> ## Content Index
> Fetch the complete content index at: https://adjacent.media/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Adjacent Brief — September 23, 2026
- URL: https://adjacent.media/briefs/2026-09-23/
- Published: 2026-09-23T14:15:03.000Z
- Updated: 2026-09-23T14:15:04.000Z
- Description: Emarketer put commerce media on track to take one in four digital ad dollars, with retail networks pushing inventory well past the store shelf into streaming and social.
- Author: Jonathan Greene
- Tags: #brief

**TL;DR:** Emarketer put commerce media on track to take one in four digital ad dollars, with retail networks pushing inventory well past the store shelf into streaming and social. Wall Street analysts began questioning the financing structures behind the data center buildout, and a group of AI startups — Harvey, Abridge, Ramp, Rogo — said they are shifting work onto open-weight or self-trained models to cut what they pay frontier labs.

## Worth Reading

- [Harvey, Abridge and Ramp are moving off the frontier labs](https://www.bloomberg.com/news/articles/2026-09-21/startups-like-harvey-embrace-open-models-to-cut-reliance-on-anthropic-openai?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc5MDAxMDkwMiwiZXhwIjoxNzkwNjE1NzAyLCJhcnRpY2xlSWQiOiJUTFEyMFpOM04wQTMwMCIsImJjb25uZWN0SWQiOiI5MDI3MkIwQzRFRjQ0MkUxOUZCQkFBRTdBODJGMjM3NSJ9.i7tIzBBJutWRddw95G5V52NMXfBwhYa2v52uZC8eJ-g&ref=adjacent.media) (paywall) — Bloomberg on the vertical AI companies that discovered their gross margin lives inside someone else's API bill.
- [Agentic commerce has a demo problem: Visa, Mastercard, PayPal and Stripe have zero live merchants](https://www.searchenginejournal.com/agentic-commerce-is-not-the-next-voice-shopping-but-it-is-still-unproven/589439/?ref=adjacent.media) — Google has three. Every other announcement is a roadmap.
- [Shopify's CEO has a name for unreviewed AI output — "slop grenades"](https://www.searchenginejournal.com/shopify-ceo-ai-slop-grenades-coworkers/590292/?ref=adjacent.media) — The cleanup labor is already showing up as paid postings on freelance marketplaces.
- [Forrester says the enterprise software business model is up for renegotiation](https://www.forrester.com/blogs/the-future-of-enterprise-software/?ref=adjacent.media) — Seat-based pricing assumed seats. Worth reading next to today's professional-services item.
- [Singapore will pay citizens two cents a day to read](https://boingboing.net/2026/09/21/fountain-pen-singapore-readsg-coins.html?ref=adjacent.media) — A state treating attention as a public health problem, in the same week BLS logged the lowest US pleasure-reading rate on record.
- [The iPhone 18 Pro's camera module is about as unrepairable as hardware gets](https://www.theregister.com/personal-tech/2026/09/21/skip-the-upgrade-iphone-18-pro-aperture-about-as-unrepairable-as-it-gets-ifixit-finds/5297970?ref=adjacent.media) — iFixit's teardown, arriving while right-to-repair rules tighten in two markets Apple sells into.
- [Sherry Turkle: chatbots are deskilling human empathy](https://www.404media.co/sherry-turkle-artificial-intimacy-podcast/?ref=adjacent.media) — AI companions aren't the problem. The practice of sitting with another person atrophies without reps.

## Brand & Growth

**Partner programs were built for a buyer who no longer shops that way**

Forrester's argument is narrow and useful: B2B partner demand programs — the co-marketing webinars, the syndicated content plays, the MDF-funded lead gen — were designed for buyers who researched across a dozen vendor touchpoints before ever talking to sales. When buyers route discovery through AI assistants, [the intermediary layer partners were selling gets compressed](https://www.forrester.com/blogs/ai-has-changed-b2b-buying-partner-demand-programs-must-follow/?ref=adjacent.media). For anyone running a channel, the lead-volume metrics that justify partner budgets are measuring an activity that's losing its place in the buying path. The partners who survive the reallocation will be the ones doing implementation and integration work — things an AI summary can describe but not perform.

**The gravity is capital and talent; trade plays a secondary role.**

Semafor resurrected Ross Perot's [giant sucking sound](https://en.wikipedia.org/wiki/Giant%5Fsucking%5Fsound?ref=adjacent.media) for a different flow. In 1992 the phrase described jobs moving south; the version being argued now is capital and talent moving toward the AI buildout and away from everything else competing for the same dollars and the same graduates. That's consistent with what we've been tracking in hiring: AI startups outbidding McKinsey and Bain for the same analyst class, and mature-market companies pivoting away from core products as growth flattens. For brand leaders outside the AI trade, the real pressure point is whether your category can still hire qualified talent, and at what premium.

## Commerce Rewired

**Retail media is becoming a media business, which means it inherits media's problems**

Emarketer's Sarah Marzano put commerce media at [one in four digital ad dollars](https://www.beet.tv/2026/09/emarketers-sarah-marzano-commerce-media-will-drive-one-in-four-digital-ad-dollars.html?ref=adjacent.media), a number that only works if the inventory stops being shelf-adjacent. And it is: retail networks are [pushing their data into CTV, social and open-web placements](https://www.beet.tv/2026/09/commerce-media-moves-beyond-the-checkout-aisle-in-omnichannel-push.html?ref=adjacent.media) rather than confining it to onsite search. For brands, the thing you were buying — closed-loop attribution against a purchase you could verify — gets weaker the further inventory travels from the transaction. Off-site commerce media is targeted media with a better first-party dataset behind it. Price it accordingly, and be suspicious of any measurement story that claims the same closed loop at the top of the funnel.

**Fixed-bid pricing is the tell in professional services**

The most concrete item out of Dreamforce wasn't a product: it was consultancies moving from time-and-materials to [fixed-bid contracts because AI changed what a delivery hour costs them](https://siliconangle.com/2026/09/21/ai-professional-services-fixed-bid-pricing-certiniaatdreamforce/?ref=adjacent.media). Billable hours functioned as a pricing convention that aligned effort with revenue. Break the alignment and the firm has two choices: hold the hour and watch margin compress as clients notice the work takes less time, or sell the outcome and keep the gap. Fixed-bid is the firms choosing to keep the gap. Anyone buying services should be renegotiating on that basis now rather than waiting until renewal.

## The New Consumer

**"That's so AI" is a brand risk with a vocabulary**

Gen Z has turned AI into a pejorative — [slop, hollow, trying too hard](https://www.theupandup.us/p/thats-so-ai-gen-z-insult-slang?ref=adjacent.media) — while using the tools constantly. Using AI privately is fine—the status rule targets output alone. Having your work read as AI-made is the insult. For marketers, that makes AI-generated creative a detectability problem rather than an ethics problem, and detectability is getting cheaper for audiences and harder for brands. The fabricated skincare brand that went viral on Gen Z platforms last week with entirely synthetic creator content is the same dynamic from the other side: it worked until someone named it.

Daniel Pink's [thirty rules for being a functional adult](https://danielpink.substack.com/p/30-cheat-codes-to-beat-every-level) sits adjacent to that: a list of small non-automatable competencies — remembering names, writing the note, showing up early — circulating as advice at the exact moment the generic version of all of it is free. Scarcity is moving toward the things that require a person to have actually been there.

**Drone delivery's binding constraint is the neighbors**

Amazon and Walmart are flying commercial drone routes over Richardson, Texas, and the resident complaints are [noise and packages dropped badly, not safety or privacy](https://www.nytimes.com/2026/09/21/us/politics/amazon-drone-delivery-texas.html?unlocked%5Farticle%5Fcode=1.C1E.QIHL.7qTJRR5Uj6n4&&ref=adjacent.media#x26;smid=bs-share) — which is worse for the operators, because noise and clumsiness get litigated at city council meetings, where a single ordinance can end a market. The FAA cleared the technical path. Nobody cleared the acoustic one. The next round of drone economics will likely be decided by rotor design and municipal decibel limits rather than by unit cost per delivery.

## Connected World

**Wall Street is questioning the financing while leaving the compute unchallenged.**

The skepticism showing up around [data center IPOs and the debt behind them](https://www.nytimes.com/2026/09/21/business/ai-data-center-ipos.The%20[story]%28html?ref=adjacent.media) (paywall) is a credit story. Investors are scrutinizing long-dated obligations underwritten by short-dated customer commitments, in a business where the tenants are private companies raising enormous sums without profitable models to show against the burn. Whoever financed the building is holding a twenty-year asset against a three-year contract. It pairs with what's happening on the technical side, where Mixture-of-Experts serving economics keep pushing inference toward commodity pricing. Commodity pricing is good for buyers and rough for anyone who underwrote a premium-revenue pro forma.

**Partial autonomy's failure mode is the human it was sold to**

Safety researchers are finding an uncomfortable pattern in advanced driver assistance: [systems good enough to trust are the ones that dull reaction time when they fail](https://www.nytimes.com/2026/09/22/business/self-driving-cars-tesla-fsd.html?ref=adjacent.media) (paywall). Complacency is the mechanism, and it scales with capability — the better the system, the less prepared the driver. That logic doesn't stay in cars. Every enterprise deploying AI with a human-in-the-loop safeguard is assuming an attentive human, and attention degrades precisely in proportion to how reliable the system usually is. If your governance model's last line of defense is review-before-send, you have designed for a reviewer who has been trained by months of good output to skim.

## Culture & Signal

**When verification fails, institutions delete the assignment**

Universities are pulling AI detection tools out of academic integrity workflows because [false positives cost more credibility than cheating does](https://www.theatlantic.com/technology/2026/09/college-professors-ai-detectors-pangram/688727/?gift=NBdGSmKfDQzLc1B6N1F-gesKt22vfyUfDVIq3Mku0a8&&ref=adjacent.media#x26;utm%5Fsource=copy-link&utm%5Fmedium=social&utm%5Fcampaign=share) — a defensible call. Less defensible is what some instructors did next: cancel the writing assignment. The Atlantic's reporting catches an institution choosing to stop measuring a skill rather than find a new way to measure it. Anyone building verification or provenance tooling should read the detector failure as market information: accuracy alone doesn't clear the bar when the cost of a false accusation lands on a named individual. The products that work here will move burden of proof; probability scores will not be enough.

**Cheap content doesn't fix a distribution problem**

McClatchy invested in AI-generated content and then [laid off more than 90 journalists anyway](https://open.substack.com/pub/simonowens/p/ai-slop-will-not-save-legacy-newspapers), which is the whole argument in one sequence. Local newspapers don't lose money because writing is expensive; they lose money because the ad market left and search referrals are drying up. Reducing the cost of the input doesn't restore demand for the output. It produces more of a thing readers already weren't paying for, on platforms that increasingly downrank it. The media companies with functioning economics right now sell membership, events, or a specific reporter's judgment. None of those get cheaper with a generation model.

## Machines & Minds

**Vertical AI companies are discovering their gross margin is someone else's revenue**

Harvey, Abridge, Ramp and Rogo are moving workloads onto [open-weight models and in-house training to cut frontier lab dependency](https://www.bloomberg.com/news/articles/2026-09-21/startups-like-harvey-embrace-open-models-to-cut-reliance-on-anthropic-openai?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc5MDAxMDkwMiwiZXhwIjoxNzkwNjE1NzAyLCJhcnRpY2xlSWQiOiJUTFEyMFpOM04wQTMwMCIsImJjb25uZWN0SWQiOiI5MDI3MkIwQzRFRjQ0MkUxOUZCQkFBRTdBODJGMjM3NSJ9.i7tIzBBJutWRddw95G5V52NMXfBwhYa2v52uZC8eJ-g&ref=adjacent.media) (paywall). These are the application-layer success stories — the companies with real revenue and real retention — and they're the ones with the most to gain from shaving inference cost, because at their volumes API spend is the difference between a software margin and a services margin. The counter-consideration is that open weights only pencil out when the task is narrow enough that a fine-tuned smaller model holds quality. That's true for contract review and clinical note generation, and less true the moment a customer asks for general reasoning. Read Stripe's move on OpenRouter in the same light: routing between models is a margin product, and someone is building the infrastructure to arbitrage it.

**The labeling pipeline has the same incentives as everyone else**

OpenAI contractors hired to hand-label training data [used AI to do the labeling and got fired for it](https://www.404media.co/people-training-openais-ai-fired-for-using-ai-to-train-the-ai/?ref=adjacent.media). Funny, and also a data integrity issue with real downstream cost: human preference data is the input that's supposed to be uncontaminated. The more instructive part is the incentive design. Workers paid per task, evaluated on throughput, given access to a tool that multiplies throughput, will use the tool. That's a piece-rate system colliding with general-purpose automation. The ethics failure is structural. Every enterprise running AI-assisted work under volume-based performance metrics is building the same trap, and Shopify's internal complaints about unreviewed output landing on coworkers' desks are the same story told from inside a company that actually likes AI.

---

*15 articles across 6 themes · 12 sources · Powered by Folo + Claude*