> ## 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 — August 25, 2026
- URL: https://adjacent.media/briefs/2026-08-25/
- Published: 2026-08-25T14:15:03.000Z
- Updated: 2026-08-25T14:15:04.000Z
- Description: Ramp card data shows Anthropic’s Fable 5, launched in June, has flattened at roughly 11% of what companies spend on Anthropic tools, with buyers rotating toward cheaper models and Opus 5 running ahead of it.
- Author: Jonathan Greene
- Tags: #brief

**TL;DR:** Ramp card data shows Anthropic's Fable 5, launched in June, has flattened at roughly 11% of what companies spend on Anthropic tools, with buyers rotating toward cheaper models and Opus 5 running ahead of it. Forrester published a related argument on the enterprise side — too many AI use cases, too little measurable impact — while The Next Web reported employers rehiring workers they cut for AI at lower salaries.

## Worth Reading

- [Anthropic's newest model stalled at 11% of its own customers' spend](https://www.ft.com/content/5ee49718-c258-4f01-aa32-7e5b76ae5245?ref=adjacent.media) (paywall) — Ramp data via the FT; the fastest repricing cycle any enterprise software category has run.
- [Public court records were always public — volunteers are making them findable](https://www.404media.co/how-a-network-of-volunteers-is-liberating-critical-court-records-for-everyone/?ref=adjacent.media) — The infrastructure of civic transparency is being rebuilt by unpaid people, again.
- [Platforms removed the logout button, and session length went up](https://open.substack.com/pub/uxco/p/we-used-to-log-off) — A small design deletion that explains a decade of behavior change.
- [Progress feels bad because attention economics set the baseline, not living standards](https://open.substack.com/pub/matthewyglesias/p/the-real-reason-nobody-feels-good) — Useful counterweight to every consumer-sentiment reading you'll see this quarter.
- [The platform layer is where connected-product margins quietly go to die](https://www.solidsmack.com/programming/connected-products-at-scale-why-the-platform-layer-becomes-your-biggest-engineering-asset-or-liability/?ref=adjacent.media) — Hardware companies keep underwriting a software business they never priced.
- [Brands are about to be graded on AI restraint](https://open.substack.com/pub/answereconomy/p/the-great-ai-backlash-has-begunand) — The backlash is a positioning opportunity before it's a risk.
- [CUDA's moat gets tested by agentic inference, not training](https://newsletter.semianalysis.com/p/agentx-inferencexv3-does-cuda-moat?ref=adjacent.media) — Where the workload shifts, the lock-in follows.

## Brand & Growth

**The org chart is hiring ahead of the evidence**

Business schools are building curricula for a chief AI officer role that most companies [still can't define the responsibilities of](https://www.bloomberg.com/news/articles/2026-08-21/how-executives-are-training-to-become-chief-ai-officers?accessToken=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.eyJzb3VyY2UiOiJTdWJzY3JpYmVyR2lmdGVkQXJ0aWNsZSIsImlhdCI6MTc4NzI5MjUxMSwiZXhwIjoxNzg3ODk3MzExLCJhcnRpY2xlSWQiOiJUSzNWQzFOM04wOVYwMCIsImJjb25uZWN0SWQiOiIwQUFENjIyQkZCODY0MjkwOTk5RkVENzQyNUJDMTI3QiJ9.79558osrv%5FMt9BbDtek10VvklRGpC-I83jF2b7hzLKc&ref=adjacent.media) (paywall), per Bloomberg. Forrester supplies the reason the job is hard to spec: enterprises have accumulated more AI use cases than they can resource, and breadth is [actively suppressing measurable impact](https://www.forrester.com/blogs/too-many-ai-use-cases-too-little-impact/?ref=adjacent.media). Twenty pilots at 3% efficiency each produce a slide, not a P&L line. The useful read for anyone hiring into this role: the CAIO's first job is subtraction — killing the use cases that exist because a vendor demo was compelling, and putting the headcount behind the two that touch revenue or a cost line someone owns. Titles are cheap right now. Portfolio discipline isn't.

**Selling a phone by promising you'll use it less**

Anti-phone positioning has moved from indie hardware into mainstream device marketing, with campaigns now [selling restraint as the feature](https://embedded.substack.com/p/anti-phone-phone-marketing-is-here) — Embedded's read is that the category has found its guilt reflex and is monetizing it. It's the same maneuver Big Tobacco's light cigarettes ran, and it works for the same reason: the product doesn't change, the relationship to it gets renamed. Read alongside the UX Collective piece in Worth Reading on the disappearing logout button — the marketing department is selling digital moderation while the growth team removes the exit.

## Commerce Rewired

**Aggregation is the endgame; the content was the pretext**

Netflix, YouTube and Amazon are each racing to become the single app where subscriptions are bought and watched, with YouTube [building an in-app aggregator to match Amazon's Channels business](https://www.nytimes.com/2026/08/24/business/media/amazon-youtube-netflix-streaming-platform.html?ref=adjacent.media) (paywall). Two decades of content spend produced this: the library is a commodity, and the valuable position is the billing relationship and the discovery surface in front of it. For anyone selling a subscription product — media or not — the strategic question is whether you'll be a merchandised SKU inside someone else's storefront or the storefront itself. Amazon has already proven the first role pays worse.

**Model loyalty now lasts about one release cycle**

Fable 5 launched in June and has [flattened at roughly 11% of spending on Anthropic's tools](https://www.ft.com/content/5ee49718-c258-4f01-aa32-7e5b76ae5245?ref=adjacent.media) (paywall), according to Ramp corporate card data reported by the FT, with Opus 5 running ahead of it and companies rotating toward cheaper alternatives. Enterprise buyers are treating frontier models as a commodity input priced per token, not as a platform commitment, which is exactly what the AI vendors' move toward token-and-API-call pricing was supposed to prevent. The switching cost they engineered is smaller than the price delta the market is offering.

**A currency nobody asked for is still a currency nobody uses**

Central banks are pushing CBDCs to defend monetary sovereignty against dollar stablecoins and [can't manufacture demand for them](https://www.bloomberg.com/opinion/articles/2026-08-23/central-bank-digital-currencies-are-no-match-for-dollar-stablecoins?ref=adjacent.media) (paywall), Bloomberg's Jonathan Levin writes. The mismatch is structural: stablecoins solve a user problem (moving dollars cheaply across borders), CBDCs solve an institutional problem (who controls the rails). Products built to defend an incumbent's position rather than serve a user tend to ship on time and get adopted by nobody.

## Connected World

**The capex is booked; the components aren't**

US corporate spending on equipment and facilities is set to rise 40% between 2021 and 2027 — [more than three times Europe's pace](https://www.ft.com/content/77b94c4a-4b4b-4983-9138-7db6926150f4?ref=adjacent.media) (paywall), per Oxford Economics, driven almost entirely by the AI buildout. The constraint underneath it: grid-scale battery installations in the US are [delayed by supply chains, not prices](https://hardware.slashdot.org/story/26/08/23/0333203/supply-chain-issues-delaying-us-grid-batteries-installation?utm%5Fsource=rss1.0mainlinkanon&utm%5Fmedium=feed), the residue of decades of not building domestic manufacturing for cells, transformers, and interconnect hardware. This is the same physical bottleneck that has been pushing PCB and ceramic capacitor prices up and squeezing automotive supply. Capital can be raised in a quarter; a transformer factory takes four years. For anyone planning around AI infrastructure timelines, the announcement date and the energized date are diverging, and the gap is where the risk lives.

**A 9.39-second 100m is a benchmark, not a labor market**

At Beijing's World Humanoid Robot Games, an Honor-built robot ran the 100m in 9.39 seconds and another Chinese machine finished the 400m in 38.16 — both inside the human world records. Impressive engineering, narrow claim: a track is a flat, obstacle-free, fully specified environment, which is the one condition robotics has always handled well. The metric that would matter — a humanoid doing eight hours of unstructured warehouse work without a teleoperator — didn't get a medal event. Read the games as a state-backed demonstration of manufacturing depth and component supply, which is the real Chinese advantage here, rather than as evidence about deployment.

## Culture & Signal

**Schools are a distribution channel with a procurement department**

Natasha Singer's reporting details the structured playbook Google, Microsoft, and OpenAI use to move into American classrooms — [partnerships, grants, teacher training, and policy advocacy](https://www.nytimes.com/2026/08/23/business/schools-big-tech-google-microsoft.html?unlocked%5Farticle%5Fcode=1.7lA.pt4L.zMFCj4l3Xj3r&&ref=adjacent.media#x26;smid=nytcore-ios-share), with mounting pushback over tools whose learning outcomes remain unproven. The mechanism is habit formation at public expense: a district that standardizes on a vendor's AI suite is producing that vendor's future users, and the district absorbs the risk of the tools not working. Educators have no benchmark for what "works" here, which is precisely why the playbook leads with free.

**The data center backlash found a Republican governor**

Greg Abbott said data center companies ["dug their own grave"](https://www.axios.com/2026/08/23/greg-abbott-texas-data-centers-ai-backlash?ref=adjacent.media) by moving into Texas communities without building local support first. When the governor of the state that has courted this buildout hardest starts describing it as self-inflicted, siting risk has stopped being a blue-state problem and become a local-politics problem everywhere. Companies planning 2027 capacity should price community affairs — water, rates, tax abatements, jobs math — as a dedicated budget line item rather than absorbing it into PR functions.

**The copyright question is unsettled, which is itself the strategy**

TechCrunch's rundown of the book-training suits lands on the honest answer: whether training on copyrighted books is legal [depends on facts courts haven't finished sorting](https://techcrunch.com/2026/08/23/is-it-legal-to-train-ai-models-on-copyrighted-books-its-complicated/?ref=adjacent.media) — acquisition method, market harm, and output substitution all cut differently. Ambiguity favors whoever has already trained. The labs buying pre-2022 book corpora are hedging both ways: clean provenance for the next model, uncontaminated data for the one after.

## The New Consumer

**Employers are repricing existing labor rather than rehiring.**

Companies that cut staff citing AI are bringing those workers back at lower pay, [often through contract arrangements that strip tenure and benefits](https://thenextweb.com/news/ai-layoff-reversals-worker-trust-works-councils?ref=adjacent.media) — and Forrester's read is that 55% of employers regret AI-attributed layoffs, with roughly half of those cuts likely to reverse. The reversal reveals that the layoff was a wage-reset mechanism that happened to have an AI narrative attached. Anyone modeling AI labor savings should assume the first-year number is inflated by rehire costs and the second-year number is really a compensation story.

**Waymo's share gains show up in driver hours before they show up in headlines**

Rideshare drivers in Waymo markets are [reporting measurably fewer hours and longer waits between fares](https://therideshareguy.com/weekly-roundup-waymo-is-taking-a-real-bite-out-of-rideshare-and-drivers-feel-it-in-their-hours/?ref=adjacent.media), which is the behavioral version of a market-share chart and arrives well before the platform earnings call confirms it. Displacement in gig work doesn't announce itself through unemployment claims. It shows up as utilization decay among people who are technically still employed.

**Scarcity is the one luxury signal that can't be generated**

GQ's read on what a hit watch looks like in 2026 lands on the [Blancpain Bathyscaphe 70th anniversary limited edition](https://www.gq.com/story/blancpain-bathyscaphe-70th-anniversary-limited-edition-new-watch-alert?ref=adjacent.media) — a numbered run, a specific mechanical lineage, a documented human maker. That's the same instinct driving the Bentley-Gallup finding that 79% of Americans have now seen an ad they believed was AI-generated: when the visual layer of commerce becomes infinitely reproducible, the premium moves to whatever carries a serial number. Brands with real provenance should be documenting it now, in public, before proof-of-human becomes a compliance checkbox rather than a differentiator.

## Machines & Minds

**A frontier-class model with no owner is a supply chain problem**

Since August 20, an unidentified party has been serving a frontier-capable coding model called Ox Alpha for free on OpenRouter, with [no claimed attribution and no clarity on where submitted code goes](https://siliconangle.com/2026/08/23/nobody-knows-who-built-ai-coding-model-ox-alpha-or-where-the-code-goes/?ref=adjacent.media). Developers are using it anyway, because it's good and it's free. Every engineering org should treat this as the AI-era equivalent of an unvetted npm dependency: the real exposure is proprietary source being pasted into an endpoint with no named operator, no terms, and no retention policy. Somebody is paying an eight-figure inference bill for that traffic. It's worth asking what they're buying.

**Benchmarks measure recall; medicine measures consequences**

a16z raises what it calls the oracle problem, a critique enterprise buyers keep rediscovering: medical AI systems post perfect benchmark scores while [lacking the supervised training pipeline that turns a knowledgeable human into a trusted clinician](https://open.substack.com/pub/a16z/p/the-oracle-problem-an-invisible-bottleneck). A resident spends years being wrong under observation, with someone accountable for the correction. No such loop exists for a model that scores 100% and then sees a patient. The bottleneck is the absence of a validation apparatus, and building one is a slower, less fundable business than building the model. Azeem Azhar's weekly data roundup circles the same gap from the macro side: the capability curves keep climbing while [the deployment evidence stays thin](https://open.substack.com/pub/exponentialview/p/data-to-start-your-week-15-june-2026). Between that and Forrester's use-case findings, the pattern worth watching over the next few months is measurement — who builds the proof layer, and whether buyers start paying for it.

---

*\*Scarcity is the one luxury signal that can't be generated\**