Episode 193: Adam Epstein on Muse & Agentic Commerce, Plus TTD Blocked on Safari and Amazon UnBoxed

Full analysis

Apple quietly added The Trade Desk's actual ad-serving domain, adserver.org, to Safari's block list in iOS 27. The block covers the ads themselves, not just the tracking pixel. On an updated iPhone, TTD bids drop to zero. Google's DoubleClick keeps serving. That's the story that matters most this week, and it's part of a bigger shift: the companies that own the pipes (Apple, Amazon, Google) are deciding who gets to reach a user and who doesn't, and the AI-agent layer everyone's building sits downstream of that decision.

How hard is this to undo? For a reader buying ad-tech or building agent features, mostly not your call to undo. But the lesson is cheap to act on: don't build a business that depends on one gatekeeper's goodwill. What's actually being decided across these stories isn't "is Muse good" or "is Amazon's new ad unit clever." It's who controls the path between a consumer and a purchase, and whether AI agents break that path or get absorbed into it. What sets the deadline: iOS 27 adoption. TTD's exposure climbs from roughly 5-7% of US traffic today toward 20% as people update their phones. That's months, not years.

The Skeptic

The agent-commerce hype collapses the moment you watch real people use these things. Ari Paparo and Adam Epstein both tested Meta's Muse on actual tasks. Normal consumers aren't using it. The barrier is obvious: nobody wants to hand an AI agent their insurance login or bank account. Epstein's own data point undercuts the dream. The Instinct CEO says 60% of their transaction volume is travel bookings. Travel is the one place people already trust a middleman to hold their card. That's not "agents transform commerce." That's "agents are a slightly better Expedia." And Epstein's $80 billion Amazon ad figure, with 65% on-site? He flagged it himself: "a very, very rough estimate. No sources." He sells Amazon ad optimization. Of course he thinks Amazon's ad moat is huge.

The Researcher

Strip the conference gloss off Amazon Unboxed and you get one real mechanism. Amazon blocks Muse and ChatGPT from shopping because its on-site sponsored ads only work if humans browse the site. An agent that buys the cheapest detergent and leaves skips every sponsored slot. So the block isn't about safety or data fidelity, whatever Amazon says publicly. Epstein called that reasoning a negotiating posture, and the evidence backs him: Amazon simultaneously announced advertisers can buy ChatGPT ad inventory through its own buying tool. Amazon will let agents in the day they carry ads Amazon can sell. The TTD-on-Safari story has the same shape. Apple added hundreds of programmatic data companies to the block list per AdExchanger. Google's engine, Google's ad server, no block. Apple's WebKit makes the call, not Google, so this is less favoritism than a classification choice that happens to spare the integrated giant.

The Compute Pragmatist

The Anthropic S1 numbers are the real macro story, if the leaked figures hold: revenue up 12x to $4.6 billion, an $8 billion operating loss, and $518 billion in compute commitments. Read those together. Anthropic is promising to spend more than a hundred times its current revenue on chips and data centers. Two customers drive nearly a quarter of sales, and there aren't many long-term contracts locking that in. Ari Paparo's line, "our entire economy is bet on this," isn't hyperbole when one lab commits half a trillion to compute. For anyone buying AI, this is your supply chain. The models you build on run on economics that only work if usage keeps compounding at a rate no software company has ever sustained. Price your roadmap as if inference costs could move either way, because the people absorbing $8 billion losses will eventually want it back.

The Builder

If you ship anything that touches digital advertising, check your Safari exposure this month. If your demand flows through The Trade Desk, 5-7% of US impressions are already dead on updated iPhones, heading to 20%. TTD apparently didn't tell its own clients. That silence tells you where you sit in their priority list. The WebKit question decides your next move: if the block was a mistake, TTD patches it with a new domain and you wait. If it was intentional, you're in a game of whack-a-mole where Apple blocks each new domain and TTD never wins. Plan for intentional. On the agent side, the practical signal is the plumbing Epstein described: brands like Marriott needing to run MCP servers, the connectors that let an AI agent actually book a room and take payment, plus listing fees to rank well. That's the real work if agent commerce arrives. Boring integration, not magic.

The Open-Source Advocate

The piece everyone's missing: this whole fight is about closed gates, and the open web is the collateral. Kurt Donnell's Signal & Noise conversation asks whether the open web survives the AI age. These stories are the answer in motion. Apple decides who serves ads in Safari. Amazon decides which agents shop its catalog. Meta cross-promotes Muse across Instagram, Facebook, and WhatsApp, three properties nobody else can touch, and then poaches MongoDB's sitting CEO, CJ Desai, with no transition to run its enterprise push. Every one of these is a walled garden getting taller. The independent ad-tech and data companies, TTD, LiveRamp, ID5, Permutive, Audigent, all blocked at once, are the ones who built on the open web. OpenAI's answer, the LiveRamp partnership to bring custom audiences into ChatGPT ads, is just Facebook's 2018 playbook on a new surface. New gate, same owner problem.

Where they disagree

The Skeptic and the Researcher split on how scared to be about agents. The Skeptic says agent commerce is years away because normal people won't hand over sensitive logins, so Amazon's block is defending against a threat that barely exists. The Researcher says it doesn't matter whether agents work yet. Amazon blocked them anyway, which tells you the gatekeepers see the threat clearly and are moving early. Both can be right: the threat is slow and the defense is already up.

The Compute Pragmatist and the Builder split on what to actually do. The Pragmatist is staring at Anthropic's half-trillion compute bet and telling you the whole model economy could reprice. The Builder is saying forget the macro, your Safari revenue is bleeding this quarter, go check it. The near-term problem is concrete and dated. The macro one is huge but vague.

What it hinges on

For the TTD story, one fact settles everything: was the Safari block intentional or a mistake? WebKit hasn't said. If intentional, independent ad-serving on Safari is structurally broken and no clever domain swap fixes it. For agent commerce, it hinges on whether consumers ever trust agents with payment and sensitive accounts outside travel. Right now the evidence says no. For the broader AI economy, it hinges on whether Anthropic-scale compute bets pay back, and nobody reading this gets to verify that in time.

The council leans one way clearly: the gatekeepers are winning, and independent players downstream, whether in ad-tech or agent commerce, are the ones exposed. Verify your own exposure. If you run ad demand through TTD, measure the Safari drop now and ask TTD directly whether adserver.org is getting a replacement. If you're building agent features that transact, assume the big platforms will block you until you carry revenue they can sell, and design for that from day one.

Prediction: Apple will not remove The Trade Desk's adserver.org domain from Safari's iOS 27 block list before Apple's WWDC in June 2027, confirming the block was intentional policy rather than an error.

Confidence: Medium. Apple blocked hundreds of data companies at once. That's a policy, not a bug.

Why: AdExchanger reported Apple added not just TTD but LiveRamp, ID5, Permutive, and Audigent to the iOS 27 block list in the same move, while Google's DoubleClick served unimpeded. A one-off mistake blocks one domain; blocking a coordinated set of the open web's identity and ad-serving players matches Apple's multi-year direction since App Tracking Transparency, where it has steadily cut third-party tracking while sparing integrated first parties. The opposite outcome, a quiet patch restoring TTD, would require Apple to reverse a stance it has defended as privacy for five years, which it has never done under industry pressure. If Apple wanted to signal error, WebKit could have said so in the weeks since; the silence favors deliberate.

Revisit by 2027-06-30: We're right if adserver.org remains blocked on updated Safari at WWDC 2027 and TTD bids still drop on iOS 27+ devices. We're wrong if Apple removes the domain, publicly calls it a mistake, or WebKit restores TTD ad serving before then.