Control Points Are Capturing Value Across the AI Economy

John Coogan treated SK Hynix’s Nasdaq debut as a measure of where value is accruing in the AI buildout: not only to model makers and accelerator designers, but to suppliers controlling scarce infrastructure around them.

The South Korean memory-chip maker raised $26.5 billion in what Coogan described as the largest first-time US share sale by a foreign company. Demand reportedly exceeded available shares by more than seven times. Its American depositary receipts opened at $170, against an offering price of $149, a 14% gain that Coogan considered relatively restrained only because of the offering’s exceptional size. The debut put SK Hynix’s market capitalization at roughly $1.03 trillion.

Coogan linked the valuation to high-bandwidth memory, used alongside advanced AI accelerators. SK Hynix, Samsung, and Micron dominate the global memory market, and he described memory as one of the AI boom’s critical bottlenecks. SK Hynix’s Korean-listed shares, he said, had risen more than 600% over the preceding year as investors sought exposure to that constraint.

The US listing matters because it gives American investors a more direct route into a company that had previously required buying through South Korea’s KOSPI market. Its initial US ticker, SKHYV, was temporary; the “V” indicated when-issued trading, allowing investors to trade the depositary receipts before the transaction formally settled and the shares were fully delivered. The permanent ticker was set to become SKHY.

The same concern with control points shaped the hosts’ discussion of Meta. Jordi Hays rejected a popular explanation that Mark Zuckerberg’s AI spending is principally a response to a permanent 30% Apple “tax.” That framing did not fit Meta’s core ad business, they argued: advertisers generally use web-based ad managers, while impressions and human attention are not app-store digital goods.

There may be narrower cases. Coogan noted that a small business boosting an Instagram post through an in-app payment flow could potentially encounter app-store economics. That is different from buying a physical product through Amazon, DoorDash, or a movie-ticket app, where Apple does not take the same cut. Digital subscriptions, virtual currency, and other in-app digital goods are the clearer cases.

But Hays argued that the strategic issue is dependence, not a particular commission. Meta is one of the world’s largest companies, yet it still operates through an ecosystem controlled by Apple—an ecosystem where, in his view, Meta cannot assume sympathy or predictability. An audience comment characterized the more consequential Apple cost as the loss of cross-platform tracking; both hosts agreed that Apple’s permissions prompt had made Meta’s advertising business less effective.

For Hays, AI and glasses are an attempt to avoid repeating that position in the next computing transition. Consumers might eventually spend less time moving among individual social platforms and more time asking an agent to retrieve information, surface content, or act on their behalf. Meta’s glasses, on that account, are a platform bet: cheap or even free devices that act as a phone and computer, mediate commerce, and create a direct relationship between Meta and the user.

He imagined glasses that could identify an item someone sees and help buy it, or surface an offer while someone walks past a restaurant. The commercial point is not just an assistant that answers questions; it is control over the interface through which recommendations, transactions, and attention flow.

Coogan said Meta’s current AI spending could also make hardware investment easier to absorb. In a capital plan discussed in terms of roughly $200 billion for data centers, an additional $8 billion on a future version of Meta Ray-Ban displays might look comparatively small.

His own behavior was, he said, an early warning for consumer platforms built around conventional screen time. He has shifted more time toward ChatGPT and deep-research reports and away from Instagram. Seeing an interesting car video may now lead him to ask an AI system questions rather than undertake the chain of searches and document gathering that research once required. Wherever people spend time, Coogan argued, consumer companies need a foothold.

The GPT-5.6 discussion centered less on benchmark-style novelty than on the growing overlap between creative tools, coding systems, image generation, and computer use. John Coogan argued that the useful question is increasingly not whether a model can produce an impressive demo, but whether users can recognize a task worth giving it and can fit the result into a controlled process.

An AI-generated recreation of an Interstellar scene demonstrated the speed with which a user could move from prompt to 3D model to rendered video. Coogan found that technically impressive but creatively incomplete. A faithful recreation of Interstellar merely leaves the creator with a version of a movie that already exists. The contribution has to be a new story, a deliberate aesthetic, or a premise that leans into the technology’s imperfect edges.

He saw more promise in workflows where creators supply structure before asking a model for fidelity. People are filming themselves delivering dialogue and using that footage as driving video for AI models. Others are blocking out driving or mechanical scenes in Blender, then using an AI video model to turn a physically coherent but low-fidelity reference into a more polished image. The reference provides deterministic motion and spatial logic, instead of requiring the model to invent every relationship from a text prompt.

A separate demonstration had a model use Blender to build a 3D cannon from primitives such as cylinders, spheres, and toruses. Hays emphasized that this was ordinary computer use, not a purpose-built Blender integration: the system was operating familiar software at a speed a person could not sustain.

Coogan used OpenAI’s example of a broccoli farmer using GPT-5.6 to make that point. Research and coding are recognizable use cases, he said. A more unusual example can still be valuable because it helps someone imagine an analogous task in their own life. The interface may remain a blank box asking what a user wants to build, know, or do; a model can improve substantially without changing that box. Users need examples that expand their idea of what belongs in it.

The systems are also becoming harder to separate by function. Packy McCormick said GPT-5.6 had created a “NOT BORING” logo while designing a website from a Cosmos link, despite not being asked to make a logo. Coogan’s explanation was that a system building a site might generate a preview image, while image models can themselves use tools and scaffold HTML. What users experience is not a clean sequence of coding model, image model, and design tool, but one blended system.

That convergence has made product consolidation seem inevitable—and made the transition difficult. OpenAI combined Codex and ChatGPT in a new desktop application while allowing users to retain the older program as “ChatGPT Classic.” Coogan said the change disrupted people who had developed workflows around the native Mac app and its integrations. The new application is built in Electron, he said, relying on web views rather than the native architecture users expected.

Hays relayed an OpenAI response acknowledging that the company had not “got everything quite right.” It had made high-compute settings too easy to invoke without clearly communicating the effect on usage limits; its reorganization made familiar features such as chats and projects harder to find; and launch language around “Work” led some Codex users to worry that Codex was being phased out. OpenAI said Codex was staying and reset usage limits twice that day to let people continue experimenting.

“I just want one place to do things,” Hays said. Coogan agreed with the destination: an omnibox capable of answering questions, writing code, building projects, and acting through tools. The difficulty is that a unified interface creates value only if it reduces user effort rather than forcing people to rebuild habits, shortcuts, and workflows.

Coogan’s own arrangement remains fragmented: Codex on the desktop, ChatGPT in a browser for conversation, and ChatGPT on a phone that can communicate with Codex. He rarely used the older desktop ChatGPT app, but he acknowledged that this was not the point. Asking users to bind a new shortcut or reconstruct a familiar workflow is still asking them to do work.

John Coogan described Bloomberg’s reporting on Phia, the shopping app co-founded by Phoebe Gates, as an affiliate-attribution dispute with a serious allegation at its center: that Phia claimed credit for online sales it did not actually drive.

Bloomberg tested Phia’s mobile browser extension across more than 50 websites and found that, during checkout, it opened a background tab without user interaction and inserted its own referral code, overriding a legitimate referral from another publisher, according to the report discussed by the hosts. Bloomberg’s findings were described as consistent with Capital One Shopping’s independent testing and code review by Edelman.

Affiliate marketing normally assigns a commission to the publisher that sends a customer to a retailer. If a review site, creator, or other publisher paid to acquire the audience, explained a product, and referred a buyer onward, that publisher expects the referral credit. The allegation is that Phia could enter at checkout and assign the commission to itself instead.

Jordi Hays stressed that the hosts did not know whether the behavior was accidental or intentional. But he said it could be more predatory than the conduct associated with Honey if it occurred without any visible interaction from the shopper.

Coogan compared Bloomberg’s account with YouTuber MegaLag’s investigation into Honey. Honey was accused of using a hidden redirect near checkout to replace creators’ affiliate tags. That created an irony: creators could promote a product and promote Honey, only for Honey to capture their commission when viewers bought that product. Coogan noted that Honey’s broader controversy included allegations of suppressing better coupons and evading affiliate-network stand-down checks.

Nothing comparably extensive had been established about Phia, he said. The crucial distinction was the interaction requirement. Honey generally created the overwrite after a shopper interacted with its checkout prompt, even if the interaction was simply dismissing it. Bloomberg’s reporting said Phia’s extension could fire with no Phia interaction at all.

Coogan illustrated the commercial harm with a hypothetical chain. A publisher he called horsereviews.com pays for Google ads, educates a shopper about the horse market, and sends the shopper to “Clydesdales Direct.” If Phia is installed and silently inserts its code at checkout, the retailer may pay Phia rather than the publisher that created the demand. The retailer still has a customer and may not care which affiliate receives the payment. The originating publisher does.

Hays speculated that if this behavior began with a December or January software update, it might have been operating for around six months. Coogan said the diffuse nature of affiliate accounting could make that hard to catch. He cited more than $12 million in attributed sales and estimated the related commission at roughly $1 million. Across many merchants and publishers, a few hundred dollars moved from one affiliate to another may not stand out—particularly to a business managing many referral relationships.

John Coogan described Hollywood’s pursuit of internet-native horror properties as another attempt to buy a proven distribution channel rather than develop one from scratch. Studios are looking for concepts that have already created recognizable images, fan communities, derivative videos, games, and fear across YouTube, Reddit, Roblox, and related online spaces.

Trevor Henderson’s Siren Head is one example. Henderson drew the tall, faceless monster with sirens for a head eight years earlier. It spread into YouTube films, knockoff Amazon merchandise, and video games, but he had not made money from the creation until now, according to the reporting discussed. The 40-year-old illustrator reportedly sold the film rights to Warner Bros. for more than $1 million.

Coogan found the path striking: an illustrator creates a strange online monster, watches it become a broad cultural object, and eventually converts that established attention into a studio deal. Warner Bros. Motion Picture Group co-chair Michael De Luca described online properties as adaptation resources “the same way we look at books and other media.”

The Mandela Catalogue, a psychological-horror YouTube series, offers a more direct indication of competition. Eleven studios reportedly bid for its film rights. Amazon-owned United Artists and Steven Spielberg’s Amblin Entertainment ultimately agreed to pay millions of dollars and let the series’ 22-year-old creator, Alex Kister, direct the film.

Hays said he has never been drawn to horror, partly because he believed avoiding it would make him less afraid while walking around at night. The studios’ bet is the inverse: online horror has already demonstrated that it can make audiences care, participate, and return. The valuable asset is not merely an idea that can be adapted, but an audience relationship that has already been built elsewhere.