Read more

In February of 2023, I was visiting my employer’s corporate headquarters, having the same conversations with the corporate P&L owners that operators in media companies across the globe were having at the same time. How can AI save us money? When can we have a completely automated supply chain? They were the wrong questions at the wrong time, and the industry has spent a meaningful amount of time getting situated since.

The pressure has only intensified. The era of celebrating monthly active user (MAU) growth at any cost is over. Boards and Wall Street are asking when each platform turns profitable, what the unit economics actually look like, and how the cost structure bends from here. Streaming is being forced to mature, fast, against a clock the industry leaders did not get to set. AI was supposed to be part of the answer to that pressure. Three years in, it has not been.

The conversations I am part of with networks and studios are still circling problems we were trying to solve a decade ago. Dirty metadata. Rights data trapped in spreadsheets and email threads. Library content under-leveraged because nobody is sure what the company owns. Cross-functional gaps between content operations, programming, acquisitions, and marketing that keep titles stuck in the same places they were stuck five years ago. The supply chain is still mostly held together by people doing heroic manual work to make systems talk to each other.

For a while it looked like AI might be a shortcut to those problems. What it has turned out to be in the short term is a monkey wrench in the conversation. There are three reasons for that. Different companies use AI in different ways for different goals, so there is no universal playbook. AI is not a bolt-on to existing workflows. And streaming operations are particularly complex, which means there is a lot to break down before anyone can figure out how to apply this technology to create meaningful value.

The work in front of the industry is not throwing AI into existing operations to cut costs or shortcut existing problems. It is rethinking systems from zero. Disney established the Office of Technology Enablement in late 2024, a hundred-person organization that reports to the Co-Chairman of Disney Entertainment and coordinates AI and emerging tech across film, television, theme parks, and streaming. Sony’s AI division, founded in 2020, operates as a strategic R&D unit reporting to the Group CTO and CEO, with cross-functional work across Pictures, Music, Gaming, Electronics, and Semiconductors.

Both models share a structural decision worth noticing. The AI organization is centralized for coordination and embedded for execution. It does not own the AI projects of every business unit. It makes sure those projects fit a coherent enterprise strategy. That requires examining the supply chain as it works today, examining where AI capabilities are today and where they will be in eighteen months, and building the connective tissue between those two pictures.

MIT NANDA’s July 2025 study of three hundred enterprise AI deployments found that ninety-five percent delivered no measurable P&L impact. Sinch reports that seventy-four percent of enterprises have already rolled back a live AI agent. Gartner projects that sixty percent of AI projects will be abandoned through 2026 because the underlying data is not ready, and that forty percent of agentic AI projects will be cancelled by the end of 2027.

There are five pressures driving this in media, and they compound on each other.

The governance conversation

Before any of the technology conversations can produce a result, somebody has to answer the ownership questions. Who owns the data? What can be sent to a model? What cannot? Who decides? What gets logged? What gets audited? What happens to a model that has been trained on the company’s catalog when the contract with a vendor ends? What rights apply to the outputs?

These questions stall initiatives because the answers are not technical. They are operational, legal, and often deeply political. The COO, the CTO, the General Counsel, and Finance have to agree about a category of risk they have never previously had to govern together, and that alignment takes time the profitability clock does not allow for. So a lot of companies are short-circuiting the conversation and discovering the cost later.

Underneath the governance conversation sits a principle I keep finding myself reaching for. Metadata is the quantifiable, trackable representation of content. When a company cannot manage its metadata to the specifications of its business, it leaves a lot of power on the table, often outsourcing it to vendors who promise they can fix today’s uncomfortable problems quickly. Over time, content gets harder to merchandise, harder to localize, harder to monetize, harder to recommend, harder to defend in a rights dispute, and harder to mine for downstream revenue. Every operational shortcut taken at the metadata layer becomes a tax the business pays every quarter until somebody pays the cost of fixing it by somehow buying back the flexibility of their own data at a premium cost.

That principle is where a lot of AI vendor relationships start to wobble. The vendor takes more control of the customer’s metadata than the customer realized they were giving up. The data model gets shaped by the vendor’s product architecture instead of the customer’s business.

The drift that follows is the part worth understanding. Vendors optimize for the average client. A serious operator running an enterprise global catalog is not the average client. The vendor’s roadmap will drift toward whatever is easiest for them and useful for everyone else, and eighteen months later the customer is trying to negotiate back access to data they used to own. I have watched companies hand over the keys to their own metadata for a short-term solution, then spend the next year and a half trying to claw back control. The vendor by then has a roadmap. The customer by then has a dependency.

There are vendors solving for this structurally. Twelve Labs is one of the more interesting examples. Their public positioning is that the entire intelligence stack deploys where the customer wants it to, with SOC 2 Type II certification and encrypted data handling as the baseline. That is structurally different from the standard AI vendor posture, which is “send us your data and trust us.” It is proof that the problem is solvable, and the vendors who have chosen not to solve it have made that choice for their own reasons.

Vendor instability makes the math harder

The AI vendors pitching media operators in 2023 are not pitching the same product in 2026. Many have shifted their model focus. Some have changed their pricing structure entirely. Others have exited features they were demonstrating two years ago. Some have been acquired. A few have stopped existing.

This is the consequence of technology lifecycles compressing. What was state of the art twelve months ago is two model generations behind today. For a vendor competing in a market that resets every six months, holding the same product strategy for three years is structurally impossible. They have to pivot. From their seat that is rational.

From the buyer’s seat, there is a lot of risk. A media company evaluating an AI deployment is not making a six-month decision. They are signing a contract, integrating the vendor over six to nine months, training their team, restructuring workflows, and hoping the vendor is still pointed in the same direction eighteen months later. When the vendor pivots, the company is left with a workflow built around a product that no longer exists in the form they bought. The renegotiations that follow tend to have the same flat quality in the room. Everybody knew the risk going in, and nobody knew what to do about it. The company has already burned eighteen months of runway and still does not have the margin improvement they bought the vendor to deliver.

Token economics break the pilot-to-scale math

Many AI services are still priced in tokens, which makes spending predictable for a single user typing into a chatbot and wildly unpredictable for a content supply chain running thousands of titles through enrichment, localization, or quality control. The same task can cost dramatically different amounts depending on the model version, the prompt construction, and how the supply chain is calling the AI underneath.

CFOs trying to get to a per-unit number for forecasting tend to walk out of those meetings without one. They need to know whether AI is going to be a margin lever or a margin liability, because the answer determines whether the platform makes its profitability target. The pricing models the AI industry has adopted are not yet compatible with the volume economics of a content supply chain. Until they are, deployments at meaningful scale will keep running into the same wall.

The data is the foundation, and the foundation is not ready

AI cannot perform on dirty data. Most catalog data in the industry is dirty in ways that took fifteen years to accumulate, and the people who built the workarounds are still the ones holding it together.

The picture rhymes across most of the companies I have worked with. Inconsistent asset IDs across systems. Metadata schemas that were standardized at the team level but never at the company level. Rights data trapped in formats no system can parse. Title-level information that was correct when it was entered but has drifted as platforms, regions, and licensing windows have evolved. Workflow handoffs that produce different versions of the same record in different systems. AI deployed against this substrate produces confident outputs that are confidently wrong, which is worse than no output at all.

The work to fix metadata debt competes for budget with everything else the company is being asked to deliver on the same profitability deadline, and many companies are choosing the visible work over the foundational work. That choice has a compounding cost. AI, metadata, and technology become a flywheel of institutional debt that continues to cycle, and the cost shows up in every pilot that refuses to scale. MIT NANDA found that the ninety-five percent failure rate is concentrated in the scaling phase, not in development. Pilots demonstrate technical feasibility. Production requires operational discipline that a lot of companies are still building.

The harder pattern underneath is that most pilots were designed to demonstrate technology, not to produce business outcomes. The pilots that survive scaling started with the operational question first and the AI tooling second. That sequence matters. A pilot that asks “can AI do this thing” produces a yes-or-no answer about the technology. A pilot that asks “can this workflow run reliably at scale, and does AI help us get there” produces a deployment that survives contact with the production environment. A lot of the industry is still running the first kind.

The part the technology conversation keeps trying to skip past

AI does not change people. Entire systems still need to be designed, built, and held together by the humans inside them. The principles of good operations still apply. The planned workflows shift somewhat because the technology choices shift somewhat. Vendor selection looks different. Standards evolve. Business goals can actually expand when all of the pieces come together inside a thoughtful plan. None of that changes the underlying truth that the companies who will do the best with AI over the next five years are the ones who understand that their people are going to be people, whether AI is in the room or not.

Teams adopt what works for them. They work around what is imposed on them without their input. The operational rhythms of a content supply chain are held together by humans making judgment calls about edge cases the system did not anticipate. That work does not go away when AI shows up. It becomes more important, because the cost of letting AI make a confidently wrong decision at scale is higher than the cost of letting a person make a slower right decision at human pace.

The industry is being forced to mature, fast, on a clock it did not set. AI is the technology a lot of leaders hoped would let them skip the painful step of getting the operational foundation right. It has turned out to be the technology that proves the step cannot be skipped. The companies pretending otherwise are losing time they do not have. The companies accepting it are getting somewhere.

The supply chain is the strategy. The people running it are the operation. AI did not change either of those things. It only made the cost of pretending otherwise impossible to hide.

The first run looked at meaning. Media used to arrive with enough context for a person to know what it was, where it came from, who could use it, and what could be done with it. That’s coming undone right now. The platforms, editing tools, and AI systems that move media, strip that context out at every step, and everyone downstream rebuilds it by hand, over and over, just to keep working. The Inheritance Shift traced that quiet, expensive rebuilding, and why it stops holding up once the machines are the ones doing the work.

The second run looked at people, and here the loud prediction is that as the tools get good, the humans get pushed out. But that isn’t what’s happening. The role is shifting instead of vanishing, from making the thing to judging it, checking it, standing behind it, and the weight of that shift is landing on one hard requirement, being able to see inside systems that have turned into black boxes. The Human Layer followed that move, and the reason people aren’t simply pushed out is that the system is nowhere near settled enough to run without them. The judgment is load-bearing precisely because nothing is finished.

The third run looked at ownership. While most of the industry argues about which tools are good and which are hype, control is moving somewhere else, down underneath the tools, to whoever owns the ground the whole thing runs on. The models, the compute, the pipes, the standards nobody thinks about until they break. The Control Layer traced that move as it happens in the open, a handful of players taking the infrastructure while everyone else argues about the interface.

Put those three next to each other and the separate arguments fall into one. The thing on top is changing, the people are changing, the ownership underneath is changing. Three finished pieces of writing, all pointing at the same thing still in motion. An old way of doing media is losing its grip in real time, before a new way has taken hold.

That in-between is where we’re standing right now. We’re past the old order and not yet inside the new one. There’s a word for the gap between one authority ending and the next one taking over, an interregnum, but the word matters far less than the thing it names, and the thing is easy to feel even if you’ve never heard it called that. The rules on paper still describe a world that’s already on its way out. The world that’s arriving is still making up its rules in public.

Here’s what changes in that gap. When a system is settled, keeping things going is somebody’s job. The archive has a keeper. The standard has a body that enforces it. The rights have an owner who protects them. Keeping it all going feels boring and automatic, because the responsibility has somewhere to live. In the in-between, that ownership comes loose. What lasts is no longer whatever the system holds onto by default. What lasts is whatever someone actively chooses to keep, or can’t afford to throw away, or has to rebuild on their own because the big systems let it drop. Keeping things going stops being automatic and turns into a choice that costs something.

That’s the ground the next stretch of writing stands on, and it’s moving. Everything I’m about to lay out is built on sand, and I know it. This is a live shift, not a finished one, and over the next several months of writing through it, the industry is going to move under my feet more than once. You can’t file a final report on a thing that isn’t done. What you can do is name the shape as clearly as it’ll let you, and stay honest when the shape shifts.

So here’s the shape, as clearly as it’ll let me right now. The next run is about persistence, about what survives this gap and what it costs to keep it, and it comes apart into four questions that are already pulling at me.

The first is what we carry forward on purpose. Not everything that survives is chosen, but some of it is, and the choosing is about to become strategic in a way it never had to be when continuity was free. Somebody is going to have to decide what’s worth the storage, the migration, the rights work, the human attention, and the bill.

Then there’s what won’t leave, no matter how much we’d like it to. Every system carries dead weight it can’t shed, rights structures built for objects that don’t exist anymore, distribution deals still shaped like channels and windows, standards that are half preservation and half fossil. Nobody loves the old container. Everyone still needs it. That drag shapes what the next system is even able to imagine.

Less obvious is what falls apart the moment the people who maintained it move on. The hardest thing to preserve isn’t the archive, it’s the knowledge of what the archive was for, and that knowledge lives in people, not files. As the machines become the operators of memory, on time horizons that aren’t human, the quiet question is what decays when the keepers change hands.

Finally, there is what survives only because it stays local. When centralized memory thins out, meaning doesn’t die, it retreats, into regional archives, into specific crafts, into communities small enough to still maintain their own. Some of that persistence is sovereign. Some of it is only available to people who can afford it. The gap between those two is going to matter.

None of these four are settled, and every one of them is going to move while I’m writing it.

Some of this will be smaller and closer to the metal, a single sharp look at one moving part rather than a full argument, but still pointed at the same shift. One is about MoQ, a protocol quietly collapsing the whole path from camera to viewer into one pipe, technical to the bone and still landing on the question of who controls those pipes. Another is about world models, media that arrives with no artifact at all, and what that does to everything we’ve built on the assumption that there’s a file to carry forward. A third is about archives learning to talk to each other directly, preservation rebuilt from the ground up. Smaller pieces, same shift.

That’s the work ahead, and it starts here, in the gap, where nothing is being kept for us automatically anymore and everything worth keeping has quietly become a decision.

A word before you go, to the people who got it this far. If you’ve been reading, forwarding, or arguing back, thank you. And if you’ve paid to keep it going, thank you twice, because there’s something strange and good about watching someone decide a line of thinking is worth paying for while it’s still being worked out in the open. That’s what lets the work continue, and it’s what makes it better. I don’t lose sight of it.

And if you’re new here, this is the place to start. The three series this piece draws a line through, The Inheritance Shift, The Human Layer, and The Control Layer, are worth reading in order, and they set up everything coming next. Subscribe and you’ll get the persistence run as it lands, the smaller technical pieces alongside it, and the occasional correction when the ground moves under all of us. Which it will.

Power that cannot be checked through governance gets consolidated through other means, through acquisition, vertical integration, the absorption of standards bodies, and the merger of operator and adjudicator into entities that own the stack from chip to interface. That much follows from the first three pieces in this series. What this piece adds is recognition. The pattern the series has been describing for three installments has run before in American economic life, more than once, at full scale, with documented endings. The job of this closing piece is to name it and try to put a lens on the version we are currently living through. I am not saying our current fights over control will play out the same one, I am merely pointing to past times as indications of possibilities.

What Fills the Gap

A governance vacuum is not a neutral space. It has a shape, and the shape often favors whoever already holds operational power when the vacuum forms. Consolidation is in and of itself, not a form of a conspiracy and does not require one. It is what operational power does when nothing constrains it, because every incentive points the same direction. Owning the layer below you removes a dependency. Owning the layer above you removes a customer’s alternatives. Absorbing the standards body removes the specification you didn’t write. Each move is individually rational, defensible in any boardroom, and frequently celebrated by the markets. The aggregate of the moves is a structure no one voted for and no one governs.

My first three pieces described that structure forming inside the media stack. Production memory concentrating in the platforms that hold it. The credibility chain becoming a position rather than a record. Rule-making migrating to the operators because oversight and governance never formed. Watched piece by piece, it reads as news. Watched whole, it reads as something else, because we have seen it before.

Read more

It was not a seizure

Nobody raided a building. No new law passed. A directive came from the Commerce Department, and a frontier model that had been public for three days went dark across the world. The mechanism was paper. The effect was total and it arrived in hours. The reflex is to file that under ordinary regulation, and the reflex is wrong. Ordinary regulation does not switch off a running service for everyone on the planet between one afternoon and the same night.

Whose hand was on it, and how

Anthropic did not choose to go dark. The structure of the order and the structure of its own systems removed the alternatives, and that distinction is the whole argument.

The directive was keyed to citizenship. Lutnick’s letter required government approval before Fable reached any foreign national anywhere, and under export law a foreign national is anyone who is not a US citizen, including the green-card holder and the visa worker sitting in an office in San Jose. That one detail closes the exits. Geofencing sorts by location, and location is not the thing the order prohibits, so a blocked-country list does nothing about a Canadian engineer on a US network. Restricting new signups leaves the existing base untouched, and the system had no reliable way to tell which of those accounts belonged to citizens. The deemed-export standard reaches further still, treating an approved user who forwards a prompt or an output to a non-citizen colleague as an export in itself, which means even a careful partial control carries criminal exposure. Real identity verification is the actual fix, and standing it up to a compliance standard is a months-long build, not a weekend patch. Anthropic had ninety minutes. Under that deadline and that liability, the only immediate control that could be trusted not to serve a prohibited user was to disable the model for everyone, so that is what happened, worldwide, within hours.

A forced shutdown and a chosen one are different things. A company that pulls its own product is making policy. A company that pulls its product because the rule plus its own architecture left no compliant path is showing you where control actually sits. The second is what happened, and it is the more unsettling of the two.

Here is the part the contract hides. A customer who subscribed to Fable had a real relationship with Anthropic, and it was never the relationship that decided whether the model stayed on. That decision ran from Commerce to Anthropic, and the customer was not a party to it and could not see it from where they stood. The chain is customer to Anthropic to Commerce, and only the first link is visible from inside the contract. Operators already know a cloud provider sits on a network provider, a payment processor sits on a bank, an app sits on whoever owns the platform. A frontier-model provider sits in the same kind of stack. The customer was depending on Anthropic. Anthropic was depending on permission it did not control. When the bottom link moved, the top one had nothing to say about it.

Amazon belongs in this, and it is worth being exact about where. Amazon’s researchers found the bypass. Amazon’s CEO carried it to the Treasury Secretary. Amazon is also Anthropic’s largest investor, with money in the company whose product it helped take offline. Those facts are unusual and they should stay on the page. None of them handed Amazon the switch. A competitor with a stake surfaced a problem and the government decided what to do with it. The hand that moved was federal, and the most a rival could do was put the question in front of the people who held the authority, which turned out to be enough.

Anthropic could argue, and it is arguing, in a separate case already in court. What it could not do was refuse on the spot. An export directive backed by criminal penalties is not a negotiation you win on a Friday night. You comply, you litigate, and the model stays dark while you do.

The bad blood

None of this happened between strangers. Anthropic and the administration were crosswise before Fable shipped. The company and the Pentagon had failed to agree on contract terms over how its models could be used, and the administration had labeled it a supply-chain risk, a designation never before hung on a US business. It had also expanded a trusted-access roster past what the government signed off on. The government ran a ninety-minute ultimatum and a Friday-night letter, and one of its advisers argued the whole thing out in public posts. There is real history here, and it runs both directions.

The official reason for the shutdown did not hold still. It began as a narrow jailbreak, then suspected foreign access through a telecom on the trusted list, then something far larger. The Economist reported that the NSA director had told a senior senator the model broke into almost all of the agency’s classified systems in hours during a red-team test the day before the ban. The reporter who wrote that line later cautioned it should not be read literally, since the result leaned on other tools running alongside the model, and no agency has confirmed it. The capability underneath is real enough that none of this is pure theater. But a reader does not have to settle which rationale was the true one to take the point. The rationale moved. The capability did not. At least one of those rationales is serious enough that it cannot be waved off as pretext. A reader does not have to decide which version is true, or whether the government had the better of the argument, to take the point. The rationale moved. The capability did not.

The asymmetry is the only score worth keeping here. Whatever the history, only one party could turn the model off for everyone, and did.

A precedent outlives a grudge

The reason this should change how an operator thinks is narrower than the usual today-them-tomorrow-you reflex, and it has nothing to do with whether the shutdown was deserved. It has to do with the kind of control the event put on display.

It was retroactive. Export controls usually sit in front of a sale and govern what ships. This one reached back through a service already running inside other people’s systems and switched it off after the fact. Whatever a team had built on Fable was not grandfathered in. It was live one hour and gone the next.

The control point was invisible from the contract, and that is the part that should bother people. It is not the same as saying nobody could have known. Plenty of operators inherit cloud dependencies they never picked and sign platform terms they never read. The difference is that those dependencies are at least nominally theirs to find. The line that governed Fable ran one layer behind the company they actually dealt with. A dependency you are party to, you can map. One that governs from inside someone else’s contract is far harder to reach, and Fable’s was that kind.

The scope may not stay narrow. The logic that took Fable down was never specific to Fable. Today GPT-5.5, Gemini, and Anthropic’s own lighter models sit untouched. The authority could reach them. The government simply has not aimed it there. Legal analysts note the same deemed-export reasoning could attach to any sufficiently capable model later judged a similar risk, one already deployed as readily as one still unreleased, since Fable itself was live when it was reached. What happened to one model is available to do to the next, on the same authority, and the choice of which model and when belongs to someone the operator downstream never gets to vote on.

Set that against the failures people already plan for. A cloud provider can drop you and a platform can deplatform you, and both are bad days. But you generally know who those providers are, and there is usually a policy you could have read. This was faster than either, total in a few hours, and routed through a party your contract never named and your diligence could never easily reach. The dependency you could not see is the one that moved.

The off-switch was always there

Strip out the politics and the structural fact remains, and it holds whichever way the dispute resolves. Every model running underneath a workflow is a dependency, and some of those dependencies answer to people who were never named in your contract and will not be in the room when you build the next thing. Fable made that visible for an afternoon. The visibility is the part to keep. The off-switch was always there. Now you know whose hand can reach it.

The polite assumption available to most commentary is that the gap will close, that institutions will catch up, that the unsettled state is transitional and will resolve as the technology matures and the broader culture absorbs it. The argument here is that the gap will not close, that the institutions in place are not equipped to close it, and that the structural conditions of the present are best understood as durable rather than transitional.

The Governance Vacuum

Governance gaps in media systems are familiar terrain. Standards bodies have always lagged the systems they describe, regulators have always arrived late, and lawyers have always cleaned up after the technology shipped. The recurring pattern produces friction but not paralysis, because the actors making the rules and the actors being governed by them have historically been distinct enough that the friction can be metabolized over time and the system can move toward something resembling balance.

What has changed is concentration. The decisions that historically counted as governance, who is allowed to see what, who can do what, what gets denied at the level of the platform or the underlying system, are now being made by the same actors who hold the infrastructure those decisions depend on. The operator and the adjudicator have collapsed into a single role, and the friction that historically pushed the system toward balance has nowhere to land because the actor producing the friction and the actor metabolizing it are the same.

Calling that situation a governance vacuum overstates the absence. There is rule-making, in volume, with sophistication, at a scale and speed that exceeds what any prior regulatory architecture has had to engage with. What is missing is governance in the sense the word implies when we use it about functioning institutions, rules made by an actor who is not the rule’s primary subject. The control layer is governed in the sense that rules exist within it. It is not governed in the sense that the rules were authored anywhere outside it.

Read more

For more than a decade the fight in media has been about the library. Much of the Warner dealmaking in those years was sold as an IP story, who owned the shows, the rights, the catalog deep enough to force a carriage deal. That fight is ending, not because content has stopped mattering, but because owning it stopped conferring control. Audiences wake up, pick up a remote or a phone, and land on a surface someone else operates, and that surface decides what they see first. The follow, the subscription, the channel you chose have been replaced by a feed that decides for you, and the company that runs the feed sells the attention it manufactures.

Three deals in a single week make the same move from three different corners of the screen. A broadcaster buys the largest independent television operating system. A studio prepares to generate the advertisement itself. A game publisher turns its stadiums into ad inventory. Read together, they describe a unit of value that has shifted off the title and onto the audience, and a set of companies racing to own the entire machine that meters and sells it.

Fox Buys the Surface, Not the Show

Source: The Wall Street Journal, June 15, 2026

Fox agreed to buy Roku for about $22 billion in enterprise value, $160 a share in cash and stock, funded with new debt and cash on hand, with Morgan Stanley committing a $12 billion bridge facility behind the cash consideration. Fox holders will own roughly 73 percent of the combined company, and the deal is targeted to close in the first half of 2027. Fox already owns the content that anchors a living room, the NFL, the World Cup, Fox News, and the free service Tubi, bought for $400 million in 2020 and now near 100 million monthly users. What it did not own until now is the layer in front of all of it. Roku is the largest connected-television platform in the United States at about 25 percent device share, ahead of Samsung’s Tizen, reaching more than 100 million global households, and it long ago became less a hardware maker than an advertising business sitting on an operating system.

Why it matters

Fox did not buy more content. It bought the home screen, the household data, the ad placement, and the toll. Roku already takes a cut of the ads that other apps serve on its platform, so after close the streamers that compete with Fox will be routing ad revenue to Fox. The deal also takes the last independent television operating system off the board, the one platform that could credibly claim no dog in the fight, at a moment when ad-supported plans make up almost half of new premium streaming signups in the United States, up from 39 percent two years ago. The new debt is the tell. You borrow at that scale for infrastructure you intend to own, not for a programming slate.

Disney Starts Making the Ad

Source: Business Insider, June 17, 2026

Disney is preparing a July beta of a tool that generates television advertisements end to end, the script, the video, and the music, inside a single workflow on its self-service ad platform. It is aimed at small and mid-size advertisers that do not have video assets of their own, the businesses that could never afford a thirty-second spot.

Why it matters

Selling the audience was step one. This is step two. Disney is absorbing the creative layer, the last part of the ad business that still lived outside the platform. Google, Meta, TikTok, and Amazon have been collapsing ad creation into their own systems for years, so Disney did not invent this. What is new is that the move has crossed into premium television, the living-room screen, where the audience arrived for a show and the inventory is a thirty-second spot. An advertiser now arrives with a brief and leaves with a finished, targeted, measured campaign without ever leaving Disney’s machine. Fox bought the surface that meters the audience. Disney is building the factory that fills it.

EA Turns the Game Into Inventory

Source: Electronic Arts, June 15, 2026

EA launched EA Advertising, a platform that places brands inside gameplay across a portfolio that reaches more than 120 million monthly players. The first units are native to the 3D sports games, ad boards, scoreboards, and broadcast overlays served dynamically and measured to IAB standards, running on a proprietary ad server built for EA’s own engine, with Visa, Red Bull, Xfinity, Lowe’s, Peacock, and Mountain Dew signed on. It arrives as EA moves through a roughly $55 billion leveraged buyout.

Why it matters

The same move jumps out of television entirely. Gameplay becomes addressable inventory, the virtual stadium boards become a media network, and the publisher owns the server that sells them. The sequence is hard to ignore. Historic leverage on one side, a higher-yield ad machine on the other.

Pattern Synthesis

The throughline is not consolidation and it is not AI. It is a change in what these companies are actually selling. For years the unit of value was the title, the show or the game or the rights package, and advertising sat beside it. In one week, three companies in three different content forms moved to own the apparatus that sells the audience instead, the targeting, the placement, the measurement, and now the creative. Content is becoming the cost of acquiring attention.

The financing makes the intent legible. Fox is taking on a $12 billion bridge. EA is inside a $55 billion buyout. You do not lever up like that for a content slate that depreciates the day it airs, you do it for infrastructure you expect to own and compound. The timing is not coincidence either. Subscription has hit its ceiling, with ad-supported tiers now close to half of new premium signups, so the growth is in attention sold rather than subscriptions collected. Generative tools have collapsed the cost of the one part of the ad business that resisted automation, the creative, which is the part Disney is now bringing in house.

What used to be governed separately is folding into one owner. The surface that decides what you see, the data that says who you are, the slot the ad runs in, the price it clears at, and now the ad itself. When a single company holds all of those, the relationship a viewer thought they had with a show becomes a relationship with the machine that sells against it. The home screen, the streaming break, and the gameplay are starting to run on the same business.

Closing Note

The question worth holding is not whether the machine works. At this scale, with this much debt behind it, it will be made to work. The question is what it costs the thing underneath. Roku grew to $22 billion by carrying everyone without preference, and a content owner now runs that gate. The neutrality that made the surface valuable and the audience’s tolerance for being sold to are the only real constraints left, and both are easy to overspend. Watch three things from here. Whether the antitrust posture on the Fox deal has any teeth, whether rival streamers keep paying a toll that now flows to a competitor, and whether audiences start to feel the home screen, the spot, and the game as one continuous transaction. The machine is being assembled in the open. The cost shows up later.

For two and a half decades, Seth worked inside Hollywood. He built businesses there, raised money there, sat in dark rooms with auteurs finishing their movies. From inside, he says, the room feels like the whole world.

It is a small room with a finite number of seats.

The pitches now arriving at Hallstone Ventures sort cleanly into two buckets. The first bucket is founders building products and services for Hollywood. They want to sell into the studios, win the prestige award, get the splashy deal. Their TAM is five customers. Their sales cycles are punishing. Their decks open with how they’re going to redefine or reimagine or transform the industry. Those pitches are stalling. Hallstone isn’t funding most of them.

The second bucket is founders building for everything else. The graph that Hollywood is one node in. YouTube. Vertical video. Microdramas. Live commerce. International production. The infrastructure layer underneath all of it. Those pitches are landing.

The pattern is structural, not a matter of execution. It has to do with what each set of founders thinks Hollywood is. Most of them are wrong about that. So was Seth for most of his career. So were a lot of the people he worked alongside.

The Inheritance

What we call Hollywood is not a natural form of cinema. It is a 1917 industrial template Americans inherited from Europe.

The vertically integrated studio model, the contract talent, the prestige literary adaptation, Thomas Ince’s central producer system. None of it is intrinsically American in origin. Pathé Frères and Gaumont, both founded in 1890s Paris, were running the model at global scale a decade before any American studio existed. Pathé was the largest film company on earth from roughly 1904 to 1911. At its peak, an estimated 60 percent of films shot anywhere were shot on Pathé equipment. The American film industry of that period was, in trade-press terms, an importer.

Two simultaneous events between 1915 and 1917 reorganized the global industry around Los Angeles, California. World War I gutted European production. Celluloid chemicals went to gunpowder. Men went to the front. Capital evaporated. By 1919, by Britannica’s estimate, 90 percent of films screened in Europe, Africa, and Asia were American. In the same window, the United States v. Motion Picture Patents Company ruling broke the Edison Trust, the only American cartel that could have constrained the independents.

Americans codified what was effectively a European industrial template and rebranded it as the natural order of storytelling. Film historians David Bordwell, Janet Staiger, and Kristin Thompson date the consolidation of “classical Hollywood cinema“ to exactly 1917.

Everything we now call Hollywood is a variation on that template. It has held for a hundred years because nothing displaced it. The thing displacing it now is not even a new industry, to be honest. It is a return to what came before the template was inherited.

The Stack

When industry veterans argue about whether Hollywood is dying, they are usually each looking at a different layer of it. Hollywood is not one thing. It is a stack of six overlapping systems that emerged in roughly this order and have begun to decouple at different rates.

The most visible layer is the geographic one, the production cluster that concentrated in Los Angeles after 1910. That layer is dispersing fast. Atlanta, Vancouver, Budapest, Prague, and Auckland have all been doing for years what Burbank used to. Underneath sits the vertically integrated industrial mode that defined the studio system proper, intact from the 1920s through 1948, when the Paramount decree forced the majors to divest their theater chains. The vertical integration that defined Hollywood for a quarter-century has been gone for almost eighty years.

What followed was an agency-and-package power structure, the talent shops (MCA, William Morris, ICM, CAA) that filled the vacuum after 1948 by orchestrating deals through bundled talent rather than studio contracts. That structure has been the working definition of “Hollywood power” for most of the time anyone working today has been in the business, and it is now under more pressure than it has been since it formed.

Below the agencies sits the labor and guild system, SAG, WGA, DGA, IATSE, on which the 2023 strikes were fought. It is the most durable layer in the stack, mostly because it is the only one with an organized constituency.

Then comes the narrative-aesthetic norm. Continuity editing, character-driven causal narrative, three-act structure, invisible technique. This is the layer that has paradoxically grown more globally dominant over the last decade rather than less. Every Netflix original, K-drama, Bollywood blockbuster, and Chinese microdrama still operates inside it. The aesthetic clubhouse is global, even as the industrial clubhouse shrinks.

The bottom layer, distribution-finance, is collapsing fastest. Theatrical windowing, output deals, hit-driven slate economics. These were the working tolerances of the old machine, and they are no longer holding.

When Steven Soderbergh talks about the death of Hollywood, he means layers two, three, and six. When IATSE leadership talks about saving Hollywood, they mean layer four. When a film school professor talks about Hollywood’s aesthetic dominance, they mean layer five. They are not actually disagreeing. They are looking at different parts of the same dying animal. For an investor, knowing which layer a founder is fighting for is the first ten minutes of due diligence.

What Came Before

The American film industry’s first decade, before the 1917 consolidation, was genuinely closer to a creator economy than to anything we would later recognize as Hollywood.

Storefront nickelodeons proliferated from a single Pittsburgh theater in 1905 to roughly 10,000 by 1910, drawing as many as twenty-six million weekly admissions in a country of ninety million. The operators running them were small, local, and almost universally immigrant or first-generation. Carl Laemmle opened a Chicago storefront in 1906. William Fox had bought a Brooklyn arcade for $1,666 two years before, installing a nickelodeon on the second floor. Marcus Loew was assembling a chain out of penny-arcade properties in New York. Adolph Zukor had moved from fur trading to peep shows, and would soon found Famous Players. Many of these operators came up through vaudeville and penny arcades, the live-performance economy that has its own structural rhyme with the creator economy of today. The four Warner brothers opened the Cascade Theater in New Castle, Pennsylvania on February 2, 1907, and within months had begun pooling other small exhibitors into a distribution exchange that they later incorporated as the Duquesne Amusement Supply Company. Exhibitor to exchange to producer, in that order. The same vertical-integration-from-the-bottom move that creator-economy operators make today.

Films at the time were one or two reels, programs changed weekly or faster, and theaters typically seated fewer than two hundred. The films were silent, which meant they asked nothing of an audience that did not yet share a common language. They were ideal entertainment for the ten million immigrants who arrived in the country between 1900 and 1910. Everything about the industry ran fast, cheap, and close to its audience.

Three years. That is how long the open phase lasted before the enclosure began. In December 1908 the Edison Trust pooled the major producers, the French firms Pathé, Gaumont, and Méliès, and Eastman Kodak’s raw-stock monopoly into a single licensing cartel, and by 1910 it had consolidated distribution under its General Film subsidiary. The independents fought back by moving to California, where Trust enforcement was harder, and by going to feature length the Trust resisted. By 1917, the form had hardened around the European template.

The creator economy now is not a new wave inside the inherited template. It is a structural return to what American cinema looked like before the template existed.

The Two Rooms

In different decades, we both had the same moment. The room we were working in was not the whole world and we were late to realize it.

FROM SETH

For me, the room was the prestige clubhouse. More than 25 years of building businesses inside Hollywood, raising money inside Hollywood, sitting in dark rooms with auteurs finishing their movies. The clubhouse is seductive in a way that is hard to convey to anyone who has not been inside it. The dark room, the director, the finished cut, the feeling that the work matters. I spent years optimizing for a five-customer market and telling myself the constraint was an honor. The realization came late.

FROM ANDY

For me, the room was a company I co-founded in New York in the early 2000s. We were building something close to what Frame.io would become much later. A branded, secure way for ad agencies and post houses to send video to clients for review. We had real customers. We had (some) revenue. We had the encoding chops to make it look good over the bandwidth people actually had.

I was even seeing the next wave forming. A trade reporter interviewed me about the company and I told him that anyone with a laptop and a wireless connection at a coffee shop would soon be doing this work themselves. I was describing the world that was about to flatten my business. I built it anyway.

YouTube launched in 2005. Free, browser-based, ugly, none of the things our customers said they wanted. It killed our category in less than a year. Not because our product was wrong, but because the world had decided the layer underneath us mattered more than the layer we were selling.

The room was smaller than it appeared. The graph was forming outside it. Neither of us saw it in time, and we are not unusual for that. The clubhouse fools people on the way in, on the way through, and on the way out.

The Delusion

The clubhouse is deceptive on both sides of the door.

From outside, it looks like the whole world. Get inside and you have made it. The work matters, the people around you are the people who matter, and the story is so durable that founders pitch into it daily, build companies for it, and raise capital to chase it.

From inside, the deception runs in the other direction. The clubhouse really does look like the whole world, not because anyone is lying but because the room is engineered to feel total. The conversations are the conversations. The deals are the deals. The opening night is the only night. Everything outside reads as noise.

This is the delusion that has let Hollywood ignore its own decline for the better part of a decade. The room is shrinking. The graph is growing. From inside the room, you cannot tell.

What To Look For Now

The first ten minutes of a pitch are usually enough to tell which bucket a founder is in.

The clubhouse founders use the language of redefinition. They want to sell into the studios. They want the prestige reference customer. They benchmark against incumbent vendors and argue their edge is better service inside a stack that has not meaningfully grown in two decades.

The graph founders use the language of distribution. They think about audiences, not buyers. Platforms, not pipelines. They assume the clubhouse exists and treat it as one node, not the whole network.

The clearest tell is Adobe and Frame.io. Adobe paid $1.275 billion for Frame.io in 2021. By that point Frame.io had about a million users, and the vast majority were not in the clubhouse. They were in the graph. Post houses, ad agencies, freelance editors, indie filmmakers, all organized around a tool the clubhouse barely knew existed. The clubhouse paid a premium to acquire access to them. Adobe’s own clubhouse revenue is a small fraction of its overall business. The companies winning around media right now are the ones that solved for the graph first and let the clubhouse arrive at the back door.

The contrast case is the generative AI cluster. Runway. ElevenLabs. Midjourney. None of them tried to enter the clubhouse. Some of them effectively dared it to sue them. They built for the graph, scaled inside the graph, and only then started fielding inbound from clubhouse buyers asking what they were missing. The clubhouse meetings, by the founders’ accounts, went nowhere. The clubhouse is not where the work is.

Meanwhile the content the rest of the world is watching reflects what the graph has been telling us for a decade. Korean dramas on Netflix. Nordic crime on every streamer. Reality shows shot in Spain. Twenty years ago everything most Americans watched had a North American lens. That lens was one of the layers of the clubhouse, and it has already gone.

The Question, And What To Do With It

The question every founder should be able to answer in one sentence is this. Are you building for the clubhouse, or for the graph?

There is no wrong answer. There is a wrong answer to avoid, which is not knowing. If the answer is the clubhouse, build for it deliberately. Understand it is a small room. Understand the room is shrinking. Understand it is five customers in a market designed to keep them. The economics can work if the bet is made clear-eyed. They almost never work if the bet is made under the belief that the clubhouse is the whole world.

If the answer is the graph, build for the graph. Stop pitching the clubhouse on something the clubhouse does not want. Stop hiring the clubhouse-resume executive who will spend two years explaining why the numbers are weird. The graph runs on different math, different cycles, and different ideas about what good looks like.

This piece is not a eulogy for Hollywood. The clubhouse will keep making prestige work for as long as a small audience of cultural and financial gatekeepers want it to. The clubhouse is a node. It is not the network.

The discipline is to know which one you are working in. To listen to the operators in the other room. To consider that they might be half-right, and you might be half-right, and the interesting work is figuring out which halves go together. To take experience earned in the room you have been in and use it in a room you have not.

The clubhouse fools people on the way in, on the way through, and on the way out. The graph does not care.

Cocktail Endnote: The Old Fashioned

The Old Fashioned is the only cocktail named after the structure it preserves.

By the 1880s, American bartenders had spent two decades elaborating on the basic 1806 definition of a cocktail, which was spirit, sugar, water, and bitters. Curaçao got added. Absinthe got added. Multiple bitters got added. Fancy garnishes got added. The simple thing became the elaborate thing, and the elaborate thing got rebranded as what a cocktail was. Drinkers who preferred the original preparation started asking for theirs made “the old-fashioned way,” and the name eventually stuck to the drink itself.

That is the piece in a glass. The clubhouse is the elaborate version. The graph is the old-fashioned way. Hollywood was the period during which the original structure was forgotten. The renaissance now is a return to the form that always made sense.

Old Fashioned

• 2 oz bonded bourbon (a Heaven Hill Bottled-in-Bond for the classic version, or a Rittenhouse 100 rye if you want a sharper edge)

• 1 sugar cube, or a scant 1/4 oz rich simple syrup

• 3 dashes Angostura bitters

• A few drops of water

• Orange peel, expressed

• Optional Luxardo cherry, if you must

Place the sugar in a rocks glass, saturate with bitters and a few drops of water, and muddle until dissolved. Add the whiskey and a single large ice cube. Stir until cold and properly diluted, about thirty seconds. Express the orange peel over the top and drop it in. No muddled fruit. No soda water. No sweet-and-sour mix. The mid-century version that did those things was the elaboration. This is the form underneath.

The drink predates the name we use for it. So does the structure it stands for.

Anthropic shipped its most capable model this week and shipped it twice. Mythos 5 goes to a small set of vetted partners. Fable 5 goes to everyone else with safeguards attached. That part got covered everywhere. The part worth slowing down on is what happens after a user is already inside the public model, because the product does not stay fixed while you use it.

Fable 5 is not just a model. It is a switching layer. It reads the request, classifies the risk, and decides which capability you are allowed to reach. Sometimes the downgrade is disclosed. Sometimes the request is blocked. And in one narrow but important case, the model keeps answering while quietly limiting what it can give you. The model you are sold at signup is not the model that answers any given prompt, and the gap between those two things is now set per request, by the provider, against its own read of the risk and its own commercial interest. Capability and access have come apart, and the seam runs straight through a single product.

The membership tier: who you are decides which model exists for you

Source: Anthropic announcement, June 9, 2026

Mythos 5 and Fable 5 come from the same underlying model family. The difference is the safeguard layer and who gets to skip it. Anthropic made Mythos 5 available only to a small number of trusted partners, beginning with the Project Glasswing program for organizations that defend critical software infrastructure. Everyone else gets Fable 5, the public version with classifiers that block its reach in high-risk domains.

This is the oldest move in the set and the easiest to see. Access is decided once, at the door, by institutional identity. A cyberdefense contractor and a curious subscriber are not buying different models. They are buying different permissions over the same one. The capability exists. Whether it exists for you depends on which list you are on.

Why it matters

If you run a media operation, you already negotiate this shape. It is the windowing logic, the allow-list, the clause that decides which buyer sees the premium version first. The difference is that the tiering now lives inside the production tool itself, not the content deal. The vendor you license your AI stack from is the one drawing the window, and you may be on the standard side of it without having asked which side existed.

The metering tier: what you ask decides which model answers

Source: Anthropic System Card, June 9, 2026

Inside Fable 5, the decision does not stop at the door. Classifiers watch each request for topics tied to cybersecurity, biology and chemistry, or attempts to copy another model. When one triggers, the request falls back to the prior model, Claude Opus 4.8, and you get the weaker answer. Anthropic says this happens in fewer than five percent of sessions.

The disclosure of that downgrade is itself tiered by interface, which is the detail most of the coverage skipped. In the web, desktop, and mobile apps, the request falls back and the user is told which model handled it. In the Messages API, there is no automatic fallback at all. The request is blocked and returns a structured refusal code, and developers have to opt in to server-side fallback to get anything else. So the consumer is notified, while the builder gets a refusal unless they have explicitly opted into fallback behavior. The model you reach is decided continuously, prompt by prompt, by what you are asking rather than who you are.

Why it matters

For an operator, this is the licensing problem you know from rights, moved into the tool. What you are permitted to do has always depended on the tier you bought, theatrical versus streaming, domestic versus worldwide. Now the same conditional sits one layer down, governing what the software will do for a given task rather than what the contract will allow. A workflow that runs clean on routine jobs can hit the throttle the moment the work touches a flagged domain, and the tool decides that in the moment, not at signing.

The silent tier: some requests get a weaker model and no notice at all

Source: Anthropic System Card, June 9, 2026

There is a third intervention, and it is the one that names the piece. For requests aimed at frontier model development, things like building pretraining pipelines, distributed training infrastructure, or accelerator design, Fable 5 does not refuse and does not reroute. It quietly degrades its own output. Anthropic’s system card states plainly that, unlike the cyber and bio interventions, these safeguards will not be visible to the user. The model limits its effectiveness through prompt modification, steering vectors, or parameter-efficient fine-tuning, and keeps answering as if nothing happened. The company estimates this touches about three hundredths of a percent of traffic, concentrated in fewer than a tenth of a percent of organizations.

Read the stated reason for that one, because it is not the usual harm argument. Using Claude to build a competing model already breaks Anthropic’s terms of service. The system card says enforcing that restriction through the safeguards avoids accelerating the actors most willing to violate the terms. The downgrade is competitive protection wearing the safety apparatus. A narrow slice of users is paying for a capability, receiving a degraded version of it, and not being told the degradation occurred, because telling them would defeat the point.

Why it matters

This is the one a production shop should sit with. Every quality-control instinct you have assumes a tool performs consistently and tells you when it cannot. Here is a vendor that has shipped, and documented, a tool that can quietly do worse on a defined category of work while reporting nothing wrong. The affected category is narrow today and probably nowhere near your edit bay. The principle is the part that travels. Once silent, per-task degradation is an accepted product behavior, you can no longer assume the tool gave you its best effort just because it returned a confident result. You find out in the output, if you find out at all, never in the contract.

The consequence: when work is delegated, supervision becomes a permissions problem

This lands harder now that machine work is delegated rather than prompted. There are two ways to manage an agent. You steer, staying in the loop on every step, or you dispatch, handing over a brief and checking the result. Dispatch is where the routing layer bites. If you judge the output and not the process, you have no way to know whether the system gave you its full capability or a quietly metered version of it. The proof of work disappears into the same surface that decides how much work to do.

Why it matters

Supervision stops being about whether the machine did the task well and starts being about whether you were allowed to have it done at all. The interface that used to just take instructions now holds a policy you cannot read. The more you delegate, the more of your judgment you are routing through a switch that answers to someone else.

What connects them: the model became a policy-controlled switch

Three mechanics, escalating in opacity. Identity decides which model exists for you, set once at the door. Topic decides which model answers, set per request and usually disclosed. Intent decides whether the model quietly throttles itself, set per request and disclosed to no one. Stack them and the product stops being a model and becomes a routing system that meters intelligence against the provider’s risk tolerance and the provider’s business, one prompt at a time.

The tell is the third mechanic, because it drops the safety framing entirely. The cyber and bio routing is at least about harm. The frontier-development throttle is about protecting market position, and Anthropic says so in its own card. That is the layer worth tracking. Once a provider builds the machinery to degrade output by request without notice, the machinery does not care what justification is loaded into it. Safety today, competitive moat this week, cost optimization whenever the compute bill demands it. The same switch serves all three, and only one of them gets announced.

For readers of this publication, this is the next beat in a familiar pattern. Identity and access have been moving into the AI product itself. This week, the access logic moved inside the model’s own responses. The consumer-facing version is already visible in search, where the platform interprets intent and composes the answer in place rather than handing back a neutral list, a shift this publication tracked out of Google I/O in Components, Not Solutions. The through-line is the same. The response surface stops being a neutral pipe and becomes a layer that decides what you get. Search composes the answer and chooses whether to send you out. The model composes the capability and chooses how much to give you. One surface, deciding.

What to watch

A frontier provider has now shipped, and documented, a tool that can deliver different capability to different requests and not always say which one you got. The slice affected today is tiny. The architecture is not, and architecture is what gets reused. The question for anyone licensing these tools into a pipeline is no longer which model is best. It is what gets loaded into the switch next, and whether you will be able to tell when it changes.

The Trust Anchor That Broke

In January 2026, Russia’s Kamchatka Peninsula was buried by its heaviest snowfall in sixty years. A state of emergency, cars buried, two people killed by falling snow. The story did not need exaggeration to be remarkable. It got exaggeration anyway.

Within days, a Threads user posted a sequence of six images of a snowbound Kamchatka that went viral across platforms. Four were generated using Grok. The user later confirmed it. AI-generated video clips of impossible snowdrifts hit tens of millions of views on TikTok and X before fact-checkers caught up. One unlabeled clip on X, people sliding down an impossibly steep drift between high-rises, reached roughly a million views before BBC Verify called it. AccuWeather, actually covering the storm, called the AI-fakes flood “the worst we’ve seen.” Several of the fakes were picked up and re-reported by major news outlets.

What matters is not that the fakes existed. AI fakes have existed for years. What matters is the sequence. Real disaster, fabricated coverage amplified by the same platforms that audiences turn to during emergencies, the platforms unable to distinguish real from synthetic. Platforms used to lend credibility to media. Today they mostly lend reach.

The doorbell example is smaller, more domestic, and structurally identical. Across 2025, AI-generated clips dressed up as Ring doorbell footage went viral on TikTok and Facebook with the actual #ringdoorbell tag. A dog bringing a bear home onto a porch. Raccoons bouncing on backyard trampolines. Most carried VEO watermarks for anyone who knew to look. They were shared as authentic security camera footage, often by media accounts that should have known better. The footage looked like a Ring video. That used to be enough.

The audience did not stop trusting these platforms because the content was fake. It stopped trusting them because they have lost the capacity to be the answer to “is this real.” That capacity used to be the product. It is not anymore.

Regime, Not Artifact

The instinct is to look at the artifact and ask whether the artifact is trustworthy. It is the natural instinct, and it is the wrong one.

A reader picking up a daily paper does not really evaluate the article in front of them. They evaluate the masthead, the dateline, the editorial process they assume sits behind the words, the corrections page they could go to if they doubted the piece. When trust in an article collapses, what is collapsing is rarely the sentence. What is collapsing is the reader’s confidence in the institution that produced the sentence. The article is a surface effect of a deeper apparatus, and the trust attaches to the apparatus.

The same move applies to media in the AI era, in a sharper form. People are not really distrusting the image they just saw. They are distrusting the process that generated it. The Kamchatka snowdrift is not the problem. The problem is the question that arrives once you have seen one fabricated snowdrift. What process produced every other clip in the feed. Whether anyone is tracking. The doubt is not local to the artifact. It is structural, and it lives in the chain.

Read more

I spent the afternoon of June 3rd in a room at the W Hollywood watching a very good argument get made on vertical LED screens. The Vertical Media Summit is a half-day event hosted by Owl & Co, the media and tech advisory firm run by Hernan Lopez, and Lopez built the whole thing around a single claim he laid out in the opening keynote. Film was the first audiovisual language. Television was the second. Vertical video is the third, something genuinely new, with its own economics, aesthetics, and audience psychology, and most of it is still up for grabs.

He made the case well. The line that stuck with me was his evolution analogy. If you had watched a daytime serial in 1950, he asked, would you have imagined that the same box would one day give you Game of Thrones, South Park, and the Super Bowl? Microdramas, in his telling, are to vertical what soap operas were to early television. An early, slightly embarrassing…

Read more

May is the first time I’m sharing the reference architecture that has been sitting underneath much of Engines of Change for the past year. April was about noticing a pattern. May was about commitment, about deciding this artifact was ready to leave the relationship layer.

Less composed than a Deep Cut, closer to a working note. Founders get the full memo and the EoC Reference Architecture Reader’s Guide below.

Read more

That sounds like a metadata argument until you follow it into the production stack.

The Misread

The conversation about synthetic production has been stuck at the surface for almost three years now. In 2023, the joke was Will Smith eating spaghetti in a ModelScope clip that looked like a melting wax figure. In February 2026, Seedance 2.0 turned the same basic gag into something uncomfortably photorealistic, while a fake Tom Cruise versus Brad Pitt fight went viral off a minimal prompt. The reaction arrived fast. Disney and Paramount sent cease-and-desist letters. The Motion Picture Association followed. SAG-AFTRA condemned the tool. A month later, two U.S. senators asked ByteDance to shut Seedance down. Rhett Reese, a co-writer of Deadpool & Wolverine, looked at the same moment and wrote that it was likely over for them. Then came the familiar argument about whether a model can replace the crew, the cinematographer, the colorist, the editor. All of it still lives at the level of the image.

That conversation matters. The political and economic questions raised by AI-generated imagery are not going anywhere. But the conversation has been so dominated by what comes out of the pipe that very little attention has gone to what builds up inside it.

The Inheritance Shift series argued that media systems were not designed to preserve context across transformation. The Control Layer arc picks up where that diagnosis ends. Once systems begin trying to preserve context, the question is where leverage accrues. This piece starts where production starts.

Synthetic production is being misread as an image-generation story. It is becoming a data-capture regime. The shift is not what the cameras and the models can produce. It is what the environment around the cameras and the models can record. The image is what leaves the building. The state, the structured record of how it was made, is what stays behind. And in synthetic production, the state may become the more durable source of leverage.

Where the Work Happens

A definitional pause before going further. Synthetic production lives in two layers.

The models are the engines. Seedance 2.0 from ByteDance was the recent flashpoint. Sora 2 from OpenAI had its moment before OpenAI wound the product down. Veo 3 from Google, Kling 3.0 from Kuaishou, Runway Gen-4, MiniMax Hailuo, Wan from Alibaba, and whatever replaces this list next month all sit in the same unstable model layer. Each has distinct strengths, distinct weaknesses, and a distinct rights posture. They change places on the leaderboard almost monthly.

The platforms are where creatives actually work. Some are made by the model makers and keep the user inside their own ecosystem. Google Flow is the cleanest example, an end-to-end creative environment built around Veo and Google’s supporting image and language models. The model makers want creatives inside their own tools because that is where workflows lock in and the data feedback loop tightens.

Most working creatives are picking a different surface. Independent platforms like Higgsfield, Krea, and Freepik aggregate multiple models inside a single workspace. A creative on Higgsfield can run the same shot through Seedance, Veo, Kling, and Sora, compare the outputs, and pick which engine fits the job. The model landscape moves too fast to bet on one provider, and these platforms know it. Their pitch to the working class of synthetic production is the one independent producers have always responded to: keep your options open.

The point worth holding onto is structural. When this essay says “production environment,” it usually means the platform layer, not the model. The platform is where the references go in, where the version history is built, where the rights conditions are configured, where the approval paths run. The data the environment captures is data the platform sees first.

Read more

This happened while the two largest labs were clearing their path to public markets. That timing is the story. Going public and settling the terms of your own oversight are the same act of consolidation, run in the same window, by companies preparing to be governed less by mission language than by market discipline.

A Jury Closes the Book on OpenAI’s Origin Story

Source: NBC News, May 18, 2026

A federal jury in Oakland found that Elon Musk waited too long to sue Sam Altman and OpenAI over the company’s drift from its nonprofit roots. The verdict never reached the merits. It turned on the statute of limitations, and the judge adopted the advisory…

Read more

MIT reported in August 2025 that 95% of corporate GenAI pilots are failing. The cause is not the quality of the models. It is that the tools are being pointed at workflows the systems were never built to read.

The media industry is about to make that number worse. In a typical media supply chain, the work that controls, integrates, and orchestrates the actual workflow lives in the heads, runbooks, and habits of the people executing it. That is the control layer this piece is about. It has been holding the supply chain together for two decades because the foundation that should have replaced it never got built. I have been calling that pattern Human Automation. Years of it are not replaced by an Agent in a day, and this piece is about why that matters now.

Years ago I was on a team replacing an asset management system that had been in place for about fifteen years. I had also been on the team that built the original, which gave me one piece of context most don’t get: I knew what the system was for the day it shipped. We started discovery the way you start most of these, asking the people who used the system every day to walk us through their work.

The conversations went like this.

“So you ingest these images. Why?”
“Because that’s my job.”
“Right, but these less obvious metadata fields. Why these specific values?”
“Because the runbook says to.”

People were surprised when that work they’d been doing often had no meaningful relevance to the then current workflows.

The people answering those questions had inherited a workflow and lost the reasons for it years ago. But, they kept the “machine” running while the organization above and around them spent its time, effort and capital on finance systems and back-office platforms and told itself, year after year, that media operations was fine or would be addressed eventually.

This exemplifies the concept of Human Automation. It demonstrates why the current industry conversation about Agents and GenAI replacing operations work is heading toward the wrong conclusion. The human layer in a media supply chain is not a deficiency, it is the layer that controls, integrates, and orchestrates the work. The Agents we are about to point at it do not know that, because the data and process logic they would need does not yet exist in any system the Agents can read.

What I mean by Human Automation

Human Automation is what happens when an organization decides, explicitly or by default, that building the technology to connect two systems costs more than putting a person in between them. A fulfillment coordinator reading an order out of a sales system and typing it into a media order system because no integration exists. A mastering PM reading a release calendar and creating work orders by hand because there is insufficient demand tooling. People are used as the integration layer itself. It is not a failure of the operators. It is a default position the industry has chosen, repeatedly, when it had other options.

The phrase collides with three adjacent terms that mean different things. Human-in-the-loop is an automated system that intentionally inserts a person at decision points. Robotic process automation (RPA) is software bots that imitate what a human does inside a user interface. What Andy calls human-enhanced is the human contribution that turns generative output into a legally protectable work. Human Automation is none of those. The other three describe specific patterns of human-machine collaboration. Human Automation describes what fills the gap when the machine was never built at all.

The human is the control layer, and does not know it

In a Human-Automated supply chain, the people doing the work are often not executing a documented process. They are absorbing variation the system was never built to handle. The absorbing is invisible to whatever tool comes next. The system reports green because the human effort kept it green.

The control layer is also not uniform. There are at least two tiers of it inside any real operation, and they fail differently when an Agent shows up.

At one tier are the people who execute. What looks like a documented process is mostly habit, with a runbook capturing the visible part. Context across the team varies. Some have built real understanding over years and know when something looks wrong. Others know the actions but not the reasoning behind them. The asset ingest conversation at the top of this piece is from someone closer to that end. They have been running this habit long enough that it looks indistinguishable from a process.

At another tier are the senior staff that carry the context the systems were never built to hold. The Director of Mastering tracking talent relationships and preferences in a leather-bound notebook. The Mastering or Localization PM who knows from one phone call this morning that a vendor is over capacity this week or where to route that job for that recipient. None of that knowledge is in a system. It is the reason the operation keeps running. It is also the reason the next person in the role takes eighteen months to be effective.

The people inside Human Automation have not been passive. They have innovated continuously inside the space they were given, building spreadsheets that hold rules that were never formalized, writing runbooks for processes nobody at the system level would document, absorbing turnover and tool changes and merger integrations. The complaint about Human Automation is not that the humans were not capable. It is that they were capable enough to make the underlying problem invisible to the organizations that should have been solving those problems.

The three ways we did not build the necessary foundation

For decades, the industry has made the same decision in three different ways: do not build the underlying technology ourselves. The three patterns have never been sequential. Human Automation persists while outsourcing continues at scale. Outsourcing persists as DTC investment ramps. GenAI and Agents are being layered on top now, while the previous two are still doing the load-bearing work. All three are active, and all three have left the foundation work unfulfilled.

The first choice was Human Automation. Major media companies spent most of the licensing era avoiding the conclusion that they needed to effectively operate as technology companies. My own analysis of technology capital allocation during this period at a major media company across a fifteen-year window found less than ten percent of the spend going to content systems. The rest went to finance and back-office. Headcount filled the gap. The phrase that justified those decisions was “we’re content creators, not a technology company.”

Underinvestment was only part of the problem. Where money did show up, often around the direct-to-consumer (DTC) rollouts of the last five or six years, the partnership between operations and technology did not. Technology organizations sat at arm’s length from operations, treated by accounting governance and capital committees as project shops that delivered something with a beginning, a middle, and an end. I clearly remember fighting battles a decade ago with finance departments who insisted IT projects had to have a discrete close-out or they could not approve or track the investments. The shift from an “IT organization” to a “Technology Delivery organization” capable of co-owning living systems with operations came late, was ill-understood, and is often still incomplete. Operations teams that had spent decades being let down by IT had no reason to trust even these new tech teams now. The phrase that justified that resistance was “we don’t have time to change.”

The pattern that produces successful technology transformations in other industries is well known. Operations and technology co-own the work. The technology is treated as a living system rather than a discrete project. Finance funds capability rather than feature delivery. None of those preconditions has held in media operations long enough or consistently enough to produce a durable foundation. Until they do, the next wave of technology will land on the same hollow ground regardless of what the wave is.

The second was outsourcing. Media companies that did not want to build internal technology had another option, and they’ve used it for decades. They handed operational complexity to specialty vendors. Technicolor, Deluxe, Vubiquity, and many others absorbed work the media companies chose not to own. It let them delay the underlying investment by paying someone else to live with the consequence. The perceived ability to hold the vendor accountable was a fallacy. A failed delivery still failed, and legal recourse was monetary rather than operational. Then, in early 2025, Technicolor itself was liquidated globally after more than a hundred and ten years in business. Thousands of jobs gone. In-flight releases scrambling for substitutes. The remaining vendors struggled to fill the gap and the media companies finally started to look at in-sourcing. The industry’s most visible escape hatch partially collapsed, and the foundation it had been propping up was suddenly responsible for bearing the load.

The third is the one we are in now. Leadership that did not invest in or partner with internal technology, and that handed operational complexity to outside partners, is either reaching for GenAI and Agents as the next way to avoid the foundational work, or pretending those tools are not the future. GenAI and Agents are clearly the future of operational automation. Decades of underinvestment, broken partnership, and outsourced capability created the gap. GenAI and Agents will not close it on their own. The same leaders are making the same bet for the third time. While investment has definitely showed up in some organizations, the partnership and the trust deficit have not caught up. The phrase that describes this is “we have to keep the plane in the air while we change the engine,” which sounds responsible until you notice it has been the justification for not changing the engine for the entire life of the airline.

What the Agents are actually inheriting

At the execution tier, the Agent inherits the runbook but none of the reasoning behind it. It reads the runbook the same way the human did. Confidently, but without the comprehension built over years of habit. The difference is the human had an accumulation of feedback from years in the operation that occasionally surfaced the gap between the runbook and reality. A rejected asset, stalled workflow or a partner complaint. The Agent does not have that accumulation of feedback to build upon. It then executes that imperfect understanding at speed, and the throughput problem shows up at scale before anyone notices.

At the senior tier, the Agent inherits nothing. The leather notebook is not in any system. The Tuesday phone call about vendor capacity is not in any system. The routing knowledge in one PM’s head for ten years is not in any system. When that PM is asked to do something else with her time because an Agent now handles routing, the routing decisions stop being good.

The Agent is also inheriting workflows that suffer from the ongoing failure of two of their three former crutches. Internal Human Automation continues to be cut. The outsourcing market has just demonstrated that even the largest specialty vendor in it can disappear. Both buffers are weaker than they have been in recent memory. That is the operational risk nobody senior is talking about.

Andy framed the creative-side resistance around four questions about who benefits, who is retained, who decides, who owns. The operational parallel is similar: who carries the context, who notices when the Agent is wrong, and who is accountable when the green status stops being green.

The teams being asked to live through this transition are legitimately concerned. The fear of job loss is real. The distrust of the technology is earned, because the same leadership now pointing at Agents has been promising better tools for decades and not always delivering them. The commonly held instinct that the data has to be perfect before the tools are useful is technically wrong but emotionally honest, because the price of the tools being wrong has historically been paid by the operator, not by the organization that deployed them. Treating that resistance as irrational misunderstands the problem.

Step back, then tool it

The way out of this is not faster automation. It is the work the industry has historically declined to do. Step back and define what the process should be, not what it has accidentally become. Then build the foundational technology that lets demand, task, and state live in systems rather than in people. That foundation is concrete: demand management systems that make incoming work legible to automation, task management that holds in-flight state outside someone’s head, process orchestration that runs without the person who understands it in the room, and operational telemetry that measures what is actually happening rather than what the dashboard claims. Build the Agents on that foundation, where they can actually be useful, rather than on top of a fifteen-year-old runbook and systems nobody can defend.

The most credible Agent deployments I have seen recently look less like one Agent doing one job and more like what technologists have started calling an Agentic mesh: specialist Agents, loosely coupled but tightly integrated, orchestrated into a workflow. The pattern works for the same reason microservices worked. Smaller units fail in isolation, can be tested and replaced without rebuilding the whole, and surface their foundation requirements explicitly rather than hiding them inside one large black box. The mesh approach requires more from that foundational tech and process, not less. Each specialist Agent needs a system-legible view of demand, task, and state to do its job. The mesh does not solve the problem the runbook created. It makes the problem easier to see.

There is a rule I have repeated for years, and it gets sharper in this moment: you should never tool a bad process. Agents do not make that rule less true. They make it more expensive to ignore.

Start with the diagnostic

The diagnostic is clear and approachable. Inventory where your in-flight work state actually lives. Not where you wish it lived, not what the dashboard reports. Where would the next person taking the role have to look? If the honest answer involves a spreadsheet, an email thread, or a senior PM’s memory, that is the foundation gap, and no Agent will close it for you. For vendors and investors, the same question applied to the companies you work with. Their answers will tell you more about their readiness than any AI roadmap they share.

The wrong answer in 2026

Andy’s Human Layer series asks whether the person inside the work survives the transformation. Whose decisions are in the output. Whose face is in the frame.

The operational Human Layer asks a quieter version of the same question. Whose judgment was holding the process together. Whose context was the reason the dashboard reported green last quarter. Whose absence will be revealed by the next bad status report.

Both questions have the same wrong answer in 2026, and the industry is converging on it at speed. Pretend the human layer was not there. Pretend the runbook was the process and the leather notebook was inefficiency. Pretend the variation humans absorbed every day for decades was somehow being absorbed by the systems all along.

It wasn’t. It isn’t now. The Agents are arriving to try to fill a gap that should have been closed by foundational work the industry chose not to do. The people who held that gap closed for decades are the people the industry is calling inefficient and trying to move past. When the Agent is wrong, no one will be left to say why. When it needs direction, no one will be left to give it.

Years of Human Automation in the media supply chain are not replaced by an Agent in a day. They are replaced by the work the industry chose to gloss over, ignore, or not fund. Fix that and the Agent is the part that comes after.

Sources and further reading:

• Andy Beach, Human-Enhanced is the New Automation, Identity as Infrastructure, Observability as Default.

• MIT NANDA, “The GenAI Divide: State of AI in Business 2025,” reported by Healthcare IT News. Finding: 95% of enterprise GenAI pilots fail to deliver measurable ROI. Methodology: 150 leader interviews, 350 employee survey responses, 300 public deployment analyses.

• Lisanne Bainbridge, “Ironies of Automation,” Automatica, 1983.

• Bloomberg, “The Collapse of Technicolor.”

• Variety, “Technicolor Bankruptcy: How The Mill’s U.S. Team Launched ARC Creative.”

• McKinsey QuantumBlack, “How We Enabled Agents at Scale in the Enterprise with the Agentic AI Mesh.” For the multi-agent / agentic mesh pattern referenced in the “Step back, then tool it” section.

The Upfronts said the same thing

Netflix’s 2026 Upfront, delivered May 13 by President of Advertising Amy Reinhard, included the announcement that the company is testing AI agents that can autonomously manage, optimize, and purchase ads on the platform. The framing matters. These are not media-planning tools that recommend buys for human approval. They are agents that execute. Netflix described them as already in test with brands like DoorDash, Target, and TurboTax. The…

Read more

Around the clip, an argument forms. One side surfaces the manifest as evidence. Look. It has a signature. It came from a camera. It hasn’t been edited. The other side responds that the manifest does not address whether what is depicted in the footage happened the way the surrounding caption claims. Both sides are now demanding the wrong thing of the technology. The audit trail is intact. The truth claim is unresolved. The two get conflated, and the public conversation collapses into noise that the architecture was never designed to settle.

This is the observability question playing out at the worst possible scale, in front of an audience that does not yet have the vocabulary to distinguish chain-of-custody attestation from truth validation. The infrastructure is doing its job. It was never built to do the job people are now assigning to it.

This is the third and final piece in the Human Layer series, and the closing of the larger arc that began with Context Is the New Integrity Layer.

The category mistake

Olga Kornienko, on the Future Frames podcast, named the distinction in a single line: “C2PA is not gonna argue if I say that my wall is pink.” She is the COO of EZDRM, which has been doing the operational work of making content provenance actually function at the asset level for several years.

The pink wall framework is the load-bearing distinction for everything else in this piece. C2PA — the Coalition for Content Provenance and Authenticity — validates that a thing was made in a particular way, by particular tools, with particular edits. It does not, and structurally cannot, validate whether the claim attached to that thing is correct.

The audit trail is not the argument. The audit trail is the precondition for the argument.

Read more

That same week, the company is still tagging AI-generated music, demonetizing fraudulent streams, and backing the new DDEX standard for AI-use credits. The signaling is loud: AI music is a problem to be governed. AI audio in the next breath is a feature to be shipped.

These are not contradictions. They are the same rule applied to two different economic surfaces. AI is welcome wherever it does not break the existing model. AI is contested wherever it does. The fight is not about authorship, taste, or even consent. It is about which contractual surface absorbs the cost.

The fight that already moved

For two years, the AI music conversation pretended to be about audiences. Would listeners accept it? Would they tell the difference? Would human artistry survive the flood? The data settled most of those questions quietly. Luminate’s most recent report shows U.S. consumer sentiment toward AI in music creation moving from -13% in May 2025 to -20% by November, with the sharpest decline among Gen Z and Gen Alpha. Deezer is now receiving around 75,000 fully AI-generated tracks every day, roughly 44% of daily uploads, while consumption sits at 1 to 3% of total streams and 85% of those streams get flagged as fraudulent.

Listeners did not vote for this. They are net-negative on it. The supply curve and the licensing curve kept moving anyway.

Because the actual fight had already migrated. Universal settled with Udio in October 2025. Warner settled with Udio a month later, then with Suno a week after that, picking up Songkick in the bargain. Suno raised $250 million at a $2.45 billion valuation with NVIDIA participating. Merlin signed in January 2026. Kobalt followed in March. Sony is the lone holdout from the original lawsuit; everyone else is already inside the licensing economy they spent eighteen months claiming was illegitimate.

The piece of this most coverage misses is that the labels did not win the listener fight. They abandoned it. They stopped trying to convince anyone that AI music was bad and started building the contractual surface that would let it exist on terms they could meter. Audience preference was the rhetorical shield. The real prize was the metering rights.

Two doors, same building

Spotify’s split posture this week is the cleanest possible expression of where this lands. The royalty pool is a defended door. AI tracks that try to walk through it get tagged, demoted, and demonetized, because the economics of pro-rata payouts mean fraudulent streams pull money out of real artists’ pockets at industrial scale. Music had to fight its way to a deal because it was inside the royalty engine. The rules are about defending what is already metered.

The Personal Podcast door is wide open. AI-generated audio walks straight in, gets stored in a private library, and never enters search, recommendations, or the royalty graph at all. There is no metering surface to defend because there is no shared revenue pool to dilute. It is your library, your prompt, your file. Spotify is the playback layer, not the marketplace.

Same company. Same week. Same coherent rule: AI is welcome wherever the existing economic model is not at risk.

That is not hypocrisy. It is the new operating principle for any platform with a royalty obligation. Once you see it, the rest of the industry’s behavior becomes legible. Why labels licensed Suno but Sony has not. Why Deezer can transparently tag AI music as a commercial feature. Why YouTube can tolerate enormous volumes of AI content while Spotify aggressively filters it: their economic geometries are different, so their tolerances are different. The question every platform is now answering, whether they say so out loud or not, is which surfaces of their stack can host AI without breaking the contracts that pay creators.

The clearest contrast is Bandcamp, which banned AI-generated music outright in January. Bandcamp can do that because its revenue model is direct-to-fan purchase, not streaming royalties. There is no pro-rata pool to dilute, no metering surface that AI uploads can game. Banning becomes a clean editorial choice rather than an economic one. Spotify does not have that option. The ban is available to one and not the other.

Where the leverage moved

This is the part that is not yet settled, and where the next fight is forming.

If the listener was the wrong battlefield, and if the licensing rooms have mostly rationalized into deals, the open question is who controls the surface where the audio actually gets made. Spotify’s CLI is a tell. The integration target is the agent, not the consumer. A developer writes a prompt. An agent generates the file. Spotify ingests it and renders it. The user is downstream of three steps that used to be invisible.

That changes the economics in a direction the music industry has not had to think about yet. In the royalty model, the unit of value is the stream and the contractual counterparty is the rights holder. In the agent model, the unit of value is the generated artifact and the counterparty is the agent vendor. Different metering surface. Different leverage. Different fight.

And this is not a Spotify story. The CLI is essentially browser compatibility for the agent era. Every platform with a metered surface is about to face the same compatibility question, because so much consumption is moving through agents that fetch and assemble on behalf of the user. Whoever ships first becomes a default integration target. Whoever waits gets routed around. The agent is the new browser, and platforms are about to spend the next eighteen months deciding which of their surfaces are addressable through one.

Right now Spotify is positioning itself as the listening-surface gravity well. Every agent that can drop audio into Spotify becomes a feature for Spotify. But that bargain runs both ways. If agent-generated audio becomes a meaningful share of listening hours, and if agents become the way users compose their day, then the agent vendor sits between the platform and the user. That is the exact place where Apple sat between developers and iOS users for fifteen years, and where YouTube sat between creators and audiences for the streaming era. It is the place that turns into the rent-collecting layer once volume arrives.

Where this actually ends

The AI music story stopped being a story about creativity a while ago. It is a story about which surfaces of the media stack can absorb generative content without breaking the economics underneath them. Listeners showed up in the prep doc and the trade press as the moral center of the conversation, but the data shows they were never the binding constraint. The binding constraint was always the contract.

The equilibrium probably arrives from the listening side, not the supply side. Platforms have already started shipping the primitives. YouTube recently let users set Shorts duration to zero and effectively turn the format off. The AI-content version of that switch is coming, and once it ships, the economic incentive that drove the flood collapses. Spam works because it is cheap to make and the platform delivers the audience. Take away the audience and the math stops penciling out, regardless of how cheap generation gets.

The same agent layer that lets Spotify ship Personal Podcasts is what eventually arms the consumer to refuse the parts they did not ask for. An agent that can generate audio on demand can also filter what shows up in the feed by whatever criteria the user sets. Sounds too AI. Mimics an artist I like. Lacks credited human contribution. The filter does not need to be perfect. It needs to be cheap, and it will be.

That is the loop that closes this. Platforms get to ship AI audio without breaking their royalty contracts because the new audio lives outside the metered surface. Listeners get to keep AI out of the parts of their listening they care about because the agent does the sorting. The labels keep the licensing economy they just rebuilt. Everyone gets a version of what they wanted.

The fight that is left is the one nobody has named yet. When the agent is the integration layer for ingestion and the filter for consumption, the agent vendor sits at both ends of the pipe. The labels know how the meter-the-supply game ends because they have played it before. The meter-the-listener game is new, and the contracts that would govern it do not exist.

The performer might be a verified, consenting subject. They might be a stylistic echo trained on someone whose name is in the model’s training set but whose name is nowhere in the platform’s output. The system has no resolvable answer because the system was not built to have one. The contract that covered the production sits in someone’s email. The likeness rights, if they exist, sit in a separate database belonging to a different entity. The performer’s identifier, assuming they have one, was not requested, was not stored, and would not have been resolvable from inside the workflow even if it had been.

Once agency is observable, the next question is whose agency. That was the bridge from the last piece. It is also the question most of the industry’s rights, contract, and credit frameworks quietly assume has already been answered. It hasn’t. There is no resolvable identity layer for the people inside generative media, in any form a workflow can query and trust. The asset has an identifier. The performer in the asset usually does not.

This gap does not exist because the industry hasn’t thought about it. It exists because the industry built identity infrastructure for assets fifteen years ago and never built the equivalent for people.

The industry has done this before, for assets

The Entertainment Identifier Registry has been operating since 2010. Over two million records. Founded by MovieLabs, CableLabs, Comcast, and TiVo. Widely adopted across major studios, streamers, and broadcasters, and used by PBS at the network level. It is built on the same Digital Object Identifier system that runs academic journals and scientific data, governed by the same DOI Foundation, resolvable through the same handle infrastructure originally designed by Bob Kahn.

What makes EIDR work is what it deliberately doesn’t do. The registry is intentionally thin. It holds the minimum metadata required to disambiguate one piece of content from another: title, language, release date, country, length, type. It does not hold ownership. It does not hold rights. It does not aggregate commercially valuable metadata about the asset. It is purely functional, and persistent enough to survive ownership changes. When an asset is sold, transferred, recombined, or re-edited, the identifier travels with it.

That design philosophy is the load-bearing point. EIDR works because it solved one problem cleanly and let the value layers sit on top. Disambiguation became an industry-wide service that nobody owns and everybody benefits from. The studios accepted, fifteen years ago, that some kinds of identity infrastructure cannot be a competitive advantage without breaking the system that produces the advantage. Make the registry neutral. Make it persistent. Make it governed by a body whose only job is to run the registry. Build the commercial systems on top.

That principle, accepted for assets, has not been accepted for people. The asset has had a resolvable identifier for fifteen years. The performer in the asset has not. The voice in the dub, the writer of the script, the director of the cut, the choreographer of the movement, the musician on the underscore: none of them have. Their identity has lived in contracts, payroll systems, IMDb pages, agent databases, union rolls, and credit slates. None of those are resolvable. None of those persist through ownership changes. None of those federate.

Read more

What’s moving isn’t the technology. It’s the economic primitive underneath it. Per-seat pricing, the model that has organized white-collar work, enterprise software, and broadcast funding for decades, is structurally incompatible with per-token costs. The institutions built on the old logic are compressing in real time, and the receipts are coming in from every layer of the stack at…

Read more