.

For streaming services, the home screen is no longer the primary acquisition surface. Catalogs become more important. Brand recognition becomes less important. The agent decides which service handles the request, and the user often does not see or care which service they ended up on.

For OEMs and CTV operating systems, the platform's job has changed. The grid is no longer the value proposition. The agent is. The platforms that ship the most useful agent, with the deepest integrations into services, the household's identity, and the moment of intent, will define the experience. And this is not something that can be bolted on after the fact. The platform must own the agent layer architecturally rather than cede it to individual apps. If the agent the user interacts with most is shipped by a streaming service, the platform has lost its position.

For advertisers and DSPs, cross-app behavioral signals become more accessible because the agent sees them. Ad inventory becomes more contextual. Frequency and attribution become cleaner because the household identity sits above the app, not inside it.

The infrastructure this requires is identity. An agent-mediated CTV experience only works if the platform has clean, cross-service identity and measurement at the household level. Without that, the agent cannot personalize across apps, cannot deduplicate across services, and cannot give advertisers the signal they need. This is the hard technical problem most of the industry is hand-waving past. Open identity standards are not a nice-to-have in the agent era. They are the load-bearing infrastructure that makes the entire model function.

This Is Already Happening

The version of this future most of the industry is debating is the dramatic one. A complete replacement of the app grid with a voice-first ambient interface shipped by a single platform.

The version actually happening is more gradual and more structural.

OS-level assistants are getting more capable across every major CTV platform. Cross-app search is becoming standard. Content metadata standards are maturing. Identity infrastructure is getting cleaner. AI-driven recommendation is moving from in-app surfaces to OS-level surfaces. LLM-based natural language interfaces are quietly being integrated into remote control voice search.

None of this is announced as a paradigm shift. Each piece feels like an incremental product release. But taken together, they are rebuilding the CTV interaction model into something agentic in everything but name. The building blocks are already on the board. The open question is whether any platform will assemble them into a coherent experience before the user stops waiting.

By the time OpenAI ships a phone with no app grid, the CTV industry will already have moved most of the way toward an agent-mediated interface. It just will not have called it that.

Where This Lands

The CTV industry has spent fifteen years optimizing a model inherited from mobile. The agent shift is the first time since the original smart TV operating systems shipped that the entire interface assumption is being rewritten.

The platforms and services that recognize this early will define the next decade of how households consume content. The ones that keep optimizing the app grid will find themselves running a perfect version of a product the user no longer interacts with.

The living room is the largest screen in the home, the most valuable advertising surface in digital media, and the next major interface to be rewritten by AI.

The app era is not ending tomorrow. But it is ending. And the platforms that start building for the agent layer now will own the surface that replaces it.

Most of the global streaming conversation is still optimized for a world that does not exist for the next billion users.

That world assumes broadband connections, credit cards, ample data plans, large screens, English interfaces, and cultural expectations shaped by Netflix, Disney, and HBO Max. It is the world established platforms were designed for and where the unit economics actually work.

The next billion users live somewhere different.

I learned this the hard way years ago. At Google, I worked on the next billion user problem from two angles. On Google Media TV Platforms and Last Mile Connectivity, the question was how to extend connectivity into homes that had never had it. On Google's Communications Apps suite, the challenge was building products that worked on slow networks, low-end devices, and across the language and cultural diversity of markets like India, Indonesia, and Brazil. Both projects forced a hard lesson. Products designed for mature markets do not translate. They have to be reimagined from the network layer up.

That lesson applies directly to streaming today. The markets driving the next decade of streaming growth are not California or London. They are India, Indonesia, the Philippines, Vietnam, Bangladesh, Nigeria, and the rest of the global south. The platforms that win there will not be the ones with the deepest content libraries. They will be the ones whose AI infrastructure is built for the realities of those markets.

The Numbers Tell the Story

The scale of the opportunity is not subtle.

Asia-Pacific screen industry revenues are projected to reach $196 billion by 2030, with all net growth coming from online video, expanding at 7% CAGR. India alone is projected to grow at 26% by 2029, China at 23%, and Japan at 15%.

But the headline numbers hide the real story. The composition of that revenue is fundamentally different from what established platforms are built around.

Advertising-funded video is now the single biggest accelerator of digital revenue in APAC. AVOD is forecast to grow 69% over the next five years and FAST revenues are projected to more than double, together generating over $25 billion by 2030. Ad-supported revenues are set to grow from 23% to 30% of total streaming revenue across the region.

In India specifically, CTV adoption hit 129.2 million users in 2025. That is an 85% year-over-year surge, with CTV ad spending racing toward $1 billion by 2026. 21% of India's 601 million OTT audience now uses CTV to stream, up from 13% in 2024.

Two things stand out. First, ad-supported is not the budget option in these markets. It is the default entry point. Ad-free is being repositioned as the premium upgrade. That is the inverse of how most established platforms still frame the choice. Second, adding 35 million new CTV viewers in a single year breaks any product roadmap built on mature-market assumptions.

What Established Platforms Keep Getting Wrong

Most international platforms enter these markets with one of three strategies. All of them underperform.

Strategy one is to drop the established product into the new market with a discount. Same UX, same content mix, lower price. This fails because the product was never designed for the constraints of the market.

Strategy two is to license local content and call it localization. Better, but still surface-level. Viewing patterns, monetization expectations, and engagement loops in these markets are structurally different. Content alone does not fix that.

Strategy three is to bundle the service into a telco mobile plan. This works for distribution but creates dependency on a single channel partner and does not solve the deeper product problem.

The platforms that will actually succeed are doing something more fundamental. They are rebuilding the product around AI as the personalization, adaptation, and monetization layer that makes the experience work in environments established products were never designed for.

The Platform Layer Is Where the Real Race Is

The real battle is not between streaming services. It is between the platforms underneath them.

Most of the conversation about emerging market streaming focuses on the services. Which app wins. Which content travels. That focus misses where the actual competition is shaping up.

Underneath every streaming service sits an operating system, an identity framework, a payment infrastructure, and an app distribution model. These layers determine what services can do, how they monetize, what data they have access to, and how the experience actually feels. In mature markets, the platform layer was largely set a decade ago. In emerging markets, it is still being written.

Three things make this moment uniquely high-stakes.

OEMs are competing for default position on hundreds of millions of new TVs. Whoever owns the operating system controls the home screen, the discovery surface, and the data flowing through it for the next seven to ten years. That control translates directly into discovery economics and the data leverage that powers everything downstream.

Identity infrastructure in these markets is fundamentally different from the cookie-based open web or the device-ID model. National digital ID systems, telco-issued identifiers, and locally-built authentication frameworks are filling the gap. Platforms that integrate cleanly with these layers will have structural advantages in measurement, attribution, and monetization that streaming services on top of them cannot replicate alone.

Payment rails are being built natively into platform layers. UPI in India enables micropayments at a granularity Western payment infrastructure cannot match. GoPay and OVO in Indonesia, GCash in the Philippines, similar systems across Vietnam and Bangladesh. When these rails sit at the OS level rather than inside individual apps, they unlock conversion and ARPU patterns that look nothing like subscription economics in mature markets.

The behavioral patterns that go with these rails are equally distinctive. India's cricket season generates concurrency spikes that routinely break infrastructure designed for Western viewing patterns. Southeast Asian users default to prepaid and sachet pricing because monthly subscriptions assume credit relationships most households do not have. Offline downloads and aggressive data-saver modes are not features in these markets. They are baseline product expectations.

This is the layer where leading CTV platform providers, and a handful of emerging local players are competing right now. Not for users. For the right to be the default platform on the next half-billion connected TVs.

Why AI Is the Unlock

In mature markets, AI in streaming is mostly an optimization layer. Better recommendations. Smarter ad targeting. Personalized thumbnails. The product works without it.

In next-billion-user markets, AI is not optimization. It is enablement.

Network adaptation in real time. Connection quality varies dramatically across the day in markets where 4G is shared, congested, or intermittent. AI-driven adaptive bitrate streaming, prefetching, and edge caching turn an unreliable network into a viewable experience. Much of this has to live at the platform layer, not inside individual apps.

Language and cultural personalization at scale. India alone has 22 official languages. Indonesia has over 700 living languages. AI-driven multilingual discovery, automated dubbing, and culturally aware recommendation are the only way to make a single platform work across this diversity.

Payment and monetization personalization. Credit card penetration in many of these markets is below 5%. Operators have responded with sachet pricing, telco bundling, and prepaid models. Running this complexity manually is impossible. AI determines which payment option, which price point, and which offer to present to which user in real time.

Data efficiency. Users watch on capped data plans. Predictive prefetching, intelligent compression, and AI-driven encoding allow viewers to consume hours of video on the data budget they actually have.

Content discovery without a content library. Most next-billion users do not have a content library at home. They discover through social media, word of mouth, and short clips. AI-driven discovery engines that surface the right content within seconds of a session start are what create engagement.

Each of these is a hard product problem. Solving them with traditional engineering would require massive headcount in every market. AI is what allows a single platform to scale across geographies without rebuilding the product from scratch.

What This Means Practically

For platform leaders thinking about emerging markets, a few practical implications matter.

AI investment is not optional. It is the precondition for the product to function at all. Building for India, Indonesia, or the Philippines without AI as a first-class capability is building a product that will be outcompeted.

The AI you need is not the same AI mature markets use. Recommendation engines built around a few dominant content categories do not work on a multilingual catalog spanning dozens of regional languages. Ad targeting models trained on cookie-era signals do not work where identity infrastructure is different. The AI infrastructure has to be built for the operating context.

Partnerships matter, but probably different ones than most platforms prioritize. Telco bundling is useful but limited. The deeper partnerships are with local payment providers, regional content producers, identity operators, and the OEMs and CTV operating systems that are setting platform-level defaults.

Timing matters too. Operators across APAC are focused on consolidation rather than expansion. The platforms making real product investments in AI infrastructure right now are doing so when the competition is paused. That window will close.

The Lesson From Google

When I worked on Google Media Platform and the Communications Apps suite, the most important lesson was that products built for the median user in a mature market simply do not work at the edges where the next billion users actually live. The constraints are different. The mental models are different. The trade-offs that engineering teams take for granted in San Francisco do not apply when your user is on a 2G connection in a tier-three Indian city or watching on a shared device in a kampung in Indonesia.

The same lesson applies to streaming today, with one important difference. AI now exists in a way it did not when we were building these products a decade ago. The personalization, adaptation, and optimization that used to require massive teams of engineers and operations staff can now be coordinated by AI infrastructure that scales across markets without scaling headcount linearly.

The platforms that understand this and invest in AI as the operating layer for emerging markets will define what streaming looks like for the half of the global population still being added to the digital economy.

The next billion users are not an extension of today's streaming market.

They are a reset.

And the companies that win will not be the ones that scale their existing products globally. They will be the ones that rebuild the stack for a different reality.

Every major technology platform eventually reaches a second S-curve.

The first curve builds the category. It proves the thesis, establishes the business model, and creates the scale that makes the market worth pursuing. The second curve changes the basis of competition. It does not replace what came before. It builds on top of it, using new capabilities to unlock value that was structurally out of reach in the first phase.

CTV is entering that second curve now.

I have watched this shift up close over the last 2 years, and the change feels more fundamental than most of the industry conversation suggests. It is not a tooling upgrade. It is a rewiring.

The first generation of smart TV platforms did something remarkable. It turned a stagnant hardware category into a scaled distribution engine for streaming, advertising, and data. A handful of platform operators built variations of the same playbook: sell hardware at thin or no margin, own the operating system, and monetize the platform over the life of the device.

For a decade, that model helped define the modern connected TV industry.

It worked because the mission was clear. Get streaming into every household. Move advertising budgets from linear to digital. Make television programmable.

That mission is largely accomplished.

What comes next is not a replacement for the first generation of CTV platforms. It is the next layer on top of it. The platforms that win from here will be AI-native in how they operate, how they monetize, and how they participate in the broader ecosystem.

The Requirements Have Changed

The first era of CTV was built around a specific set of platform requirements: scale the installed base, build an advertising marketplace, capture a share of streaming consumption, and establish a meaningful data position.

Those were enormous challenges, and the platforms that solved them earned the positions they hold today.

But the requirements for the next phase are different.

The market is no longer mostly greenfield. Penetration is high in mature markets. Advertising inventory is abundant, not scarce. Streaming apps have built their own direct viewer relationships, their own ad businesses, and increasingly their own measurement environments. Viewers are loyal to apps, not to device ecosystems. Privacy frameworks and identity shifts have also changed what "owning the viewer" really means.

I run a quiet test whenever I meet someone outside of streaming and adtech. They can name five streaming apps before they can name the TV they are watching on. That gap tells you everything about where the viewer relationship actually lives now.

None of that signals weakness in CTV. It signals maturity.

But maturity changes what matters.

The next era does not need more scale for its own sake. It needs platforms that can operate intelligently across a fragmented, fast-moving, interoperable ecosystem where value moves between apps, publishers, advertisers, and viewers at speeds no human organization can coordinate manually.

That is the shift AI-native infrastructure makes possible.

What AI-Native Actually Means

AI-native is an overused phrase, so it is worth being precise.

In CTV, AI-native does not mean adding a recommendation engine. Every platform already has one.

It does not mean putting a chatbot on the home screen. That may be useful, but it is not the real transformation.

It does not mean sprinkling machine learning into the ad stack. That has been standard practice for years.

Honestly, the amount of product marketing that slaps "AI-native" onto feature releases right now is a problem. It obscures what the real shift actually is. I find myself correcting this in almost every meeting lately, which is why I want to be careful with the term here.

AI-native means intelligence is built into the operating model of the platform itself. Decisions are made continuously across layers. Signals from hardware performance, viewing behavior, app engagement, ad response, and ecosystem partners are not analyzed in isolated workflows. They are fed into systems that can coordinate them in real time.

The platform does not just collect signals. It learns across functions, adapts continuously, and optimizes the system as a whole.

That is the practical difference.

A legacy platform treats hardware, platform, and data as separate operating domains, each with its own team, tools, incentives, and reporting cadence. An AI native platform treats them as one coordinated system, where intelligence moves between them continuously.

That matters because more and more of the value in CTV now lives in the connections between those layers, not inside any one layer by itself.

Three Layers, One Intelligence

The clearest way to understand the second era of CTV is to look at the three major platform layers and what changes when intelligence becomes the connective tissue between them.

Hardware as efficient distribution

In the next era, hardware remains essential, but its role becomes clearer.

The television is the most effective way to get platform software into the living room. That means the hardware business should be optimized around distribution: scale, retail efficiency, speed to market, and global footprint.

AI plays an important role here, even if it is mostly invisible to consumers. Demand forecasting, regional SKU optimization, dynamic sourcing, and supply chain intelligence all help hardware businesses operate efficiently at very low margin while preserving capital for the platform layers above them.

This may not be the glamorous side of AI, but it is foundational. Efficient distribution is what makes the rest of the system economically sustainable.

The platform layer as shared intelligence for the ecosystem

This is the biggest unlock of the second era.

The platform layer no longer needs to win by owning every viewer interaction end to end. It wins by becoming the place where the ecosystem performs better.

That means cross-app discovery that helps viewers navigate a fragmented streaming landscape. Recommendation systems that work across services, not just inside closed walls. OS-level assistants that can surface, compare, and eventually transact across multiple apps. Advertising and monetization systems that help publishers and advertisers produce better outcomes than they could achieve off-platform.

This is where AI native architecture becomes a real differentiator.

The clearest analogy is what happened in mobile. Developers did not flock to iOS because Apple forced them. They flocked because the platform made their apps better. CTV is at a moment where the same dynamic is available, if the platform layer chooses to earn it rather than compel it.

An OEM platform whose intelligence improves outcomes for apps, publishers, advertisers, and users creates gravity. Apps want to be there because they perform better. Publishers want to be there because monetization works better. Advertisers want to be there because the signals are cleaner and the outcomes are easier to measure.

That is a very different platform thesis from trying to capture every impression, every workflow, and every relationship for the OEM itself.

The identity and data layer as interoperable intelligence

This is the most counterintuitive shift, and perhaps the most important one.

In the first era of CTV, the instinct was to treat device-level data as a proprietary moat. In the next era, the value of a data position depends less on exclusivity and more on usefulness inside a broader ecosystem.

Signals trapped inside a single platform become less valuable over time. Signals that can participate cleanly in identity frameworks, privacy-safe collaboration environments, publisher partnerships, and measurement systems become more valuable.

AI is what changes the economics of that layer.

Raw viewer signals are becoming easier to collect and harder to differentiate. What matters is the intelligence derived from them: signals around content affinity, churn risk, ad receptivity, engagement patterns, and cross-device behavior.

That is what publishers value. That is what advertisers value. That is what creates durable strategic leverage.

The platforms that move earliest toward AI-enriched, interoperable identity infrastructure will likely benefit disproportionately, not because they own the most raw data, but because the intelligence flowing through them is the most useful to the rest of the ecosystem.

Here is what I keep coming back to. The platforms that are most defensive about protecting first-era advantages are the ones most at risk. The ones already moving, even if it means cannibalizing some of what they built, are the ones that will still matter in five years.

Why This Is Not a Zero-Sum Transition

The second curve of CTV is not a rejection of the first.

The first generation of platforms built the installed base. It proved the business model. It created the distribution scale that makes the next phase possible. Those advantages do not disappear.

What changes is the center of gravity.

The first era was defined by distribution and scale. The next era will be defined by intelligence and interoperability.

The platforms best positioned to win are not necessarily new entrants starting from zero. More likely, they are the platforms that successfully evolve their architecture while preserving what the first era got right.

The risk is not having first-era strengths. The risk is overprotecting them.

Platforms that continue to rely on closed capture models, over-tax ecosystem participants, or treat the home screen as the primary source of durable leverage may find those advantages eroding faster than expected. But the ecosystem now has more ways to route around friction than it did in the first phase of the market.

That is the pattern mature platform markets tend to follow. Early winners build the category. Later winners redefine how value is created inside it.

CTV is arriving at that moment now.

The Shift That Matters

The change underway in CTV is less about which company wins and more about how platforms need to operate to stay relevant.

The first era rewarded endpoint control. Owning the screen, the home row, the default app, the billing relationship. That posture worked when the market was still being built and distribution was the scarce resource.

The next era rewards something different. Ecosystem gravity. The ability to make apps, publishers, advertisers, and viewers all perform better by being on the platform rather than around it.

This is not a clean binary. Most platforms will sit somewhere between the two. But the direction of travel is clear, and AI is what makes the second posture structurally possible.

Without AI-native infrastructure, coordinating value across a fragmented ecosystem at machine speed is too complex for traditional software and organizational structures. With it, the model starts to work. The platform becomes more useful as more participants build through it. The data becomes more valuable as it participates more intelligently and more openly.

That is what second-curve advantage looks like. Not a win condition for any one platform. A new set of design principles for the category as a whole.

The Front Door Still Matters

The OEM business was never really about televisions alone. It was about owning the front door to the living room.

That front door still matters.

The installed base built in the first era is real. The distribution advantage is real. The position is real.

But the lock on that door is changing.

In the first decade of CTV, the lock was distribution. Whoever reached the living room first had the advantage.

In the next decade, the lock will be intelligence.

If I had to bet my product thesis on one thing in CTV right now, it would be this one. The platforms that understand it are already moving. The ones that do not will keep optimizing the wrong variables until the market has quietly moved past them.

The platforms that can coordinate hardware, platform, and data into a coherent AI-native architecture, while participating openly and effectively in the broader CTV ecosystem, will define what the next era of connected TV looks like.

Everyone else will keep running first-era playbooks and wonder why the market moved on.